linux/drivers/target/target_core_user.c

1731 lines
43 KiB
C
Raw Normal View History

target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
/*
* Copyright (C) 2013 Shaohua Li <shli@kernel.org>
* Copyright (C) 2014 Red Hat, Inc.
* Copyright (C) 2015 Arrikto, Inc.
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
* Copyright (C) 2017 Chinamobile, Inc.
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
*
* This program is free software; you can redistribute it and/or modify it
* under the terms and conditions of the GNU General Public License,
* version 2, as published by the Free Software Foundation.
*
* This program is distributed in the hope it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
* more details.
*
* You should have received a copy of the GNU General Public License along with
* this program; if not, write to the Free Software Foundation, Inc.,
* 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
*/
#include <linux/spinlock.h>
#include <linux/module.h>
#include <linux/idr.h>
#include <linux/kernel.h>
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
#include <linux/timer.h>
#include <linux/parser.h>
#include <linux/vmalloc.h>
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
#include <linux/uio_driver.h>
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
#include <linux/radix-tree.h>
#include <linux/stringify.h>
#include <linux/bitops.h>
#include <linux/highmem.h>
#include <linux/configfs.h>
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
#include <linux/mutex.h>
#include <linux/kthread.h>
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
#include <net/genetlink.h>
#include <scsi/scsi_common.h>
#include <scsi/scsi_proto.h>
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
#include <target/target_core_base.h>
#include <target/target_core_fabric.h>
#include <target/target_core_backend.h>
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
#include <linux/target_core_user.h>
/*
* Define a shared-memory interface for LIO to pass SCSI commands and
* data to userspace for processing. This is to allow backends that
* are too complex for in-kernel support to be possible.
*
* It uses the UIO framework to do a lot of the device-creation and
* introspection work for us.
*
* See the .h file for how the ring is laid out. Note that while the
* command ring is defined, the particulars of the data area are
* not. Offset values in the command entry point to other locations
* internal to the mmap()ed area. There is separate space outside the
* command ring for data buffers. This leaves maximum flexibility for
* moving buffer allocations, or even page flipping or other
* allocation techniques, without altering the command ring layout.
*
* SECURITY:
* The user process must be assumed to be malicious. There's no way to
* prevent it breaking the command ring protocol if it wants, but in
* order to prevent other issues we must only ever read *data* from
* the shared memory area, not offsets or sizes. This applies to
* command ring entries as well as the mailbox. Extra code needed for
* this may have a 'UAM' comment.
*/
#define TCMU_TIME_OUT (30 * MSEC_PER_SEC)
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
/* For cmd area, the size is fixed 8MB */
#define CMDR_SIZE (8 * 1024 * 1024)
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
/*
* For data area, the block size is PAGE_SIZE and
* the total size is 256K * PAGE_SIZE.
*/
#define DATA_BLOCK_SIZE PAGE_SIZE
#define DATA_BLOCK_BITS (256 * 1024)
#define DATA_SIZE (DATA_BLOCK_BITS * DATA_BLOCK_SIZE)
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
#define DATA_BLOCK_INIT_BITS 128
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
/* The total size of the ring is 8M + 256K * PAGE_SIZE */
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
#define TCMU_RING_SIZE (CMDR_SIZE + DATA_SIZE)
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
/* Default maximum of the global data blocks(512K * PAGE_SIZE) */
#define TCMU_GLOBAL_MAX_BLOCKS (512 * 1024)
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
static struct device *tcmu_root_device;
struct tcmu_hba {
u32 host_id;
};
#define TCMU_CONFIG_LEN 256
struct tcmu_dev {
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
struct list_head node;
struct kref kref;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
struct se_device se_dev;
char *name;
struct se_hba *hba;
#define TCMU_DEV_BIT_OPEN 0
#define TCMU_DEV_BIT_BROKEN 1
unsigned long flags;
struct uio_info uio_info;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
struct inode *inode;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
struct tcmu_mailbox *mb_addr;
size_t dev_size;
u32 cmdr_size;
u32 cmdr_last_cleaned;
/* Offset of data area from start of mb */
/* Must add data_off and mb_addr to get the address */
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
size_t data_off;
size_t data_size;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
wait_queue_head_t wait_cmdr;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
struct mutex cmdr_lock;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
bool waiting_global;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
uint32_t dbi_max;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
uint32_t dbi_thresh;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
DECLARE_BITMAP(data_bitmap, DATA_BLOCK_BITS);
struct radix_tree_root data_blocks;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
struct idr commands;
spinlock_t commands_lock;
struct timer_list timeout;
unsigned int cmd_time_out;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
char dev_config[TCMU_CONFIG_LEN];
};
#define TCMU_DEV(_se_dev) container_of(_se_dev, struct tcmu_dev, se_dev)
#define CMDR_OFF sizeof(struct tcmu_mailbox)
struct tcmu_cmd {
struct se_cmd *se_cmd;
struct tcmu_dev *tcmu_dev;
uint16_t cmd_id;
/* Can't use se_cmd when cleaning up expired cmds, because if
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
cmd has been completed then accessing se_cmd is off limits */
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
uint32_t dbi_cnt;
uint32_t dbi_cur;
uint32_t *dbi;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
unsigned long deadline;
#define TCMU_CMD_BIT_EXPIRED 0
unsigned long flags;
};
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
static struct task_struct *unmap_thread;
static wait_queue_head_t unmap_wait;
static DEFINE_MUTEX(root_udev_mutex);
static LIST_HEAD(root_udev);
static atomic_t global_db_count = ATOMIC_INIT(0);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
static struct kmem_cache *tcmu_cmd_cache;
/* multicast group */
enum tcmu_multicast_groups {
TCMU_MCGRP_CONFIG,
};
static const struct genl_multicast_group tcmu_mcgrps[] = {
[TCMU_MCGRP_CONFIG] = { .name = "config", },
};
/* Our generic netlink family */
static struct genl_family tcmu_genl_family __ro_after_init = {
.module = THIS_MODULE,
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
.hdrsize = 0,
.name = "TCM-USER",
.version = 1,
.maxattr = TCMU_ATTR_MAX,
.mcgrps = tcmu_mcgrps,
.n_mcgrps = ARRAY_SIZE(tcmu_mcgrps),
.netnsok = true,
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
};
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
#define tcmu_cmd_set_dbi_cur(cmd, index) ((cmd)->dbi_cur = (index))
#define tcmu_cmd_reset_dbi_cur(cmd) tcmu_cmd_set_dbi_cur(cmd, 0)
#define tcmu_cmd_set_dbi(cmd, index) ((cmd)->dbi[(cmd)->dbi_cur++] = (index))
#define tcmu_cmd_get_dbi(cmd) ((cmd)->dbi[(cmd)->dbi_cur++])
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
static void tcmu_cmd_free_data(struct tcmu_cmd *tcmu_cmd, uint32_t len)
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
{
struct tcmu_dev *udev = tcmu_cmd->tcmu_dev;
uint32_t i;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
for (i = 0; i < len; i++)
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
clear_bit(tcmu_cmd->dbi[i], udev->data_bitmap);
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
static inline bool tcmu_get_empty_block(struct tcmu_dev *udev,
struct tcmu_cmd *tcmu_cmd)
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
{
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
struct page *page;
int ret, dbi;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
dbi = find_first_zero_bit(udev->data_bitmap, udev->dbi_thresh);
if (dbi == udev->dbi_thresh)
return false;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
page = radix_tree_lookup(&udev->data_blocks, dbi);
if (!page) {
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
if (atomic_add_return(1, &global_db_count) >
TCMU_GLOBAL_MAX_BLOCKS) {
atomic_dec(&global_db_count);
return false;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
/* try to get new page from the mm */
page = alloc_page(GFP_KERNEL);
if (!page)
return false;
ret = radix_tree_insert(&udev->data_blocks, dbi, page);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
if (ret) {
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
__free_page(page);
return false;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
if (dbi > udev->dbi_max)
udev->dbi_max = dbi;
set_bit(dbi, udev->data_bitmap);
tcmu_cmd_set_dbi(tcmu_cmd, dbi);
return true;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
static bool tcmu_get_empty_blocks(struct tcmu_dev *udev,
struct tcmu_cmd *tcmu_cmd)
{
int i;
udev->waiting_global = false;
for (i = tcmu_cmd->dbi_cur; i < tcmu_cmd->dbi_cnt; i++) {
if (!tcmu_get_empty_block(udev, tcmu_cmd))
goto err;
}
return true;
err:
udev->waiting_global = true;
/* Try to wake up the unmap thread */
wake_up(&unmap_wait);
return false;
}
static inline struct page *
tcmu_get_block_page(struct tcmu_dev *udev, uint32_t dbi)
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
{
return radix_tree_lookup(&udev->data_blocks, dbi);
}
static inline void tcmu_free_cmd(struct tcmu_cmd *tcmu_cmd)
{
kfree(tcmu_cmd->dbi);
kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
}
static inline size_t tcmu_cmd_get_data_length(struct tcmu_cmd *tcmu_cmd)
{
struct se_cmd *se_cmd = tcmu_cmd->se_cmd;
size_t data_length = round_up(se_cmd->data_length, DATA_BLOCK_SIZE);
if (se_cmd->se_cmd_flags & SCF_BIDI) {
BUG_ON(!(se_cmd->t_bidi_data_sg && se_cmd->t_bidi_data_nents));
data_length += round_up(se_cmd->t_bidi_data_sg->length,
DATA_BLOCK_SIZE);
}
return data_length;
}
static inline uint32_t tcmu_cmd_get_block_cnt(struct tcmu_cmd *tcmu_cmd)
{
size_t data_length = tcmu_cmd_get_data_length(tcmu_cmd);
return data_length / DATA_BLOCK_SIZE;
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
static struct tcmu_cmd *tcmu_alloc_cmd(struct se_cmd *se_cmd)
{
struct se_device *se_dev = se_cmd->se_dev;
struct tcmu_dev *udev = TCMU_DEV(se_dev);
struct tcmu_cmd *tcmu_cmd;
int cmd_id;
tcmu_cmd = kmem_cache_zalloc(tcmu_cmd_cache, GFP_KERNEL);
if (!tcmu_cmd)
return NULL;
tcmu_cmd->se_cmd = se_cmd;
tcmu_cmd->tcmu_dev = udev;
if (udev->cmd_time_out)
tcmu_cmd->deadline = jiffies +
msecs_to_jiffies(udev->cmd_time_out);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
tcmu_cmd_reset_dbi_cur(tcmu_cmd);
tcmu_cmd->dbi_cnt = tcmu_cmd_get_block_cnt(tcmu_cmd);
tcmu_cmd->dbi = kcalloc(tcmu_cmd->dbi_cnt, sizeof(uint32_t),
GFP_KERNEL);
if (!tcmu_cmd->dbi) {
kmem_cache_free(tcmu_cmd_cache, tcmu_cmd);
return NULL;
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
idr_preload(GFP_KERNEL);
spin_lock_irq(&udev->commands_lock);
cmd_id = idr_alloc(&udev->commands, tcmu_cmd, 0,
USHRT_MAX, GFP_NOWAIT);
spin_unlock_irq(&udev->commands_lock);
idr_preload_end();
if (cmd_id < 0) {
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
tcmu_free_cmd(tcmu_cmd);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
return NULL;
}
tcmu_cmd->cmd_id = cmd_id;
return tcmu_cmd;
}
static inline void tcmu_flush_dcache_range(void *vaddr, size_t size)
{
unsigned long offset = offset_in_page(vaddr);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
size = round_up(size+offset, PAGE_SIZE);
vaddr -= offset;
while (size) {
flush_dcache_page(virt_to_page(vaddr));
size -= PAGE_SIZE;
}
}
/*
* Some ring helper functions. We don't assume size is a power of 2 so
* we can't use circ_buf.h.
*/
static inline size_t spc_used(size_t head, size_t tail, size_t size)
{
int diff = head - tail;
if (diff >= 0)
return diff;
else
return size + diff;
}
static inline size_t spc_free(size_t head, size_t tail, size_t size)
{
/* Keep 1 byte unused or we can't tell full from empty */
return (size - spc_used(head, tail, size) - 1);
}
static inline size_t head_to_end(size_t head, size_t size)
{
return size - head;
}
static inline void new_iov(struct iovec **iov, int *iov_cnt,
struct tcmu_dev *udev)
{
struct iovec *iovec;
if (*iov_cnt != 0)
(*iov)++;
(*iov_cnt)++;
iovec = *iov;
memset(iovec, 0, sizeof(struct iovec));
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
#define UPDATE_HEAD(head, used, size) smp_store_release(&head, ((head % size) + used) % size)
/* offset is relative to mb_addr */
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
static inline size_t get_block_offset_user(struct tcmu_dev *dev,
int dbi, int remaining)
{
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
return dev->data_off + dbi * DATA_BLOCK_SIZE +
DATA_BLOCK_SIZE - remaining;
}
static inline size_t iov_tail(struct tcmu_dev *udev, struct iovec *iov)
{
return (size_t)iov->iov_base + iov->iov_len;
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
static int scatter_data_area(struct tcmu_dev *udev,
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
struct tcmu_cmd *tcmu_cmd, struct scatterlist *data_sg,
unsigned int data_nents, struct iovec **iov,
int *iov_cnt, bool copy_data)
{
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
int i, dbi;
int block_remaining = 0;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
void *from, *to = NULL;
size_t copy_bytes, to_offset, offset;
struct scatterlist *sg;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
struct page *page;
for_each_sg(data_sg, sg, data_nents, i) {
int sg_remaining = sg->length;
from = kmap_atomic(sg_page(sg)) + sg->offset;
while (sg_remaining > 0) {
if (block_remaining == 0) {
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
if (to)
kunmap_atomic(to);
block_remaining = DATA_BLOCK_SIZE;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
dbi = tcmu_cmd_get_dbi(tcmu_cmd);
page = tcmu_get_block_page(udev, dbi);
to = kmap_atomic(page);
}
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
copy_bytes = min_t(size_t, sg_remaining,
block_remaining);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
to_offset = get_block_offset_user(udev, dbi,
block_remaining);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
offset = DATA_BLOCK_SIZE - block_remaining;
to = (void *)(unsigned long)to + offset;
if (*iov_cnt != 0 &&
to_offset == iov_tail(udev, *iov)) {
(*iov)->iov_len += copy_bytes;
} else {
new_iov(iov, iov_cnt, udev);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
(*iov)->iov_base = (void __user *)to_offset;
(*iov)->iov_len = copy_bytes;
}
if (copy_data) {
memcpy(to, from + sg->length - sg_remaining,
copy_bytes);
tcmu_flush_dcache_range(to, copy_bytes);
}
sg_remaining -= copy_bytes;
block_remaining -= copy_bytes;
}
kunmap_atomic(from - sg->offset);
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
if (to)
kunmap_atomic(to);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
return 0;
}
static void gather_data_area(struct tcmu_dev *udev, struct tcmu_cmd *cmd,
bool bidi)
{
struct se_cmd *se_cmd = cmd->se_cmd;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
int i, dbi;
int block_remaining = 0;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
void *from = NULL, *to;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
size_t copy_bytes, offset;
struct scatterlist *sg, *data_sg;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
struct page *page;
unsigned int data_nents;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
uint32_t count = 0;
if (!bidi) {
data_sg = se_cmd->t_data_sg;
data_nents = se_cmd->t_data_nents;
} else {
/*
* For bidi case, the first count blocks are for Data-Out
* buffer blocks, and before gathering the Data-In buffer
* the Data-Out buffer blocks should be discarded.
*/
count = DIV_ROUND_UP(se_cmd->data_length, DATA_BLOCK_SIZE);
data_sg = se_cmd->t_bidi_data_sg;
data_nents = se_cmd->t_bidi_data_nents;
}
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
tcmu_cmd_set_dbi_cur(cmd, count);
for_each_sg(data_sg, sg, data_nents, i) {
int sg_remaining = sg->length;
to = kmap_atomic(sg_page(sg)) + sg->offset;
while (sg_remaining > 0) {
if (block_remaining == 0) {
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
if (from)
kunmap_atomic(from);
block_remaining = DATA_BLOCK_SIZE;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
dbi = tcmu_cmd_get_dbi(cmd);
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
page = tcmu_get_block_page(udev, dbi);
from = kmap_atomic(page);
}
copy_bytes = min_t(size_t, sg_remaining,
block_remaining);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
offset = DATA_BLOCK_SIZE - block_remaining;
from = (void *)(unsigned long)from + offset;
tcmu_flush_dcache_range(from, copy_bytes);
memcpy(to + sg->length - sg_remaining, from,
copy_bytes);
sg_remaining -= copy_bytes;
block_remaining -= copy_bytes;
}
kunmap_atomic(to - sg->offset);
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
if (from)
kunmap_atomic(from);
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
static inline size_t spc_bitmap_free(unsigned long *bitmap, uint32_t thresh)
{
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
return DATA_BLOCK_SIZE * (thresh - bitmap_weight(bitmap, thresh));
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
/*
* We can't queue a command until we have space available on the cmd ring *and*
* space available on the data area.
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
*
* Called with ring lock held.
*/
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
static bool is_ring_space_avail(struct tcmu_dev *udev, struct tcmu_cmd *cmd,
size_t cmd_size, size_t data_needed)
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
{
struct tcmu_mailbox *mb = udev->mb_addr;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
uint32_t blocks_needed = (data_needed + DATA_BLOCK_SIZE - 1)
/ DATA_BLOCK_SIZE;
size_t space, cmd_needed;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
u32 cmd_head;
tcmu_flush_dcache_range(mb, sizeof(*mb));
cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
/*
* If cmd end-of-ring space is too small then we need space for a NOP plus
* original cmd - cmds are internally contiguous.
*/
if (head_to_end(cmd_head, udev->cmdr_size) >= cmd_size)
cmd_needed = cmd_size;
else
cmd_needed = cmd_size + head_to_end(cmd_head, udev->cmdr_size);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
space = spc_free(cmd_head, udev->cmdr_last_cleaned, udev->cmdr_size);
if (space < cmd_needed) {
pr_debug("no cmd space: %u %u %u\n", cmd_head,
udev->cmdr_last_cleaned, udev->cmdr_size);
return false;
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
/* try to check and get the data blocks as needed */
space = spc_bitmap_free(udev->data_bitmap, udev->dbi_thresh);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
if (space < data_needed) {
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
unsigned long blocks_left = DATA_BLOCK_BITS - udev->dbi_thresh;
unsigned long grow;
if (blocks_left < blocks_needed) {
pr_debug("no data space: only %lu available, but ask for %zu\n",
blocks_left * DATA_BLOCK_SIZE,
data_needed);
return false;
}
/* Try to expand the thresh */
if (!udev->dbi_thresh) {
/* From idle state */
uint32_t init_thresh = DATA_BLOCK_INIT_BITS;
udev->dbi_thresh = max(blocks_needed, init_thresh);
} else {
/*
* Grow the data area by max(blocks needed,
* dbi_thresh / 2), but limited to the max
* DATA_BLOCK_BITS size.
*/
grow = max(blocks_needed, udev->dbi_thresh / 2);
udev->dbi_thresh += grow;
if (udev->dbi_thresh > DATA_BLOCK_BITS)
udev->dbi_thresh = DATA_BLOCK_BITS;
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
if (!tcmu_get_empty_blocks(udev, cmd))
return false;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
return true;
}
static inline size_t tcmu_cmd_get_base_cmd_size(size_t iov_cnt)
{
return max(offsetof(struct tcmu_cmd_entry, req.iov[iov_cnt]),
sizeof(struct tcmu_cmd_entry));
}
static inline size_t tcmu_cmd_get_cmd_size(struct tcmu_cmd *tcmu_cmd,
size_t base_command_size)
{
struct se_cmd *se_cmd = tcmu_cmd->se_cmd;
size_t command_size;
command_size = base_command_size +
round_up(scsi_command_size(se_cmd->t_task_cdb),
TCMU_OP_ALIGN_SIZE);
WARN_ON(command_size & (TCMU_OP_ALIGN_SIZE-1));
return command_size;
}
static sense_reason_t
tcmu_queue_cmd_ring(struct tcmu_cmd *tcmu_cmd)
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
{
struct tcmu_dev *udev = tcmu_cmd->tcmu_dev;
struct se_cmd *se_cmd = tcmu_cmd->se_cmd;
size_t base_command_size, command_size;
struct tcmu_mailbox *mb;
struct tcmu_cmd_entry *entry;
struct iovec *iov;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
int iov_cnt, ret;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
uint32_t cmd_head;
uint64_t cdb_off;
bool copy_to_data_area;
size_t data_length = tcmu_cmd_get_data_length(tcmu_cmd);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags))
return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
/*
* Must be a certain minimum size for response sense info, but
* also may be larger if the iov array is large.
*
* We prepare as many iovs as possbile for potential uses here,
* because it's expensive to tell how many regions are freed in
* the bitmap & global data pool, as the size calculated here
* will only be used to do the checks.
*
* The size will be recalculated later as actually needed to save
* cmd area memories.
*/
base_command_size = tcmu_cmd_get_base_cmd_size(tcmu_cmd->dbi_cnt);
command_size = tcmu_cmd_get_cmd_size(tcmu_cmd, base_command_size);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
mutex_lock(&udev->cmdr_lock);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
mb = udev->mb_addr;
cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
if ((command_size > (udev->cmdr_size / 2)) ||
data_length > udev->data_size) {
pr_warn("TCMU: Request of size %zu/%zu is too big for %u/%zu "
"cmd ring/data area\n", command_size, data_length,
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
udev->cmdr_size, udev->data_size);
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
mutex_unlock(&udev->cmdr_lock);
return TCM_INVALID_CDB_FIELD;
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
while (!is_ring_space_avail(udev, tcmu_cmd, command_size, data_length)) {
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
int ret;
DEFINE_WAIT(__wait);
prepare_to_wait(&udev->wait_cmdr, &__wait, TASK_INTERRUPTIBLE);
pr_debug("sleeping for ring space\n");
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
mutex_unlock(&udev->cmdr_lock);
if (udev->cmd_time_out)
ret = schedule_timeout(
msecs_to_jiffies(udev->cmd_time_out));
else
ret = schedule_timeout(msecs_to_jiffies(TCMU_TIME_OUT));
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
finish_wait(&udev->wait_cmdr, &__wait);
if (!ret) {
pr_warn("tcmu: command timed out\n");
return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
mutex_lock(&udev->cmdr_lock);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
/* We dropped cmdr_lock, cmd_head is stale */
cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
}
/* Insert a PAD if end-of-ring space is too small */
if (head_to_end(cmd_head, udev->cmdr_size) < command_size) {
size_t pad_size = head_to_end(cmd_head, udev->cmdr_size);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
entry = (void *) mb + CMDR_OFF + cmd_head;
tcmu_flush_dcache_range(entry, sizeof(*entry));
target: Version 2 of TCMU ABI The initial version of TCMU (in 3.18) does not properly handle bidirectional SCSI commands -- those with both an in and out buffer. In looking to fix this it also became clear that TCMU's support for adding new types of entries (opcodes) to the command ring was broken. We need to fix this now, so that future issues can be handled properly by adding new opcodes. We make the most of this ABI break by enabling bidi cmd handling within TCMP_OP_CMD opcode. Add an iov_bidi_cnt field to tcmu_cmd_entry.req. This enables TCMU to describe bidi commands, but further kernel work is needed for full bidi support. Enlarge tcmu_cmd_entry_hdr by 32 bits by pulling in cmd_id and __pad1. Turn __pad1 into two 8 bit flags fields, for kernel-set and userspace-set flags, "kflags" and "uflags" respectively. Update version fields so userspace can tell the interface is changed. Update tcmu-design.txt with details of how new stuff works: - Specify an additional requirement for userspace to set UNKNOWN_OP (bit 0) in hdr.uflags for unknown/unhandled opcodes. - Define how Data-In and Data-Out fields are described in req.iov[] Changed in v2: - Change name of SKIPPED bit to UNKNOWN bit - PAD op does not set the bit any more - Change len_op helper functions to take just len_op, not the whole struct - Change version to 2 in missed spots, and use defines - Add 16 unused bytes to cmd_entry.req, in case additional SAM cmd parameters need to be included - Add iov_dif_cnt field to specify buffers used for DIF info in iov[] - Rearrange fields to naturally align cdb_off - Handle if userspace sets UNKNOWN_OP by indicating failure of the cmd - Wrap some overly long UPDATE_HEAD lines (Add missing req.iov_bidi_cnt + req.iov_dif_cnt zeroing - Ilias) Signed-off-by: Andy Grover <agrover@redhat.com> Reviewed-by: Ilias Tsitsimpis <iliastsi@arrikto.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2015-04-15 03:30:04 +03:00
tcmu_hdr_set_op(&entry->hdr.len_op, TCMU_OP_PAD);
tcmu_hdr_set_len(&entry->hdr.len_op, pad_size);
entry->hdr.cmd_id = 0; /* not used for PAD */
entry->hdr.kflags = 0;
entry->hdr.uflags = 0;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
UPDATE_HEAD(mb->cmd_head, pad_size, udev->cmdr_size);
cmd_head = mb->cmd_head % udev->cmdr_size; /* UAM */
WARN_ON(cmd_head != 0);
}
entry = (void *) mb + CMDR_OFF + cmd_head;
tcmu_flush_dcache_range(entry, sizeof(*entry));
target: Version 2 of TCMU ABI The initial version of TCMU (in 3.18) does not properly handle bidirectional SCSI commands -- those with both an in and out buffer. In looking to fix this it also became clear that TCMU's support for adding new types of entries (opcodes) to the command ring was broken. We need to fix this now, so that future issues can be handled properly by adding new opcodes. We make the most of this ABI break by enabling bidi cmd handling within TCMP_OP_CMD opcode. Add an iov_bidi_cnt field to tcmu_cmd_entry.req. This enables TCMU to describe bidi commands, but further kernel work is needed for full bidi support. Enlarge tcmu_cmd_entry_hdr by 32 bits by pulling in cmd_id and __pad1. Turn __pad1 into two 8 bit flags fields, for kernel-set and userspace-set flags, "kflags" and "uflags" respectively. Update version fields so userspace can tell the interface is changed. Update tcmu-design.txt with details of how new stuff works: - Specify an additional requirement for userspace to set UNKNOWN_OP (bit 0) in hdr.uflags for unknown/unhandled opcodes. - Define how Data-In and Data-Out fields are described in req.iov[] Changed in v2: - Change name of SKIPPED bit to UNKNOWN bit - PAD op does not set the bit any more - Change len_op helper functions to take just len_op, not the whole struct - Change version to 2 in missed spots, and use defines - Add 16 unused bytes to cmd_entry.req, in case additional SAM cmd parameters need to be included - Add iov_dif_cnt field to specify buffers used for DIF info in iov[] - Rearrange fields to naturally align cdb_off - Handle if userspace sets UNKNOWN_OP by indicating failure of the cmd - Wrap some overly long UPDATE_HEAD lines (Add missing req.iov_bidi_cnt + req.iov_dif_cnt zeroing - Ilias) Signed-off-by: Andy Grover <agrover@redhat.com> Reviewed-by: Ilias Tsitsimpis <iliastsi@arrikto.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2015-04-15 03:30:04 +03:00
tcmu_hdr_set_op(&entry->hdr.len_op, TCMU_OP_CMD);
entry->hdr.cmd_id = tcmu_cmd->cmd_id;
entry->hdr.kflags = 0;
entry->hdr.uflags = 0;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
/* Handle allocating space from the data area */
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
tcmu_cmd_reset_dbi_cur(tcmu_cmd);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
iov = &entry->req.iov[0];
iov_cnt = 0;
copy_to_data_area = (se_cmd->data_direction == DMA_TO_DEVICE
|| se_cmd->se_cmd_flags & SCF_BIDI);
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
ret = scatter_data_area(udev, tcmu_cmd, se_cmd->t_data_sg,
se_cmd->t_data_nents, &iov, &iov_cnt,
copy_to_data_area);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
if (ret) {
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
tcmu_cmd_free_data(tcmu_cmd, tcmu_cmd->dbi_cnt);
mutex_unlock(&udev->cmdr_lock);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
pr_err("tcmu: alloc and scatter data failed\n");
return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
entry->req.iov_cnt = iov_cnt;
target: Version 2 of TCMU ABI The initial version of TCMU (in 3.18) does not properly handle bidirectional SCSI commands -- those with both an in and out buffer. In looking to fix this it also became clear that TCMU's support for adding new types of entries (opcodes) to the command ring was broken. We need to fix this now, so that future issues can be handled properly by adding new opcodes. We make the most of this ABI break by enabling bidi cmd handling within TCMP_OP_CMD opcode. Add an iov_bidi_cnt field to tcmu_cmd_entry.req. This enables TCMU to describe bidi commands, but further kernel work is needed for full bidi support. Enlarge tcmu_cmd_entry_hdr by 32 bits by pulling in cmd_id and __pad1. Turn __pad1 into two 8 bit flags fields, for kernel-set and userspace-set flags, "kflags" and "uflags" respectively. Update version fields so userspace can tell the interface is changed. Update tcmu-design.txt with details of how new stuff works: - Specify an additional requirement for userspace to set UNKNOWN_OP (bit 0) in hdr.uflags for unknown/unhandled opcodes. - Define how Data-In and Data-Out fields are described in req.iov[] Changed in v2: - Change name of SKIPPED bit to UNKNOWN bit - PAD op does not set the bit any more - Change len_op helper functions to take just len_op, not the whole struct - Change version to 2 in missed spots, and use defines - Add 16 unused bytes to cmd_entry.req, in case additional SAM cmd parameters need to be included - Add iov_dif_cnt field to specify buffers used for DIF info in iov[] - Rearrange fields to naturally align cdb_off - Handle if userspace sets UNKNOWN_OP by indicating failure of the cmd - Wrap some overly long UPDATE_HEAD lines (Add missing req.iov_bidi_cnt + req.iov_dif_cnt zeroing - Ilias) Signed-off-by: Andy Grover <agrover@redhat.com> Reviewed-by: Ilias Tsitsimpis <iliastsi@arrikto.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2015-04-15 03:30:04 +03:00
entry->req.iov_dif_cnt = 0;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
/* Handle BIDI commands */
if (se_cmd->se_cmd_flags & SCF_BIDI) {
iov_cnt = 0;
iov++;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
ret = scatter_data_area(udev, tcmu_cmd,
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
se_cmd->t_bidi_data_sg,
se_cmd->t_bidi_data_nents,
&iov, &iov_cnt, false);
if (ret) {
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
tcmu_cmd_free_data(tcmu_cmd, tcmu_cmd->dbi_cnt);
mutex_unlock(&udev->cmdr_lock);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
pr_err("tcmu: alloc and scatter bidi data failed\n");
return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
}
entry->req.iov_bidi_cnt = iov_cnt;
}
/*
* Recalaulate the command's base size and size according
* to the actual needs
*/
base_command_size = tcmu_cmd_get_base_cmd_size(entry->req.iov_cnt +
entry->req.iov_bidi_cnt);
command_size = tcmu_cmd_get_cmd_size(tcmu_cmd, base_command_size);
tcmu_hdr_set_len(&entry->hdr.len_op, command_size);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
/* All offsets relative to mb_addr, not start of entry! */
cdb_off = CMDR_OFF + cmd_head + base_command_size;
memcpy((void *) mb + cdb_off, se_cmd->t_task_cdb, scsi_command_size(se_cmd->t_task_cdb));
entry->req.cdb_off = cdb_off;
tcmu_flush_dcache_range(entry, sizeof(*entry));
UPDATE_HEAD(mb->cmd_head, command_size, udev->cmdr_size);
tcmu_flush_dcache_range(mb, sizeof(*mb));
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
mutex_unlock(&udev->cmdr_lock);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
/* TODO: only if FLUSH and FUA? */
uio_event_notify(&udev->uio_info);
if (udev->cmd_time_out)
mod_timer(&udev->timeout, round_jiffies_up(jiffies +
msecs_to_jiffies(udev->cmd_time_out)));
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
return TCM_NO_SENSE;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
}
static sense_reason_t
tcmu_queue_cmd(struct se_cmd *se_cmd)
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
{
struct se_device *se_dev = se_cmd->se_dev;
struct tcmu_dev *udev = TCMU_DEV(se_dev);
struct tcmu_cmd *tcmu_cmd;
sense_reason_t ret;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
tcmu_cmd = tcmu_alloc_cmd(se_cmd);
if (!tcmu_cmd)
return TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
ret = tcmu_queue_cmd_ring(tcmu_cmd);
if (ret != TCM_NO_SENSE) {
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
pr_err("TCMU: Could not queue command\n");
spin_lock_irq(&udev->commands_lock);
idr_remove(&udev->commands, tcmu_cmd->cmd_id);
spin_unlock_irq(&udev->commands_lock);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
tcmu_free_cmd(tcmu_cmd);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
}
return ret;
}
static void tcmu_handle_completion(struct tcmu_cmd *cmd, struct tcmu_cmd_entry *entry)
{
struct se_cmd *se_cmd = cmd->se_cmd;
struct tcmu_dev *udev = cmd->tcmu_dev;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
/*
* cmd has been completed already from timeout, just reclaim
* data area space and free cmd
*/
if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags))
goto out;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
tcmu_cmd_reset_dbi_cur(cmd);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
target: Version 2 of TCMU ABI The initial version of TCMU (in 3.18) does not properly handle bidirectional SCSI commands -- those with both an in and out buffer. In looking to fix this it also became clear that TCMU's support for adding new types of entries (opcodes) to the command ring was broken. We need to fix this now, so that future issues can be handled properly by adding new opcodes. We make the most of this ABI break by enabling bidi cmd handling within TCMP_OP_CMD opcode. Add an iov_bidi_cnt field to tcmu_cmd_entry.req. This enables TCMU to describe bidi commands, but further kernel work is needed for full bidi support. Enlarge tcmu_cmd_entry_hdr by 32 bits by pulling in cmd_id and __pad1. Turn __pad1 into two 8 bit flags fields, for kernel-set and userspace-set flags, "kflags" and "uflags" respectively. Update version fields so userspace can tell the interface is changed. Update tcmu-design.txt with details of how new stuff works: - Specify an additional requirement for userspace to set UNKNOWN_OP (bit 0) in hdr.uflags for unknown/unhandled opcodes. - Define how Data-In and Data-Out fields are described in req.iov[] Changed in v2: - Change name of SKIPPED bit to UNKNOWN bit - PAD op does not set the bit any more - Change len_op helper functions to take just len_op, not the whole struct - Change version to 2 in missed spots, and use defines - Add 16 unused bytes to cmd_entry.req, in case additional SAM cmd parameters need to be included - Add iov_dif_cnt field to specify buffers used for DIF info in iov[] - Rearrange fields to naturally align cdb_off - Handle if userspace sets UNKNOWN_OP by indicating failure of the cmd - Wrap some overly long UPDATE_HEAD lines (Add missing req.iov_bidi_cnt + req.iov_dif_cnt zeroing - Ilias) Signed-off-by: Andy Grover <agrover@redhat.com> Reviewed-by: Ilias Tsitsimpis <iliastsi@arrikto.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2015-04-15 03:30:04 +03:00
if (entry->hdr.uflags & TCMU_UFLAG_UNKNOWN_OP) {
pr_warn("TCMU: Userspace set UNKNOWN_OP flag on se_cmd %p\n",
cmd->se_cmd);
entry->rsp.scsi_status = SAM_STAT_CHECK_CONDITION;
} else if (entry->rsp.scsi_status == SAM_STAT_CHECK_CONDITION) {
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
memcpy(se_cmd->sense_buffer, entry->rsp.sense_buffer,
se_cmd->scsi_sense_length);
} else if (se_cmd->se_cmd_flags & SCF_BIDI) {
/* Get Data-In buffer before clean up */
gather_data_area(udev, cmd, true);
} else if (se_cmd->data_direction == DMA_FROM_DEVICE) {
gather_data_area(udev, cmd, false);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
} else if (se_cmd->data_direction == DMA_TO_DEVICE) {
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
/* TODO: */
} else if (se_cmd->data_direction != DMA_NONE) {
pr_warn("TCMU: data direction was %d!\n",
se_cmd->data_direction);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
}
target_complete_cmd(cmd->se_cmd, entry->rsp.scsi_status);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
out:
cmd->se_cmd = NULL;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
tcmu_cmd_free_data(cmd, cmd->dbi_cnt);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
tcmu_free_cmd(cmd);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
}
static unsigned int tcmu_handle_completions(struct tcmu_dev *udev)
{
struct tcmu_mailbox *mb;
int handled = 0;
if (test_bit(TCMU_DEV_BIT_BROKEN, &udev->flags)) {
pr_err("ring broken, not handling completions\n");
return 0;
}
mb = udev->mb_addr;
tcmu_flush_dcache_range(mb, sizeof(*mb));
while (udev->cmdr_last_cleaned != ACCESS_ONCE(mb->cmd_tail)) {
struct tcmu_cmd_entry *entry = (void *) mb + CMDR_OFF + udev->cmdr_last_cleaned;
struct tcmu_cmd *cmd;
tcmu_flush_dcache_range(entry, sizeof(*entry));
target: Version 2 of TCMU ABI The initial version of TCMU (in 3.18) does not properly handle bidirectional SCSI commands -- those with both an in and out buffer. In looking to fix this it also became clear that TCMU's support for adding new types of entries (opcodes) to the command ring was broken. We need to fix this now, so that future issues can be handled properly by adding new opcodes. We make the most of this ABI break by enabling bidi cmd handling within TCMP_OP_CMD opcode. Add an iov_bidi_cnt field to tcmu_cmd_entry.req. This enables TCMU to describe bidi commands, but further kernel work is needed for full bidi support. Enlarge tcmu_cmd_entry_hdr by 32 bits by pulling in cmd_id and __pad1. Turn __pad1 into two 8 bit flags fields, for kernel-set and userspace-set flags, "kflags" and "uflags" respectively. Update version fields so userspace can tell the interface is changed. Update tcmu-design.txt with details of how new stuff works: - Specify an additional requirement for userspace to set UNKNOWN_OP (bit 0) in hdr.uflags for unknown/unhandled opcodes. - Define how Data-In and Data-Out fields are described in req.iov[] Changed in v2: - Change name of SKIPPED bit to UNKNOWN bit - PAD op does not set the bit any more - Change len_op helper functions to take just len_op, not the whole struct - Change version to 2 in missed spots, and use defines - Add 16 unused bytes to cmd_entry.req, in case additional SAM cmd parameters need to be included - Add iov_dif_cnt field to specify buffers used for DIF info in iov[] - Rearrange fields to naturally align cdb_off - Handle if userspace sets UNKNOWN_OP by indicating failure of the cmd - Wrap some overly long UPDATE_HEAD lines (Add missing req.iov_bidi_cnt + req.iov_dif_cnt zeroing - Ilias) Signed-off-by: Andy Grover <agrover@redhat.com> Reviewed-by: Ilias Tsitsimpis <iliastsi@arrikto.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2015-04-15 03:30:04 +03:00
if (tcmu_hdr_get_op(entry->hdr.len_op) == TCMU_OP_PAD) {
UPDATE_HEAD(udev->cmdr_last_cleaned,
tcmu_hdr_get_len(entry->hdr.len_op),
udev->cmdr_size);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
continue;
}
target: Version 2 of TCMU ABI The initial version of TCMU (in 3.18) does not properly handle bidirectional SCSI commands -- those with both an in and out buffer. In looking to fix this it also became clear that TCMU's support for adding new types of entries (opcodes) to the command ring was broken. We need to fix this now, so that future issues can be handled properly by adding new opcodes. We make the most of this ABI break by enabling bidi cmd handling within TCMP_OP_CMD opcode. Add an iov_bidi_cnt field to tcmu_cmd_entry.req. This enables TCMU to describe bidi commands, but further kernel work is needed for full bidi support. Enlarge tcmu_cmd_entry_hdr by 32 bits by pulling in cmd_id and __pad1. Turn __pad1 into two 8 bit flags fields, for kernel-set and userspace-set flags, "kflags" and "uflags" respectively. Update version fields so userspace can tell the interface is changed. Update tcmu-design.txt with details of how new stuff works: - Specify an additional requirement for userspace to set UNKNOWN_OP (bit 0) in hdr.uflags for unknown/unhandled opcodes. - Define how Data-In and Data-Out fields are described in req.iov[] Changed in v2: - Change name of SKIPPED bit to UNKNOWN bit - PAD op does not set the bit any more - Change len_op helper functions to take just len_op, not the whole struct - Change version to 2 in missed spots, and use defines - Add 16 unused bytes to cmd_entry.req, in case additional SAM cmd parameters need to be included - Add iov_dif_cnt field to specify buffers used for DIF info in iov[] - Rearrange fields to naturally align cdb_off - Handle if userspace sets UNKNOWN_OP by indicating failure of the cmd - Wrap some overly long UPDATE_HEAD lines (Add missing req.iov_bidi_cnt + req.iov_dif_cnt zeroing - Ilias) Signed-off-by: Andy Grover <agrover@redhat.com> Reviewed-by: Ilias Tsitsimpis <iliastsi@arrikto.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2015-04-15 03:30:04 +03:00
WARN_ON(tcmu_hdr_get_op(entry->hdr.len_op) != TCMU_OP_CMD);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
spin_lock(&udev->commands_lock);
cmd = idr_remove(&udev->commands, entry->hdr.cmd_id);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
spin_unlock(&udev->commands_lock);
if (!cmd) {
pr_err("cmd_id not found, ring is broken\n");
set_bit(TCMU_DEV_BIT_BROKEN, &udev->flags);
break;
}
tcmu_handle_completion(cmd, entry);
target: Version 2 of TCMU ABI The initial version of TCMU (in 3.18) does not properly handle bidirectional SCSI commands -- those with both an in and out buffer. In looking to fix this it also became clear that TCMU's support for adding new types of entries (opcodes) to the command ring was broken. We need to fix this now, so that future issues can be handled properly by adding new opcodes. We make the most of this ABI break by enabling bidi cmd handling within TCMP_OP_CMD opcode. Add an iov_bidi_cnt field to tcmu_cmd_entry.req. This enables TCMU to describe bidi commands, but further kernel work is needed for full bidi support. Enlarge tcmu_cmd_entry_hdr by 32 bits by pulling in cmd_id and __pad1. Turn __pad1 into two 8 bit flags fields, for kernel-set and userspace-set flags, "kflags" and "uflags" respectively. Update version fields so userspace can tell the interface is changed. Update tcmu-design.txt with details of how new stuff works: - Specify an additional requirement for userspace to set UNKNOWN_OP (bit 0) in hdr.uflags for unknown/unhandled opcodes. - Define how Data-In and Data-Out fields are described in req.iov[] Changed in v2: - Change name of SKIPPED bit to UNKNOWN bit - PAD op does not set the bit any more - Change len_op helper functions to take just len_op, not the whole struct - Change version to 2 in missed spots, and use defines - Add 16 unused bytes to cmd_entry.req, in case additional SAM cmd parameters need to be included - Add iov_dif_cnt field to specify buffers used for DIF info in iov[] - Rearrange fields to naturally align cdb_off - Handle if userspace sets UNKNOWN_OP by indicating failure of the cmd - Wrap some overly long UPDATE_HEAD lines (Add missing req.iov_bidi_cnt + req.iov_dif_cnt zeroing - Ilias) Signed-off-by: Andy Grover <agrover@redhat.com> Reviewed-by: Ilias Tsitsimpis <iliastsi@arrikto.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2015-04-15 03:30:04 +03:00
UPDATE_HEAD(udev->cmdr_last_cleaned,
tcmu_hdr_get_len(entry->hdr.len_op),
udev->cmdr_size);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
handled++;
}
if (mb->cmd_tail == mb->cmd_head)
del_timer(&udev->timeout); /* no more pending cmds */
wake_up(&udev->wait_cmdr);
return handled;
}
static int tcmu_check_expired_cmd(int id, void *p, void *data)
{
struct tcmu_cmd *cmd = p;
if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags))
return 0;
if (!time_after(jiffies, cmd->deadline))
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
return 0;
set_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags);
target_complete_cmd(cmd->se_cmd, SAM_STAT_CHECK_CONDITION);
cmd->se_cmd = NULL;
return 0;
}
static void tcmu_device_timedout(unsigned long data)
{
struct tcmu_dev *udev = (struct tcmu_dev *)data;
unsigned long flags;
spin_lock_irqsave(&udev->commands_lock, flags);
idr_for_each(&udev->commands, tcmu_check_expired_cmd, NULL);
spin_unlock_irqrestore(&udev->commands_lock, flags);
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
/* Try to wake up the ummap thread */
wake_up(&unmap_wait);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
/*
* We don't need to wakeup threads on wait_cmdr since they have their
* own timeout.
*/
}
static int tcmu_attach_hba(struct se_hba *hba, u32 host_id)
{
struct tcmu_hba *tcmu_hba;
tcmu_hba = kzalloc(sizeof(struct tcmu_hba), GFP_KERNEL);
if (!tcmu_hba)
return -ENOMEM;
tcmu_hba->host_id = host_id;
hba->hba_ptr = tcmu_hba;
return 0;
}
static void tcmu_detach_hba(struct se_hba *hba)
{
kfree(hba->hba_ptr);
hba->hba_ptr = NULL;
}
static struct se_device *tcmu_alloc_device(struct se_hba *hba, const char *name)
{
struct tcmu_dev *udev;
udev = kzalloc(sizeof(struct tcmu_dev), GFP_KERNEL);
if (!udev)
return NULL;
kref_init(&udev->kref);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
udev->name = kstrdup(name, GFP_KERNEL);
if (!udev->name) {
kfree(udev);
return NULL;
}
udev->hba = hba;
udev->cmd_time_out = TCMU_TIME_OUT;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
init_waitqueue_head(&udev->wait_cmdr);
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
mutex_init(&udev->cmdr_lock);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
idr_init(&udev->commands);
spin_lock_init(&udev->commands_lock);
setup_timer(&udev->timeout, tcmu_device_timedout,
(unsigned long)udev);
return &udev->se_dev;
}
static int tcmu_irqcontrol(struct uio_info *info, s32 irq_on)
{
struct tcmu_dev *tcmu_dev = container_of(info, struct tcmu_dev, uio_info);
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
mutex_lock(&tcmu_dev->cmdr_lock);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
tcmu_handle_completions(tcmu_dev);
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
mutex_unlock(&tcmu_dev->cmdr_lock);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
return 0;
}
/*
* mmap code from uio.c. Copied here because we want to hook mmap()
* and this stuff must come along.
*/
static int tcmu_find_mem_index(struct vm_area_struct *vma)
{
struct tcmu_dev *udev = vma->vm_private_data;
struct uio_info *info = &udev->uio_info;
if (vma->vm_pgoff < MAX_UIO_MAPS) {
if (info->mem[vma->vm_pgoff].size == 0)
return -1;
return (int)vma->vm_pgoff;
}
return -1;
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
static struct page *tcmu_try_get_block_page(struct tcmu_dev *udev, uint32_t dbi)
{
struct page *page;
int ret;
mutex_lock(&udev->cmdr_lock);
page = tcmu_get_block_page(udev, dbi);
if (likely(page)) {
mutex_unlock(&udev->cmdr_lock);
return page;
}
/*
* Normally it shouldn't be here:
* Only when the userspace has touched the blocks which
* are out of the tcmu_cmd's data iov[], and will return
* one zeroed page.
*/
pr_warn("Block(%u) out of cmd's iov[] has been touched!\n", dbi);
pr_warn("Mostly it will be a bug of userspace, please have a check!\n");
if (dbi >= udev->dbi_thresh) {
/* Extern the udev->dbi_thresh to dbi + 1 */
udev->dbi_thresh = dbi + 1;
udev->dbi_max = dbi;
}
page = radix_tree_lookup(&udev->data_blocks, dbi);
if (!page) {
page = alloc_page(GFP_KERNEL | __GFP_ZERO);
if (!page) {
mutex_unlock(&udev->cmdr_lock);
return NULL;
}
ret = radix_tree_insert(&udev->data_blocks, dbi, page);
if (ret) {
mutex_unlock(&udev->cmdr_lock);
__free_page(page);
return NULL;
}
/*
* Since this case is rare in page fault routine, here we
* will allow the global_db_count >= TCMU_GLOBAL_MAX_BLOCKS
* to reduce possible page fault call trace.
*/
atomic_inc(&global_db_count);
}
mutex_unlock(&udev->cmdr_lock);
return page;
}
static int tcmu_vma_fault(struct vm_fault *vmf)
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
{
struct tcmu_dev *udev = vmf->vma->vm_private_data;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
struct uio_info *info = &udev->uio_info;
struct page *page;
unsigned long offset;
void *addr;
int mi = tcmu_find_mem_index(vmf->vma);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
if (mi < 0)
return VM_FAULT_SIGBUS;
/*
* We need to subtract mi because userspace uses offset = N*PAGE_SIZE
* to use mem[N].
*/
offset = (vmf->pgoff - mi) << PAGE_SHIFT;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
if (offset < udev->data_off) {
/* For the vmalloc()ed cmd area pages */
addr = (void *)(unsigned long)info->mem[mi].addr + offset;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
page = vmalloc_to_page(addr);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
} else {
uint32_t dbi;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
/* For the dynamically growing data area pages */
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
dbi = (offset - udev->data_off) / DATA_BLOCK_SIZE;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
page = tcmu_try_get_block_page(udev, dbi);
if (!page)
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
return VM_FAULT_NOPAGE;
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
get_page(page);
vmf->page = page;
return 0;
}
static const struct vm_operations_struct tcmu_vm_ops = {
.fault = tcmu_vma_fault,
};
static int tcmu_mmap(struct uio_info *info, struct vm_area_struct *vma)
{
struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
vma->vm_ops = &tcmu_vm_ops;
vma->vm_private_data = udev;
/* Ensure the mmap is exactly the right size */
if (vma_pages(vma) != (TCMU_RING_SIZE >> PAGE_SHIFT))
return -EINVAL;
return 0;
}
static int tcmu_open(struct uio_info *info, struct inode *inode)
{
struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
/* O_EXCL not supported for char devs, so fake it? */
if (test_and_set_bit(TCMU_DEV_BIT_OPEN, &udev->flags))
return -EBUSY;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
udev->inode = inode;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
pr_debug("open\n");
return 0;
}
static void tcmu_dev_call_rcu(struct rcu_head *p)
{
struct se_device *dev = container_of(p, struct se_device, rcu_head);
struct tcmu_dev *udev = TCMU_DEV(dev);
kfree(udev->uio_info.name);
kfree(udev->name);
kfree(udev);
}
static void tcmu_dev_kref_release(struct kref *kref)
{
struct tcmu_dev *udev = container_of(kref, struct tcmu_dev, kref);
struct se_device *dev = &udev->se_dev;
call_rcu(&dev->rcu_head, tcmu_dev_call_rcu);
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
static int tcmu_release(struct uio_info *info, struct inode *inode)
{
struct tcmu_dev *udev = container_of(info, struct tcmu_dev, uio_info);
clear_bit(TCMU_DEV_BIT_OPEN, &udev->flags);
pr_debug("close\n");
/* release ref from configure */
kref_put(&udev->kref, tcmu_dev_kref_release);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
return 0;
}
static int tcmu_netlink_event(enum tcmu_genl_cmd cmd, const char *name, int minor)
{
struct sk_buff *skb;
void *msg_header;
int ret = -ENOMEM;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
skb = genlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
if (!skb)
return ret;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
msg_header = genlmsg_put(skb, 0, 0, &tcmu_genl_family, 0, cmd);
if (!msg_header)
goto free_skb;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
ret = nla_put_string(skb, TCMU_ATTR_DEVICE, name);
if (ret < 0)
goto free_skb;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
ret = nla_put_u32(skb, TCMU_ATTR_MINOR, minor);
if (ret < 0)
goto free_skb;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
netlink: make nlmsg_end() and genlmsg_end() void Contrary to common expectations for an "int" return, these functions return only a positive value -- if used correctly they cannot even return 0 because the message header will necessarily be in the skb. This makes the very common pattern of if (genlmsg_end(...) < 0) { ... } be a whole bunch of dead code. Many places also simply do return nlmsg_end(...); and the caller is expected to deal with it. This also commonly (at least for me) causes errors, because it is very common to write if (my_function(...)) /* error condition */ and if my_function() does "return nlmsg_end()" this is of course wrong. Additionally, there's not a single place in the kernel that actually needs the message length returned, and if anyone needs it later then it'll be very easy to just use skb->len there. Remove this, and make the functions void. This removes a bunch of dead code as described above. The patch adds lines because I did - return nlmsg_end(...); + nlmsg_end(...); + return 0; I could have preserved all the function's return values by returning skb->len, but instead I've audited all the places calling the affected functions and found that none cared. A few places actually compared the return value with <= 0 in dump functionality, but that could just be changed to < 0 with no change in behaviour, so I opted for the more efficient version. One instance of the error I've made numerous times now is also present in net/phonet/pn_netlink.c in the route_dumpit() function - it didn't check for <0 or <=0 and thus broke out of the loop every single time. I've preserved this since it will (I think) have caused the messages to userspace to be formatted differently with just a single message for every SKB returned to userspace. It's possible that this isn't needed for the tools that actually use this, but I don't even know what they are so couldn't test that changing this behaviour would be acceptable. Signed-off-by: Johannes Berg <johannes.berg@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2015-01-17 00:09:00 +03:00
genlmsg_end(skb, msg_header);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
ret = genlmsg_multicast_allns(&tcmu_genl_family, skb, 0,
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
TCMU_MCGRP_CONFIG, GFP_KERNEL);
/* We don't care if no one is listening */
if (ret == -ESRCH)
ret = 0;
return ret;
free_skb:
nlmsg_free(skb);
return ret;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
}
static int tcmu_configure_device(struct se_device *dev)
{
struct tcmu_dev *udev = TCMU_DEV(dev);
struct tcmu_hba *hba = udev->hba->hba_ptr;
struct uio_info *info;
struct tcmu_mailbox *mb;
size_t size;
size_t used;
int ret = 0;
char *str;
info = &udev->uio_info;
size = snprintf(NULL, 0, "tcm-user/%u/%s/%s", hba->host_id, udev->name,
udev->dev_config);
size += 1; /* for \0 */
str = kmalloc(size, GFP_KERNEL);
if (!str)
return -ENOMEM;
used = snprintf(str, size, "tcm-user/%u/%s", hba->host_id, udev->name);
if (udev->dev_config[0])
snprintf(str + used, size - used, "/%s", udev->dev_config);
info->name = str;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
udev->mb_addr = vzalloc(CMDR_SIZE);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
if (!udev->mb_addr) {
ret = -ENOMEM;
goto err_vzalloc;
}
/* mailbox fits in first part of CMDR space */
udev->cmdr_size = CMDR_SIZE - CMDR_OFF;
udev->data_off = CMDR_SIZE;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
udev->data_size = DATA_SIZE;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
udev->dbi_thresh = 0; /* Default in Idle state */
udev->waiting_global = false;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
/* Initialise the mailbox of the ring buffer */
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
mb = udev->mb_addr;
target: Version 2 of TCMU ABI The initial version of TCMU (in 3.18) does not properly handle bidirectional SCSI commands -- those with both an in and out buffer. In looking to fix this it also became clear that TCMU's support for adding new types of entries (opcodes) to the command ring was broken. We need to fix this now, so that future issues can be handled properly by adding new opcodes. We make the most of this ABI break by enabling bidi cmd handling within TCMP_OP_CMD opcode. Add an iov_bidi_cnt field to tcmu_cmd_entry.req. This enables TCMU to describe bidi commands, but further kernel work is needed for full bidi support. Enlarge tcmu_cmd_entry_hdr by 32 bits by pulling in cmd_id and __pad1. Turn __pad1 into two 8 bit flags fields, for kernel-set and userspace-set flags, "kflags" and "uflags" respectively. Update version fields so userspace can tell the interface is changed. Update tcmu-design.txt with details of how new stuff works: - Specify an additional requirement for userspace to set UNKNOWN_OP (bit 0) in hdr.uflags for unknown/unhandled opcodes. - Define how Data-In and Data-Out fields are described in req.iov[] Changed in v2: - Change name of SKIPPED bit to UNKNOWN bit - PAD op does not set the bit any more - Change len_op helper functions to take just len_op, not the whole struct - Change version to 2 in missed spots, and use defines - Add 16 unused bytes to cmd_entry.req, in case additional SAM cmd parameters need to be included - Add iov_dif_cnt field to specify buffers used for DIF info in iov[] - Rearrange fields to naturally align cdb_off - Handle if userspace sets UNKNOWN_OP by indicating failure of the cmd - Wrap some overly long UPDATE_HEAD lines (Add missing req.iov_bidi_cnt + req.iov_dif_cnt zeroing - Ilias) Signed-off-by: Andy Grover <agrover@redhat.com> Reviewed-by: Ilias Tsitsimpis <iliastsi@arrikto.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2015-04-15 03:30:04 +03:00
mb->version = TCMU_MAILBOX_VERSION;
mb->flags = TCMU_MAILBOX_FLAG_CAP_OOOC;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
mb->cmdr_off = CMDR_OFF;
mb->cmdr_size = udev->cmdr_size;
WARN_ON(!PAGE_ALIGNED(udev->data_off));
WARN_ON(udev->data_size % PAGE_SIZE);
WARN_ON(udev->data_size % DATA_BLOCK_SIZE);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
INIT_RADIX_TREE(&udev->data_blocks, GFP_KERNEL);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
info->version = __stringify(TCMU_MAILBOX_VERSION);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
info->mem[0].name = "tcm-user command & data buffer";
info->mem[0].addr = (phys_addr_t)(uintptr_t)udev->mb_addr;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
info->mem[0].size = TCMU_RING_SIZE;
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
info->mem[0].memtype = UIO_MEM_NONE;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
info->irqcontrol = tcmu_irqcontrol;
info->irq = UIO_IRQ_CUSTOM;
info->mmap = tcmu_mmap;
info->open = tcmu_open;
info->release = tcmu_release;
ret = uio_register_device(tcmu_root_device, info);
if (ret)
goto err_register;
/* User can set hw_block_size before enable the device */
if (dev->dev_attrib.hw_block_size == 0)
dev->dev_attrib.hw_block_size = 512;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
/* Other attributes can be configured in userspace */
if (!dev->dev_attrib.hw_max_sectors)
dev->dev_attrib.hw_max_sectors = 128;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
dev->dev_attrib.hw_queue_depth = 128;
/*
* Get a ref incase userspace does a close on the uio device before
* LIO has initiated tcmu_free_device.
*/
kref_get(&udev->kref);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
ret = tcmu_netlink_event(TCMU_CMD_ADDED_DEVICE, udev->uio_info.name,
udev->uio_info.uio_dev->minor);
if (ret)
goto err_netlink;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
mutex_lock(&root_udev_mutex);
list_add(&udev->node, &root_udev);
mutex_unlock(&root_udev_mutex);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
return 0;
err_netlink:
kref_put(&udev->kref, tcmu_dev_kref_release);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
uio_unregister_device(&udev->uio_info);
err_register:
vfree(udev->mb_addr);
err_vzalloc:
kfree(info->name);
info->name = NULL;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
return ret;
}
static int tcmu_check_and_free_pending_cmd(struct tcmu_cmd *cmd)
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
{
if (test_bit(TCMU_CMD_BIT_EXPIRED, &cmd->flags)) {
kmem_cache_free(tcmu_cmd_cache, cmd);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
return 0;
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
return -EINVAL;
}
static bool tcmu_dev_configured(struct tcmu_dev *udev)
{
return udev->uio_info.uio_dev ? true : false;
}
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
static void tcmu_blocks_release(struct tcmu_dev *udev)
{
int i;
struct page *page;
/* Try to release all block pages */
mutex_lock(&udev->cmdr_lock);
for (i = 0; i <= udev->dbi_max; i++) {
page = radix_tree_delete(&udev->data_blocks, i);
if (page) {
__free_page(page);
atomic_dec(&global_db_count);
}
}
mutex_unlock(&udev->cmdr_lock);
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
static void tcmu_free_device(struct se_device *dev)
{
struct tcmu_dev *udev = TCMU_DEV(dev);
struct tcmu_cmd *cmd;
bool all_expired = true;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
int i;
del_timer_sync(&udev->timeout);
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
mutex_lock(&root_udev_mutex);
list_del(&udev->node);
mutex_unlock(&root_udev_mutex);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
vfree(udev->mb_addr);
/* Upper layer should drain all requests before calling this */
spin_lock_irq(&udev->commands_lock);
idr_for_each_entry(&udev->commands, cmd, i) {
if (tcmu_check_and_free_pending_cmd(cmd) != 0)
all_expired = false;
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
idr_destroy(&udev->commands);
spin_unlock_irq(&udev->commands_lock);
WARN_ON(!all_expired);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
tcmu_blocks_release(udev);
tcmu: Add dynamic growing data area feature support Currently for the TCMU, the ring buffer size is fixed to 64K cmd area + 1M data area, and this will be bottlenecks for high iops. The struct tcmu_cmd_entry {} size is fixed about 112 bytes with iovec[N] & N <= 4, and the size of struct iovec is about 16 bytes. If N == 0, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : 0, no data area is need. If 0 < N <=4, the ratio will be sizeof(cmd entry) : sizeof(datas) == 112Bytes : (N * 4096)Bytes = 28 : (N * 1024), so the max will be 28 : 1024. If N > 4, the sizeof(cmd entry) will be [(N - 4) *16 + 112] bytes, and its corresponding data size will be [N * 4096], so the ratio of sizeof(cmd entry) : sizeof(datas) == [(N - 4) * 16 + 112)Bytes : (N * 4096)Bytes == 4/1024 - 12/(N * 1024), so the max is about 4 : 1024. When N is bigger, the ratio will be smaller. As the initial patch, we will set the cmd area size to 2M, and the cmd area size to 32M. The TCMU will dynamically grows the data area from 0 to max 32M size as needed. The cmd area memory will be allocated through vmalloc(), and the data area's blocks will be allocated individually later when needed. The allocated data area block memory will be managed via radix tree. For now the bitmap still be the most efficient way to search and manage the block index, this could be update later. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:05 +03:00
if (tcmu_dev_configured(udev)) {
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
tcmu_netlink_event(TCMU_CMD_REMOVED_DEVICE, udev->uio_info.name,
udev->uio_info.uio_dev->minor);
uio_unregister_device(&udev->uio_info);
}
/* release ref from init */
kref_put(&udev->kref, tcmu_dev_kref_release);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
}
enum {
Opt_dev_config, Opt_dev_size, Opt_hw_block_size, Opt_hw_max_sectors,
Opt_err,
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
};
static match_table_t tokens = {
{Opt_dev_config, "dev_config=%s"},
{Opt_dev_size, "dev_size=%u"},
{Opt_hw_block_size, "hw_block_size=%u"},
{Opt_hw_max_sectors, "hw_max_sectors=%u"},
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
{Opt_err, NULL}
};
static int tcmu_set_dev_attrib(substring_t *arg, u32 *dev_attrib)
{
unsigned long tmp_ul;
char *arg_p;
int ret;
arg_p = match_strdup(arg);
if (!arg_p)
return -ENOMEM;
ret = kstrtoul(arg_p, 0, &tmp_ul);
kfree(arg_p);
if (ret < 0) {
pr_err("kstrtoul() failed for dev attrib\n");
return ret;
}
if (!tmp_ul) {
pr_err("dev attrib must be nonzero\n");
return -EINVAL;
}
*dev_attrib = tmp_ul;
return 0;
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
static ssize_t tcmu_set_configfs_dev_params(struct se_device *dev,
const char *page, ssize_t count)
{
struct tcmu_dev *udev = TCMU_DEV(dev);
char *orig, *ptr, *opts, *arg_p;
substring_t args[MAX_OPT_ARGS];
int ret = 0, token;
opts = kstrdup(page, GFP_KERNEL);
if (!opts)
return -ENOMEM;
orig = opts;
while ((ptr = strsep(&opts, ",\n")) != NULL) {
if (!*ptr)
continue;
token = match_token(ptr, tokens, args);
switch (token) {
case Opt_dev_config:
if (match_strlcpy(udev->dev_config, &args[0],
TCMU_CONFIG_LEN) == 0) {
ret = -EINVAL;
break;
}
pr_debug("TCMU: Referencing Path: %s\n", udev->dev_config);
break;
case Opt_dev_size:
arg_p = match_strdup(&args[0]);
if (!arg_p) {
ret = -ENOMEM;
break;
}
ret = kstrtoul(arg_p, 0, (unsigned long *) &udev->dev_size);
kfree(arg_p);
if (ret < 0)
pr_err("kstrtoul() failed for dev_size=\n");
break;
case Opt_hw_block_size:
ret = tcmu_set_dev_attrib(&args[0],
&(dev->dev_attrib.hw_block_size));
break;
case Opt_hw_max_sectors:
ret = tcmu_set_dev_attrib(&args[0],
&(dev->dev_attrib.hw_max_sectors));
break;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
default:
break;
}
if (ret)
break;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
}
kfree(orig);
return (!ret) ? count : ret;
}
static ssize_t tcmu_show_configfs_dev_params(struct se_device *dev, char *b)
{
struct tcmu_dev *udev = TCMU_DEV(dev);
ssize_t bl = 0;
bl = sprintf(b + bl, "Config: %s ",
udev->dev_config[0] ? udev->dev_config : "NULL");
bl += sprintf(b + bl, "Size: %zu\n", udev->dev_size);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
return bl;
}
static sector_t tcmu_get_blocks(struct se_device *dev)
{
struct tcmu_dev *udev = TCMU_DEV(dev);
return div_u64(udev->dev_size - dev->dev_attrib.block_size,
dev->dev_attrib.block_size);
}
static sense_reason_t
tcmu_parse_cdb(struct se_cmd *cmd)
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
{
return passthrough_parse_cdb(cmd, tcmu_queue_cmd);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
}
static ssize_t tcmu_cmd_time_out_show(struct config_item *item, char *page)
{
struct se_dev_attrib *da = container_of(to_config_group(item),
struct se_dev_attrib, da_group);
struct tcmu_dev *udev = container_of(da->da_dev,
struct tcmu_dev, se_dev);
return snprintf(page, PAGE_SIZE, "%lu\n", udev->cmd_time_out / MSEC_PER_SEC);
}
static ssize_t tcmu_cmd_time_out_store(struct config_item *item, const char *page,
size_t count)
{
struct se_dev_attrib *da = container_of(to_config_group(item),
struct se_dev_attrib, da_group);
struct tcmu_dev *udev = container_of(da->da_dev,
struct tcmu_dev, se_dev);
u32 val;
int ret;
if (da->da_dev->export_count) {
pr_err("Unable to set tcmu cmd_time_out while exports exist\n");
return -EINVAL;
}
ret = kstrtou32(page, 0, &val);
if (ret < 0)
return ret;
udev->cmd_time_out = val * MSEC_PER_SEC;
return count;
}
CONFIGFS_ATTR(tcmu_, cmd_time_out);
static struct configfs_attribute **tcmu_attrs;
static struct target_backend_ops tcmu_ops = {
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
.name = "user",
.owner = THIS_MODULE,
.transport_flags = TRANSPORT_FLAG_PASSTHROUGH,
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
.attach_hba = tcmu_attach_hba,
.detach_hba = tcmu_detach_hba,
.alloc_device = tcmu_alloc_device,
.configure_device = tcmu_configure_device,
.free_device = tcmu_free_device,
.parse_cdb = tcmu_parse_cdb,
.set_configfs_dev_params = tcmu_set_configfs_dev_params,
.show_configfs_dev_params = tcmu_show_configfs_dev_params,
.get_device_type = sbc_get_device_type,
.get_blocks = tcmu_get_blocks,
.tb_dev_attrib_attrs = NULL,
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
};
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
static int unmap_thread_fn(void *data)
{
struct tcmu_dev *udev;
loff_t off;
uint32_t start, end, block;
struct page *page;
int i;
while (1) {
DEFINE_WAIT(__wait);
prepare_to_wait(&unmap_wait, &__wait, TASK_INTERRUPTIBLE);
schedule();
finish_wait(&unmap_wait, &__wait);
if (kthread_should_stop())
break;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
mutex_lock(&root_udev_mutex);
list_for_each_entry(udev, &root_udev, node) {
mutex_lock(&udev->cmdr_lock);
/* Try to complete the finished commands first */
tcmu_handle_completions(udev);
/* Skip the udevs waiting the global pool or in idle */
if (udev->waiting_global || !udev->dbi_thresh) {
mutex_unlock(&udev->cmdr_lock);
continue;
}
end = udev->dbi_max + 1;
block = find_last_bit(udev->data_bitmap, end);
if (block == udev->dbi_max) {
/*
* The last bit is dbi_max, so there is
* no need to shrink any blocks.
*/
mutex_unlock(&udev->cmdr_lock);
continue;
} else if (block == end) {
/* The current udev will goto idle state */
udev->dbi_thresh = start = 0;
udev->dbi_max = 0;
} else {
udev->dbi_thresh = start = block + 1;
udev->dbi_max = block;
}
/* Here will truncate the data area from off */
off = udev->data_off + start * DATA_BLOCK_SIZE;
unmap_mapping_range(udev->inode->i_mapping, off, 0, 1);
/* Release the block pages */
for (i = start; i < end; i++) {
page = radix_tree_delete(&udev->data_blocks, i);
if (page) {
__free_page(page);
atomic_dec(&global_db_count);
}
}
mutex_unlock(&udev->cmdr_lock);
}
/*
* Try to wake up the udevs who are waiting
* for the global data pool.
*/
list_for_each_entry(udev, &root_udev, node) {
if (udev->waiting_global)
wake_up(&udev->wait_cmdr);
}
mutex_unlock(&root_udev_mutex);
}
return 0;
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
static int __init tcmu_module_init(void)
{
int ret, i, len = 0;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
BUILD_BUG_ON((sizeof(struct tcmu_cmd_entry) % TCMU_OP_ALIGN_SIZE) != 0);
tcmu_cmd_cache = kmem_cache_create("tcmu_cmd_cache",
sizeof(struct tcmu_cmd),
__alignof__(struct tcmu_cmd),
0, NULL);
if (!tcmu_cmd_cache)
return -ENOMEM;
tcmu_root_device = root_device_register("tcm_user");
if (IS_ERR(tcmu_root_device)) {
ret = PTR_ERR(tcmu_root_device);
goto out_free_cache;
}
ret = genl_register_family(&tcmu_genl_family);
if (ret < 0) {
goto out_unreg_device;
}
for (i = 0; passthrough_attrib_attrs[i] != NULL; i++) {
len += sizeof(struct configfs_attribute *);
}
len += sizeof(struct configfs_attribute *) * 2;
tcmu_attrs = kzalloc(len, GFP_KERNEL);
if (!tcmu_attrs) {
ret = -ENOMEM;
goto out_unreg_genl;
}
for (i = 0; passthrough_attrib_attrs[i] != NULL; i++) {
tcmu_attrs[i] = passthrough_attrib_attrs[i];
}
tcmu_attrs[i] = &tcmu_attr_cmd_time_out;
tcmu_ops.tb_dev_attrib_attrs = tcmu_attrs;
ret = transport_backend_register(&tcmu_ops);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
if (ret)
goto out_attrs;
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
init_waitqueue_head(&unmap_wait);
unmap_thread = kthread_run(unmap_thread_fn, NULL, "tcmu_unmap");
if (IS_ERR(unmap_thread)) {
ret = PTR_ERR(unmap_thread);
goto out_unreg_transport;
}
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
return 0;
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
out_unreg_transport:
target_backend_unregister(&tcmu_ops);
out_attrs:
kfree(tcmu_attrs);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
out_unreg_genl:
genl_unregister_family(&tcmu_genl_family);
out_unreg_device:
root_device_unregister(tcmu_root_device);
out_free_cache:
kmem_cache_destroy(tcmu_cmd_cache);
return ret;
}
static void __exit tcmu_module_exit(void)
{
tcmu: Add global data block pool support For each target there will be one ring, when the target number grows larger and larger, it could eventually runs out of the system memories. In this patch for each target ring, currently for the cmd area the size will be fixed to 8MB and for the data area the size will grow from 0 to max 256K * PAGE_SIZE(1G for 4K page size). For all the targets' data areas, they will get empty blocks from the "global data block pool", which has limited to 512K * PAGE_SIZE(2G for 4K page size) for now. When the "global data block pool" has been used up, then any target could wake up the unmap thread routine to shrink other targets' data area memories. And the unmap thread routine will always try to truncate the ring vma from the last using block offset. When user space has touched the data blocks out of tcmu_cmd iov[], the tcmu_page_fault() will try to return one zeroed blocks. Here we move the timeout's tcmu_handle_completions() into unmap thread routine, that's to say when the timeout fired, it will only do the tcmu_check_expired_cmd() and then wake up the unmap thread to do the completions() and then try to shrink its idle memories. Then the cmdr_lock could be a mutex and could simplify this patch because the unmap_mapping_range() or zap_* may go to sleep. Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com> Signed-off-by: Jianfei Hu <hujianfei@cmss.chinamobile.com> Acked-by: Mike Christie <mchristi@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2017-05-02 06:38:06 +03:00
kthread_stop(unmap_thread);
target_backend_unregister(&tcmu_ops);
kfree(tcmu_attrs);
target: Add a user-passthrough backstore Add a LIO storage engine that presents commands to userspace for execution. This would allow more complex backstores to be implemented out-of-kernel, and also make experimentation a-la FUSE (but at the SCSI level -- "SUSE"?) possible. It uses a mmap()able UIO device per LUN to share a command ring and data area. The commands are raw SCSI CDBs and iovs for in/out data. The command ring is also reused for returning scsi command status and optional sense data. This implementation is based on Shaohua Li's earlier version but heavily modified. Differences include: * Shared memory allocated by kernel, not locked-down user pages * Single ring for command request and response * Offsets instead of embedded pointers * Generic SCSI CDB passthrough instead of per-cmd specialization in ring format. * Uses UIO device instead of anon_file passed in mailbox. * Optional in-kernel handling of some commands. The main reason for these differences is to permit greater resiliency if the user process dies or hangs. Things not yet implemented (on purpose): * Zero copy. The data area is flexible enough to allow page flipping or backend-allocated pages to be used by fabrics, but it's not clear these are performance wins. Can come later. * Out-of-order command completion by userspace. Possible to add by just allowing userspace to change cmd_id in rsp cmd entries, but currently not supported. * No locks between kernel cmd submission and completion routines. Sounds like it's possible, but this can come later. * Sparse allocation of mmaped area. Current code vmallocs the whole thing. If the mapped area was larger and not fully mapped then the driver would have more freedom to change cmd and data area sizes based on demand. Current code open issues: * The use of idrs may be overkill -- we maybe can replace them with a simple counter to generate cmd_ids, and a hash table to get a cmd_id's associated pointer. * Use of a free-running counter for cmd ring instead of explicit modulo math. This would require power-of-2 cmd ring size. (Add kconfig depends NET - Randy) Signed-off-by: Andy Grover <agrover@redhat.com> Signed-off-by: Nicholas Bellinger <nab@linux-iscsi.org>
2014-10-02 03:07:05 +04:00
genl_unregister_family(&tcmu_genl_family);
root_device_unregister(tcmu_root_device);
kmem_cache_destroy(tcmu_cmd_cache);
}
MODULE_DESCRIPTION("TCM USER subsystem plugin");
MODULE_AUTHOR("Shaohua Li <shli@kernel.org>");
MODULE_AUTHOR("Andy Grover <agrover@redhat.com>");
MODULE_LICENSE("GPL");
module_init(tcmu_module_init);
module_exit(tcmu_module_exit);