android_kernel_xiaomi_sm8350/fs/autofs4/root.c

918 lines
24 KiB
C
Raw Normal View History

/* -*- c -*- --------------------------------------------------------------- *
*
* linux/fs/autofs/root.c
*
* Copyright 1997-1998 Transmeta Corporation -- All Rights Reserved
* Copyright 1999-2000 Jeremy Fitzhardinge <jeremy@goop.org>
* Copyright 2001-2006 Ian Kent <raven@themaw.net>
*
* This file is part of the Linux kernel and is made available under
* the terms of the GNU General Public License, version 2, or at your
* option, any later version, incorporated herein by reference.
*
* ------------------------------------------------------------------------- */
#include <linux/capability.h>
#include <linux/errno.h>
#include <linux/stat.h>
#include <linux/param.h>
#include <linux/time.h>
#include "autofs_i.h"
static int autofs4_dir_symlink(struct inode *,struct dentry *,const char *);
static int autofs4_dir_unlink(struct inode *,struct dentry *);
static int autofs4_dir_rmdir(struct inode *,struct dentry *);
static int autofs4_dir_mkdir(struct inode *,struct dentry *,int);
static int autofs4_root_ioctl(struct inode *, struct file *,unsigned int,unsigned long);
static int autofs4_dir_open(struct inode *inode, struct file *file);
static struct dentry *autofs4_lookup(struct inode *,struct dentry *, struct nameidata *);
static void *autofs4_follow_link(struct dentry *, struct nameidata *);
autofs4: use lookup intent flags to trigger mounts When an open(2) call is made on an autofs mount point directory that already exists and the O_DIRECTORY flag is not used the needed mount callback to the daemon is not done. This leads to the path walk continuing resulting in a callback to the daemon with an incorrect key. open(2) is called without O_DIRECTORY by the "find" utility but this should be handled properly anyway. This happens because autofs needs to use the lookup flags to decide when to callback to the daemon to perform a mount to prevent mount storms. For example, an autofs indirect mount map that has the "browse" option will have the mount point directories are pre-created and the stat(2) call made by a color ls against each directory will cause all these directories to be mounted. It is unfortunate we need to resort to this but mount maps can be quite large. Additionally, if a user manually umounts an autofs indirect mount the directory isn't removed which also leads to this situation. To resolve this autofs needs to use the lookup intent flags to enable it to make this decision. This patch adds this check and triggers a call back if any of the lookup intent flags are set as all these calls warrant a mount attempt be requested. I know that external VFS code which uses the lookup flags is something that the VFS would like to eliminate but I have no choice as I can't see any other way to do this. A VFS dentry or inode operation callback which returns the lookup "type" (requires a definition) would be sufficient. But this change is needed now and I'm not aware of the form that coming VFS changes will take so I'm not willing to propose anything along these lines. If anyone can provide an alternate method I would be happy to use it. [akpm@linux-foundation.org: fix build for concurrent VFS changes] Signed-off-by: Ian Kent <raven@themaw.net> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Jeff Moyer <jmoyer@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:15 -04:00
#define TRIGGER_FLAGS (LOOKUP_CONTINUE | LOOKUP_DIRECTORY)
#define TRIGGER_INTENTS (LOOKUP_OPEN | LOOKUP_CREATE)
const struct file_operations autofs4_root_operations = {
.open = dcache_dir_open,
.release = dcache_dir_close,
.read = generic_read_dir,
.readdir = dcache_readdir,
.llseek = dcache_dir_lseek,
.ioctl = autofs4_root_ioctl,
};
const struct file_operations autofs4_dir_operations = {
.open = autofs4_dir_open,
.release = dcache_dir_close,
.read = generic_read_dir,
.readdir = dcache_readdir,
.llseek = dcache_dir_lseek,
};
const struct inode_operations autofs4_indirect_root_inode_operations = {
.lookup = autofs4_lookup,
.unlink = autofs4_dir_unlink,
.symlink = autofs4_dir_symlink,
.mkdir = autofs4_dir_mkdir,
.rmdir = autofs4_dir_rmdir,
};
const struct inode_operations autofs4_direct_root_inode_operations = {
.lookup = autofs4_lookup,
.unlink = autofs4_dir_unlink,
.mkdir = autofs4_dir_mkdir,
.rmdir = autofs4_dir_rmdir,
.follow_link = autofs4_follow_link,
};
const struct inode_operations autofs4_dir_inode_operations = {
.lookup = autofs4_lookup,
.unlink = autofs4_dir_unlink,
.symlink = autofs4_dir_symlink,
.mkdir = autofs4_dir_mkdir,
.rmdir = autofs4_dir_rmdir,
};
static int autofs4_dir_open(struct inode *inode, struct file *file)
{
struct dentry *dentry = file->f_path.dentry;
struct autofs_sb_info *sbi = autofs4_sbi(dentry->d_sb);
DPRINTK("file=%p dentry=%p %.*s",
file, dentry, dentry->d_name.len, dentry->d_name.name);
if (autofs4_oz_mode(sbi))
goto out;
/*
* An empty directory in an autofs file system is always a
* mount point. The daemon must have failed to mount this
* during lookup so it doesn't exist. This can happen, for
* example, if user space returns an incorrect status for a
* mount request. Otherwise we're doing a readdir on the
* autofs file system so just let the libfs routines handle
* it.
*/
spin_lock(&dcache_lock);
if (!d_mountpoint(dentry) && __simple_empty(dentry)) {
spin_unlock(&dcache_lock);
return -ENOENT;
}
spin_unlock(&dcache_lock);
out:
return dcache_dir_open(inode, file);
}
static int try_to_fill_dentry(struct dentry *dentry, int flags)
{
struct autofs_sb_info *sbi = autofs4_sbi(dentry->d_sb);
struct autofs_info *ino = autofs4_dentry_ino(dentry);
int status;
DPRINTK("dentry=%p %.*s ino=%p",
dentry, dentry->d_name.len, dentry->d_name.name, dentry->d_inode);
/*
* Wait for a pending mount, triggering one if there
* isn't one already
*/
if (dentry->d_inode == NULL) {
DPRINTK("waiting for mount name=%.*s",
dentry->d_name.len, dentry->d_name.name);
status = autofs4_wait(sbi, dentry, NFY_MOUNT);
DPRINTK("mount done status=%d", status);
/* Turn this into a real negative dentry? */
if (status == -ENOENT) {
spin_lock(&dentry->d_lock);
dentry->d_flags &= ~DCACHE_AUTOFS_PENDING;
spin_unlock(&dentry->d_lock);
return status;
} else if (status) {
/* Return a negative dentry, but leave it "pending" */
return status;
}
/* Trigger mount for path component or follow link */
} else if (dentry->d_flags & DCACHE_AUTOFS_PENDING ||
flags & (TRIGGER_FLAGS | TRIGGER_INTENTS) ||
current->link_count) {
DPRINTK("waiting for mount name=%.*s",
dentry->d_name.len, dentry->d_name.name);
spin_lock(&dentry->d_lock);
dentry->d_flags |= DCACHE_AUTOFS_PENDING;
spin_unlock(&dentry->d_lock);
status = autofs4_wait(sbi, dentry, NFY_MOUNT);
DPRINTK("mount done status=%d", status);
if (status) {
spin_lock(&dentry->d_lock);
dentry->d_flags &= ~DCACHE_AUTOFS_PENDING;
spin_unlock(&dentry->d_lock);
return status;
}
}
/* Initialize expiry counter after successful mount */
if (ino)
ino->last_used = jiffies;
spin_lock(&dentry->d_lock);
dentry->d_flags &= ~DCACHE_AUTOFS_PENDING;
spin_unlock(&dentry->d_lock);
autofs4: fix execution order race in mount request code Jeff Moyer has identified a race in due to an execution order dependency in the autofs4 function root.c:try_to_fill_dentry(). Jeff's description of this race is: "P1 does a lookup of /mount/submount/foo. Since the VFS can't find an entry for "foo" under /mount/submount, it calls into the autofs4 kernel module to allocate a new dentry, D1. The kernel creates a new waitq for this lookup and calls the daemon to perform the mount. The daemon performs a mkdir of the "foo" directory under /mount/submount, which ends up creating a *new* dentry, D2. Then, P2 does a lookup of /mount/submount/foo. The VFS path walking logic finds a dentry in the dcache, D2, and calls the revalidate function with this. In the autofs4 revalidate code, we then trigger a mount, since the dentry is an empty directory that isn't a mountpoint, and so set DCACHE_AUTOFS_PENDING and call into the wait code to trigger the mount. The wait code finds our existing waitq entry (since it is keyed off of the directory name) and adds itself to the list of waiters. After the daemon finishes the mount, it calls back into the kernel to release the waiters. When this happens, P1 is woken up and goes about clearing the DCACHE_AUTOFS_PENDING flag, but it does this in D1! So, given that P1 in our case is a program that will immediately try to access a file under /mount/submount/foo, we end up finding the dentry D2 which still has the pending flag set, and we set out to wait for a mount *again*! So, one way to address this is to re-do the lookup at the end of try_to_fill_dentry, and to clear the pending flag on the hashed dentry. This seems a sane approach to me." And Jeff's patch does this. Signed-off-by: Jeff Moyer <jmoyer@redhat.com> Signed-off-by-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-05-01 07:35:08 -04:00
return 0;
}
/* For autofs direct mounts the follow link triggers the mount */
static void *autofs4_follow_link(struct dentry *dentry, struct nameidata *nd)
{
struct autofs_sb_info *sbi = autofs4_sbi(dentry->d_sb);
struct autofs_info *ino = autofs4_dentry_ino(dentry);
int oz_mode = autofs4_oz_mode(sbi);
unsigned int lookup_type;
int status;
DPRINTK("dentry=%p %.*s oz_mode=%d nd->flags=%d",
dentry, dentry->d_name.len, dentry->d_name.name, oz_mode,
nd->flags);
/*
* For an expire of a covered direct or offset mount we need
* to break out of follow_down() at the autofs mount trigger
* (d_mounted--), so we can see the expiring flag, and manage
* the blocking and following here until the expire is completed.
*/
if (oz_mode) {
spin_lock(&sbi->fs_lock);
if (ino->flags & AUTOFS_INF_EXPIRING) {
spin_unlock(&sbi->fs_lock);
/* Follow down to our covering mount. */
if (!follow_down(&nd->path))
goto done;
goto follow;
}
spin_unlock(&sbi->fs_lock);
goto done;
}
/* If an expire request is pending everyone must wait. */
autofs4_expire_wait(dentry);
/* We trigger a mount for almost all flags */
lookup_type = nd->flags & (TRIGGER_FLAGS | TRIGGER_INTENTS);
if (!(lookup_type || dentry->d_flags & DCACHE_AUTOFS_PENDING))
goto follow;
/*
* If the dentry contains directories then it is an autofs
* multi-mount with no root mount offset. So don't try to
* mount it again.
*/
spin_lock(&dcache_lock);
if (dentry->d_flags & DCACHE_AUTOFS_PENDING ||
(!d_mountpoint(dentry) && __simple_empty(dentry))) {
spin_unlock(&dcache_lock);
status = try_to_fill_dentry(dentry, 0);
if (status)
goto out_error;
goto follow;
}
spin_unlock(&dcache_lock);
follow:
/*
* If there is no root mount it must be an autofs
* multi-mount with no root offset so we don't need
* to follow it.
*/
if (d_mountpoint(dentry)) {
if (!autofs4_follow_mount(&nd->path)) {
status = -ENOENT;
goto out_error;
}
}
done:
return NULL;
out_error:
path_put(&nd->path);
return ERR_PTR(status);
}
/*
* Revalidate is called on every cache lookup. Some of those
* cache lookups may actually happen while the dentry is not
* yet completely filled in, and revalidate has to delay such
* lookups..
*/
static int autofs4_revalidate(struct dentry *dentry, struct nameidata *nd)
{
struct inode *dir = dentry->d_parent->d_inode;
struct autofs_sb_info *sbi = autofs4_sbi(dir->i_sb);
int oz_mode = autofs4_oz_mode(sbi);
int flags = nd ? nd->flags : 0;
int status = 1;
/* Pending dentry */
spin_lock(&sbi->fs_lock);
if (autofs4_ispending(dentry)) {
/* The daemon never causes a mount to trigger */
spin_unlock(&sbi->fs_lock);
if (oz_mode)
return 1;
/*
* If the directory has gone away due to an expire
* we have been called as ->d_revalidate() and so
* we need to return false and proceed to ->lookup().
*/
if (autofs4_expire_wait(dentry) == -EAGAIN)
return 0;
/*
* A zero status is success otherwise we have a
* negative error code.
*/
status = try_to_fill_dentry(dentry, flags);
if (status == 0)
return 1;
return status;
}
spin_unlock(&sbi->fs_lock);
/* Negative dentry.. invalidate if "old" */
if (dentry->d_inode == NULL)
return 0;
/* Check for a non-mountpoint directory with no contents */
spin_lock(&dcache_lock);
if (S_ISDIR(dentry->d_inode->i_mode) &&
!d_mountpoint(dentry) &&
__simple_empty(dentry)) {
DPRINTK("dentry=%p %.*s, emptydir",
dentry, dentry->d_name.len, dentry->d_name.name);
spin_unlock(&dcache_lock);
/* The daemon never causes a mount to trigger */
if (oz_mode)
return 1;
/*
* A zero status is success otherwise we have a
* negative error code.
*/
status = try_to_fill_dentry(dentry, flags);
if (status == 0)
return 1;
return status;
}
spin_unlock(&dcache_lock);
return 1;
}
void autofs4_dentry_release(struct dentry *de)
{
struct autofs_info *inf;
DPRINTK("releasing %p", de);
inf = autofs4_dentry_ino(de);
de->d_fsdata = NULL;
if (inf) {
struct autofs_sb_info *sbi = autofs4_sbi(de->d_sb);
if (sbi) {
spin_lock(&sbi->lookup_lock);
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
if (!list_empty(&inf->active))
list_del(&inf->active);
if (!list_empty(&inf->expiring))
list_del(&inf->expiring);
spin_unlock(&sbi->lookup_lock);
}
inf->dentry = NULL;
inf->inode = NULL;
autofs4_free_ino(inf);
}
}
/* For dentries of directories in the root dir */
static const struct dentry_operations autofs4_root_dentry_operations = {
.d_revalidate = autofs4_revalidate,
.d_release = autofs4_dentry_release,
};
/* For other dentries */
static const struct dentry_operations autofs4_dentry_operations = {
.d_revalidate = autofs4_revalidate,
.d_release = autofs4_dentry_release,
};
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
static struct dentry *autofs4_lookup_active(struct autofs_sb_info *sbi, struct dentry *parent, struct qstr *name)
{
unsigned int len = name->len;
unsigned int hash = name->hash;
const unsigned char *str = name->name;
struct list_head *p, *head;
spin_lock(&dcache_lock);
spin_lock(&sbi->lookup_lock);
head = &sbi->active_list;
list_for_each(p, head) {
struct autofs_info *ino;
struct dentry *dentry;
struct qstr *qstr;
ino = list_entry(p, struct autofs_info, active);
dentry = ino->dentry;
spin_lock(&dentry->d_lock);
/* Already gone? */
if (atomic_read(&dentry->d_count) == 0)
goto next;
qstr = &dentry->d_name;
if (dentry->d_name.hash != hash)
goto next;
if (dentry->d_parent != parent)
goto next;
if (qstr->len != len)
goto next;
if (memcmp(qstr->name, str, len))
goto next;
if (d_unhashed(dentry)) {
dget(dentry);
spin_unlock(&dentry->d_lock);
spin_unlock(&sbi->lookup_lock);
spin_unlock(&dcache_lock);
return dentry;
}
next:
spin_unlock(&dentry->d_lock);
}
spin_unlock(&sbi->lookup_lock);
spin_unlock(&dcache_lock);
return NULL;
}
static struct dentry *autofs4_lookup_expiring(struct autofs_sb_info *sbi, struct dentry *parent, struct qstr *name)
{
unsigned int len = name->len;
unsigned int hash = name->hash;
const unsigned char *str = name->name;
struct list_head *p, *head;
spin_lock(&dcache_lock);
spin_lock(&sbi->lookup_lock);
head = &sbi->expiring_list;
list_for_each(p, head) {
struct autofs_info *ino;
struct dentry *dentry;
struct qstr *qstr;
ino = list_entry(p, struct autofs_info, expiring);
dentry = ino->dentry;
spin_lock(&dentry->d_lock);
/* Bad luck, we've already been dentry_iput */
if (!dentry->d_inode)
goto next;
qstr = &dentry->d_name;
if (dentry->d_name.hash != hash)
goto next;
if (dentry->d_parent != parent)
goto next;
if (qstr->len != len)
goto next;
if (memcmp(qstr->name, str, len))
goto next;
if (d_unhashed(dentry)) {
dget(dentry);
spin_unlock(&dentry->d_lock);
spin_unlock(&sbi->lookup_lock);
spin_unlock(&dcache_lock);
return dentry;
}
next:
spin_unlock(&dentry->d_lock);
}
spin_unlock(&sbi->lookup_lock);
spin_unlock(&dcache_lock);
return NULL;
}
/* Lookups in the root directory */
static struct dentry *autofs4_lookup(struct inode *dir, struct dentry *dentry, struct nameidata *nd)
{
struct autofs_sb_info *sbi;
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
struct autofs_info *ino;
struct dentry *expiring, *unhashed;
int oz_mode;
DPRINTK("name = %.*s",
dentry->d_name.len, dentry->d_name.name);
/* File name too long to exist */
if (dentry->d_name.len > NAME_MAX)
return ERR_PTR(-ENAMETOOLONG);
sbi = autofs4_sbi(dir->i_sb);
oz_mode = autofs4_oz_mode(sbi);
DPRINTK("pid = %u, pgrp = %u, catatonic = %d, oz_mode = %d",
current->pid, task_pgrp_nr(current), sbi->catatonic, oz_mode);
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
unhashed = autofs4_lookup_active(sbi, dentry->d_parent, &dentry->d_name);
if (unhashed)
dentry = unhashed;
else {
/*
* Mark the dentry incomplete but don't hash it. We do this
* to serialize our inode creation operations (symlink and
* mkdir) which prevents deadlock during the callback to
* the daemon. Subsequent user space lookups for the same
* dentry are placed on the wait queue while the daemon
* itself is allowed passage unresticted so the create
* operation itself can then hash the dentry. Finally,
* we check for the hashed dentry and return the newly
* hashed dentry.
*/
dentry->d_op = &autofs4_root_dentry_operations;
/*
* And we need to ensure that the same dentry is used for
* all following lookup calls until it is hashed so that
* the dentry flags are persistent throughout the request.
*/
ino = autofs4_init_ino(NULL, sbi, 0555);
if (!ino)
return ERR_PTR(-ENOMEM);
dentry->d_fsdata = ino;
ino->dentry = dentry;
spin_lock(&sbi->lookup_lock);
list_add(&ino->active, &sbi->active_list);
spin_unlock(&sbi->lookup_lock);
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
d_instantiate(dentry, NULL);
}
if (!oz_mode) {
mutex_unlock(&dir->i_mutex);
expiring = autofs4_lookup_expiring(sbi,
dentry->d_parent,
&dentry->d_name);
if (expiring) {
/*
* If we are racing with expire the request might not
* be quite complete but the directory has been removed
* so it must have been successful, so just wait for it.
*/
ino = autofs4_dentry_ino(expiring);
autofs4_expire_wait(expiring);
spin_lock(&sbi->lookup_lock);
if (!list_empty(&ino->expiring))
list_del_init(&ino->expiring);
spin_unlock(&sbi->lookup_lock);
dput(expiring);
}
spin_lock(&dentry->d_lock);
dentry->d_flags |= DCACHE_AUTOFS_PENDING;
spin_unlock(&dentry->d_lock);
if (dentry->d_op && dentry->d_op->d_revalidate)
(dentry->d_op->d_revalidate)(dentry, nd);
mutex_lock(&dir->i_mutex);
}
/*
* If we are still pending, check if we had to handle
* a signal. If so we can force a restart..
*/
if (dentry->d_flags & DCACHE_AUTOFS_PENDING) {
/* See if we were interrupted */
if (signal_pending(current)) {
sigset_t *sigset = &current->pending.signal;
if (sigismember (sigset, SIGKILL) ||
sigismember (sigset, SIGQUIT) ||
sigismember (sigset, SIGINT)) {
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
if (unhashed)
dput(unhashed);
return ERR_PTR(-ERESTARTNOINTR);
}
}
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
if (!oz_mode) {
spin_lock(&dentry->d_lock);
dentry->d_flags &= ~DCACHE_AUTOFS_PENDING;
spin_unlock(&dentry->d_lock);
}
}
/*
* If this dentry is unhashed, then we shouldn't honour this
* lookup. Returning ENOENT here doesn't do the right thing
* for all system calls, but it should be OK for the operations
* we permit from an autofs.
*/
if (!oz_mode && d_unhashed(dentry)) {
/*
* A user space application can (and has done in the past)
* remove and re-create this directory during the callback.
* This can leave us with an unhashed dentry, but a
* successful mount! So we need to perform another
* cached lookup in case the dentry now exists.
*/
struct dentry *parent = dentry->d_parent;
struct dentry *new = d_lookup(parent, &dentry->d_name);
if (new != NULL)
dentry = new;
else
dentry = ERR_PTR(-ENOENT);
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
if (unhashed)
dput(unhashed);
return dentry;
}
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
if (unhashed)
return unhashed;
return NULL;
}
static int autofs4_dir_symlink(struct inode *dir,
struct dentry *dentry,
const char *symname)
{
struct autofs_sb_info *sbi = autofs4_sbi(dir->i_sb);
struct autofs_info *ino = autofs4_dentry_ino(dentry);
struct autofs_info *p_ino;
struct inode *inode;
char *cp;
DPRINTK("%s <- %.*s", symname,
dentry->d_name.len, dentry->d_name.name);
if (!autofs4_oz_mode(sbi))
return -EACCES;
ino = autofs4_init_ino(ino, sbi, S_IFLNK | 0555);
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
if (!ino)
return -ENOMEM;
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
spin_lock(&sbi->lookup_lock);
if (!list_empty(&ino->active))
list_del_init(&ino->active);
spin_unlock(&sbi->lookup_lock);
ino->size = strlen(symname);
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
cp = kmalloc(ino->size + 1, GFP_KERNEL);
if (!cp) {
if (!dentry->d_fsdata)
kfree(ino);
return -ENOMEM;
}
strcpy(cp, symname);
inode = autofs4_get_inode(dir->i_sb, ino);
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
if (!inode) {
kfree(cp);
if (!dentry->d_fsdata)
kfree(ino);
return -ENOMEM;
}
d_add(dentry, inode);
if (dir == dir->i_sb->s_root->d_inode)
dentry->d_op = &autofs4_root_dentry_operations;
else
dentry->d_op = &autofs4_dentry_operations;
dentry->d_fsdata = ino;
ino->dentry = dget(dentry);
atomic_inc(&ino->count);
p_ino = autofs4_dentry_ino(dentry->d_parent);
if (p_ino && dentry->d_parent != dentry)
atomic_inc(&p_ino->count);
ino->inode = inode;
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
ino->u.symlink = cp;
dir->i_mtime = CURRENT_TIME;
return 0;
}
/*
* NOTE!
*
* Normal filesystems would do a "d_delete()" to tell the VFS dcache
* that the file no longer exists. However, doing that means that the
* VFS layer can turn the dentry into a negative dentry. We don't want
* this, because the unlink is probably the result of an expire.
* We simply d_drop it and add it to a expiring list in the super block,
* which allows the dentry lookup to check for an incomplete expire.
*
* If a process is blocked on the dentry waiting for the expire to finish,
* it will invalidate the dentry and try to mount with a new one.
*
* Also see autofs4_dir_rmdir()..
*/
static int autofs4_dir_unlink(struct inode *dir, struct dentry *dentry)
{
struct autofs_sb_info *sbi = autofs4_sbi(dir->i_sb);
struct autofs_info *ino = autofs4_dentry_ino(dentry);
struct autofs_info *p_ino;
/* This allows root to remove symlinks */
if (!autofs4_oz_mode(sbi) && !capable(CAP_SYS_ADMIN))
return -EACCES;
if (atomic_dec_and_test(&ino->count)) {
p_ino = autofs4_dentry_ino(dentry->d_parent);
if (p_ino && dentry->d_parent != dentry)
atomic_dec(&p_ino->count);
}
dput(ino->dentry);
dentry->d_inode->i_size = 0;
clear_nlink(dentry->d_inode);
dir->i_mtime = CURRENT_TIME;
spin_lock(&dcache_lock);
spin_lock(&sbi->lookup_lock);
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
if (list_empty(&ino->expiring))
list_add(&ino->expiring, &sbi->expiring_list);
spin_unlock(&sbi->lookup_lock);
spin_lock(&dentry->d_lock);
__d_drop(dentry);
spin_unlock(&dentry->d_lock);
spin_unlock(&dcache_lock);
return 0;
}
static int autofs4_dir_rmdir(struct inode *dir, struct dentry *dentry)
{
struct autofs_sb_info *sbi = autofs4_sbi(dir->i_sb);
struct autofs_info *ino = autofs4_dentry_ino(dentry);
struct autofs_info *p_ino;
DPRINTK("dentry %p, removing %.*s",
dentry, dentry->d_name.len, dentry->d_name.name);
if (!autofs4_oz_mode(sbi))
return -EACCES;
spin_lock(&dcache_lock);
if (!list_empty(&dentry->d_subdirs)) {
spin_unlock(&dcache_lock);
return -ENOTEMPTY;
}
spin_lock(&sbi->lookup_lock);
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
if (list_empty(&ino->expiring))
list_add(&ino->expiring, &sbi->expiring_list);
spin_unlock(&sbi->lookup_lock);
spin_lock(&dentry->d_lock);
__d_drop(dentry);
spin_unlock(&dentry->d_lock);
spin_unlock(&dcache_lock);
if (atomic_dec_and_test(&ino->count)) {
p_ino = autofs4_dentry_ino(dentry->d_parent);
if (p_ino && dentry->d_parent != dentry)
atomic_dec(&p_ino->count);
}
dput(ino->dentry);
dentry->d_inode->i_size = 0;
clear_nlink(dentry->d_inode);
if (dir->i_nlink)
drop_nlink(dir);
return 0;
}
static int autofs4_dir_mkdir(struct inode *dir, struct dentry *dentry, int mode)
{
struct autofs_sb_info *sbi = autofs4_sbi(dir->i_sb);
struct autofs_info *ino = autofs4_dentry_ino(dentry);
struct autofs_info *p_ino;
struct inode *inode;
if (!autofs4_oz_mode(sbi))
return -EACCES;
DPRINTK("dentry %p, creating %.*s",
dentry, dentry->d_name.len, dentry->d_name.name);
ino = autofs4_init_ino(ino, sbi, S_IFDIR | 0555);
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
if (!ino)
return -ENOMEM;
spin_lock(&sbi->lookup_lock);
if (!list_empty(&ino->active))
list_del_init(&ino->active);
spin_unlock(&sbi->lookup_lock);
inode = autofs4_get_inode(dir->i_sb, ino);
autofs4: use look aside list for lookups A while ago a patch to resolve a deadlock during directory creation was merged. This delayed the hashing of lookup dentrys until the ->mkdir() (or ->symlink()) operation completed to ensure we always went through ->lookup() instead of also having processes go through ->revalidate() so our VFS locking remained consistent. Now we are seeing a couple of side affects of that change in situations with heavy mount activity. Two cases have been identified: 1) When a mount request is triggered, due to the delayed hashing, the directory created by user space for the mount point doesn't have the DCACHE_AUTOFS_PENDING flag set. In the case of an autofs multi-mount where a tree of mount point directories are created this can lead to the path walk continuing rather than the dentry being sent to the wait queue to wait for request completion. This is because, if the pending flag isn't set, the criteria for deciding this is a mount in progress fails to hold, namely that the dentry is not a mount point and has no subdirectories. 2) A mount request dentry is initially created negative and unhashed. It remains this way until the ->mkdir() callback completes. Since it is unhashed a fresh dentry is used when the user space mount request creates the mount point directory. This leaves the original dentry negative and unhashed. But revalidate has no way to tell the VFS that the dentry has changed, other than to force another ->lookup() by returning false, which is at best wastefull and at worst not possible. This results in an -ENOENT return from the original path walk when in fact the mount succeeded. To resolve this we need to ensure that the same dentry is used in all calls to ->lookup() during the course of a mount request. This patch achieves that by adding the initial dentry to a look aside list and removes it at ->mkdir() or ->symlink() completion (or when the dentry is released), since these are the only create operations autofs4 supports. Signed-off-by: Ian Kent <raven@themaw.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-24 00:30:12 -04:00
if (!inode) {
if (!dentry->d_fsdata)
kfree(ino);
return -ENOMEM;
}
d_add(dentry, inode);
if (dir == dir->i_sb->s_root->d_inode)
dentry->d_op = &autofs4_root_dentry_operations;
else
dentry->d_op = &autofs4_dentry_operations;
dentry->d_fsdata = ino;
ino->dentry = dget(dentry);
atomic_inc(&ino->count);
p_ino = autofs4_dentry_ino(dentry->d_parent);
if (p_ino && dentry->d_parent != dentry)
atomic_inc(&p_ino->count);
ino->inode = inode;
inc_nlink(dir);
dir->i_mtime = CURRENT_TIME;
return 0;
}
/* Get/set timeout ioctl() operation */
static inline int autofs4_get_set_timeout(struct autofs_sb_info *sbi,
unsigned long __user *p)
{
int rv;
unsigned long ntimeout;
if ((rv = get_user(ntimeout, p)) ||
(rv = put_user(sbi->exp_timeout/HZ, p)))
return rv;
if (ntimeout > ULONG_MAX/HZ)
sbi->exp_timeout = 0;
else
sbi->exp_timeout = ntimeout * HZ;
return 0;
}
/* Return protocol version */
static inline int autofs4_get_protover(struct autofs_sb_info *sbi, int __user *p)
{
return put_user(sbi->version, p);
}
/* Return protocol sub version */
static inline int autofs4_get_protosubver(struct autofs_sb_info *sbi, int __user *p)
{
return put_user(sbi->sub_version, p);
}
/*
* Tells the daemon whether it can umount the autofs mount.
*/
static inline int autofs4_ask_umount(struct vfsmount *mnt, int __user *p)
{
int status = 0;
if (may_umount(mnt))
status = 1;
DPRINTK("returning %d", status);
status = put_user(status, p);
return status;
}
/* Identify autofs4_dentries - this is so we can tell if there's
an extra dentry refcount or not. We only hold a refcount on the
dentry if its non-negative (ie, d_inode != NULL)
*/
int is_autofs4_dentry(struct dentry *dentry)
{
return dentry && dentry->d_inode &&
(dentry->d_op == &autofs4_root_dentry_operations ||
dentry->d_op == &autofs4_dentry_operations) &&
dentry->d_fsdata != NULL;
}
/*
* ioctl()'s on the root directory is the chief method for the daemon to
* generate kernel reactions
*/
static int autofs4_root_ioctl(struct inode *inode, struct file *filp,
unsigned int cmd, unsigned long arg)
{
struct autofs_sb_info *sbi = autofs4_sbi(inode->i_sb);
void __user *p = (void __user *)arg;
DPRINTK("cmd = 0x%08x, arg = 0x%08lx, sbi = %p, pgrp = %u",
cmd,arg,sbi,task_pgrp_nr(current));
if (_IOC_TYPE(cmd) != _IOC_TYPE(AUTOFS_IOC_FIRST) ||
_IOC_NR(cmd) - _IOC_NR(AUTOFS_IOC_FIRST) >= AUTOFS_IOC_COUNT)
return -ENOTTY;
if (!autofs4_oz_mode(sbi) && !capable(CAP_SYS_ADMIN))
return -EPERM;
switch(cmd) {
case AUTOFS_IOC_READY: /* Wait queue: go ahead and retry */
return autofs4_wait_release(sbi,(autofs_wqt_t)arg,0);
case AUTOFS_IOC_FAIL: /* Wait queue: fail with ENOENT */
return autofs4_wait_release(sbi,(autofs_wqt_t)arg,-ENOENT);
case AUTOFS_IOC_CATATONIC: /* Enter catatonic mode (daemon shutdown) */
autofs4_catatonic_mode(sbi);
return 0;
case AUTOFS_IOC_PROTOVER: /* Get protocol version */
return autofs4_get_protover(sbi, p);
case AUTOFS_IOC_PROTOSUBVER: /* Get protocol sub version */
return autofs4_get_protosubver(sbi, p);
case AUTOFS_IOC_SETTIMEOUT:
return autofs4_get_set_timeout(sbi, p);
case AUTOFS_IOC_ASKUMOUNT:
return autofs4_ask_umount(filp->f_path.mnt, p);
/* return a single thing to expire */
case AUTOFS_IOC_EXPIRE:
return autofs4_expire_run(inode->i_sb,filp->f_path.mnt,sbi, p);
/* same as above, but can send multiple expires through pipe */
case AUTOFS_IOC_EXPIRE_MULTI:
return autofs4_expire_multi(inode->i_sb,filp->f_path.mnt,sbi, p);
default:
return -ENOSYS;
}
}