1
0
mirror of https://github.com/samba-team/samba.git synced 2024-12-25 23:21:54 +03:00
samba-mirror/source3/modules/onefs_dir.c

623 lines
18 KiB
C
Raw Normal View History

/*
* Unix SMB/CIFS implementation.
*
* Support for OneFS bulk directory enumeration API
*
* Copyright (C) Steven Danneman, 2009
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 3 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, see <http://www.gnu.org/licenses/>.
*/
#include "includes.h"
#include "smbd/smbd.h"
#include "onefs.h"
#include "onefs_config.h"
#include <ifs/ifs_syscalls.h>
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
#include <isi_util/isi_dir.h>
/* The OneFS filesystem provides a readdirplus() syscall, equivalent to the
* NFSv3 PDU, which retrieves bulk directory listings with stat information
* in a single syscall.
*
* This file hides this bulk interface underneath Samba's very POSIX like
* opendir/readdir/telldir VFS interface. This is done to provide a
* significant performance improvement when listing the contents of large
* directories, which also require file meta information. ie a typical
* Windows Explorer request.
*/
#define RDP_RESUME_KEY_START 0x1
#define RDP_BATCH_SIZE 128
#define RDP_DIRENTRIES_SIZE ((size_t)(RDP_BATCH_SIZE * sizeof(struct dirent)))
static char *rdp_direntries = NULL;
static struct stat *rdp_stats = NULL;
static uint64_t *rdp_cookies = NULL;
struct rdp_dir_state {
struct rdp_dir_state *next, *prev;
SMB_STRUCT_DIR *dirp;
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
char *direntries_cursor; /* cursor to last returned direntry in cache */
size_t stat_count; /* number of entries stored in the cache */
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
size_t stat_cursor; /* cursor to last returned stat in the cache */
uint64_t resume_cookie; /* cookie from the last entry returned from the
cache */
};
static struct rdp_dir_state *dirstatelist = NULL;
SMB_STRUCT_DIR *rdp_last_dirp = NULL;
/**
* Given a DIR pointer, return our internal state.
*
* This function also tells us whether the given DIR is the same as we saw
* during the last call. Because we use a single globally allocated buffer
* for readdirplus entries we must check every call into this API to see if
* it's for the same directory listing, or a new one. If it's the same we can
* maintain our current cached entries, otherwise we must go to the kernel.
*
* @return 0 on success, 1 on failure
*/
static int
rdp_retrieve_dir_state(SMB_STRUCT_DIR *dirp, struct rdp_dir_state **dir_state,
bool *same_as_last)
{
struct rdp_dir_state *dsp;
/* Is this directory the same as the last call */
*same_as_last = (dirp == rdp_last_dirp);
for(dsp = dirstatelist; dsp; dsp = dsp->next)
if (dsp->dirp == dirp) {
*dir_state = dsp;
return 0;
}
/* Couldn't find existing dir_state for the given directory
* pointer. */
return 1;
}
/**
* Initialize the global readdirplus buffers.
*
* These same buffers are used for all calls into readdirplus.
*
* @return 0 on success, errno value on failure
*/
static int
rdp_init(struct rdp_dir_state *dsp)
{
/* Unfortunately, there is no good way to free these buffers. If we
* allocated and freed for every DIR handle performance would be
* adversely affected. For now these buffers will be leaked and only
* freed when the smbd process dies. */
if (!rdp_direntries) {
rdp_direntries = SMB_MALLOC(RDP_DIRENTRIES_SIZE);
if (!rdp_direntries)
return ENOMEM;
}
if (!rdp_stats) {
rdp_stats =
SMB_MALLOC(RDP_BATCH_SIZE * sizeof(struct stat));
if (!rdp_stats)
return ENOMEM;
}
if (!rdp_cookies) {
rdp_cookies = SMB_MALLOC(RDP_BATCH_SIZE * sizeof(uint64_t));
if (!rdp_cookies)
return ENOMEM;
}
dsp->direntries_cursor = rdp_direntries + RDP_DIRENTRIES_SIZE;
dsp->stat_count = RDP_BATCH_SIZE;
dsp->stat_cursor = RDP_BATCH_SIZE;
dsp->resume_cookie = RDP_RESUME_KEY_START;
return 0;
}
/**
* Call into readdirplus() to refill our global dirent cache.
*
* This function also resets all cursors back to the beginning of the cache.
* All stat buffers are retrieved by following symlinks.
*
* @return number of entries retrieved, -1 on error
*/
static int
rdp_fill_cache(struct rdp_dir_state *dsp)
{
int nread, dirfd;
dirfd = dirfd(dsp->dirp);
if (dirfd < 0) {
DEBUG(1, ("Could not retrieve fd for DIR\n"));
return -1;
}
/* Resize the stat_count to grab as many entries as possible */
dsp->stat_count = RDP_BATCH_SIZE;
DEBUG(9, ("Calling readdirplus() with DIR %p, dirfd: %d, "
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
"resume_cookie %#llx, size_to_read: %zu, "
"direntries_size: %zu, stat_count: %u\n",
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
dsp->dirp, dirfd, dsp->resume_cookie, RDP_BATCH_SIZE,
RDP_DIRENTRIES_SIZE, dsp->stat_count));
nread = readdirplus(dirfd,
RDP_FOLLOW,
&dsp->resume_cookie,
RDP_BATCH_SIZE,
rdp_direntries,
RDP_DIRENTRIES_SIZE,
&dsp->stat_count,
rdp_stats,
rdp_cookies);
if (nread < 0) {
DEBUG(1, ("Error calling readdirplus(): %s\n",
strerror(errno)));
return -1;
}
DEBUG(9, ("readdirplus() returned %u entries from DIR %p\n",
dsp->stat_count, dsp->dirp));
dsp->direntries_cursor = rdp_direntries;
dsp->stat_cursor = 0;
return nread;
}
/**
* Create a dir_state to track an open directory that we're enumerating.
*
* This utility function is globally accessible for use by other parts of the
* onefs.so module to initialize a dir_state when a directory is opened through
* a path other than the VFS layer.
*
* @return 0 on success and errno on failure
*
* @note: Callers of this function MUST cleanup the dir_state through a proper
* call to VFS_CLOSEDIR().
*/
int
onefs_rdp_add_dir_state(connection_struct *conn, SMB_STRUCT_DIR *dirp)
{
int ret = 0;
struct rdp_dir_state *dsp = NULL;
/* No-op if readdirplus is disabled */
if (!lp_parm_bool(SNUM(conn), PARM_ONEFS_TYPE,
PARM_USE_READDIRPLUS, PARM_USE_READDIRPLUS_DEFAULT))
{
return 0;
}
/* Create a struct dir_state */
dsp = SMB_MALLOC_P(struct rdp_dir_state);
if (!dsp) {
DEBUG(0, ("Error allocating struct rdp_dir_state.\n"));
return ENOMEM;
}
/* Initialize the dir_state structure and add it to the list */
ret = rdp_init(dsp);
if (ret) {
DEBUG(0, ("Error initializing readdirplus() buffers: %s\n",
strerror(ret)));
return ret;
}
/* Set the SMB_STRUCT_DIR in the dsp */
dsp->dirp = dirp;
DLIST_ADD(dirstatelist, dsp);
return 0;
}
/**
* Open a directory for enumeration.
*
* Create a state struct to track the state of this directory for the life
* of this open.
*
* @param[in] handle vfs handle given in most VFS calls
* @param[in] fname filename of the directory to open
* @param[in] mask unused
* @param[in] attr unused
*
* @return DIR pointer, NULL if directory does not exist, NULL on error
*/
SMB_STRUCT_DIR *
onefs_opendir(vfs_handle_struct *handle, const char *fname, const char *mask,
uint32 attr)
{
int ret = 0;
SMB_STRUCT_DIR *ret_dirp;
/* Fallback to default system routines if readdirplus is disabled */
if (!lp_parm_bool(SNUM(handle->conn), PARM_ONEFS_TYPE,
PARM_USE_READDIRPLUS, PARM_USE_READDIRPLUS_DEFAULT))
{
return SMB_VFS_NEXT_OPENDIR(handle, fname, mask, attr);
}
/* Open the directory */
ret_dirp = SMB_VFS_NEXT_OPENDIR(handle, fname, mask, attr);
if (!ret_dirp) {
DEBUG(3, ("Unable to open directory: %s\n", fname));
return NULL;
}
/* Create the dir_state struct and add it to the list */
ret = onefs_rdp_add_dir_state(handle->conn, ret_dirp);
if (ret) {
DEBUG(0, ("Error adding dir_state to the list\n"));
return NULL;
}
DEBUG(9, ("Opened handle on directory: \"%s\", DIR %p\n",
fname, ret_dirp));
return ret_dirp;
}
/**
* Retrieve one direntry and optional stat buffer from our readdir cache.
*
* Increment the internal resume cookie, and refresh the cache from the
* kernel if necessary.
*
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
* The cache cursor tracks the last entry which was successfully returned
* to a caller of onefs_readdir(). When a new entry is requested, this
* function first increments the cursor, then returns that entry.
*
* @param[in] handle vfs handle given in most VFS calls
* @param[in] dirp system DIR handle to retrieve direntries from
* @param[in/out] sbuf optional stat buffer to fill, this can be NULL
*
* @return dirent structure, NULL if at the end of the directory, NULL on error
*/
SMB_STRUCT_DIRENT *
onefs_readdir(vfs_handle_struct *handle, SMB_STRUCT_DIR *dirp,
SMB_STRUCT_STAT *sbuf)
{
struct rdp_dir_state *dsp = NULL;
SMB_STRUCT_DIRENT *ret_direntp;
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
bool same_as_last, filled_cache = false;
int ret = -1;
/* Set stat invalid in-case we error out */
if (sbuf)
SET_STAT_INVALID(*sbuf);
/* Fallback to default system routines if readdirplus is disabled */
if (!lp_parm_bool(SNUM(handle->conn), PARM_ONEFS_TYPE,
PARM_USE_READDIRPLUS, PARM_USE_READDIRPLUS_DEFAULT))
{
return sys_readdir(dirp);
}
/* Retrieve state based off DIR handle */
ret = rdp_retrieve_dir_state(dirp, &dsp, &same_as_last);
if (ret) {
DEBUG(1, ("Could not retrieve dir_state struct for "
"SMB_STRUCT_DIR pointer.\n"));
ret_direntp = NULL;
goto end;
}
/* DIR is the same, current buffer and cursors are valid.
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
* Check if there are any entries left in our current cache. */
if (same_as_last) {
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
if (dsp->stat_cursor == dsp->stat_count - 1) {
/* Cache is empty, refill from kernel */
ret = rdp_fill_cache(dsp);
if (ret <= 0) {
ret_direntp = NULL;
goto end;
}
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
filled_cache = true;
}
} else {
/* DIR is different from last call, reset all buffers and
* cursors, and refill the global cache from the new DIR */
ret = rdp_fill_cache(dsp);
if (ret <= 0) {
ret_direntp = NULL;
goto end;
}
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
filled_cache = true;
DEBUG(8, ("Switched global rdp cache to new DIR entry.\n"));
}
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
/* If we just filled the cache we treat that action as the cursor
* increment as the resume cookie used belonged to the previous
* directory entry. If the cache has not changed we first increment
* our cursor, then return the next entry */
if (!filled_cache) {
dsp->direntries_cursor +=
((SMB_STRUCT_DIRENT *)dsp->direntries_cursor)->d_reclen;
dsp->stat_cursor++;
}
/* The resume_cookie stored here purposely differs based on whether we
* just filled the cache. The resume cookie stored must always provide
* the next direntry, in case the cache is reloaded on every
* onefs_readdir() */
dsp->resume_cookie = rdp_cookies[dsp->stat_cursor];
/* Return an entry from cache */
ret_direntp = ((SMB_STRUCT_DIRENT *)dsp->direntries_cursor);
if (sbuf) {
struct stat onefs_sbuf;
onefs_sbuf = rdp_stats[dsp->stat_cursor];
init_stat_ex_from_onefs_stat(sbuf, &onefs_sbuf);
/* readdirplus() sets st_ino field to 0, if it was
* unable to retrieve stat information for that
* particular directory entry. */
if (sbuf->st_ex_ino == 0)
SET_STAT_INVALID(*sbuf);
}
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
DEBUG(9, ("Read from DIR %p, direntry: \"%s\", resume cookie: %#llx, "
"cache cursor: %zu, cache count: %zu\n",
dsp->dirp, ret_direntp->d_name, dsp->resume_cookie,
dsp->stat_cursor, dsp->stat_count));
/* FALLTHROUGH */
end:
/* Set rdp_last_dirp at the end of every VFS call where the cache was
* reloaded */
rdp_last_dirp = dirp;
return ret_direntp;
}
/**
* Set the location of the next direntry to be read via onefs_readdir().
*
* This function should only pass in locations retrieved from onefs_telldir().
*
* @param[in] handle vfs handle given in most VFS calls
* @param[in] dirp system DIR handle to set offset on
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
* @param[in] offset into the directory to resume reading from
*
* @return no return value
*/
void
onefs_seekdir(vfs_handle_struct *handle, SMB_STRUCT_DIR *dirp, long offset)
{
struct rdp_dir_state *dsp = NULL;
bool same_as_last;
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
uint64_t resume_cookie = 0;
int ret = -1;
/* Fallback to default system routines if readdirplus is disabled */
if (!lp_parm_bool(SNUM(handle->conn), PARM_ONEFS_TYPE,
PARM_USE_READDIRPLUS, PARM_USE_READDIRPLUS_DEFAULT))
{
return sys_seekdir(dirp, offset);
}
/* Validate inputs */
if (offset < 0) {
DEBUG(1, ("Invalid offset %ld passed.\n", offset));
return;
}
/* Retrieve state based off DIR handle */
ret = rdp_retrieve_dir_state(dirp, &dsp, &same_as_last);
if (ret) {
DEBUG(1, ("Could not retrieve dir_state struct for "
"SMB_STRUCT_DIR pointer.\n"));
/* XXX: we can't return an error, should we ABORT rather than
* return without actually seeking? */
return;
}
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
/* Convert offset to resume_cookie */
resume_cookie = rdp_offset31_to_cookie63(offset);
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
DEBUG(9, ("Seek DIR %p, offset: %ld, resume_cookie: %#llx\n",
dsp->dirp, offset, resume_cookie));
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
/* TODO: We could check if the resume_cookie is already in the cache
* through a linear search. This would allow us to avoid the cost of
* flushing the cache. Frequently, the seekdir offset will only be
* one entry before the current cache cursor. However, usually
* VFS_SEEKDIR() is only called at the end of a TRAND2_FIND read and
* we'll flush the cache at the beginning of the next PDU anyway. Some
* analysis should be done to see if this enhancement would provide
* better performance. */
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
/* Set the resume cookie and indicate that the cache should be reloaded
* on next read */
dsp->resume_cookie = resume_cookie;
rdp_last_dirp = NULL;
return;
}
/**
* Returns the location of the next direntry to be read via onefs_readdir().
*
* This value can be passed into onefs_seekdir().
*
* @param[in] handle vfs handle given in most VFS calls
* @param[in] dirp system DIR handle to set offset on
*
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
* @return offset into the directory to resume reading from
*/
long
onefs_telldir(vfs_handle_struct *handle, SMB_STRUCT_DIR *dirp)
{
struct rdp_dir_state *dsp = NULL;
bool same_as_last;
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
long offset;
int ret = -1;
/* Fallback to default system routines if readdirplus is disabled */
if (!lp_parm_bool(SNUM(handle->conn), PARM_ONEFS_TYPE,
PARM_USE_READDIRPLUS, PARM_USE_READDIRPLUS_DEFAULT))
{
return sys_telldir(dirp);
}
/* Retrieve state based off DIR handle */
ret = rdp_retrieve_dir_state(dirp, &dsp, &same_as_last);
if (ret) {
DEBUG(1, ("Could not retrieve dir_state struct for "
"SMB_STRUCT_DIR pointer.\n"));
return -1;
}
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
/* Convert resume_cookie to offset */
offset = rdp_cookie63_to_offset31(dsp->resume_cookie);
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
if (offset < 0) {
DEBUG(1, ("Unable to convert resume_cookie: %#llx to a "
"suitable 32-bit offset value. Error: %s\n",
dsp->resume_cookie, strerror(errno)));
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
return -1;
}
DEBUG(9, ("Seek DIR %p, offset: %ld, resume_cookie: %#llx\n",
dsp->dirp, offset, dsp->resume_cookie));
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
return offset;
}
/**
* Set the next direntry to be read via onefs_readdir() to the beginning of the
* directory.
*
* @param[in] handle vfs handle given in most VFS calls
* @param[in] dirp system DIR handle to set offset on
*
* @return no return value
*/
void
onefs_rewinddir(vfs_handle_struct *handle, SMB_STRUCT_DIR *dirp)
{
struct rdp_dir_state *dsp = NULL;
bool same_as_last;
int ret = -1;
/* Fallback to default system routines if readdirplus is disabled */
if (!lp_parm_bool(SNUM(handle->conn), PARM_ONEFS_TYPE,
PARM_USE_READDIRPLUS, PARM_USE_READDIRPLUS_DEFAULT))
{
return sys_rewinddir(dirp);
}
/* Retrieve state based off DIR handle */
ret = rdp_retrieve_dir_state(dirp, &dsp, &same_as_last);
if (ret) {
DEBUG(1, ("Could not retrieve dir_state struct for "
"SMB_STRUCT_DIR pointer.\n"));
return;
}
/* Reset location and resume key to beginning */
ret = rdp_init(dsp);
if (ret) {
DEBUG(0, ("Error re-initializing rdp cursors: %s\n",
strerror(ret)));
return;
}
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
DEBUG(9, ("Rewind DIR: %p, to resume_cookie: %#llx\n", dsp->dirp,
dsp->resume_cookie));
return;
}
/**
* Close DIR pointer and remove all state for that directory open.
*
* @param[in] handle vfs handle given in most VFS calls
* @param[in] dirp system DIR handle to set offset on
*
* @return -1 on failure, setting errno
*/
int
onefs_closedir(vfs_handle_struct *handle, SMB_STRUCT_DIR *dirp)
{
struct rdp_dir_state *dsp = NULL;
bool same_as_last;
int ret_val = -1;
int ret = -1;
/* Fallback to default system routines if readdirplus is disabled */
if (!lp_parm_bool(SNUM(handle->conn), PARM_ONEFS_TYPE,
PARM_USE_READDIRPLUS, PARM_USE_READDIRPLUS_DEFAULT))
{
return SMB_VFS_NEXT_CLOSEDIR(handle, dirp);
}
/* Retrieve state based off DIR handle */
ret = rdp_retrieve_dir_state(dirp, &dsp, &same_as_last);
if (ret) {
DEBUG(1, ("Could not retrieve dir_state struct for "
"SMB_STRUCT_DIR pointer.\n"));
errno = ENOENT;
return -1;
}
/* Close DIR pointer */
ret_val = SMB_VFS_NEXT_CLOSEDIR(handle, dsp->dirp);
DEBUG(9, ("Closed handle on DIR %p\n", dsp->dirp));
/* Tear down state struct */
DLIST_REMOVE(dirstatelist, dsp);
SAFE_FREE(dsp);
/* Set lastp to NULL, as cache is no longer valid */
rdp_last_dirp = NULL;
return ret_val;
}
/**
* Initialize cache data at the beginning of every SMB search operation
*
* Since filesystem operations, such as delete files or meta data
* updates can occur to files in the directory we're searching
* between FIND_FIRST and FIND_NEXT calls we must refresh the cache
* from the kernel on every new search SMB.
*
* @param[in] handle vfs handle given in most VFS calls
* @param[in] dirp system DIR handle for the current search
*
* @return nothing
*/
void
onefs_init_search_op(vfs_handle_struct *handle, SMB_STRUCT_DIR *dirp)
{
s3:onefs.so fix issue with missing entries when enumerating directories This bug prompted several, fairly large changes to the of OneFS's readdirplus() within Samba. One fundamental problem is that we kept our cache cursor pointed at the next entry to be returned from onefs_readdir(), while the resume cookie needed to refill the cache such that our cursor would be on this entry, was located in the previous cache entry. This meant that to correctly handle seekdir() cases which could be found within the existing cache, and cases where a cache reload was needed, required that the cache always hold at least two entries: the entry we wished to return, and the previous entry which held the resume cookie. Since the readdirplus() syscall gives us no guarantee that it will always return these two direntries, there was a fundamental problem with this design. To fix this problem, I have rearchitected the onefs_readdir() path to keep its pointer on the entry which contains the resume_cookie, not the entry which will be returned next. Essentially, I changed onefs_readdir() from a "return an entry then increment the cursor" model to "increment the cursor then return an entry". By doing this, we only require that a single entry be within the cache: the entry containing the resume cookie. Second, there have been numerous off-by-one bugs in my implementation of onefs_seekdir() which did a mapping between the 64-bit resume cookie returned by readdirplus() and its own monotonically increasing "location" offset. Furthermore, this design caused a somewhat frequent waste of cycles, as in some cases we'd need to re-enumerate the entire directory to recover the current "location" from an old resume cookie. As this code was somewhat difficult to understand, prone to bugs, and innefficient in some cases I decided it was better to wholesale replace it now, rather than later. It is possible to algorithmically map the 64-bit resume cookies from readdirplus() into 32-bit offset values which SMB requires. The onefs.so module now calls into a system library to do this conversion. This greatly simplifies both the seekdir() and telldir() paths and is more efficient.
2009-05-05 02:02:17 +04:00
/* Setting the rdp_last_dirp to NULL will cause the next readdir
* operation to refill the cache. */
rdp_last_dirp = NULL;
return;
}