linux/net/smc/smc_core.c

2615 lines
68 KiB
C
Raw Normal View History

License cleanup: add SPDX GPL-2.0 license identifier to files with no license Many source files in the tree are missing licensing information, which makes it harder for compliance tools to determine the correct license. By default all files without license information are under the default license of the kernel, which is GPL version 2. Update the files which contain no license information with the 'GPL-2.0' SPDX license identifier. The SPDX identifier is a legally binding shorthand, which can be used instead of the full boiler plate text. This patch is based on work done by Thomas Gleixner and Kate Stewart and Philippe Ombredanne. How this work was done: Patches were generated and checked against linux-4.14-rc6 for a subset of the use cases: - file had no licensing information it it. - file was a */uapi/* one with no licensing information in it, - file was a */uapi/* one with existing licensing information, Further patches will be generated in subsequent months to fix up cases where non-standard license headers were used, and references to license had to be inferred by heuristics based on keywords. The analysis to determine which SPDX License Identifier to be applied to a file was done in a spreadsheet of side by side results from of the output of two independent scanners (ScanCode & Windriver) producing SPDX tag:value files created by Philippe Ombredanne. Philippe prepared the base worksheet, and did an initial spot review of a few 1000 files. The 4.13 kernel was the starting point of the analysis with 60,537 files assessed. Kate Stewart did a file by file comparison of the scanner results in the spreadsheet to determine which SPDX license identifier(s) to be applied to the file. She confirmed any determination that was not immediately clear with lawyers working with the Linux Foundation. Criteria used to select files for SPDX license identifier tagging was: - Files considered eligible had to be source code files. - Make and config files were included as candidates if they contained >5 lines of source - File already had some variant of a license header in it (even if <5 lines). All documentation files were explicitly excluded. The following heuristics were used to determine which SPDX license identifiers to apply. - when both scanners couldn't find any license traces, file was considered to have no license information in it, and the top level COPYING file license applied. For non */uapi/* files that summary was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 11139 and resulted in the first patch in this series. If that file was a */uapi/* path one, it was "GPL-2.0 WITH Linux-syscall-note" otherwise it was "GPL-2.0". Results of that was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 WITH Linux-syscall-note 930 and resulted in the second patch in this series. - if a file had some form of licensing information in it, and was one of the */uapi/* ones, it was denoted with the Linux-syscall-note if any GPL family license was found in the file or had no licensing in it (per prior point). Results summary: SPDX license identifier # files ---------------------------------------------------|------ GPL-2.0 WITH Linux-syscall-note 270 GPL-2.0+ WITH Linux-syscall-note 169 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-2-Clause) 21 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-3-Clause) 17 LGPL-2.1+ WITH Linux-syscall-note 15 GPL-1.0+ WITH Linux-syscall-note 14 ((GPL-2.0+ WITH Linux-syscall-note) OR BSD-3-Clause) 5 LGPL-2.0+ WITH Linux-syscall-note 4 LGPL-2.1 WITH Linux-syscall-note 3 ((GPL-2.0 WITH Linux-syscall-note) OR MIT) 3 ((GPL-2.0 WITH Linux-syscall-note) AND MIT) 1 and that resulted in the third patch in this series. - when the two scanners agreed on the detected license(s), that became the concluded license(s). - when there was disagreement between the two scanners (one detected a license but the other didn't, or they both detected different licenses) a manual inspection of the file occurred. - In most cases a manual inspection of the information in the file resulted in a clear resolution of the license that should apply (and which scanner probably needed to revisit its heuristics). - When it was not immediately clear, the license identifier was confirmed with lawyers working with the Linux Foundation. - If there was any question as to the appropriate license identifier, the file was flagged for further research and to be revisited later in time. In total, over 70 hours of logged manual review was done on the spreadsheet to determine the SPDX license identifiers to apply to the source files by Kate, Philippe, Thomas and, in some cases, confirmation by lawyers working with the Linux Foundation. Kate also obtained a third independent scan of the 4.13 code base from FOSSology, and compared selected files where the other two scanners disagreed against that SPDX file, to see if there was new insights. The Windriver scanner is based on an older version of FOSSology in part, so they are related. Thomas did random spot checks in about 500 files from the spreadsheets for the uapi headers and agreed with SPDX license identifier in the files he inspected. For the non-uapi files Thomas did random spot checks in about 15000 files. In initial set of patches against 4.14-rc6, 3 files were found to have copy/paste license identifier errors, and have been fixed to reflect the correct identifier. Additionally Philippe spent 10 hours this week doing a detailed manual inspection and review of the 12,461 patched files from the initial patch version early this week with: - a full scancode scan run, collecting the matched texts, detected license ids and scores - reviewing anything where there was a license detected (about 500+ files) to ensure that the applied SPDX license was correct - reviewing anything where there was no detection but the patch license was not GPL-2.0 WITH Linux-syscall-note to ensure that the applied SPDX license was correct This produced a worksheet with 20 files needing minor correction. This worksheet was then exported into 3 different .csv files for the different types of files to be modified. These .csv files were then reviewed by Greg. Thomas wrote a script to parse the csv files and add the proper SPDX tag to the file, in the format that the file expected. This script was further refined by Greg based on the output to detect more types of files automatically and to distinguish between header and source .c files (which need different comment types.) Finally Greg ran the script using the .csv files to generate the patches. Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org> Reviewed-by: Philippe Ombredanne <pombredanne@nexb.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2017-11-01 17:07:57 +03:00
// SPDX-License-Identifier: GPL-2.0
/*
* Shared Memory Communications over RDMA (SMC-R) and RoCE
*
* Basic Transport Functions exploiting Infiniband API
*
* Copyright IBM Corp. 2016
*
* Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
*/
#include <linux/socket.h>
#include <linux/if_vlan.h>
#include <linux/random.h>
#include <linux/workqueue.h>
#include <linux/wait.h>
#include <linux/reboot.h>
net/smc: fix sleep bug in smc_pnet_find_roce_resource() Tests showed this BUG: [572555.252867] BUG: sleeping function called from invalid context at kernel/locking/mutex.c:935 [572555.252876] in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 131031, name: smcapp [572555.252879] INFO: lockdep is turned off. [572555.252883] CPU: 1 PID: 131031 Comm: smcapp Tainted: G O 5.7.0-rc3uschi+ #356 [572555.252885] Hardware name: IBM 3906 M03 703 (LPAR) [572555.252887] Call Trace: [572555.252896] [<00000000ac364554>] show_stack+0x94/0xe8 [572555.252901] [<00000000aca1f400>] dump_stack+0xa0/0xe0 [572555.252906] [<00000000ac3c8c10>] ___might_sleep+0x260/0x280 [572555.252910] [<00000000acdc0c98>] __mutex_lock+0x48/0x940 [572555.252912] [<00000000acdc15c2>] mutex_lock_nested+0x32/0x40 [572555.252975] [<000003ff801762d0>] mlx5_lag_get_roce_netdev+0x30/0xc0 [mlx5_core] [572555.252996] [<000003ff801fb3aa>] mlx5_ib_get_netdev+0x3a/0xe0 [mlx5_ib] [572555.253007] [<000003ff80063848>] smc_pnet_find_roce_resource+0x1d8/0x310 [smc] [572555.253011] [<000003ff800602f0>] __smc_connect+0x1f0/0x3e0 [smc] [572555.253015] [<000003ff80060634>] smc_connect+0x154/0x190 [smc] [572555.253022] [<00000000acbed8d4>] __sys_connect+0x94/0xd0 [572555.253025] [<00000000acbef620>] __s390x_sys_socketcall+0x170/0x360 [572555.253028] [<00000000acdc6800>] system_call+0x298/0x2b8 [572555.253030] INFO: lockdep is turned off. Function smc_pnet_find_rdma_dev() might be called from smc_pnet_find_roce_resource(). It holds the smc_ib_devices list spinlock while calling infiniband op get_netdev(). At least for mlx5 the get_netdev operation wants mutex serialization, which conflicts with the smc_ib_devices spinlock. This patch switches the smc_ib_devices spinlock into a mutex to allow sleeping when calling get_netdev(). Fixes: a4cf0443c414 ("smc: introduce SMC as an IB-client") Signed-off-by: Ursula Braun <ubraun@linux.ibm.com> Signed-off-by: Karsten Graul <kgraul@linux.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-07-08 18:05:13 +03:00
#include <linux/mutex.h>
#include <linux/list.h>
#include <linux/smc.h>
#include <net/tcp.h>
#include <net/sock.h>
#include <rdma/ib_verbs.h>
#include <rdma/ib_cache.h>
#include "smc.h"
#include "smc_clc.h"
#include "smc_core.h"
#include "smc_ib.h"
#include "smc_wr.h"
#include "smc_llc.h"
#include "smc_cdc.h"
#include "smc_close.h"
#include "smc_ism.h"
#include "smc_netlink.h"
#include "smc_stats.h"
#include "smc_tracepoint.h"
#define SMC_LGR_NUM_INCR 256
#define SMC_LGR_FREE_DELAY_SERV (600 * HZ)
#define SMC_LGR_FREE_DELAY_CLNT (SMC_LGR_FREE_DELAY_SERV + 10 * HZ)
struct smc_lgr_list smc_lgr_list = { /* established link groups */
.lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
.list = LIST_HEAD_INIT(smc_lgr_list.list),
.num = 0,
};
static atomic_t lgr_cnt = ATOMIC_INIT(0); /* number of existing link groups */
static DECLARE_WAIT_QUEUE_HEAD(lgrs_deleted);
static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
struct smc_buf_desc *buf_desc);
static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft);
static void smc_link_down_work(struct work_struct *work);
/* return head of link group list and its lock for a given link group */
static inline struct list_head *smc_lgr_list_head(struct smc_link_group *lgr,
spinlock_t **lgr_lock)
{
if (lgr->is_smcd) {
*lgr_lock = &lgr->smcd->lgr_lock;
return &lgr->smcd->lgr_list;
}
*lgr_lock = &smc_lgr_list.lock;
return &smc_lgr_list.list;
}
static void smc_ibdev_cnt_inc(struct smc_link *lnk)
{
atomic_inc(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
}
static void smc_ibdev_cnt_dec(struct smc_link *lnk)
{
atomic_dec(&lnk->smcibdev->lnk_cnt_by_port[lnk->ibport - 1]);
}
static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
{
/* client link group creation always follows the server link group
* creation. For client use a somewhat higher removal delay time,
* otherwise there is a risk of out-of-sync link groups.
*/
if (!lgr->freeing) {
mod_delayed_work(system_wq, &lgr->free_work,
(!lgr->is_smcd && lgr->role == SMC_CLNT) ?
SMC_LGR_FREE_DELAY_CLNT :
SMC_LGR_FREE_DELAY_SERV);
}
}
/* Register connection's alert token in our lookup structure.
* To use rbtrees we have to implement our own insert core.
* Requires @conns_lock
* @smc connection to register
* Returns 0 on success, != otherwise.
*/
static void smc_lgr_add_alert_token(struct smc_connection *conn)
{
struct rb_node **link, *parent = NULL;
u32 token = conn->alert_token_local;
link = &conn->lgr->conns_all.rb_node;
while (*link) {
struct smc_connection *cur = rb_entry(*link,
struct smc_connection, alert_node);
parent = *link;
if (cur->alert_token_local > token)
link = &parent->rb_left;
else
link = &parent->rb_right;
}
/* Put the new node there */
rb_link_node(&conn->alert_node, parent, link);
rb_insert_color(&conn->alert_node, &conn->lgr->conns_all);
}
/* assign an SMC-R link to the connection */
static int smcr_lgr_conn_assign_link(struct smc_connection *conn, bool first)
{
enum smc_link_state expected = first ? SMC_LNK_ACTIVATING :
SMC_LNK_ACTIVE;
int i, j;
/* do link balancing */
for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
struct smc_link *lnk = &conn->lgr->lnk[i];
if (lnk->state != expected || lnk->link_is_asym)
continue;
if (conn->lgr->role == SMC_CLNT) {
conn->lnk = lnk; /* temporary, SMC server assigns link*/
break;
}
if (conn->lgr->conns_num % 2) {
for (j = i + 1; j < SMC_LINKS_PER_LGR_MAX; j++) {
struct smc_link *lnk2;
lnk2 = &conn->lgr->lnk[j];
if (lnk2->state == expected &&
!lnk2->link_is_asym) {
conn->lnk = lnk2;
break;
}
}
}
if (!conn->lnk)
conn->lnk = lnk;
break;
}
if (!conn->lnk)
return SMC_CLC_DECL_NOACTLINK;
atomic_inc(&conn->lnk->conn_cnt);
return 0;
}
/* Register connection in link group by assigning an alert token
* registered in a search tree.
* Requires @conns_lock
* Note that '0' is a reserved value and not assigned.
*/
static int smc_lgr_register_conn(struct smc_connection *conn, bool first)
{
struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
static atomic_t nexttoken = ATOMIC_INIT(0);
int rc;
if (!conn->lgr->is_smcd) {
rc = smcr_lgr_conn_assign_link(conn, first);
if (rc) {
conn->lgr = NULL;
return rc;
}
}
/* find a new alert_token_local value not yet used by some connection
* in this link group
*/
sock_hold(&smc->sk); /* sock_put in smc_lgr_unregister_conn() */
while (!conn->alert_token_local) {
conn->alert_token_local = atomic_inc_return(&nexttoken);
if (smc_lgr_find_conn(conn->alert_token_local, conn->lgr))
conn->alert_token_local = 0;
}
smc_lgr_add_alert_token(conn);
conn->lgr->conns_num++;
return 0;
}
/* Unregister connection and reset the alert token of the given connection<
*/
static void __smc_lgr_unregister_conn(struct smc_connection *conn)
{
struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
struct smc_link_group *lgr = conn->lgr;
rb_erase(&conn->alert_node, &lgr->conns_all);
if (conn->lnk)
atomic_dec(&conn->lnk->conn_cnt);
lgr->conns_num--;
conn->alert_token_local = 0;
sock_put(&smc->sk); /* sock_hold in smc_lgr_register_conn() */
}
/* Unregister connection from lgr
*/
static void smc_lgr_unregister_conn(struct smc_connection *conn)
{
struct smc_link_group *lgr = conn->lgr;
if (!smc_conn_lgr_valid(conn))
return;
write_lock_bh(&lgr->conns_lock);
if (conn->alert_token_local) {
__smc_lgr_unregister_conn(conn);
}
write_unlock_bh(&lgr->conns_lock);
}
int smc_nl_get_sys_info(struct sk_buff *skb, struct netlink_callback *cb)
{
struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
char hostname[SMC_MAX_HOSTNAME_LEN + 1];
char smc_seid[SMC_MAX_EID_LEN + 1];
struct nlattr *attrs;
u8 *seid = NULL;
u8 *host = NULL;
void *nlh;
nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
&smc_gen_nl_family, NLM_F_MULTI,
SMC_NETLINK_GET_SYS_INFO);
if (!nlh)
goto errmsg;
if (cb_ctx->pos[0])
goto errout;
attrs = nla_nest_start(skb, SMC_GEN_SYS_INFO);
if (!attrs)
goto errout;
if (nla_put_u8(skb, SMC_NLA_SYS_VER, SMC_V2))
goto errattr;
if (nla_put_u8(skb, SMC_NLA_SYS_REL, SMC_RELEASE))
goto errattr;
if (nla_put_u8(skb, SMC_NLA_SYS_IS_ISM_V2, smc_ism_is_v2_capable()))
goto errattr;
if (nla_put_u8(skb, SMC_NLA_SYS_IS_SMCR_V2, true))
goto errattr;
smc_clc_get_hostname(&host);
if (host) {
memcpy(hostname, host, SMC_MAX_HOSTNAME_LEN);
hostname[SMC_MAX_HOSTNAME_LEN] = 0;
if (nla_put_string(skb, SMC_NLA_SYS_LOCAL_HOST, hostname))
goto errattr;
}
if (smc_ism_is_v2_capable()) {
smc_ism_get_system_eid(&seid);
memcpy(smc_seid, seid, SMC_MAX_EID_LEN);
smc_seid[SMC_MAX_EID_LEN] = 0;
if (nla_put_string(skb, SMC_NLA_SYS_SEID, smc_seid))
goto errattr;
}
nla_nest_end(skb, attrs);
genlmsg_end(skb, nlh);
cb_ctx->pos[0] = 1;
return skb->len;
errattr:
nla_nest_cancel(skb, attrs);
errout:
genlmsg_cancel(skb, nlh);
errmsg:
return skb->len;
}
/* Fill SMC_NLA_LGR_D_V2_COMMON/SMC_NLA_LGR_R_V2_COMMON nested attributes */
static int smc_nl_fill_lgr_v2_common(struct smc_link_group *lgr,
struct sk_buff *skb,
struct netlink_callback *cb,
struct nlattr *v2_attrs)
{
char smc_host[SMC_MAX_HOSTNAME_LEN + 1];
char smc_eid[SMC_MAX_EID_LEN + 1];
if (nla_put_u8(skb, SMC_NLA_LGR_V2_VER, lgr->smc_version))
goto errv2attr;
if (nla_put_u8(skb, SMC_NLA_LGR_V2_REL, lgr->peer_smc_release))
goto errv2attr;
if (nla_put_u8(skb, SMC_NLA_LGR_V2_OS, lgr->peer_os))
goto errv2attr;
memcpy(smc_host, lgr->peer_hostname, SMC_MAX_HOSTNAME_LEN);
smc_host[SMC_MAX_HOSTNAME_LEN] = 0;
if (nla_put_string(skb, SMC_NLA_LGR_V2_PEER_HOST, smc_host))
goto errv2attr;
memcpy(smc_eid, lgr->negotiated_eid, SMC_MAX_EID_LEN);
smc_eid[SMC_MAX_EID_LEN] = 0;
if (nla_put_string(skb, SMC_NLA_LGR_V2_NEG_EID, smc_eid))
goto errv2attr;
nla_nest_end(skb, v2_attrs);
return 0;
errv2attr:
nla_nest_cancel(skb, v2_attrs);
return -EMSGSIZE;
}
static int smc_nl_fill_smcr_lgr_v2(struct smc_link_group *lgr,
struct sk_buff *skb,
struct netlink_callback *cb)
{
struct nlattr *v2_attrs;
v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_R_V2);
if (!v2_attrs)
goto errattr;
if (nla_put_u8(skb, SMC_NLA_LGR_R_V2_DIRECT, !lgr->uses_gateway))
goto errv2attr;
nla_nest_end(skb, v2_attrs);
return 0;
errv2attr:
nla_nest_cancel(skb, v2_attrs);
errattr:
return -EMSGSIZE;
}
static int smc_nl_fill_lgr(struct smc_link_group *lgr,
struct sk_buff *skb,
struct netlink_callback *cb)
{
char smc_target[SMC_MAX_PNETID_LEN + 1];
struct nlattr *attrs, *v2_attrs;
attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCR);
if (!attrs)
goto errout;
if (nla_put_u32(skb, SMC_NLA_LGR_R_ID, *((u32 *)&lgr->id)))
goto errattr;
if (nla_put_u32(skb, SMC_NLA_LGR_R_CONNS_NUM, lgr->conns_num))
goto errattr;
if (nla_put_u8(skb, SMC_NLA_LGR_R_ROLE, lgr->role))
goto errattr;
if (nla_put_u8(skb, SMC_NLA_LGR_R_TYPE, lgr->type))
goto errattr;
if (nla_put_u8(skb, SMC_NLA_LGR_R_BUF_TYPE, lgr->buf_type))
goto errattr;
if (nla_put_u8(skb, SMC_NLA_LGR_R_VLAN_ID, lgr->vlan_id))
goto errattr;
if (nla_put_u64_64bit(skb, SMC_NLA_LGR_R_NET_COOKIE,
lgr->net->net_cookie, SMC_NLA_LGR_R_PAD))
goto errattr;
memcpy(smc_target, lgr->pnet_id, SMC_MAX_PNETID_LEN);
smc_target[SMC_MAX_PNETID_LEN] = 0;
if (nla_put_string(skb, SMC_NLA_LGR_R_PNETID, smc_target))
goto errattr;
if (lgr->smc_version > SMC_V1) {
v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_R_V2_COMMON);
if (!v2_attrs)
goto errattr;
if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
goto errattr;
if (smc_nl_fill_smcr_lgr_v2(lgr, skb, cb))
goto errattr;
}
nla_nest_end(skb, attrs);
return 0;
errattr:
nla_nest_cancel(skb, attrs);
errout:
return -EMSGSIZE;
}
static int smc_nl_fill_lgr_link(struct smc_link_group *lgr,
struct smc_link *link,
struct sk_buff *skb,
struct netlink_callback *cb)
{
char smc_ibname[IB_DEVICE_NAME_MAX];
u8 smc_gid_target[41];
struct nlattr *attrs;
u32 link_uid = 0;
void *nlh;
nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
&smc_gen_nl_family, NLM_F_MULTI,
SMC_NETLINK_GET_LINK_SMCR);
if (!nlh)
goto errmsg;
attrs = nla_nest_start(skb, SMC_GEN_LINK_SMCR);
if (!attrs)
goto errout;
if (nla_put_u8(skb, SMC_NLA_LINK_ID, link->link_id))
goto errattr;
if (nla_put_u32(skb, SMC_NLA_LINK_STATE, link->state))
goto errattr;
if (nla_put_u32(skb, SMC_NLA_LINK_CONN_CNT,
atomic_read(&link->conn_cnt)))
goto errattr;
if (nla_put_u8(skb, SMC_NLA_LINK_IB_PORT, link->ibport))
goto errattr;
if (nla_put_u32(skb, SMC_NLA_LINK_NET_DEV, link->ndev_ifidx))
goto errattr;
snprintf(smc_ibname, sizeof(smc_ibname), "%s", link->ibname);
if (nla_put_string(skb, SMC_NLA_LINK_IB_DEV, smc_ibname))
goto errattr;
memcpy(&link_uid, link->link_uid, sizeof(link_uid));
if (nla_put_u32(skb, SMC_NLA_LINK_UID, link_uid))
goto errattr;
memcpy(&link_uid, link->peer_link_uid, sizeof(link_uid));
if (nla_put_u32(skb, SMC_NLA_LINK_PEER_UID, link_uid))
goto errattr;
memset(smc_gid_target, 0, sizeof(smc_gid_target));
smc_gid_be16_convert(smc_gid_target, link->gid);
if (nla_put_string(skb, SMC_NLA_LINK_GID, smc_gid_target))
goto errattr;
memset(smc_gid_target, 0, sizeof(smc_gid_target));
smc_gid_be16_convert(smc_gid_target, link->peer_gid);
if (nla_put_string(skb, SMC_NLA_LINK_PEER_GID, smc_gid_target))
goto errattr;
nla_nest_end(skb, attrs);
genlmsg_end(skb, nlh);
return 0;
errattr:
nla_nest_cancel(skb, attrs);
errout:
genlmsg_cancel(skb, nlh);
errmsg:
return -EMSGSIZE;
}
static int smc_nl_handle_lgr(struct smc_link_group *lgr,
struct sk_buff *skb,
struct netlink_callback *cb,
bool list_links)
{
void *nlh;
int i;
nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
&smc_gen_nl_family, NLM_F_MULTI,
SMC_NETLINK_GET_LGR_SMCR);
if (!nlh)
goto errmsg;
if (smc_nl_fill_lgr(lgr, skb, cb))
goto errout;
genlmsg_end(skb, nlh);
if (!list_links)
goto out;
for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
if (!smc_link_usable(&lgr->lnk[i]))
continue;
if (smc_nl_fill_lgr_link(lgr, &lgr->lnk[i], skb, cb))
goto errout;
}
out:
return 0;
errout:
genlmsg_cancel(skb, nlh);
errmsg:
return -EMSGSIZE;
}
static void smc_nl_fill_lgr_list(struct smc_lgr_list *smc_lgr,
struct sk_buff *skb,
struct netlink_callback *cb,
bool list_links)
{
struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
struct smc_link_group *lgr;
int snum = cb_ctx->pos[0];
int num = 0;
spin_lock_bh(&smc_lgr->lock);
list_for_each_entry(lgr, &smc_lgr->list, list) {
if (num < snum)
goto next;
if (smc_nl_handle_lgr(lgr, skb, cb, list_links))
goto errout;
next:
num++;
}
errout:
spin_unlock_bh(&smc_lgr->lock);
cb_ctx->pos[0] = num;
}
static int smc_nl_fill_smcd_lgr(struct smc_link_group *lgr,
struct sk_buff *skb,
struct netlink_callback *cb)
{
char smc_pnet[SMC_MAX_PNETID_LEN + 1];
struct nlattr *attrs;
void *nlh;
nlh = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
&smc_gen_nl_family, NLM_F_MULTI,
SMC_NETLINK_GET_LGR_SMCD);
if (!nlh)
goto errmsg;
attrs = nla_nest_start(skb, SMC_GEN_LGR_SMCD);
if (!attrs)
goto errout;
if (nla_put_u32(skb, SMC_NLA_LGR_D_ID, *((u32 *)&lgr->id)))
goto errattr;
if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_GID, lgr->smcd->local_gid,
SMC_NLA_LGR_D_PAD))
goto errattr;
if (nla_put_u64_64bit(skb, SMC_NLA_LGR_D_PEER_GID, lgr->peer_gid,
SMC_NLA_LGR_D_PAD))
goto errattr;
if (nla_put_u8(skb, SMC_NLA_LGR_D_VLAN_ID, lgr->vlan_id))
goto errattr;
if (nla_put_u32(skb, SMC_NLA_LGR_D_CONNS_NUM, lgr->conns_num))
goto errattr;
if (nla_put_u32(skb, SMC_NLA_LGR_D_CHID, smc_ism_get_chid(lgr->smcd)))
goto errattr;
memcpy(smc_pnet, lgr->smcd->pnetid, SMC_MAX_PNETID_LEN);
smc_pnet[SMC_MAX_PNETID_LEN] = 0;
if (nla_put_string(skb, SMC_NLA_LGR_D_PNETID, smc_pnet))
goto errattr;
if (lgr->smc_version > SMC_V1) {
struct nlattr *v2_attrs;
v2_attrs = nla_nest_start(skb, SMC_NLA_LGR_D_V2_COMMON);
if (!v2_attrs)
goto errattr;
if (smc_nl_fill_lgr_v2_common(lgr, skb, cb, v2_attrs))
goto errattr;
}
nla_nest_end(skb, attrs);
genlmsg_end(skb, nlh);
return 0;
errattr:
nla_nest_cancel(skb, attrs);
errout:
genlmsg_cancel(skb, nlh);
errmsg:
return -EMSGSIZE;
}
static int smc_nl_handle_smcd_lgr(struct smcd_dev *dev,
struct sk_buff *skb,
struct netlink_callback *cb)
{
struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
struct smc_link_group *lgr;
int snum = cb_ctx->pos[1];
int rc = 0, num = 0;
spin_lock_bh(&dev->lgr_lock);
list_for_each_entry(lgr, &dev->lgr_list, list) {
if (!lgr->is_smcd)
continue;
if (num < snum)
goto next;
rc = smc_nl_fill_smcd_lgr(lgr, skb, cb);
if (rc)
goto errout;
next:
num++;
}
errout:
spin_unlock_bh(&dev->lgr_lock);
cb_ctx->pos[1] = num;
return rc;
}
static int smc_nl_fill_smcd_dev(struct smcd_dev_list *dev_list,
struct sk_buff *skb,
struct netlink_callback *cb)
{
struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
struct smcd_dev *smcd_dev;
int snum = cb_ctx->pos[0];
int rc = 0, num = 0;
mutex_lock(&dev_list->mutex);
list_for_each_entry(smcd_dev, &dev_list->list, list) {
if (list_empty(&smcd_dev->lgr_list))
continue;
if (num < snum)
goto next;
rc = smc_nl_handle_smcd_lgr(smcd_dev, skb, cb);
if (rc)
goto errout;
next:
num++;
}
errout:
mutex_unlock(&dev_list->mutex);
cb_ctx->pos[0] = num;
return rc;
}
int smcr_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
{
bool list_links = false;
smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
return skb->len;
}
int smcr_nl_get_link(struct sk_buff *skb, struct netlink_callback *cb)
{
bool list_links = true;
smc_nl_fill_lgr_list(&smc_lgr_list, skb, cb, list_links);
return skb->len;
}
int smcd_nl_get_lgr(struct sk_buff *skb, struct netlink_callback *cb)
{
smc_nl_fill_smcd_dev(&smcd_dev_list, skb, cb);
return skb->len;
}
void smc_lgr_cleanup_early(struct smc_link_group *lgr)
{
spinlock_t *lgr_lock;
if (!lgr)
return;
net/smc: fix wrong list_del in smc_lgr_cleanup_early smc_lgr_cleanup_early() meant to delete the link group from the link group list, but it deleted the list head by mistake. This may cause memory corruption since we didn't remove the real link group from the list and later memseted the link group structure. We got a list corruption panic when testing: [  231.277259] list_del corruption. prev->next should be ffff8881398a8000, but was 0000000000000000 [  231.278222] ------------[ cut here ]------------ [  231.278726] kernel BUG at lib/list_debug.c:53! [  231.279326] invalid opcode: 0000 [#1] SMP NOPTI [  231.279803] CPU: 0 PID: 5 Comm: kworker/0:0 Not tainted 5.10.46+ #435 [  231.280466] Hardware name: Alibaba Cloud ECS, BIOS 8c24b4c 04/01/2014 [  231.281248] Workqueue: events smc_link_down_work [  231.281732] RIP: 0010:__list_del_entry_valid+0x70/0x90 [  231.282258] Code: 4c 60 82 e8 7d cc 6a 00 0f 0b 48 89 fe 48 c7 c7 88 4c 60 82 e8 6c cc 6a 00 0f 0b 48 89 fe 48 c7 c7 c0 4c 60 82 e8 5b cc 6a 00 <0f> 0b 48 89 fe 48 c7 c7 00 4d 60 82 e8 4a cc 6a 00 0f 0b cc cc cc [  231.284146] RSP: 0018:ffffc90000033d58 EFLAGS: 00010292 [  231.284685] RAX: 0000000000000054 RBX: ffff8881398a8000 RCX: 0000000000000000 [  231.285415] RDX: 0000000000000001 RSI: ffff88813bc18040 RDI: ffff88813bc18040 [  231.286141] RBP: ffffffff8305ad40 R08: 0000000000000003 R09: 0000000000000001 [  231.286873] R10: ffffffff82803da0 R11: ffffc90000033b90 R12: 0000000000000001 [  231.287606] R13: 0000000000000000 R14: ffff8881398a8000 R15: 0000000000000003 [  231.288337] FS:  0000000000000000(0000) GS:ffff88813bc00000(0000) knlGS:0000000000000000 [  231.289160] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [  231.289754] CR2: 0000000000e72058 CR3: 000000010fa96006 CR4: 00000000003706f0 [  231.290485] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [  231.291211] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [  231.291940] Call Trace: [  231.292211]  smc_lgr_terminate_sched+0x53/0xa0 [  231.292677]  smc_switch_conns+0x75/0x6b0 [  231.293085]  ? update_load_avg+0x1a6/0x590 [  231.293517]  ? ttwu_do_wakeup+0x17/0x150 [  231.293907]  ? update_load_avg+0x1a6/0x590 [  231.294317]  ? newidle_balance+0xca/0x3d0 [  231.294716]  smcr_link_down+0x50/0x1a0 [  231.295090]  ? __wake_up_common_lock+0x77/0x90 [  231.295534]  smc_link_down_work+0x46/0x60 [  231.295933]  process_one_work+0x18b/0x350 Fixes: a0a62ee15a829 ("net/smc: separate locks for SMCD and SMCR link group lists") Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Acked-by: Karsten Graul <kgraul@linux.ibm.com> Reviewed-by: Tony Lu <tonylu@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2021-12-01 06:02:30 +03:00
smc_lgr_list_head(lgr, &lgr_lock);
spin_lock_bh(lgr_lock);
/* do not use this link group for new connections */
net/smc: fix wrong list_del in smc_lgr_cleanup_early smc_lgr_cleanup_early() meant to delete the link group from the link group list, but it deleted the list head by mistake. This may cause memory corruption since we didn't remove the real link group from the list and later memseted the link group structure. We got a list corruption panic when testing: [  231.277259] list_del corruption. prev->next should be ffff8881398a8000, but was 0000000000000000 [  231.278222] ------------[ cut here ]------------ [  231.278726] kernel BUG at lib/list_debug.c:53! [  231.279326] invalid opcode: 0000 [#1] SMP NOPTI [  231.279803] CPU: 0 PID: 5 Comm: kworker/0:0 Not tainted 5.10.46+ #435 [  231.280466] Hardware name: Alibaba Cloud ECS, BIOS 8c24b4c 04/01/2014 [  231.281248] Workqueue: events smc_link_down_work [  231.281732] RIP: 0010:__list_del_entry_valid+0x70/0x90 [  231.282258] Code: 4c 60 82 e8 7d cc 6a 00 0f 0b 48 89 fe 48 c7 c7 88 4c 60 82 e8 6c cc 6a 00 0f 0b 48 89 fe 48 c7 c7 c0 4c 60 82 e8 5b cc 6a 00 <0f> 0b 48 89 fe 48 c7 c7 00 4d 60 82 e8 4a cc 6a 00 0f 0b cc cc cc [  231.284146] RSP: 0018:ffffc90000033d58 EFLAGS: 00010292 [  231.284685] RAX: 0000000000000054 RBX: ffff8881398a8000 RCX: 0000000000000000 [  231.285415] RDX: 0000000000000001 RSI: ffff88813bc18040 RDI: ffff88813bc18040 [  231.286141] RBP: ffffffff8305ad40 R08: 0000000000000003 R09: 0000000000000001 [  231.286873] R10: ffffffff82803da0 R11: ffffc90000033b90 R12: 0000000000000001 [  231.287606] R13: 0000000000000000 R14: ffff8881398a8000 R15: 0000000000000003 [  231.288337] FS:  0000000000000000(0000) GS:ffff88813bc00000(0000) knlGS:0000000000000000 [  231.289160] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [  231.289754] CR2: 0000000000e72058 CR3: 000000010fa96006 CR4: 00000000003706f0 [  231.290485] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [  231.291211] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [  231.291940] Call Trace: [  231.292211]  smc_lgr_terminate_sched+0x53/0xa0 [  231.292677]  smc_switch_conns+0x75/0x6b0 [  231.293085]  ? update_load_avg+0x1a6/0x590 [  231.293517]  ? ttwu_do_wakeup+0x17/0x150 [  231.293907]  ? update_load_avg+0x1a6/0x590 [  231.294317]  ? newidle_balance+0xca/0x3d0 [  231.294716]  smcr_link_down+0x50/0x1a0 [  231.295090]  ? __wake_up_common_lock+0x77/0x90 [  231.295534]  smc_link_down_work+0x46/0x60 [  231.295933]  process_one_work+0x18b/0x350 Fixes: a0a62ee15a829 ("net/smc: separate locks for SMCD and SMCR link group lists") Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Acked-by: Karsten Graul <kgraul@linux.ibm.com> Reviewed-by: Tony Lu <tonylu@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2021-12-01 06:02:30 +03:00
if (!list_empty(&lgr->list))
list_del_init(&lgr->list);
spin_unlock_bh(lgr_lock);
__smc_lgr_terminate(lgr, true);
}
static void smcr_lgr_link_deactivate_all(struct smc_link_group *lgr)
{
int i;
for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
struct smc_link *lnk = &lgr->lnk[i];
if (smc_link_sendable(lnk))
lnk->state = SMC_LNK_INACTIVE;
}
wake_up_all(&lgr->llc_msg_waiter);
wake_up_all(&lgr->llc_flow_waiter);
}
static void smc_lgr_free(struct smc_link_group *lgr);
static void smc_lgr_free_work(struct work_struct *work)
{
struct smc_link_group *lgr = container_of(to_delayed_work(work),
struct smc_link_group,
free_work);
spinlock_t *lgr_lock;
bool conns;
smc_lgr_list_head(lgr, &lgr_lock);
spin_lock_bh(lgr_lock);
if (lgr->freeing) {
spin_unlock_bh(lgr_lock);
return;
}
read_lock_bh(&lgr->conns_lock);
conns = RB_EMPTY_ROOT(&lgr->conns_all);
read_unlock_bh(&lgr->conns_lock);
if (!conns) { /* number of lgr connections is no longer zero */
spin_unlock_bh(lgr_lock);
return;
}
list_del_init(&lgr->list); /* remove from smc_lgr_list */
lgr->freeing = 1; /* this instance does the freeing, no new schedule */
spin_unlock_bh(lgr_lock);
cancel_delayed_work(&lgr->free_work);
if (!lgr->is_smcd && !lgr->terminating)
smc_llc_send_link_delete_all(lgr, true,
SMC_LLC_DEL_PROG_INIT_TERM);
if (lgr->is_smcd && !lgr->terminating)
smc_ism_signal_shutdown(lgr);
if (!lgr->is_smcd)
smcr_lgr_link_deactivate_all(lgr);
smc_lgr_free(lgr);
}
static void smc_lgr_terminate_work(struct work_struct *work)
{
struct smc_link_group *lgr = container_of(work, struct smc_link_group,
terminate_work);
__smc_lgr_terminate(lgr, true);
}
/* return next unique link id for the lgr */
static u8 smcr_next_link_id(struct smc_link_group *lgr)
{
u8 link_id;
int i;
while (1) {
again:
link_id = ++lgr->next_link_id;
if (!link_id) /* skip zero as link_id */
link_id = ++lgr->next_link_id;
for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
if (smc_link_usable(&lgr->lnk[i]) &&
lgr->lnk[i].link_id == link_id)
goto again;
}
break;
}
return link_id;
}
static void smcr_copy_dev_info_to_link(struct smc_link *link)
{
struct smc_ib_device *smcibdev = link->smcibdev;
snprintf(link->ibname, sizeof(link->ibname), "%s",
smcibdev->ibdev->name);
link->ndev_ifidx = smcibdev->ndev_ifidx[link->ibport - 1];
}
int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
u8 link_idx, struct smc_init_info *ini)
{
struct smc_ib_device *smcibdev;
u8 rndvec[3];
int rc;
if (lgr->smc_version == SMC_V2) {
lnk->smcibdev = ini->smcrv2.ib_dev_v2;
lnk->ibport = ini->smcrv2.ib_port_v2;
} else {
lnk->smcibdev = ini->ib_dev;
lnk->ibport = ini->ib_port;
}
get_device(&lnk->smcibdev->ibdev->dev);
atomic_inc(&lnk->smcibdev->lnk_cnt);
net/smc: Resolve the race between SMC-R link access and clear We encountered some crashes caused by the race between SMC-R link access and link clear that triggered by abnormal link group termination, such as port error. Here is an example of this kind of crashes: BUG: kernel NULL pointer dereference, address: 0000000000000000 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_llc_flow_initiate+0x44/0x190 [smc] Call Trace: <TASK> ? __smc_buf_create+0x75a/0x950 [smc] smcr_lgr_reg_rmbs+0x2a/0xbf [smc] smc_listen_work+0xf72/0x1230 [smc] ? process_one_work+0x25c/0x600 process_one_work+0x25c/0x600 worker_thread+0x4f/0x3a0 ? process_one_work+0x600/0x600 kthread+0x15d/0x1a0 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() __smc_lgr_terminate() --------------------------------------------------------------- | smc_lgr_free() | |- smcr_link_clear() | |- memset(lnk, 0) smc_listen_rdma_reg() | |- smcr_lgr_reg_rmbs() | |- smc_llc_flow_initiate() | |- access lnk->lgr (panic) | These crashes are similarly caused by clearing SMC-R link resources when some functions is still accessing to them. This patch tries to fix the issue by introducing reference count of SMC-R links and ensuring that the sensitive resources of links won't be cleared until reference count reaches zero. The operation to the SMC-R link reference count can be concluded as follows: object [hold or initialized as 1] [put] -------------------------------------------------------------------- links smcr_link_init() smcr_link_clear() connections smc_conn_create() smc_conn_free() Through this way, the clear of SMC-R links is later than the free of all the smc connections above it, thus avoiding the unsafe reference to SMC-R links. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:42 +03:00
refcount_set(&lnk->refcnt, 1); /* link refcnt is set to 1 */
lnk->clearing = 0;
lnk->path_mtu = lnk->smcibdev->pattr[lnk->ibport - 1].active_mtu;
lnk->link_id = smcr_next_link_id(lgr);
lnk->lgr = lgr;
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
smc_lgr_hold(lgr); /* lgr_put in smcr_link_clear() */
lnk->link_idx = link_idx;
net/smc: Fix possible access to freed memory in link clear After modifying the QP to the Error state, all RX WR would be completed with WC in IB_WC_WR_FLUSH_ERR status. Current implementation does not wait for it is done, but destroy the QP and free the link group directly. So there is a risk that accessing the freed memory in tasklet context. Here is a crash example: BUG: unable to handle page fault for address: ffffffff8f220860 #PF: supervisor write access in kernel mode #PF: error_code(0x0002) - not-present page PGD f7300e067 P4D f7300e067 PUD f7300f063 PMD 8c4e45063 PTE 800ffff08c9df060 Oops: 0002 [#1] SMP PTI CPU: 1 PID: 0 Comm: swapper/1 Kdump: loaded Tainted: G S OE 5.10.0-0607+ #23 Hardware name: Inspur NF5280M4/YZMB-00689-101, BIOS 4.1.20 07/09/2018 RIP: 0010:native_queued_spin_lock_slowpath+0x176/0x1b0 Code: f3 90 48 8b 32 48 85 f6 74 f6 eb d5 c1 ee 12 83 e0 03 83 ee 01 48 c1 e0 05 48 63 f6 48 05 00 c8 02 00 48 03 04 f5 00 09 98 8e <48> 89 10 8b 42 08 85 c0 75 09 f3 90 8b 42 08 85 c0 74 f7 48 8b 32 RSP: 0018:ffffb3b6c001ebd8 EFLAGS: 00010086 RAX: ffffffff8f220860 RBX: 0000000000000246 RCX: 0000000000080000 RDX: ffff91db1f86c800 RSI: 000000000000173c RDI: ffff91db62bace00 RBP: ffff91db62bacc00 R08: 0000000000000000 R09: c00000010000028b R10: 0000000000055198 R11: ffffb3b6c001ea58 R12: ffff91db80e05010 R13: 000000000000000a R14: 0000000000000006 R15: 0000000000000040 FS: 0000000000000000(0000) GS:ffff91db1f840000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: ffffffff8f220860 CR3: 00000001f9580004 CR4: 00000000003706e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 Call Trace: <IRQ> _raw_spin_lock_irqsave+0x30/0x40 mlx5_ib_poll_cq+0x4c/0xc50 [mlx5_ib] smc_wr_rx_tasklet_fn+0x56/0xa0 [smc] tasklet_action_common.isra.21+0x66/0x100 __do_softirq+0xd5/0x29c asm_call_irq_on_stack+0x12/0x20 </IRQ> do_softirq_own_stack+0x37/0x40 irq_exit_rcu+0x9d/0xa0 sysvec_call_function_single+0x34/0x80 asm_sysvec_call_function_single+0x12/0x20 Fixes: bd4ad57718cc ("smc: initialize IB transport incl. PD, MR, QP, CQ, event, WR") Signed-off-by: Yacan Liu <liuyacan@corp.netease.com> Reviewed-by: Tony Lu <tonylu@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-09-06 16:01:39 +03:00
lnk->wr_rx_id_compl = 0;
smc_ibdev_cnt_inc(lnk);
smcr_copy_dev_info_to_link(lnk);
atomic_set(&lnk->conn_cnt, 0);
smc_llc_link_set_uid(lnk);
INIT_WORK(&lnk->link_down_wrk, smc_link_down_work);
if (!lnk->smcibdev->initialized) {
rc = (int)smc_ib_setup_per_ibdev(lnk->smcibdev);
if (rc)
goto out;
}
get_random_bytes(rndvec, sizeof(rndvec));
lnk->psn_initial = rndvec[0] + (rndvec[1] << 8) +
(rndvec[2] << 16);
rc = smc_ib_determine_gid(lnk->smcibdev, lnk->ibport,
ini->vlan_id, lnk->gid, &lnk->sgid_index,
lgr->smc_version == SMC_V2 ?
&ini->smcrv2 : NULL);
if (rc)
goto out;
rc = smc_llc_link_init(lnk);
if (rc)
goto out;
rc = smc_wr_alloc_link_mem(lnk);
if (rc)
goto clear_llc_lnk;
rc = smc_ib_create_protection_domain(lnk);
if (rc)
goto free_link_mem;
rc = smc_ib_create_queue_pair(lnk);
if (rc)
goto dealloc_pd;
rc = smc_wr_create_link(lnk);
if (rc)
goto destroy_qp;
lnk->state = SMC_LNK_ACTIVATING;
return 0;
destroy_qp:
smc_ib_destroy_queue_pair(lnk);
dealloc_pd:
smc_ib_dealloc_protection_domain(lnk);
free_link_mem:
smc_wr_free_link_mem(lnk);
clear_llc_lnk:
smc_llc_link_clear(lnk, false);
out:
smc_ibdev_cnt_dec(lnk);
put_device(&lnk->smcibdev->ibdev->dev);
smcibdev = lnk->smcibdev;
memset(lnk, 0, sizeof(struct smc_link));
lnk->state = SMC_LNK_UNUSED;
if (!atomic_dec_return(&smcibdev->lnk_cnt))
wake_up(&smcibdev->lnks_deleted);
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
smc_lgr_put(lgr); /* lgr_hold above */
return rc;
}
/* create a new SMC link group */
static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
{
struct smc_link_group *lgr;
struct list_head *lgr_list;
struct smc_link *lnk;
spinlock_t *lgr_lock;
u8 link_idx;
int rc = 0;
int i;
if (ini->is_smcd && ini->vlan_id) {
if (smc_ism_get_vlan(ini->ism_dev[ini->ism_selected],
ini->vlan_id)) {
rc = SMC_CLC_DECL_ISMVLANERR;
goto out;
}
}
lgr = kzalloc(sizeof(*lgr), GFP_KERNEL);
if (!lgr) {
rc = SMC_CLC_DECL_MEM;
goto ism_put_vlan;
}
lgr->tx_wq = alloc_workqueue("smc_tx_wq-%*phN", 0, 0,
SMC_LGR_ID_SIZE, &lgr->id);
if (!lgr->tx_wq) {
rc = -ENOMEM;
goto free_lgr;
}
lgr->is_smcd = ini->is_smcd;
lgr->sync_err = 0;
lgr->terminating = 0;
lgr->freeing = 0;
lgr->vlan_id = ini->vlan_id;
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
refcount_set(&lgr->refcnt, 1); /* set lgr refcnt to 1 */
mutex_init(&lgr->sndbufs_lock);
mutex_init(&lgr->rmbs_lock);
rwlock_init(&lgr->conns_lock);
for (i = 0; i < SMC_RMBE_SIZES; i++) {
INIT_LIST_HEAD(&lgr->sndbufs[i]);
INIT_LIST_HEAD(&lgr->rmbs[i]);
}
lgr->next_link_id = 0;
smc_lgr_list.num += SMC_LGR_NUM_INCR;
memcpy(&lgr->id, (u8 *)&smc_lgr_list.num, SMC_LGR_ID_SIZE);
INIT_DELAYED_WORK(&lgr->free_work, smc_lgr_free_work);
INIT_WORK(&lgr->terminate_work, smc_lgr_terminate_work);
lgr->conns_all = RB_ROOT;
if (ini->is_smcd) {
/* SMC-D specific settings */
get_device(&ini->ism_dev[ini->ism_selected]->dev);
lgr->peer_gid = ini->ism_peer_gid[ini->ism_selected];
lgr->smcd = ini->ism_dev[ini->ism_selected];
lgr_list = &ini->ism_dev[ini->ism_selected]->lgr_list;
lgr_lock = &lgr->smcd->lgr_lock;
lgr->smc_version = ini->smcd_version;
lgr->peer_shutdown = 0;
atomic_inc(&ini->ism_dev[ini->ism_selected]->lgr_cnt);
} else {
/* SMC-R specific settings */
struct smc_ib_device *ibdev;
int ibport;
lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
lgr->smc_version = ini->smcr_version;
memcpy(lgr->peer_systemid, ini->peer_systemid,
SMC_SYSTEMID_LEN);
if (lgr->smc_version == SMC_V2) {
ibdev = ini->smcrv2.ib_dev_v2;
ibport = ini->smcrv2.ib_port_v2;
lgr->saddr = ini->smcrv2.saddr;
lgr->uses_gateway = ini->smcrv2.uses_gateway;
memcpy(lgr->nexthop_mac, ini->smcrv2.nexthop_mac,
ETH_ALEN);
} else {
ibdev = ini->ib_dev;
ibport = ini->ib_port;
}
memcpy(lgr->pnet_id, ibdev->pnetid[ibport - 1],
SMC_MAX_PNETID_LEN);
if (smc_wr_alloc_lgr_mem(lgr))
goto free_wq;
smc_llc_lgr_init(lgr, smc);
link_idx = SMC_SINGLE_LINK;
lnk = &lgr->lnk[link_idx];
rc = smcr_link_init(lgr, lnk, link_idx, ini);
if (rc) {
smc_wr_free_lgr_mem(lgr);
goto free_wq;
}
lgr->net = smc_ib_net(lnk->smcibdev);
lgr_list = &smc_lgr_list.list;
lgr_lock = &smc_lgr_list.lock;
lgr->buf_type = lgr->net->smc.sysctl_smcr_buf_type;
atomic_inc(&lgr_cnt);
}
smc->conn.lgr = lgr;
spin_lock_bh(lgr_lock);
net/smc: fix dmb buffer shortage There is a current limit of 1920 registered dmb buffers per ISM device for smc-d. One link group can contain 255 connections, each connection is using one dmb buffer. When the connection is closed then the registered buffer is held in a queue and is reused by the next connection. When a link group is 'full' then another link group is created and uses an own buffer pool. The link groups are added to a list using list_add() which puts a new link group to the first position in the list. In the situation that many connections are opened (>1920) and a few of them stay open while others are closed quickly we end up with at least 8 link groups. For a new connection a matching link group is looked up, iterating over the list of link groups. The trailing 7 link groups all have registered dmb buffers which could be reused, while the first link group has only a few dmb buffers and then hit the 1920 limit. Because the first link group is not full (255 connection limit not reached) it is chosen and finally the connection falls back to TCP because there is no dmb buffer available in this link group. There are multiple ways to fix that: using list_add_tail() allows to scan older link groups first for free buffers which ensures that buffers are reused first. This fixes the problem for smc-r link groups as well. For smc-d there is an even better way to address this problem because smc-d does not have the 255 connections per link group limit. So fix the problem for smc-d by allowing large link groups. Fixes: c6ba7c9ba43d ("net/smc: add base infrastructure for SMC-D and ISM") Reviewed-by: Ursula Braun <ubraun@linux.ibm.com> Signed-off-by: Karsten Graul <kgraul@linux.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-07-20 17:24:29 +03:00
list_add_tail(&lgr->list, lgr_list);
spin_unlock_bh(lgr_lock);
return 0;
free_wq:
destroy_workqueue(lgr->tx_wq);
free_lgr:
kfree(lgr);
ism_put_vlan:
if (ini->is_smcd && ini->vlan_id)
smc_ism_put_vlan(ini->ism_dev[ini->ism_selected], ini->vlan_id);
out:
if (rc < 0) {
if (rc == -ENOMEM)
rc = SMC_CLC_DECL_MEM;
else
rc = SMC_CLC_DECL_INTERR;
}
return rc;
}
static int smc_write_space(struct smc_connection *conn)
{
int buffer_len = conn->peer_rmbe_size;
union smc_host_cursor prod;
union smc_host_cursor cons;
int space;
smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
/* determine rx_buf space */
space = buffer_len - smc_curs_diff(buffer_len, &cons, &prod);
return space;
}
static int smc_switch_cursor(struct smc_sock *smc, struct smc_cdc_tx_pend *pend,
struct smc_wr_buf *wr_buf)
{
struct smc_connection *conn = &smc->conn;
union smc_host_cursor cons, fin;
int rc = 0;
int diff;
smc_curs_copy(&conn->tx_curs_sent, &conn->tx_curs_fin, conn);
smc_curs_copy(&fin, &conn->local_tx_ctrl_fin, conn);
/* set prod cursor to old state, enforce tx_rdma_writes() */
smc_curs_copy(&conn->local_tx_ctrl.prod, &fin, conn);
smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
if (smc_curs_comp(conn->peer_rmbe_size, &cons, &fin) < 0) {
/* cons cursor advanced more than fin, and prod was set
* fin above, so now prod is smaller than cons. Fix that.
*/
diff = smc_curs_diff(conn->peer_rmbe_size, &fin, &cons);
smc_curs_add(conn->sndbuf_desc->len,
&conn->tx_curs_sent, diff);
smc_curs_add(conn->sndbuf_desc->len,
&conn->tx_curs_fin, diff);
smp_mb__before_atomic();
atomic_add(diff, &conn->sndbuf_space);
smp_mb__after_atomic();
smc_curs_add(conn->peer_rmbe_size,
&conn->local_tx_ctrl.prod, diff);
smc_curs_add(conn->peer_rmbe_size,
&conn->local_tx_ctrl_fin, diff);
}
/* recalculate, value is used by tx_rdma_writes() */
atomic_set(&smc->conn.peer_rmbe_space, smc_write_space(conn));
if (smc->sk.sk_state != SMC_INIT &&
smc->sk.sk_state != SMC_CLOSED) {
rc = smcr_cdc_msg_send_validation(conn, pend, wr_buf);
if (!rc) {
queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work, 0);
smc->sk.sk_data_ready(&smc->sk);
}
} else {
smc_wr_tx_put_slot(conn->lnk,
(struct smc_wr_tx_pend_priv *)pend);
}
return rc;
}
void smc_switch_link_and_count(struct smc_connection *conn,
struct smc_link *to_lnk)
{
atomic_dec(&conn->lnk->conn_cnt);
net/smc: Resolve the race between SMC-R link access and clear We encountered some crashes caused by the race between SMC-R link access and link clear that triggered by abnormal link group termination, such as port error. Here is an example of this kind of crashes: BUG: kernel NULL pointer dereference, address: 0000000000000000 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_llc_flow_initiate+0x44/0x190 [smc] Call Trace: <TASK> ? __smc_buf_create+0x75a/0x950 [smc] smcr_lgr_reg_rmbs+0x2a/0xbf [smc] smc_listen_work+0xf72/0x1230 [smc] ? process_one_work+0x25c/0x600 process_one_work+0x25c/0x600 worker_thread+0x4f/0x3a0 ? process_one_work+0x600/0x600 kthread+0x15d/0x1a0 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() __smc_lgr_terminate() --------------------------------------------------------------- | smc_lgr_free() | |- smcr_link_clear() | |- memset(lnk, 0) smc_listen_rdma_reg() | |- smcr_lgr_reg_rmbs() | |- smc_llc_flow_initiate() | |- access lnk->lgr (panic) | These crashes are similarly caused by clearing SMC-R link resources when some functions is still accessing to them. This patch tries to fix the issue by introducing reference count of SMC-R links and ensuring that the sensitive resources of links won't be cleared until reference count reaches zero. The operation to the SMC-R link reference count can be concluded as follows: object [hold or initialized as 1] [put] -------------------------------------------------------------------- links smcr_link_init() smcr_link_clear() connections smc_conn_create() smc_conn_free() Through this way, the clear of SMC-R links is later than the free of all the smc connections above it, thus avoiding the unsafe reference to SMC-R links. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:42 +03:00
/* link_hold in smc_conn_create() */
smcr_link_put(conn->lnk);
conn->lnk = to_lnk;
atomic_inc(&conn->lnk->conn_cnt);
net/smc: Resolve the race between SMC-R link access and clear We encountered some crashes caused by the race between SMC-R link access and link clear that triggered by abnormal link group termination, such as port error. Here is an example of this kind of crashes: BUG: kernel NULL pointer dereference, address: 0000000000000000 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_llc_flow_initiate+0x44/0x190 [smc] Call Trace: <TASK> ? __smc_buf_create+0x75a/0x950 [smc] smcr_lgr_reg_rmbs+0x2a/0xbf [smc] smc_listen_work+0xf72/0x1230 [smc] ? process_one_work+0x25c/0x600 process_one_work+0x25c/0x600 worker_thread+0x4f/0x3a0 ? process_one_work+0x600/0x600 kthread+0x15d/0x1a0 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() __smc_lgr_terminate() --------------------------------------------------------------- | smc_lgr_free() | |- smcr_link_clear() | |- memset(lnk, 0) smc_listen_rdma_reg() | |- smcr_lgr_reg_rmbs() | |- smc_llc_flow_initiate() | |- access lnk->lgr (panic) | These crashes are similarly caused by clearing SMC-R link resources when some functions is still accessing to them. This patch tries to fix the issue by introducing reference count of SMC-R links and ensuring that the sensitive resources of links won't be cleared until reference count reaches zero. The operation to the SMC-R link reference count can be concluded as follows: object [hold or initialized as 1] [put] -------------------------------------------------------------------- links smcr_link_init() smcr_link_clear() connections smc_conn_create() smc_conn_free() Through this way, the clear of SMC-R links is later than the free of all the smc connections above it, thus avoiding the unsafe reference to SMC-R links. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:42 +03:00
/* link_put in smc_conn_free() */
smcr_link_hold(conn->lnk);
}
struct smc_link *smc_switch_conns(struct smc_link_group *lgr,
struct smc_link *from_lnk, bool is_dev_err)
{
struct smc_link *to_lnk = NULL;
struct smc_cdc_tx_pend *pend;
struct smc_connection *conn;
struct smc_wr_buf *wr_buf;
struct smc_sock *smc;
struct rb_node *node;
int i, rc = 0;
/* link is inactive, wake up tx waiters */
smc_wr_wakeup_tx_wait(from_lnk);
for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
if (!smc_link_active(&lgr->lnk[i]) || i == from_lnk->link_idx)
continue;
if (is_dev_err && from_lnk->smcibdev == lgr->lnk[i].smcibdev &&
from_lnk->ibport == lgr->lnk[i].ibport) {
continue;
}
to_lnk = &lgr->lnk[i];
break;
}
if (!to_lnk || !smc_wr_tx_link_hold(to_lnk)) {
smc_lgr_terminate_sched(lgr);
return NULL;
}
again:
read_lock_bh(&lgr->conns_lock);
for (node = rb_first(&lgr->conns_all); node; node = rb_next(node)) {
conn = rb_entry(node, struct smc_connection, alert_node);
if (conn->lnk != from_lnk)
continue;
smc = container_of(conn, struct smc_sock, conn);
/* conn->lnk not yet set in SMC_INIT state */
if (smc->sk.sk_state == SMC_INIT)
continue;
if (smc->sk.sk_state == SMC_CLOSED ||
smc->sk.sk_state == SMC_PEERCLOSEWAIT1 ||
smc->sk.sk_state == SMC_PEERCLOSEWAIT2 ||
smc->sk.sk_state == SMC_APPFINCLOSEWAIT ||
smc->sk.sk_state == SMC_APPCLOSEWAIT1 ||
smc->sk.sk_state == SMC_APPCLOSEWAIT2 ||
smc->sk.sk_state == SMC_PEERFINCLOSEWAIT ||
smc->sk.sk_state == SMC_PEERABORTWAIT ||
smc->sk.sk_state == SMC_PROCESSABORT) {
spin_lock_bh(&conn->send_lock);
smc_switch_link_and_count(conn, to_lnk);
spin_unlock_bh(&conn->send_lock);
continue;
}
sock_hold(&smc->sk);
read_unlock_bh(&lgr->conns_lock);
/* pre-fetch buffer outside of send_lock, might sleep */
rc = smc_cdc_get_free_slot(conn, to_lnk, &wr_buf, NULL, &pend);
if (rc)
goto err_out;
/* avoid race with smcr_tx_sndbuf_nonempty() */
spin_lock_bh(&conn->send_lock);
smc_switch_link_and_count(conn, to_lnk);
rc = smc_switch_cursor(smc, pend, wr_buf);
spin_unlock_bh(&conn->send_lock);
sock_put(&smc->sk);
if (rc)
goto err_out;
goto again;
}
read_unlock_bh(&lgr->conns_lock);
smc_wr_tx_link_put(to_lnk);
return to_lnk;
err_out:
smcr_link_down_cond_sched(to_lnk);
smc_wr_tx_link_put(to_lnk);
return NULL;
}
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
static void smcr_buf_unuse(struct smc_buf_desc *buf_desc, bool is_rmb,
struct smc_link_group *lgr)
{
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
struct mutex *lock; /* lock buffer list */
int rc;
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
if (is_rmb && buf_desc->is_conf_rkey && !list_empty(&lgr->list)) {
/* unregister rmb with peer */
rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
if (!rc) {
/* protect against smc_llc_cli_rkey_exchange() */
mutex_lock(&lgr->llc_conf_mutex);
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
smc_llc_do_delete_rkey(lgr, buf_desc);
buf_desc->is_conf_rkey = false;
mutex_unlock(&lgr->llc_conf_mutex);
smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
}
}
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
if (buf_desc->is_reg_err) {
/* buf registration failed, reuse not possible */
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
lock = is_rmb ? &lgr->rmbs_lock :
&lgr->sndbufs_lock;
mutex_lock(lock);
list_del(&buf_desc->list);
mutex_unlock(lock);
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
smc_buf_free(lgr, is_rmb, buf_desc);
} else {
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
buf_desc->used = 0;
memset(buf_desc->cpu_addr, 0, buf_desc->len);
}
}
static void smc_buf_unuse(struct smc_connection *conn,
struct smc_link_group *lgr)
{
if (conn->sndbuf_desc) {
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
if (!lgr->is_smcd && conn->sndbuf_desc->is_vm) {
smcr_buf_unuse(conn->sndbuf_desc, false, lgr);
} else {
conn->sndbuf_desc->used = 0;
memset(conn->sndbuf_desc->cpu_addr, 0,
conn->sndbuf_desc->len);
}
}
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
if (conn->rmb_desc) {
if (!lgr->is_smcd) {
smcr_buf_unuse(conn->rmb_desc, true, lgr);
} else {
conn->rmb_desc->used = 0;
memset(conn->rmb_desc->cpu_addr, 0,
conn->rmb_desc->len +
sizeof(struct smcd_cdc_msg));
}
}
}
/* remove a finished connection from its link group */
void smc_conn_free(struct smc_connection *conn)
{
struct smc_link_group *lgr = conn->lgr;
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
if (!lgr || conn->freed)
/* Connection has never been registered in a
* link group, or has already been freed.
*/
return;
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
conn->freed = 1;
if (!smc_conn_lgr_valid(conn))
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
/* Connection has already unregistered from
* link group.
*/
goto lgr_put;
if (lgr->is_smcd) {
if (!list_empty(&lgr->list))
smc_ism_unset_conn(conn);
tasklet_kill(&conn->rx_tsklet);
} else {
net/smc: fix kernel panic caused by race of smc_sock A crash occurs when smc_cdc_tx_handler() tries to access smc_sock but smc_release() has already freed it. [ 4570.695099] BUG: unable to handle page fault for address: 000000002eae9e88 [ 4570.696048] #PF: supervisor write access in kernel mode [ 4570.696728] #PF: error_code(0x0002) - not-present page [ 4570.697401] PGD 0 P4D 0 [ 4570.697716] Oops: 0002 [#1] PREEMPT SMP NOPTI [ 4570.698228] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.16.0-rc4+ #111 [ 4570.699013] Hardware name: Alibaba Cloud Alibaba Cloud ECS, BIOS 8c24b4c 04/0 [ 4570.699933] RIP: 0010:_raw_spin_lock+0x1a/0x30 <...> [ 4570.711446] Call Trace: [ 4570.711746] <IRQ> [ 4570.711992] smc_cdc_tx_handler+0x41/0xc0 [ 4570.712470] smc_wr_tx_tasklet_fn+0x213/0x560 [ 4570.712981] ? smc_cdc_tx_dismisser+0x10/0x10 [ 4570.713489] tasklet_action_common.isra.17+0x66/0x140 [ 4570.714083] __do_softirq+0x123/0x2f4 [ 4570.714521] irq_exit_rcu+0xc4/0xf0 [ 4570.714934] common_interrupt+0xba/0xe0 Though smc_cdc_tx_handler() checked the existence of smc connection, smc_release() may have already dismissed and released the smc socket before smc_cdc_tx_handler() further visits it. smc_cdc_tx_handler() |smc_release() if (!conn) | | |smc_cdc_tx_dismiss_slots() | smc_cdc_tx_dismisser() | |sock_put(&smc->sk) <- last sock_put, | smc_sock freed bh_lock_sock(&smc->sk) (panic) | To make sure we won't receive any CDC messages after we free the smc_sock, add a refcount on the smc_connection for inflight CDC message(posted to the QP but haven't received related CQE), and don't release the smc_connection until all the inflight CDC messages haven been done, for both success or failed ones. Using refcount on CDC messages brings another problem: when the link is going to be destroyed, smcr_link_clear() will reset the QP, which then remove all the pending CQEs related to the QP in the CQ. To make sure all the CQEs will always come back so the refcount on the smc_connection can always reach 0, smc_ib_modify_qp_reset() was replaced by smc_ib_modify_qp_error(). And remove the timeout in smc_wr_tx_wait_no_pending_sends() since we need to wait for all pending WQEs done, or we may encounter use-after- free when handling CQEs. For IB device removal routine, we need to wait for all the QPs on that device been destroyed before we can destroy CQs on the device, or the refcount on smc_connection won't reach 0 and smc_sock cannot be released. Fixes: 5f08318f617b ("smc: connection data control (CDC)") Reported-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2021-12-28 12:03:25 +03:00
smc_cdc_wait_pend_tx_wr(conn);
if (current_work() != &conn->abort_work)
cancel_work_sync(&conn->abort_work);
}
if (!list_empty(&lgr->list)) {
smc_buf_unuse(conn, lgr); /* allow buffer reuse */
smc_lgr_unregister_conn(conn);
}
if (!lgr->conns_num)
smc_lgr_schedule_free_work(lgr);
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
lgr_put:
net/smc: Resolve the race between SMC-R link access and clear We encountered some crashes caused by the race between SMC-R link access and link clear that triggered by abnormal link group termination, such as port error. Here is an example of this kind of crashes: BUG: kernel NULL pointer dereference, address: 0000000000000000 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_llc_flow_initiate+0x44/0x190 [smc] Call Trace: <TASK> ? __smc_buf_create+0x75a/0x950 [smc] smcr_lgr_reg_rmbs+0x2a/0xbf [smc] smc_listen_work+0xf72/0x1230 [smc] ? process_one_work+0x25c/0x600 process_one_work+0x25c/0x600 worker_thread+0x4f/0x3a0 ? process_one_work+0x600/0x600 kthread+0x15d/0x1a0 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() __smc_lgr_terminate() --------------------------------------------------------------- | smc_lgr_free() | |- smcr_link_clear() | |- memset(lnk, 0) smc_listen_rdma_reg() | |- smcr_lgr_reg_rmbs() | |- smc_llc_flow_initiate() | |- access lnk->lgr (panic) | These crashes are similarly caused by clearing SMC-R link resources when some functions is still accessing to them. This patch tries to fix the issue by introducing reference count of SMC-R links and ensuring that the sensitive resources of links won't be cleared until reference count reaches zero. The operation to the SMC-R link reference count can be concluded as follows: object [hold or initialized as 1] [put] -------------------------------------------------------------------- links smcr_link_init() smcr_link_clear() connections smc_conn_create() smc_conn_free() Through this way, the clear of SMC-R links is later than the free of all the smc connections above it, thus avoiding the unsafe reference to SMC-R links. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:42 +03:00
if (!lgr->is_smcd)
smcr_link_put(conn->lnk); /* link_hold in smc_conn_create() */
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
smc_lgr_put(lgr); /* lgr_hold in smc_conn_create() */
}
/* unregister a link from a buf_desc */
static void smcr_buf_unmap_link(struct smc_buf_desc *buf_desc, bool is_rmb,
struct smc_link *lnk)
{
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
if (is_rmb || buf_desc->is_vm)
buf_desc->is_reg_mr[lnk->link_idx] = false;
if (!buf_desc->is_map_ib[lnk->link_idx])
return;
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
if ((is_rmb || buf_desc->is_vm) &&
buf_desc->mr[lnk->link_idx]) {
smc_ib_put_memory_region(buf_desc->mr[lnk->link_idx]);
buf_desc->mr[lnk->link_idx] = NULL;
}
if (is_rmb)
smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_FROM_DEVICE);
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
else
smc_ib_buf_unmap_sg(lnk, buf_desc, DMA_TO_DEVICE);
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
sg_free_table(&buf_desc->sgt[lnk->link_idx]);
buf_desc->is_map_ib[lnk->link_idx] = false;
}
/* unmap all buffers of lgr for a deleted link */
static void smcr_buf_unmap_lgr(struct smc_link *lnk)
{
struct smc_link_group *lgr = lnk->lgr;
struct smc_buf_desc *buf_desc, *bf;
int i;
for (i = 0; i < SMC_RMBE_SIZES; i++) {
mutex_lock(&lgr->rmbs_lock);
list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list)
smcr_buf_unmap_link(buf_desc, true, lnk);
mutex_unlock(&lgr->rmbs_lock);
mutex_lock(&lgr->sndbufs_lock);
list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i],
list)
smcr_buf_unmap_link(buf_desc, false, lnk);
mutex_unlock(&lgr->sndbufs_lock);
}
}
static void smcr_rtoken_clear_link(struct smc_link *lnk)
{
struct smc_link_group *lgr = lnk->lgr;
int i;
for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
lgr->rtokens[i][lnk->link_idx].rkey = 0;
lgr->rtokens[i][lnk->link_idx].dma_addr = 0;
}
}
net/smc: Resolve the race between SMC-R link access and clear We encountered some crashes caused by the race between SMC-R link access and link clear that triggered by abnormal link group termination, such as port error. Here is an example of this kind of crashes: BUG: kernel NULL pointer dereference, address: 0000000000000000 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_llc_flow_initiate+0x44/0x190 [smc] Call Trace: <TASK> ? __smc_buf_create+0x75a/0x950 [smc] smcr_lgr_reg_rmbs+0x2a/0xbf [smc] smc_listen_work+0xf72/0x1230 [smc] ? process_one_work+0x25c/0x600 process_one_work+0x25c/0x600 worker_thread+0x4f/0x3a0 ? process_one_work+0x600/0x600 kthread+0x15d/0x1a0 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() __smc_lgr_terminate() --------------------------------------------------------------- | smc_lgr_free() | |- smcr_link_clear() | |- memset(lnk, 0) smc_listen_rdma_reg() | |- smcr_lgr_reg_rmbs() | |- smc_llc_flow_initiate() | |- access lnk->lgr (panic) | These crashes are similarly caused by clearing SMC-R link resources when some functions is still accessing to them. This patch tries to fix the issue by introducing reference count of SMC-R links and ensuring that the sensitive resources of links won't be cleared until reference count reaches zero. The operation to the SMC-R link reference count can be concluded as follows: object [hold or initialized as 1] [put] -------------------------------------------------------------------- links smcr_link_init() smcr_link_clear() connections smc_conn_create() smc_conn_free() Through this way, the clear of SMC-R links is later than the free of all the smc connections above it, thus avoiding the unsafe reference to SMC-R links. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:42 +03:00
static void __smcr_link_clear(struct smc_link *lnk)
{
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
struct smc_link_group *lgr = lnk->lgr;
struct smc_ib_device *smcibdev;
smc_wr_free_link_mem(lnk);
smc_ibdev_cnt_dec(lnk);
put_device(&lnk->smcibdev->ibdev->dev);
smcibdev = lnk->smcibdev;
memset(lnk, 0, sizeof(struct smc_link));
lnk->state = SMC_LNK_UNUSED;
if (!atomic_dec_return(&smcibdev->lnk_cnt))
wake_up(&smcibdev->lnks_deleted);
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
smc_lgr_put(lgr); /* lgr_hold in smcr_link_init() */
}
net/smc: Resolve the race between SMC-R link access and clear We encountered some crashes caused by the race between SMC-R link access and link clear that triggered by abnormal link group termination, such as port error. Here is an example of this kind of crashes: BUG: kernel NULL pointer dereference, address: 0000000000000000 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_llc_flow_initiate+0x44/0x190 [smc] Call Trace: <TASK> ? __smc_buf_create+0x75a/0x950 [smc] smcr_lgr_reg_rmbs+0x2a/0xbf [smc] smc_listen_work+0xf72/0x1230 [smc] ? process_one_work+0x25c/0x600 process_one_work+0x25c/0x600 worker_thread+0x4f/0x3a0 ? process_one_work+0x600/0x600 kthread+0x15d/0x1a0 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() __smc_lgr_terminate() --------------------------------------------------------------- | smc_lgr_free() | |- smcr_link_clear() | |- memset(lnk, 0) smc_listen_rdma_reg() | |- smcr_lgr_reg_rmbs() | |- smc_llc_flow_initiate() | |- access lnk->lgr (panic) | These crashes are similarly caused by clearing SMC-R link resources when some functions is still accessing to them. This patch tries to fix the issue by introducing reference count of SMC-R links and ensuring that the sensitive resources of links won't be cleared until reference count reaches zero. The operation to the SMC-R link reference count can be concluded as follows: object [hold or initialized as 1] [put] -------------------------------------------------------------------- links smcr_link_init() smcr_link_clear() connections smc_conn_create() smc_conn_free() Through this way, the clear of SMC-R links is later than the free of all the smc connections above it, thus avoiding the unsafe reference to SMC-R links. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:42 +03:00
/* must be called under lgr->llc_conf_mutex lock */
void smcr_link_clear(struct smc_link *lnk, bool log)
{
if (!lnk->lgr || lnk->clearing ||
lnk->state == SMC_LNK_UNUSED)
return;
lnk->clearing = 1;
lnk->peer_qpn = 0;
smc_llc_link_clear(lnk, log);
smcr_buf_unmap_lgr(lnk);
smcr_rtoken_clear_link(lnk);
smc_ib_modify_qp_error(lnk);
smc_wr_free_link(lnk);
smc_ib_destroy_queue_pair(lnk);
smc_ib_dealloc_protection_domain(lnk);
smcr_link_put(lnk); /* theoretically last link_put */
}
void smcr_link_hold(struct smc_link *lnk)
{
refcount_inc(&lnk->refcnt);
}
void smcr_link_put(struct smc_link *lnk)
{
if (refcount_dec_and_test(&lnk->refcnt))
__smcr_link_clear(lnk);
}
static void smcr_buf_free(struct smc_link_group *lgr, bool is_rmb,
struct smc_buf_desc *buf_desc)
{
int i;
for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
smcr_buf_unmap_link(buf_desc, is_rmb, &lgr->lnk[i]);
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
if (!buf_desc->is_vm && buf_desc->pages)
__free_pages(buf_desc->pages, buf_desc->order);
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
else if (buf_desc->is_vm && buf_desc->cpu_addr)
vfree(buf_desc->cpu_addr);
kfree(buf_desc);
}
static void smcd_buf_free(struct smc_link_group *lgr, bool is_dmb,
struct smc_buf_desc *buf_desc)
{
if (is_dmb) {
/* restore original buf len */
buf_desc->len += sizeof(struct smcd_cdc_msg);
smc_ism_unregister_dmb(lgr->smcd, buf_desc);
} else {
kfree(buf_desc->cpu_addr);
}
kfree(buf_desc);
}
static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
struct smc_buf_desc *buf_desc)
{
if (lgr->is_smcd)
smcd_buf_free(lgr, is_rmb, buf_desc);
else
smcr_buf_free(lgr, is_rmb, buf_desc);
}
static void __smc_lgr_free_bufs(struct smc_link_group *lgr, bool is_rmb)
{
struct smc_buf_desc *buf_desc, *bf_desc;
struct list_head *buf_list;
int i;
for (i = 0; i < SMC_RMBE_SIZES; i++) {
if (is_rmb)
buf_list = &lgr->rmbs[i];
else
buf_list = &lgr->sndbufs[i];
list_for_each_entry_safe(buf_desc, bf_desc, buf_list,
list) {
list_del(&buf_desc->list);
smc_buf_free(lgr, is_rmb, buf_desc);
}
}
}
static void smc_lgr_free_bufs(struct smc_link_group *lgr)
{
/* free send buffers */
__smc_lgr_free_bufs(lgr, false);
/* free rmbs */
__smc_lgr_free_bufs(lgr, true);
}
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
/* won't be freed until no one accesses to lgr anymore */
static void __smc_lgr_free(struct smc_link_group *lgr)
{
smc_lgr_free_bufs(lgr);
if (lgr->is_smcd) {
if (!atomic_dec_return(&lgr->smcd->lgr_cnt))
wake_up(&lgr->smcd->lgrs_deleted);
} else {
smc_wr_free_lgr_mem(lgr);
if (!atomic_dec_return(&lgr_cnt))
wake_up(&lgrs_deleted);
}
kfree(lgr);
}
/* remove a link group */
static void smc_lgr_free(struct smc_link_group *lgr)
{
int i;
if (!lgr->is_smcd) {
mutex_lock(&lgr->llc_conf_mutex);
for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
if (lgr->lnk[i].state != SMC_LNK_UNUSED)
smcr_link_clear(&lgr->lnk[i], false);
}
mutex_unlock(&lgr->llc_conf_mutex);
smc_llc_lgr_clear(lgr);
}
destroy_workqueue(lgr->tx_wq);
if (lgr->is_smcd) {
smc_ism_put_vlan(lgr->smcd, lgr->vlan_id);
put_device(&lgr->smcd->dev);
}
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
smc_lgr_put(lgr); /* theoretically last lgr_put */
}
void smc_lgr_hold(struct smc_link_group *lgr)
{
refcount_inc(&lgr->refcnt);
}
void smc_lgr_put(struct smc_link_group *lgr)
{
if (refcount_dec_and_test(&lgr->refcnt))
__smc_lgr_free(lgr);
}
static void smc_sk_wake_ups(struct smc_sock *smc)
{
smc->sk.sk_write_space(&smc->sk);
smc->sk.sk_data_ready(&smc->sk);
smc->sk.sk_state_change(&smc->sk);
}
/* kill a connection */
static void smc_conn_kill(struct smc_connection *conn, bool soft)
{
struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
if (conn->lgr->is_smcd && conn->lgr->peer_shutdown)
conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
else
smc_close_abort(conn);
conn->killed = 1;
smc->sk.sk_err = ECONNABORTED;
smc_sk_wake_ups(smc);
if (conn->lgr->is_smcd) {
smc_ism_unset_conn(conn);
if (soft)
tasklet_kill(&conn->rx_tsklet);
else
tasklet_unlock_wait(&conn->rx_tsklet);
} else {
net/smc: fix kernel panic caused by race of smc_sock A crash occurs when smc_cdc_tx_handler() tries to access smc_sock but smc_release() has already freed it. [ 4570.695099] BUG: unable to handle page fault for address: 000000002eae9e88 [ 4570.696048] #PF: supervisor write access in kernel mode [ 4570.696728] #PF: error_code(0x0002) - not-present page [ 4570.697401] PGD 0 P4D 0 [ 4570.697716] Oops: 0002 [#1] PREEMPT SMP NOPTI [ 4570.698228] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.16.0-rc4+ #111 [ 4570.699013] Hardware name: Alibaba Cloud Alibaba Cloud ECS, BIOS 8c24b4c 04/0 [ 4570.699933] RIP: 0010:_raw_spin_lock+0x1a/0x30 <...> [ 4570.711446] Call Trace: [ 4570.711746] <IRQ> [ 4570.711992] smc_cdc_tx_handler+0x41/0xc0 [ 4570.712470] smc_wr_tx_tasklet_fn+0x213/0x560 [ 4570.712981] ? smc_cdc_tx_dismisser+0x10/0x10 [ 4570.713489] tasklet_action_common.isra.17+0x66/0x140 [ 4570.714083] __do_softirq+0x123/0x2f4 [ 4570.714521] irq_exit_rcu+0xc4/0xf0 [ 4570.714934] common_interrupt+0xba/0xe0 Though smc_cdc_tx_handler() checked the existence of smc connection, smc_release() may have already dismissed and released the smc socket before smc_cdc_tx_handler() further visits it. smc_cdc_tx_handler() |smc_release() if (!conn) | | |smc_cdc_tx_dismiss_slots() | smc_cdc_tx_dismisser() | |sock_put(&smc->sk) <- last sock_put, | smc_sock freed bh_lock_sock(&smc->sk) (panic) | To make sure we won't receive any CDC messages after we free the smc_sock, add a refcount on the smc_connection for inflight CDC message(posted to the QP but haven't received related CQE), and don't release the smc_connection until all the inflight CDC messages haven been done, for both success or failed ones. Using refcount on CDC messages brings another problem: when the link is going to be destroyed, smcr_link_clear() will reset the QP, which then remove all the pending CQEs related to the QP in the CQ. To make sure all the CQEs will always come back so the refcount on the smc_connection can always reach 0, smc_ib_modify_qp_reset() was replaced by smc_ib_modify_qp_error(). And remove the timeout in smc_wr_tx_wait_no_pending_sends() since we need to wait for all pending WQEs done, or we may encounter use-after- free when handling CQEs. For IB device removal routine, we need to wait for all the QPs on that device been destroyed before we can destroy CQs on the device, or the refcount on smc_connection won't reach 0 and smc_sock cannot be released. Fixes: 5f08318f617b ("smc: connection data control (CDC)") Reported-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2021-12-28 12:03:25 +03:00
smc_cdc_wait_pend_tx_wr(conn);
}
smc_lgr_unregister_conn(conn);
smc_close_active_abort(smc);
}
static void smc_lgr_cleanup(struct smc_link_group *lgr)
{
if (lgr->is_smcd) {
smc_ism_signal_shutdown(lgr);
} else {
u32 rsn = lgr->llc_termination_rsn;
if (!rsn)
rsn = SMC_LLC_DEL_PROG_INIT_TERM;
smc_llc_send_link_delete_all(lgr, false, rsn);
smcr_lgr_link_deactivate_all(lgr);
}
}
/* terminate link group
* @soft: true if link group shutdown can take its time
* false if immediate link group shutdown is required
*/
static void __smc_lgr_terminate(struct smc_link_group *lgr, bool soft)
{
struct smc_connection *conn;
struct smc_sock *smc;
struct rb_node *node;
if (lgr->terminating)
return; /* lgr already terminating */
/* cancel free_work sync, will terminate when lgr->freeing is set */
cancel_delayed_work_sync(&lgr->free_work);
lgr->terminating = 1;
/* kill remaining link group connections */
read_lock_bh(&lgr->conns_lock);
node = rb_first(&lgr->conns_all);
while (node) {
read_unlock_bh(&lgr->conns_lock);
conn = rb_entry(node, struct smc_connection, alert_node);
smc = container_of(conn, struct smc_sock, conn);
sock_hold(&smc->sk); /* sock_put below */
lock_sock(&smc->sk);
smc_conn_kill(conn, soft);
release_sock(&smc->sk);
sock_put(&smc->sk); /* sock_hold above */
read_lock_bh(&lgr->conns_lock);
node = rb_first(&lgr->conns_all);
}
read_unlock_bh(&lgr->conns_lock);
smc_lgr_cleanup(lgr);
smc_lgr_free(lgr);
}
/* unlink link group and schedule termination */
void smc_lgr_terminate_sched(struct smc_link_group *lgr)
{
spinlock_t *lgr_lock;
smc_lgr_list_head(lgr, &lgr_lock);
spin_lock_bh(lgr_lock);
if (list_empty(&lgr->list) || lgr->terminating || lgr->freeing) {
spin_unlock_bh(lgr_lock);
return; /* lgr already terminating */
}
list_del_init(&lgr->list);
lgr->freeing = 1;
spin_unlock_bh(lgr_lock);
schedule_work(&lgr->terminate_work);
}
/* Called when peer lgr shutdown (regularly or abnormally) is received */
void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
{
struct smc_link_group *lgr, *l;
LIST_HEAD(lgr_free_list);
/* run common cleanup function and build free list */
spin_lock_bh(&dev->lgr_lock);
list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
if ((!peer_gid || lgr->peer_gid == peer_gid) &&
(vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
if (peer_gid) /* peer triggered termination */
lgr->peer_shutdown = 1;
list_move(&lgr->list, &lgr_free_list);
lgr->freeing = 1;
}
}
spin_unlock_bh(&dev->lgr_lock);
/* cancel the regular free workers and actually free lgrs */
list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
list_del_init(&lgr->list);
schedule_work(&lgr->terminate_work);
}
}
/* Called when an SMCD device is removed or the smc module is unloaded */
void smc_smcd_terminate_all(struct smcd_dev *smcd)
{
struct smc_link_group *lgr, *lg;
LIST_HEAD(lgr_free_list);
spin_lock_bh(&smcd->lgr_lock);
list_splice_init(&smcd->lgr_list, &lgr_free_list);
list_for_each_entry(lgr, &lgr_free_list, list)
lgr->freeing = 1;
spin_unlock_bh(&smcd->lgr_lock);
list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
list_del_init(&lgr->list);
__smc_lgr_terminate(lgr, false);
}
if (atomic_read(&smcd->lgr_cnt))
wait_event(smcd->lgrs_deleted, !atomic_read(&smcd->lgr_cnt));
}
/* Called when an SMCR device is removed or the smc module is unloaded.
* If smcibdev is given, all SMCR link groups using this device are terminated.
* If smcibdev is NULL, all SMCR link groups are terminated.
*/
void smc_smcr_terminate_all(struct smc_ib_device *smcibdev)
{
struct smc_link_group *lgr, *lg;
LIST_HEAD(lgr_free_list);
int i;
spin_lock_bh(&smc_lgr_list.lock);
if (!smcibdev) {
list_splice_init(&smc_lgr_list.list, &lgr_free_list);
list_for_each_entry(lgr, &lgr_free_list, list)
lgr->freeing = 1;
} else {
list_for_each_entry_safe(lgr, lg, &smc_lgr_list.list, list) {
for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
if (lgr->lnk[i].smcibdev == smcibdev)
smcr_link_down_cond_sched(&lgr->lnk[i]);
}
}
}
spin_unlock_bh(&smc_lgr_list.lock);
list_for_each_entry_safe(lgr, lg, &lgr_free_list, list) {
list_del_init(&lgr->list);
smc_llc_set_termination_rsn(lgr, SMC_LLC_DEL_OP_INIT_TERM);
__smc_lgr_terminate(lgr, false);
net/smc: fix kernel panic caused by race of smc_sock A crash occurs when smc_cdc_tx_handler() tries to access smc_sock but smc_release() has already freed it. [ 4570.695099] BUG: unable to handle page fault for address: 000000002eae9e88 [ 4570.696048] #PF: supervisor write access in kernel mode [ 4570.696728] #PF: error_code(0x0002) - not-present page [ 4570.697401] PGD 0 P4D 0 [ 4570.697716] Oops: 0002 [#1] PREEMPT SMP NOPTI [ 4570.698228] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.16.0-rc4+ #111 [ 4570.699013] Hardware name: Alibaba Cloud Alibaba Cloud ECS, BIOS 8c24b4c 04/0 [ 4570.699933] RIP: 0010:_raw_spin_lock+0x1a/0x30 <...> [ 4570.711446] Call Trace: [ 4570.711746] <IRQ> [ 4570.711992] smc_cdc_tx_handler+0x41/0xc0 [ 4570.712470] smc_wr_tx_tasklet_fn+0x213/0x560 [ 4570.712981] ? smc_cdc_tx_dismisser+0x10/0x10 [ 4570.713489] tasklet_action_common.isra.17+0x66/0x140 [ 4570.714083] __do_softirq+0x123/0x2f4 [ 4570.714521] irq_exit_rcu+0xc4/0xf0 [ 4570.714934] common_interrupt+0xba/0xe0 Though smc_cdc_tx_handler() checked the existence of smc connection, smc_release() may have already dismissed and released the smc socket before smc_cdc_tx_handler() further visits it. smc_cdc_tx_handler() |smc_release() if (!conn) | | |smc_cdc_tx_dismiss_slots() | smc_cdc_tx_dismisser() | |sock_put(&smc->sk) <- last sock_put, | smc_sock freed bh_lock_sock(&smc->sk) (panic) | To make sure we won't receive any CDC messages after we free the smc_sock, add a refcount on the smc_connection for inflight CDC message(posted to the QP but haven't received related CQE), and don't release the smc_connection until all the inflight CDC messages haven been done, for both success or failed ones. Using refcount on CDC messages brings another problem: when the link is going to be destroyed, smcr_link_clear() will reset the QP, which then remove all the pending CQEs related to the QP in the CQ. To make sure all the CQEs will always come back so the refcount on the smc_connection can always reach 0, smc_ib_modify_qp_reset() was replaced by smc_ib_modify_qp_error(). And remove the timeout in smc_wr_tx_wait_no_pending_sends() since we need to wait for all pending WQEs done, or we may encounter use-after- free when handling CQEs. For IB device removal routine, we need to wait for all the QPs on that device been destroyed before we can destroy CQs on the device, or the refcount on smc_connection won't reach 0 and smc_sock cannot be released. Fixes: 5f08318f617b ("smc: connection data control (CDC)") Reported-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2021-12-28 12:03:25 +03:00
}
if (smcibdev) {
if (atomic_read(&smcibdev->lnk_cnt))
wait_event(smcibdev->lnks_deleted,
!atomic_read(&smcibdev->lnk_cnt));
} else {
if (atomic_read(&lgr_cnt))
wait_event(lgrs_deleted, !atomic_read(&lgr_cnt));
}
}
/* set new lgr type and clear all asymmetric link tagging */
void smcr_lgr_set_type(struct smc_link_group *lgr, enum smc_lgr_type new_type)
{
char *lgr_type = "";
int i;
for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++)
if (smc_link_usable(&lgr->lnk[i]))
lgr->lnk[i].link_is_asym = false;
if (lgr->type == new_type)
return;
lgr->type = new_type;
switch (lgr->type) {
case SMC_LGR_NONE:
lgr_type = "NONE";
break;
case SMC_LGR_SINGLE:
lgr_type = "SINGLE";
break;
case SMC_LGR_SYMMETRIC:
lgr_type = "SYMMETRIC";
break;
case SMC_LGR_ASYMMETRIC_PEER:
lgr_type = "ASYMMETRIC_PEER";
break;
case SMC_LGR_ASYMMETRIC_LOCAL:
lgr_type = "ASYMMETRIC_LOCAL";
break;
}
pr_warn_ratelimited("smc: SMC-R lg %*phN net %llu state changed: "
"%s, pnetid %.16s\n", SMC_LGR_ID_SIZE, &lgr->id,
lgr->net->net_cookie, lgr_type, lgr->pnet_id);
}
/* set new lgr type and tag a link as asymmetric */
void smcr_lgr_set_type_asym(struct smc_link_group *lgr,
enum smc_lgr_type new_type, int asym_lnk_idx)
{
smcr_lgr_set_type(lgr, new_type);
lgr->lnk[asym_lnk_idx].link_is_asym = true;
}
/* abort connection, abort_work scheduled from tasklet context */
static void smc_conn_abort_work(struct work_struct *work)
{
struct smc_connection *conn = container_of(work,
struct smc_connection,
abort_work);
struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
lock_sock(&smc->sk);
smc_conn_kill(conn, true);
release_sock(&smc->sk);
sock_put(&smc->sk); /* sock_hold done by schedulers of abort_work */
}
void smcr_port_add(struct smc_ib_device *smcibdev, u8 ibport)
{
struct smc_link_group *lgr, *n;
list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
struct smc_link *link;
if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
SMC_MAX_PNETID_LEN) ||
lgr->type == SMC_LGR_SYMMETRIC ||
lgr->type == SMC_LGR_ASYMMETRIC_PEER ||
!rdma_dev_access_netns(smcibdev->ibdev, lgr->net))
continue;
/* trigger local add link processing */
link = smc_llc_usable_link(lgr);
if (link)
smc_llc_add_link_local(link);
}
}
/* link is down - switch connections to alternate link,
* must be called under lgr->llc_conf_mutex lock
*/
static void smcr_link_down(struct smc_link *lnk)
{
struct smc_link_group *lgr = lnk->lgr;
struct smc_link *to_lnk;
int del_link_id;
if (!lgr || lnk->state == SMC_LNK_UNUSED || list_empty(&lgr->list))
return;
to_lnk = smc_switch_conns(lgr, lnk, true);
if (!to_lnk) { /* no backup link available */
smcr_link_clear(lnk, true);
return;
}
smcr_lgr_set_type(lgr, SMC_LGR_SINGLE);
del_link_id = lnk->link_id;
if (lgr->role == SMC_SERV) {
/* trigger local delete link processing */
smc_llc_srv_delete_link_local(to_lnk, del_link_id);
} else {
if (lgr->llc_flow_lcl.type != SMC_LLC_FLOW_NONE) {
/* another llc task is ongoing */
mutex_unlock(&lgr->llc_conf_mutex);
wait_event_timeout(lgr->llc_flow_waiter,
(list_empty(&lgr->list) ||
lgr->llc_flow_lcl.type == SMC_LLC_FLOW_NONE),
SMC_LLC_WAIT_TIME);
mutex_lock(&lgr->llc_conf_mutex);
}
if (!list_empty(&lgr->list)) {
smc_llc_send_delete_link(to_lnk, del_link_id,
SMC_LLC_REQ, true,
SMC_LLC_DEL_LOST_PATH);
smcr_link_clear(lnk, true);
}
wake_up(&lgr->llc_flow_waiter); /* wake up next waiter */
}
}
/* must be called under lgr->llc_conf_mutex lock */
void smcr_link_down_cond(struct smc_link *lnk)
{
if (smc_link_downing(&lnk->state)) {
trace_smcr_link_down(lnk, __builtin_return_address(0));
smcr_link_down(lnk);
}
}
/* will get the lgr->llc_conf_mutex lock */
void smcr_link_down_cond_sched(struct smc_link *lnk)
{
if (smc_link_downing(&lnk->state)) {
trace_smcr_link_down(lnk, __builtin_return_address(0));
schedule_work(&lnk->link_down_wrk);
}
}
void smcr_port_err(struct smc_ib_device *smcibdev, u8 ibport)
{
struct smc_link_group *lgr, *n;
int i;
list_for_each_entry_safe(lgr, n, &smc_lgr_list.list, list) {
if (strncmp(smcibdev->pnetid[ibport - 1], lgr->pnet_id,
SMC_MAX_PNETID_LEN))
continue; /* lgr is not affected */
if (list_empty(&lgr->list))
continue;
for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
struct smc_link *lnk = &lgr->lnk[i];
if (smc_link_usable(lnk) &&
lnk->smcibdev == smcibdev && lnk->ibport == ibport)
smcr_link_down_cond_sched(lnk);
}
}
}
static void smc_link_down_work(struct work_struct *work)
{
struct smc_link *link = container_of(work, struct smc_link,
link_down_wrk);
struct smc_link_group *lgr = link->lgr;
if (list_empty(&lgr->list))
return;
wake_up_all(&lgr->llc_msg_waiter);
mutex_lock(&lgr->llc_conf_mutex);
smcr_link_down(link);
mutex_unlock(&lgr->llc_conf_mutex);
}
static int smc_vlan_by_tcpsk_walk(struct net_device *lower_dev,
struct netdev_nested_priv *priv)
{
unsigned short *vlan_id = (unsigned short *)priv->data;
if (is_vlan_dev(lower_dev)) {
*vlan_id = vlan_dev_vlan_id(lower_dev);
return 1;
}
return 0;
}
/* Determine vlan of internal TCP socket. */
int smc_vlan_by_tcpsk(struct socket *clcsock, struct smc_init_info *ini)
{
struct dst_entry *dst = sk_dst_get(clcsock->sk);
struct netdev_nested_priv priv;
struct net_device *ndev;
int rc = 0;
ini->vlan_id = 0;
if (!dst) {
rc = -ENOTCONN;
goto out;
}
if (!dst->dev) {
rc = -ENODEV;
goto out_rel;
}
ndev = dst->dev;
if (is_vlan_dev(ndev)) {
ini->vlan_id = vlan_dev_vlan_id(ndev);
goto out_rel;
}
priv.data = (void *)&ini->vlan_id;
rtnl_lock();
netdev_walk_all_lower_dev(ndev, smc_vlan_by_tcpsk_walk, &priv);
rtnl_unlock();
out_rel:
dst_release(dst);
out:
return rc;
}
static bool smcr_lgr_match(struct smc_link_group *lgr, u8 smcr_version,
u8 peer_systemid[],
u8 peer_gid[],
u8 peer_mac_v1[],
enum smc_lgr_role role, u32 clcqpn,
struct net *net)
{
struct smc_link *lnk;
int i;
if (memcmp(lgr->peer_systemid, peer_systemid, SMC_SYSTEMID_LEN) ||
lgr->role != role)
return false;
for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
lnk = &lgr->lnk[i];
if (!smc_link_active(lnk))
continue;
/* use verbs API to check netns, instead of lgr->net */
if (!rdma_dev_access_netns(lnk->smcibdev->ibdev, net))
return false;
if ((lgr->role == SMC_SERV || lnk->peer_qpn == clcqpn) &&
!memcmp(lnk->peer_gid, peer_gid, SMC_GID_SIZE) &&
(smcr_version == SMC_V2 ||
!memcmp(lnk->peer_mac, peer_mac_v1, ETH_ALEN)))
return true;
}
return false;
}
static bool smcd_lgr_match(struct smc_link_group *lgr,
struct smcd_dev *smcismdev, u64 peer_gid)
{
return lgr->peer_gid == peer_gid && lgr->smcd == smcismdev;
}
/* create a new SMC connection (and a new link group if necessary) */
int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
{
struct smc_connection *conn = &smc->conn;
struct net *net = sock_net(&smc->sk);
struct list_head *lgr_list;
struct smc_link_group *lgr;
enum smc_lgr_role role;
spinlock_t *lgr_lock;
int rc = 0;
lgr_list = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_list :
&smc_lgr_list.list;
lgr_lock = ini->is_smcd ? &ini->ism_dev[ini->ism_selected]->lgr_lock :
&smc_lgr_list.lock;
ini->first_contact_local = 1;
role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
if (role == SMC_CLNT && ini->first_contact_peer)
/* create new link group as well */
goto create;
/* determine if an existing link group can be reused */
spin_lock_bh(lgr_lock);
list_for_each_entry(lgr, lgr_list, list) {
write_lock_bh(&lgr->conns_lock);
if ((ini->is_smcd ?
smcd_lgr_match(lgr, ini->ism_dev[ini->ism_selected],
ini->ism_peer_gid[ini->ism_selected]) :
smcr_lgr_match(lgr, ini->smcr_version,
ini->peer_systemid,
ini->peer_gid, ini->peer_mac, role,
ini->ib_clcqpn, net)) &&
!lgr->sync_err &&
(ini->smcd_version == SMC_V2 ||
lgr->vlan_id == ini->vlan_id) &&
net/smc: fix dmb buffer shortage There is a current limit of 1920 registered dmb buffers per ISM device for smc-d. One link group can contain 255 connections, each connection is using one dmb buffer. When the connection is closed then the registered buffer is held in a queue and is reused by the next connection. When a link group is 'full' then another link group is created and uses an own buffer pool. The link groups are added to a list using list_add() which puts a new link group to the first position in the list. In the situation that many connections are opened (>1920) and a few of them stay open while others are closed quickly we end up with at least 8 link groups. For a new connection a matching link group is looked up, iterating over the list of link groups. The trailing 7 link groups all have registered dmb buffers which could be reused, while the first link group has only a few dmb buffers and then hit the 1920 limit. Because the first link group is not full (255 connection limit not reached) it is chosen and finally the connection falls back to TCP because there is no dmb buffer available in this link group. There are multiple ways to fix that: using list_add_tail() allows to scan older link groups first for free buffers which ensures that buffers are reused first. This fixes the problem for smc-r link groups as well. For smc-d there is an even better way to address this problem because smc-d does not have the 255 connections per link group limit. So fix the problem for smc-d by allowing large link groups. Fixes: c6ba7c9ba43d ("net/smc: add base infrastructure for SMC-D and ISM") Reviewed-by: Ursula Braun <ubraun@linux.ibm.com> Signed-off-by: Karsten Graul <kgraul@linux.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-07-20 17:24:29 +03:00
(role == SMC_CLNT || ini->is_smcd ||
(lgr->conns_num < SMC_RMBS_PER_LGR_MAX &&
!bitmap_full(lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX)))) {
/* link group found */
ini->first_contact_local = 0;
conn->lgr = lgr;
rc = smc_lgr_register_conn(conn, false);
write_unlock_bh(&lgr->conns_lock);
if (!rc && delayed_work_pending(&lgr->free_work))
cancel_delayed_work(&lgr->free_work);
break;
}
write_unlock_bh(&lgr->conns_lock);
}
spin_unlock_bh(lgr_lock);
if (rc)
return rc;
if (role == SMC_CLNT && !ini->first_contact_peer &&
ini->first_contact_local) {
/* Server reuses a link group, but Client wants to start
* a new one
* send out_of_sync decline, reason synchr. error
*/
return SMC_CLC_DECL_SYNCERR;
}
create:
if (ini->first_contact_local) {
rc = smc_lgr_create(smc, ini);
if (rc)
goto out;
lgr = conn->lgr;
write_lock_bh(&lgr->conns_lock);
rc = smc_lgr_register_conn(conn, true);
write_unlock_bh(&lgr->conns_lock);
if (rc) {
smc_lgr_cleanup_early(lgr);
goto out;
}
}
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
smc_lgr_hold(conn->lgr); /* lgr_put in smc_conn_free() */
net/smc: Resolve the race between SMC-R link access and clear We encountered some crashes caused by the race between SMC-R link access and link clear that triggered by abnormal link group termination, such as port error. Here is an example of this kind of crashes: BUG: kernel NULL pointer dereference, address: 0000000000000000 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_llc_flow_initiate+0x44/0x190 [smc] Call Trace: <TASK> ? __smc_buf_create+0x75a/0x950 [smc] smcr_lgr_reg_rmbs+0x2a/0xbf [smc] smc_listen_work+0xf72/0x1230 [smc] ? process_one_work+0x25c/0x600 process_one_work+0x25c/0x600 worker_thread+0x4f/0x3a0 ? process_one_work+0x600/0x600 kthread+0x15d/0x1a0 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() __smc_lgr_terminate() --------------------------------------------------------------- | smc_lgr_free() | |- smcr_link_clear() | |- memset(lnk, 0) smc_listen_rdma_reg() | |- smcr_lgr_reg_rmbs() | |- smc_llc_flow_initiate() | |- access lnk->lgr (panic) | These crashes are similarly caused by clearing SMC-R link resources when some functions is still accessing to them. This patch tries to fix the issue by introducing reference count of SMC-R links and ensuring that the sensitive resources of links won't be cleared until reference count reaches zero. The operation to the SMC-R link reference count can be concluded as follows: object [hold or initialized as 1] [put] -------------------------------------------------------------------- links smcr_link_init() smcr_link_clear() connections smc_conn_create() smc_conn_free() Through this way, the clear of SMC-R links is later than the free of all the smc connections above it, thus avoiding the unsafe reference to SMC-R links. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:42 +03:00
if (!conn->lgr->is_smcd)
smcr_link_hold(conn->lnk); /* link_put in smc_conn_free() */
net/smc: Resolve the race between link group access and termination We encountered some crashes caused by the race between the access and the termination of link groups. Here are some of panic stacks we met: 1) Race between smc_clc_wait_msg() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002f0 Workqueue: smc_hs_wq smc_listen_work [smc] RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc] Call Trace: <TASK> ? smc_clc_send_accept+0x45/0xa0 [smc] ? smc_clc_send_accept+0x45/0xa0 [smc] smc_listen_work+0x783/0x1220 [smc] ? finish_task_switch+0xc4/0x2e0 ? process_one_work+0x1ad/0x3c0 process_one_work+0x1ad/0x3c0 worker_thread+0x4c/0x390 ? rescuer_thread+0x320/0x320 kthread+0x149/0x190 ? set_kthread_struct+0x40/0x40 ret_from_fork+0x1f/0x30 </TASK> smc_listen_work() abnormal case like port error --------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL smc_clc_wait_msg() | |- access conn->lgr (panic) | 2) Race between smc_setsockopt() and __smc_lgr_terminate() BUG: kernel NULL pointer dereference, address: 00000000000002e8 RIP: 0010:smc_setsockopt+0x17a/0x280 [smc] Call Trace: <TASK> __sys_setsockopt+0xfc/0x190 __x64_sys_setsockopt+0x20/0x30 do_syscall_64+0x34/0x90 entry_SYSCALL_64_after_hwframe+0x44/0xae </TASK> smc_setsockopt() abnormal case like port error -------------------------------------------------------------- | __smc_lgr_terminate() | |- smc_conn_kill() | |- smc_lgr_unregister_conn() | |- set conn->lgr = NULL mod_delayed_work() | |- access conn->lgr (panic) | There are some other panic places and they are caused by the similar reason as described above, which is accessing link group after termination, thus getting a NULL pointer or invalid resource. Currently, there seems to be no synchronization between the link group access and a sudden termination of it. This patch tries to fix this by introducing reference count of link group and not freeing link group until reference count is zero. Link group might be referred to by links or smc connections. So the operation to the link group reference count can be concluded as follows: object [hold or initialized as 1] [put] ------------------------------------------------------------------- link group smc_lgr_create() smc_lgr_free() connections smc_conn_create() smc_conn_free() links smcr_link_init() smcr_link_clear() Througth this way, we extend the life cycle of link group and ensure it is longer than the life cycle of connections and links above it, so that avoid invalid access to link group after its termination. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-01-13 11:36:40 +03:00
conn->freed = 0;
conn->local_tx_ctrl.common.type = SMC_CDC_MSG_TYPE;
conn->local_tx_ctrl.len = SMC_WR_TX_SIZE;
conn->urg_state = SMC_URG_READ;
net/smc: fix kernel panic caused by race of smc_sock A crash occurs when smc_cdc_tx_handler() tries to access smc_sock but smc_release() has already freed it. [ 4570.695099] BUG: unable to handle page fault for address: 000000002eae9e88 [ 4570.696048] #PF: supervisor write access in kernel mode [ 4570.696728] #PF: error_code(0x0002) - not-present page [ 4570.697401] PGD 0 P4D 0 [ 4570.697716] Oops: 0002 [#1] PREEMPT SMP NOPTI [ 4570.698228] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.16.0-rc4+ #111 [ 4570.699013] Hardware name: Alibaba Cloud Alibaba Cloud ECS, BIOS 8c24b4c 04/0 [ 4570.699933] RIP: 0010:_raw_spin_lock+0x1a/0x30 <...> [ 4570.711446] Call Trace: [ 4570.711746] <IRQ> [ 4570.711992] smc_cdc_tx_handler+0x41/0xc0 [ 4570.712470] smc_wr_tx_tasklet_fn+0x213/0x560 [ 4570.712981] ? smc_cdc_tx_dismisser+0x10/0x10 [ 4570.713489] tasklet_action_common.isra.17+0x66/0x140 [ 4570.714083] __do_softirq+0x123/0x2f4 [ 4570.714521] irq_exit_rcu+0xc4/0xf0 [ 4570.714934] common_interrupt+0xba/0xe0 Though smc_cdc_tx_handler() checked the existence of smc connection, smc_release() may have already dismissed and released the smc socket before smc_cdc_tx_handler() further visits it. smc_cdc_tx_handler() |smc_release() if (!conn) | | |smc_cdc_tx_dismiss_slots() | smc_cdc_tx_dismisser() | |sock_put(&smc->sk) <- last sock_put, | smc_sock freed bh_lock_sock(&smc->sk) (panic) | To make sure we won't receive any CDC messages after we free the smc_sock, add a refcount on the smc_connection for inflight CDC message(posted to the QP but haven't received related CQE), and don't release the smc_connection until all the inflight CDC messages haven been done, for both success or failed ones. Using refcount on CDC messages brings another problem: when the link is going to be destroyed, smcr_link_clear() will reset the QP, which then remove all the pending CQEs related to the QP in the CQ. To make sure all the CQEs will always come back so the refcount on the smc_connection can always reach 0, smc_ib_modify_qp_reset() was replaced by smc_ib_modify_qp_error(). And remove the timeout in smc_wr_tx_wait_no_pending_sends() since we need to wait for all pending WQEs done, or we may encounter use-after- free when handling CQEs. For IB device removal routine, we need to wait for all the QPs on that device been destroyed before we can destroy CQs on the device, or the refcount on smc_connection won't reach 0 and smc_sock cannot be released. Fixes: 5f08318f617b ("smc: connection data control (CDC)") Reported-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2021-12-28 12:03:25 +03:00
init_waitqueue_head(&conn->cdc_pend_tx_wq);
INIT_WORK(&smc->conn.abort_work, smc_conn_abort_work);
if (ini->is_smcd) {
conn->rx_off = sizeof(struct smcd_cdc_msg);
smcd_cdc_rx_init(conn); /* init tasklet for this conn */
} else {
conn->rx_off = 0;
}
#ifndef KERNEL_HAS_ATOMIC64
spin_lock_init(&conn->acurs_lock);
#endif
out:
return rc;
}
#define SMCD_DMBE_SIZES 6 /* 0 -> 16KB, 1 -> 32KB, .. 6 -> 1MB */
#define SMCR_RMBE_SIZES 5 /* 0 -> 16KB, 1 -> 32KB, .. 5 -> 512KB */
/* convert the RMB size into the compressed notation (minimum 16K, see
* SMCD/R_DMBE_SIZES.
* In contrast to plain ilog2, this rounds towards the next power of 2,
* so the socket application gets at least its desired sndbuf / rcvbuf size.
*/
static u8 smc_compress_bufsize(int size, bool is_smcd, bool is_rmb)
{
const unsigned int max_scat = SG_MAX_SINGLE_ALLOC * PAGE_SIZE;
u8 compressed;
if (size <= SMC_BUF_MIN_SIZE)
return 0;
size = (size - 1) >> 14; /* convert to 16K multiple */
compressed = min_t(u8, ilog2(size) + 1,
is_smcd ? SMCD_DMBE_SIZES : SMCR_RMBE_SIZES);
if (!is_smcd && is_rmb)
/* RMBs are backed by & limited to max size of scatterlists */
compressed = min_t(u8, compressed, ilog2(max_scat >> 14));
return compressed;
}
/* convert the RMB size from compressed notation into integer */
int smc_uncompress_bufsize(u8 compressed)
{
u32 size;
size = 0x00000001 << (((int)compressed) + 14);
return (int)size;
}
/* try to reuse a sndbuf or rmb description slot for a certain
* buffer size; if not available, return NULL
*/
static struct smc_buf_desc *smc_buf_get_slot(int compressed_bufsize,
struct mutex *lock,
struct list_head *buf_list)
{
struct smc_buf_desc *buf_slot;
mutex_lock(lock);
list_for_each_entry(buf_slot, buf_list, list) {
if (cmpxchg(&buf_slot->used, 0, 1) == 0) {
mutex_unlock(lock);
return buf_slot;
}
}
mutex_unlock(lock);
return NULL;
}
/* one of the conditions for announcing a receiver's current window size is
* that it "results in a minimum increase in the window size of 10% of the
* receive buffer space" [RFC7609]
*/
static inline int smc_rmb_wnd_update_limit(int rmbe_size)
{
return max_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2);
}
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
/* map an buf to a link */
static int smcr_buf_map_link(struct smc_buf_desc *buf_desc, bool is_rmb,
struct smc_link *lnk)
{
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
int rc, i, nents, offset, buf_size, size, access_flags;
struct scatterlist *sg;
void *buf;
if (buf_desc->is_map_ib[lnk->link_idx])
return 0;
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
if (buf_desc->is_vm) {
buf = buf_desc->cpu_addr;
buf_size = buf_desc->len;
offset = offset_in_page(buf_desc->cpu_addr);
nents = PAGE_ALIGN(buf_size + offset) / PAGE_SIZE;
} else {
nents = 1;
}
rc = sg_alloc_table(&buf_desc->sgt[lnk->link_idx], nents, GFP_KERNEL);
if (rc)
return rc;
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
if (buf_desc->is_vm) {
/* virtually contiguous buffer */
for_each_sg(buf_desc->sgt[lnk->link_idx].sgl, sg, nents, i) {
size = min_t(int, PAGE_SIZE - offset, buf_size);
sg_set_page(sg, vmalloc_to_page(buf), size, offset);
buf += size / sizeof(*buf);
buf_size -= size;
offset = 0;
}
} else {
/* physically contiguous buffer */
sg_set_buf(buf_desc->sgt[lnk->link_idx].sgl,
buf_desc->cpu_addr, buf_desc->len);
}
/* map sg table to DMA address */
rc = smc_ib_buf_map_sg(lnk, buf_desc,
is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
/* SMC protocol depends on mapping to one DMA address only */
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
if (rc != nents) {
rc = -EAGAIN;
goto free_table;
}
buf_desc->is_dma_need_sync |=
smc_ib_is_sg_need_sync(lnk, buf_desc) << lnk->link_idx;
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
if (is_rmb || buf_desc->is_vm) {
/* create a new memory region for the RMB or vzalloced sndbuf */
access_flags = is_rmb ?
IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
IB_ACCESS_LOCAL_WRITE;
rc = smc_ib_get_memory_region(lnk->roce_pd, access_flags,
buf_desc, lnk->link_idx);
if (rc)
goto buf_unmap;
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
smc_ib_sync_sg_for_device(lnk, buf_desc,
is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
}
buf_desc->is_map_ib[lnk->link_idx] = true;
return 0;
buf_unmap:
smc_ib_buf_unmap_sg(lnk, buf_desc,
is_rmb ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
free_table:
sg_free_table(&buf_desc->sgt[lnk->link_idx]);
return rc;
}
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
/* register a new buf on IB device, rmb or vzalloced sndbuf
* must be called under lgr->llc_conf_mutex lock
*/
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
int smcr_link_reg_buf(struct smc_link *link, struct smc_buf_desc *buf_desc)
{
if (list_empty(&link->lgr->list))
return -ENOLINK;
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
if (!buf_desc->is_reg_mr[link->link_idx]) {
/* register memory region for new buf */
if (buf_desc->is_vm)
buf_desc->mr[link->link_idx]->iova =
(uintptr_t)buf_desc->cpu_addr;
if (smc_wr_reg_send(link, buf_desc->mr[link->link_idx])) {
buf_desc->is_reg_err = true;
return -EFAULT;
}
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
buf_desc->is_reg_mr[link->link_idx] = true;
}
return 0;
}
static int _smcr_buf_map_lgr(struct smc_link *lnk, struct mutex *lock,
struct list_head *lst, bool is_rmb)
{
struct smc_buf_desc *buf_desc, *bf;
int rc = 0;
mutex_lock(lock);
list_for_each_entry_safe(buf_desc, bf, lst, list) {
if (!buf_desc->used)
continue;
rc = smcr_buf_map_link(buf_desc, is_rmb, lnk);
if (rc)
goto out;
}
out:
mutex_unlock(lock);
return rc;
}
/* map all used buffers of lgr for a new link */
int smcr_buf_map_lgr(struct smc_link *lnk)
{
struct smc_link_group *lgr = lnk->lgr;
int i, rc = 0;
for (i = 0; i < SMC_RMBE_SIZES; i++) {
rc = _smcr_buf_map_lgr(lnk, &lgr->rmbs_lock,
&lgr->rmbs[i], true);
if (rc)
return rc;
rc = _smcr_buf_map_lgr(lnk, &lgr->sndbufs_lock,
&lgr->sndbufs[i], false);
if (rc)
return rc;
}
return 0;
}
/* register all used buffers of lgr for a new link,
* must be called under lgr->llc_conf_mutex lock
*/
int smcr_buf_reg_lgr(struct smc_link *lnk)
{
struct smc_link_group *lgr = lnk->lgr;
struct smc_buf_desc *buf_desc, *bf;
int i, rc = 0;
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
/* reg all RMBs for a new link */
mutex_lock(&lgr->rmbs_lock);
for (i = 0; i < SMC_RMBE_SIZES; i++) {
list_for_each_entry_safe(buf_desc, bf, &lgr->rmbs[i], list) {
if (!buf_desc->used)
continue;
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
rc = smcr_link_reg_buf(lnk, buf_desc);
if (rc) {
mutex_unlock(&lgr->rmbs_lock);
return rc;
}
}
}
mutex_unlock(&lgr->rmbs_lock);
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
return rc;
/* reg all vzalloced sndbufs for a new link */
mutex_lock(&lgr->sndbufs_lock);
for (i = 0; i < SMC_RMBE_SIZES; i++) {
list_for_each_entry_safe(buf_desc, bf, &lgr->sndbufs[i], list) {
if (!buf_desc->used || !buf_desc->is_vm)
continue;
rc = smcr_link_reg_buf(lnk, buf_desc);
if (rc) {
mutex_unlock(&lgr->sndbufs_lock);
return rc;
}
}
}
mutex_unlock(&lgr->sndbufs_lock);
return rc;
}
static struct smc_buf_desc *smcr_new_buf_create(struct smc_link_group *lgr,
bool is_rmb, int bufsize)
{
struct smc_buf_desc *buf_desc;
/* try to alloc a new buffer */
buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
if (!buf_desc)
return ERR_PTR(-ENOMEM);
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
switch (lgr->buf_type) {
case SMCR_PHYS_CONT_BUFS:
case SMCR_MIXED_BUFS:
buf_desc->order = get_order(bufsize);
buf_desc->pages = alloc_pages(GFP_KERNEL | __GFP_NOWARN |
__GFP_NOMEMALLOC | __GFP_COMP |
__GFP_NORETRY | __GFP_ZERO,
buf_desc->order);
if (buf_desc->pages) {
buf_desc->cpu_addr =
(void *)page_address(buf_desc->pages);
buf_desc->len = bufsize;
buf_desc->is_vm = false;
break;
}
if (lgr->buf_type == SMCR_PHYS_CONT_BUFS)
goto out;
fallthrough; // try virtually continguous buf
case SMCR_VIRT_CONT_BUFS:
buf_desc->order = get_order(bufsize);
buf_desc->cpu_addr = vzalloc(PAGE_SIZE << buf_desc->order);
if (!buf_desc->cpu_addr)
goto out;
buf_desc->pages = NULL;
buf_desc->len = bufsize;
buf_desc->is_vm = true;
break;
}
return buf_desc;
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
out:
kfree(buf_desc);
return ERR_PTR(-EAGAIN);
}
/* map buf_desc on all usable links,
* unused buffers stay mapped as long as the link is up
*/
static int smcr_buf_map_usable_links(struct smc_link_group *lgr,
struct smc_buf_desc *buf_desc, bool is_rmb)
{
int i, rc = 0, cnt = 0;
/* protect against parallel link reconfiguration */
mutex_lock(&lgr->llc_conf_mutex);
for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
struct smc_link *lnk = &lgr->lnk[i];
if (!smc_link_usable(lnk))
continue;
if (smcr_buf_map_link(buf_desc, is_rmb, lnk)) {
rc = -ENOMEM;
goto out;
}
cnt++;
}
out:
mutex_unlock(&lgr->llc_conf_mutex);
if (!rc && !cnt)
rc = -EINVAL;
return rc;
}
static struct smc_buf_desc *smcd_new_buf_create(struct smc_link_group *lgr,
bool is_dmb, int bufsize)
{
struct smc_buf_desc *buf_desc;
int rc;
/* try to alloc a new DMB */
buf_desc = kzalloc(sizeof(*buf_desc), GFP_KERNEL);
if (!buf_desc)
return ERR_PTR(-ENOMEM);
if (is_dmb) {
rc = smc_ism_register_dmb(lgr, bufsize, buf_desc);
if (rc) {
kfree(buf_desc);
if (rc == -ENOMEM)
return ERR_PTR(-EAGAIN);
if (rc == -ENOSPC)
return ERR_PTR(-ENOSPC);
return ERR_PTR(-EIO);
}
buf_desc->pages = virt_to_page(buf_desc->cpu_addr);
/* CDC header stored in buf. So, pretend it was smaller */
buf_desc->len = bufsize - sizeof(struct smcd_cdc_msg);
} else {
buf_desc->cpu_addr = kzalloc(bufsize, GFP_KERNEL |
__GFP_NOWARN | __GFP_NORETRY |
__GFP_NOMEMALLOC);
if (!buf_desc->cpu_addr) {
kfree(buf_desc);
return ERR_PTR(-EAGAIN);
}
buf_desc->len = bufsize;
}
return buf_desc;
}
static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
{
struct smc_buf_desc *buf_desc = ERR_PTR(-ENOMEM);
struct smc_connection *conn = &smc->conn;
struct smc_link_group *lgr = conn->lgr;
struct list_head *buf_list;
int bufsize, bufsize_short;
bool is_dgraded = false;
struct mutex *lock; /* lock buffer list */
int sk_buf_size;
if (is_rmb)
/* use socket recv buffer size (w/o overhead) as start value */
sk_buf_size = smc->sk.sk_rcvbuf;
else
/* use socket send buffer size (w/o overhead) as start value */
sk_buf_size = smc->sk.sk_sndbuf;
for (bufsize_short = smc_compress_bufsize(sk_buf_size, is_smcd, is_rmb);
bufsize_short >= 0; bufsize_short--) {
if (is_rmb) {
lock = &lgr->rmbs_lock;
buf_list = &lgr->rmbs[bufsize_short];
} else {
lock = &lgr->sndbufs_lock;
buf_list = &lgr->sndbufs[bufsize_short];
}
bufsize = smc_uncompress_bufsize(bufsize_short);
/* check for reusable slot in the link group */
buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
if (buf_desc) {
buf_desc->is_dma_need_sync = 0;
SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
SMC_STAT_BUF_REUSE(smc, is_smcd, is_rmb);
break; /* found reusable slot */
}
if (is_smcd)
buf_desc = smcd_new_buf_create(lgr, is_rmb, bufsize);
else
buf_desc = smcr_new_buf_create(lgr, is_rmb, bufsize);
if (PTR_ERR(buf_desc) == -ENOMEM)
break;
if (IS_ERR(buf_desc)) {
if (!is_dgraded) {
is_dgraded = true;
SMC_STAT_RMB_DOWNGRADED(smc, is_smcd, is_rmb);
}
continue;
}
SMC_STAT_RMB_ALLOC(smc, is_smcd, is_rmb);
SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
buf_desc->used = 1;
mutex_lock(lock);
list_add(&buf_desc->list, buf_list);
mutex_unlock(lock);
break; /* found */
}
if (IS_ERR(buf_desc))
return PTR_ERR(buf_desc);
if (!is_smcd) {
if (smcr_buf_map_usable_links(lgr, buf_desc, is_rmb)) {
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 12:44:04 +03:00
smcr_buf_unuse(buf_desc, is_rmb, lgr);
return -ENOMEM;
}
}
if (is_rmb) {
conn->rmb_desc = buf_desc;
conn->rmbe_size_short = bufsize_short;
smc->sk.sk_rcvbuf = bufsize;
atomic_set(&conn->bytes_to_rcv, 0);
conn->rmbe_update_limit =
smc_rmb_wnd_update_limit(buf_desc->len);
if (is_smcd)
smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
} else {
conn->sndbuf_desc = buf_desc;
smc->sk.sk_sndbuf = bufsize;
atomic_set(&conn->sndbuf_space, bufsize);
}
return 0;
}
void smc_sndbuf_sync_sg_for_device(struct smc_connection *conn)
{
if (!conn->sndbuf_desc->is_dma_need_sync)
return;
if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd ||
!smc_link_active(conn->lnk))
return;
smc_ib_sync_sg_for_device(conn->lnk, conn->sndbuf_desc, DMA_TO_DEVICE);
}
void smc_rmb_sync_sg_for_cpu(struct smc_connection *conn)
{
int i;
if (!conn->rmb_desc->is_dma_need_sync)
return;
if (!smc_conn_lgr_valid(conn) || conn->lgr->is_smcd)
return;
for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
if (!smc_link_active(&conn->lgr->lnk[i]))
continue;
smc_ib_sync_sg_for_cpu(&conn->lgr->lnk[i], conn->rmb_desc,
DMA_FROM_DEVICE);
}
}
/* create the send and receive buffer for an SMC socket;
* receive buffers are called RMBs;
* (even though the SMC protocol allows more than one RMB-element per RMB,
* the Linux implementation uses just one RMB-element per RMB, i.e. uses an
* extra RMB for every connection in a link group
*/
int smc_buf_create(struct smc_sock *smc, bool is_smcd)
{
int rc;
/* create send buffer */
rc = __smc_buf_create(smc, is_smcd, false);
if (rc)
return rc;
/* create rmb */
rc = __smc_buf_create(smc, is_smcd, true);
if (rc) {
mutex_lock(&smc->conn.lgr->sndbufs_lock);
list_del(&smc->conn.sndbuf_desc->list);
mutex_unlock(&smc->conn.lgr->sndbufs_lock);
smc_buf_free(smc->conn.lgr, false, smc->conn.sndbuf_desc);
net/smc: reset sndbuf_desc if freed When an SMC connection is created, and there is a problem to create an RMB or DMB, the previously created send buffer is thrown away as well including buffer descriptor freeing. Make sure the connection no longer references the freed buffer descriptor, otherwise bugs like this are possible: [71556.835148] ============================================================================= [71556.835168] BUG kmalloc-128 (Tainted: G B OE ): Poison overwritten [71556.835172] ----------------------------------------------------------------------------- [71556.835179] INFO: 0x00000000d20894be-0x00000000aaef63e9 @offset=2724. First byte 0x0 instead of 0x6b [71556.835215] INFO: Allocated in __smc_buf_create+0x184/0x578 [smc] age=0 cpu=5 pid=46726 [71556.835234] ___slab_alloc+0x5a4/0x690 [71556.835239] __slab_alloc.constprop.0+0x70/0xb0 [71556.835243] kmem_cache_alloc_trace+0x38e/0x3f8 [71556.835250] __smc_buf_create+0x184/0x578 [smc] [71556.835257] smc_buf_create+0x2e/0xe8 [smc] [71556.835264] smc_listen_work+0x516/0x6a0 [smc] [71556.835275] process_one_work+0x280/0x478 [71556.835280] worker_thread+0x66/0x368 [71556.835287] kthread+0x17a/0x1a0 [71556.835294] ret_from_fork+0x28/0x2c [71556.835301] INFO: Freed in smc_buf_create+0xd8/0xe8 [smc] age=0 cpu=5 pid=46726 [71556.835307] __slab_free+0x246/0x560 [71556.835311] kfree+0x398/0x3f8 [71556.835318] smc_buf_create+0xd8/0xe8 [smc] [71556.835324] smc_listen_work+0x516/0x6a0 [smc] [71556.835328] process_one_work+0x280/0x478 [71556.835332] worker_thread+0x66/0x368 [71556.835337] kthread+0x17a/0x1a0 [71556.835344] ret_from_fork+0x28/0x2c [71556.835348] INFO: Slab 0x00000000a0744551 objects=51 used=51 fp=0x0000000000000000 flags=0x1ffff00000010200 [71556.835352] INFO: Object 0x00000000563480a1 @offset=2688 fp=0x00000000289567b2 [71556.835359] Redzone 000000006783cde2: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb ................ [71556.835363] Redzone 00000000e35b876e: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb ................ [71556.835367] Redzone 0000000023074562: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb ................ [71556.835372] Redzone 00000000b9564b8c: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb ................ [71556.835376] Redzone 00000000810c6362: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb ................ [71556.835380] Redzone 0000000065ef52c3: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb ................ [71556.835384] Redzone 00000000c5dd6984: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb ................ [71556.835388] Redzone 000000004c480f8f: bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb bb ................ [71556.835392] Object 00000000563480a1: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [71556.835397] Object 000000009c479d06: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [71556.835401] Object 000000006e1dce92: 6b 6b 6b 6b 00 00 00 00 6b 6b 6b 6b 6b 6b 6b 6b kkkk....kkkkkkkk [71556.835405] Object 00000000227f7cf8: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [71556.835410] Object 000000009a701215: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [71556.835414] Object 000000003731ce76: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [71556.835418] Object 00000000f7085967: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [71556.835422] Object 0000000007f99927: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b a5 kkkkkkkkkkkkkkk. [71556.835427] Redzone 00000000579c4913: bb bb bb bb bb bb bb bb ........ [71556.835431] Padding 00000000305aef82: 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZZZZZZZZZ [71556.835435] Padding 00000000b1cdd722: 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZZZZZZZZZ [71556.835438] Padding 00000000c7568199: 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZZZZZZZZZ [71556.835442] Padding 00000000fad4c4d4: 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a 5a ZZZZZZZZZZZZZZZZ [71556.835451] CPU: 0 PID: 47939 Comm: kworker/0:15 Tainted: G B OE 5.9.0-rc1uschi+ #54 [71556.835456] Hardware name: IBM 3906 M03 703 (LPAR) [71556.835464] Workqueue: events smc_listen_work [smc] [71556.835470] Call Trace: [71556.835478] [<00000000d5eaeb10>] show_stack+0x90/0xf8 [71556.835493] [<00000000d66fc0f8>] dump_stack+0xa8/0xe8 [71556.835499] [<00000000d61a511c>] check_bytes_and_report+0x104/0x130 [71556.835504] [<00000000d61a57b2>] check_object+0x26a/0x2e0 [71556.835509] [<00000000d61a59bc>] alloc_debug_processing+0x194/0x238 [71556.835514] [<00000000d61a8c14>] ___slab_alloc+0x5a4/0x690 [71556.835519] [<00000000d61a9170>] __slab_alloc.constprop.0+0x70/0xb0 [71556.835524] [<00000000d61aaf66>] kmem_cache_alloc_trace+0x38e/0x3f8 [71556.835530] [<000003ff80549bbc>] __smc_buf_create+0x184/0x578 [smc] [71556.835538] [<000003ff8054a396>] smc_buf_create+0x2e/0xe8 [smc] [71556.835545] [<000003ff80540c16>] smc_listen_work+0x516/0x6a0 [smc] [71556.835549] [<00000000d5f0f448>] process_one_work+0x280/0x478 [71556.835554] [<00000000d5f0f6a6>] worker_thread+0x66/0x368 [71556.835559] [<00000000d5f18692>] kthread+0x17a/0x1a0 [71556.835563] [<00000000d6abf3b8>] ret_from_fork+0x28/0x2c [71556.835569] INFO: lockdep is turned off. [71556.835573] FIX kmalloc-128: Restoring 0x00000000d20894be-0x00000000aaef63e9=0x6b [71556.835577] FIX kmalloc-128: Marking all objects used Fixes: fd7f3a746582 ("net/smc: remove freed buffer from list") Reviewed-by: Karsten Graul <kgraul@linux.ibm.com> Signed-off-by: Ursula Braun <ubraun@linux.ibm.com> Signed-off-by: Karsten Graul <kgraul@linux.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-09-03 22:53:17 +03:00
smc->conn.sndbuf_desc = NULL;
}
return rc;
}
static inline int smc_rmb_reserve_rtoken_idx(struct smc_link_group *lgr)
{
int i;
for_each_clear_bit(i, lgr->rtokens_used_mask, SMC_RMBS_PER_LGR_MAX) {
if (!test_and_set_bit(i, lgr->rtokens_used_mask))
return i;
}
return -ENOSPC;
}
static int smc_rtoken_find_by_link(struct smc_link_group *lgr, int lnk_idx,
u32 rkey)
{
int i;
for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
if (test_bit(i, lgr->rtokens_used_mask) &&
lgr->rtokens[i][lnk_idx].rkey == rkey)
return i;
}
return -ENOENT;
}
/* set rtoken for a new link to an existing rmb */
void smc_rtoken_set(struct smc_link_group *lgr, int link_idx, int link_idx_new,
__be32 nw_rkey_known, __be64 nw_vaddr, __be32 nw_rkey)
{
int rtok_idx;
rtok_idx = smc_rtoken_find_by_link(lgr, link_idx, ntohl(nw_rkey_known));
if (rtok_idx == -ENOENT)
return;
lgr->rtokens[rtok_idx][link_idx_new].rkey = ntohl(nw_rkey);
lgr->rtokens[rtok_idx][link_idx_new].dma_addr = be64_to_cpu(nw_vaddr);
}
/* set rtoken for a new link whose link_id is given */
void smc_rtoken_set2(struct smc_link_group *lgr, int rtok_idx, int link_id,
__be64 nw_vaddr, __be32 nw_rkey)
{
u64 dma_addr = be64_to_cpu(nw_vaddr);
u32 rkey = ntohl(nw_rkey);
bool found = false;
int link_idx;
for (link_idx = 0; link_idx < SMC_LINKS_PER_LGR_MAX; link_idx++) {
if (lgr->lnk[link_idx].link_id == link_id) {
found = true;
break;
}
}
if (!found)
return;
lgr->rtokens[rtok_idx][link_idx].rkey = rkey;
lgr->rtokens[rtok_idx][link_idx].dma_addr = dma_addr;
}
/* add a new rtoken from peer */
int smc_rtoken_add(struct smc_link *lnk, __be64 nw_vaddr, __be32 nw_rkey)
{
struct smc_link_group *lgr = smc_get_lgr(lnk);
u64 dma_addr = be64_to_cpu(nw_vaddr);
u32 rkey = ntohl(nw_rkey);
int i;
for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
lgr->rtokens[i][lnk->link_idx].dma_addr == dma_addr &&
test_bit(i, lgr->rtokens_used_mask)) {
/* already in list */
return i;
}
}
i = smc_rmb_reserve_rtoken_idx(lgr);
if (i < 0)
return i;
lgr->rtokens[i][lnk->link_idx].rkey = rkey;
lgr->rtokens[i][lnk->link_idx].dma_addr = dma_addr;
return i;
}
/* delete an rtoken from all links */
int smc_rtoken_delete(struct smc_link *lnk, __be32 nw_rkey)
{
struct smc_link_group *lgr = smc_get_lgr(lnk);
u32 rkey = ntohl(nw_rkey);
int i, j;
for (i = 0; i < SMC_RMBS_PER_LGR_MAX; i++) {
if (lgr->rtokens[i][lnk->link_idx].rkey == rkey &&
test_bit(i, lgr->rtokens_used_mask)) {
for (j = 0; j < SMC_LINKS_PER_LGR_MAX; j++) {
lgr->rtokens[i][j].rkey = 0;
lgr->rtokens[i][j].dma_addr = 0;
}
clear_bit(i, lgr->rtokens_used_mask);
return 0;
}
}
return -ENOENT;
}
/* save rkey and dma_addr received from peer during clc handshake */
int smc_rmb_rtoken_handling(struct smc_connection *conn,
struct smc_link *lnk,
struct smc_clc_msg_accept_confirm *clc)
{
conn->rtoken_idx = smc_rtoken_add(lnk, clc->r0.rmb_dma_addr,
clc->r0.rmb_rkey);
if (conn->rtoken_idx < 0)
return conn->rtoken_idx;
return 0;
}
static void smc_core_going_away(void)
{
struct smc_ib_device *smcibdev;
struct smcd_dev *smcd;
net/smc: fix sleep bug in smc_pnet_find_roce_resource() Tests showed this BUG: [572555.252867] BUG: sleeping function called from invalid context at kernel/locking/mutex.c:935 [572555.252876] in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 131031, name: smcapp [572555.252879] INFO: lockdep is turned off. [572555.252883] CPU: 1 PID: 131031 Comm: smcapp Tainted: G O 5.7.0-rc3uschi+ #356 [572555.252885] Hardware name: IBM 3906 M03 703 (LPAR) [572555.252887] Call Trace: [572555.252896] [<00000000ac364554>] show_stack+0x94/0xe8 [572555.252901] [<00000000aca1f400>] dump_stack+0xa0/0xe0 [572555.252906] [<00000000ac3c8c10>] ___might_sleep+0x260/0x280 [572555.252910] [<00000000acdc0c98>] __mutex_lock+0x48/0x940 [572555.252912] [<00000000acdc15c2>] mutex_lock_nested+0x32/0x40 [572555.252975] [<000003ff801762d0>] mlx5_lag_get_roce_netdev+0x30/0xc0 [mlx5_core] [572555.252996] [<000003ff801fb3aa>] mlx5_ib_get_netdev+0x3a/0xe0 [mlx5_ib] [572555.253007] [<000003ff80063848>] smc_pnet_find_roce_resource+0x1d8/0x310 [smc] [572555.253011] [<000003ff800602f0>] __smc_connect+0x1f0/0x3e0 [smc] [572555.253015] [<000003ff80060634>] smc_connect+0x154/0x190 [smc] [572555.253022] [<00000000acbed8d4>] __sys_connect+0x94/0xd0 [572555.253025] [<00000000acbef620>] __s390x_sys_socketcall+0x170/0x360 [572555.253028] [<00000000acdc6800>] system_call+0x298/0x2b8 [572555.253030] INFO: lockdep is turned off. Function smc_pnet_find_rdma_dev() might be called from smc_pnet_find_roce_resource(). It holds the smc_ib_devices list spinlock while calling infiniband op get_netdev(). At least for mlx5 the get_netdev operation wants mutex serialization, which conflicts with the smc_ib_devices spinlock. This patch switches the smc_ib_devices spinlock into a mutex to allow sleeping when calling get_netdev(). Fixes: a4cf0443c414 ("smc: introduce SMC as an IB-client") Signed-off-by: Ursula Braun <ubraun@linux.ibm.com> Signed-off-by: Karsten Graul <kgraul@linux.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-07-08 18:05:13 +03:00
mutex_lock(&smc_ib_devices.mutex);
list_for_each_entry(smcibdev, &smc_ib_devices.list, list) {
int i;
for (i = 0; i < SMC_MAX_PORTS; i++)
set_bit(i, smcibdev->ports_going_away);
}
net/smc: fix sleep bug in smc_pnet_find_roce_resource() Tests showed this BUG: [572555.252867] BUG: sleeping function called from invalid context at kernel/locking/mutex.c:935 [572555.252876] in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 131031, name: smcapp [572555.252879] INFO: lockdep is turned off. [572555.252883] CPU: 1 PID: 131031 Comm: smcapp Tainted: G O 5.7.0-rc3uschi+ #356 [572555.252885] Hardware name: IBM 3906 M03 703 (LPAR) [572555.252887] Call Trace: [572555.252896] [<00000000ac364554>] show_stack+0x94/0xe8 [572555.252901] [<00000000aca1f400>] dump_stack+0xa0/0xe0 [572555.252906] [<00000000ac3c8c10>] ___might_sleep+0x260/0x280 [572555.252910] [<00000000acdc0c98>] __mutex_lock+0x48/0x940 [572555.252912] [<00000000acdc15c2>] mutex_lock_nested+0x32/0x40 [572555.252975] [<000003ff801762d0>] mlx5_lag_get_roce_netdev+0x30/0xc0 [mlx5_core] [572555.252996] [<000003ff801fb3aa>] mlx5_ib_get_netdev+0x3a/0xe0 [mlx5_ib] [572555.253007] [<000003ff80063848>] smc_pnet_find_roce_resource+0x1d8/0x310 [smc] [572555.253011] [<000003ff800602f0>] __smc_connect+0x1f0/0x3e0 [smc] [572555.253015] [<000003ff80060634>] smc_connect+0x154/0x190 [smc] [572555.253022] [<00000000acbed8d4>] __sys_connect+0x94/0xd0 [572555.253025] [<00000000acbef620>] __s390x_sys_socketcall+0x170/0x360 [572555.253028] [<00000000acdc6800>] system_call+0x298/0x2b8 [572555.253030] INFO: lockdep is turned off. Function smc_pnet_find_rdma_dev() might be called from smc_pnet_find_roce_resource(). It holds the smc_ib_devices list spinlock while calling infiniband op get_netdev(). At least for mlx5 the get_netdev operation wants mutex serialization, which conflicts with the smc_ib_devices spinlock. This patch switches the smc_ib_devices spinlock into a mutex to allow sleeping when calling get_netdev(). Fixes: a4cf0443c414 ("smc: introduce SMC as an IB-client") Signed-off-by: Ursula Braun <ubraun@linux.ibm.com> Signed-off-by: Karsten Graul <kgraul@linux.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-07-08 18:05:13 +03:00
mutex_unlock(&smc_ib_devices.mutex);
mutex_lock(&smcd_dev_list.mutex);
list_for_each_entry(smcd, &smcd_dev_list.list, list) {
smcd->going_away = 1;
}
mutex_unlock(&smcd_dev_list.mutex);
}
/* Clean up all SMC link groups */
static void smc_lgrs_shutdown(void)
{
struct smcd_dev *smcd;
smc_core_going_away();
smc_smcr_terminate_all(NULL);
mutex_lock(&smcd_dev_list.mutex);
list_for_each_entry(smcd, &smcd_dev_list.list, list)
smc_smcd_terminate_all(smcd);
mutex_unlock(&smcd_dev_list.mutex);
}
static int smc_core_reboot_event(struct notifier_block *this,
unsigned long event, void *ptr)
{
smc_lgrs_shutdown();
smc_ib_unregister_client();
return 0;
}
static struct notifier_block smc_reboot_notifier = {
.notifier_call = smc_core_reboot_event,
};
int __init smc_core_init(void)
{
return register_reboot_notifier(&smc_reboot_notifier);
}
/* Called (from smc_exit) when module is removed */
void smc_core_exit(void)
{
unregister_reboot_notifier(&smc_reboot_notifier);
smc_lgrs_shutdown();
}