| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
|
|
|
| |
inotify will WARN() if it finds that the idr and the fsnotify internals
somehow got out of sync. It was only supposed to do this once but due
to this stupid bug it would warn every single time a problem was
detected.
Signed-off-by: Eric Paris <eparis@redhat.com>
Cc: stable@kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Since commit 7e790dd5fc937bc8d2400c30a05e32a9e9eef276 ("inotify: fix
error paths in inotify_update_watch") inotify changed the manor in which
it gave watch descriptors back to userspace. Previous to this commit
inotify acted like the following:
inotify_add_watch(X, Y, Z) = 1
inotify_rm_watch(X, 1);
inotify_add_watch(X, Y, Z) = 2
but after this patch inotify would return watch descriptors like so:
inotify_add_watch(X, Y, Z) = 1
inotify_rm_watch(X, 1);
inotify_add_watch(X, Y, Z) = 1
which I saw as equivalent to opening an fd where
open(file) = 1;
close(1);
open(file) = 1;
seemed perfectly reasonable. The issue is that quite a bit of userspace
apparently relies on the behavior in which watch descriptors will not be
quickly reused. KDE relies on it, I know some selinux packages rely on
it, and I have heard complaints from other random sources such as debian
bug 558981.
Although the man page implies what we do is ok, we broke userspace so
this patch almost reverts us to the old behavior. It is still slightly
racey and I have patches that would fix that, but they are rather large
and this will fix it for all real world cases. The race is as follows:
- task1 creates a watch and blocks in idr_new_watch() before it updates
the hint.
- task2 creates a watch and updates the hint.
- task1 updates the hint with it's older wd
- task removes the watch created by task2
- task adds a new watch and will reuse the wd originally given to task2
it requires moving some locking around the hint (last_wd) but this should
solve it for the real world and be -stable safe.
As a side effect this patch papers over a bug in the lib/idr code which
is causing a large number WARN's to pop on people's system and many
reports in kerneloops.org. I'm working on the root cause of that idr
bug seperately but this should make inotify immune to that issue.
Signed-off-by: Eric Paris <eparis@redhat.com>
Cc: stable@kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
|
|
|
|
|
| |
... and have the caller grab both mnt and dentry; kill
leak in infiniband, while we are at it.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
|
|
|
|
| |
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
|
|\
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
git://git.kernel.org/pub/scm/linux/kernel/git/jikos/trivial
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jikos/trivial: (42 commits)
tree-wide: fix misspelling of "definition" in comments
reiserfs: fix misspelling of "journaled"
doc: Fix a typo in slub.txt.
inotify: remove superfluous return code check
hdlc: spelling fix in find_pvc() comment
doc: fix regulator docs cut-and-pasteism
mtd: Fix comment in Kconfig
doc: Fix IRQ chip docs
tree-wide: fix assorted typos all over the place
drivers/ata/libata-sff.c: comment spelling fixes
fix typos/grammos in Documentation/edac.txt
sysctl: add missing comments
fs/debugfs/inode.c: fix comment typos
sgivwfb: Make use of ARRAY_SIZE.
sky2: fix sky2_link_down copy/paste comment error
tree-wide: fix typos "couter" -> "counter"
tree-wide: fix typos "offest" -> "offset"
fix kerneldoc for set_irq_msi()
spidev: fix double "of of" in comment
comment typo fix: sybsystem -> subsystem
...
|
| |
| |
| |
| |
| | |
Signed-off-by: Giuseppe Scrivano <gscrivano@gnu.org>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
For consistency drop & in front of every proc_handler. Explicity
taking the address is unnecessary and it prevents optimizations
like stubbing the proc_handlers to NULL.
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Joe Perches <joe@perches.com>
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
|
|/
|
|
|
|
|
|
| |
Now that sys_sysctl is a generic wrapper around /proc/sys .ctl_name
and .strategy members of sysctl tables are dead code. Remove them.
Cc: Jan Harkes <jaharkes@cs.cmu.edu>
Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Mask off FS_EVENT_ON_CHILD in dnotify_handle_event(). Otherwise, when there
is more than one watch on a directory and dnotify_should_send_event()
succeeds, events with FS_EVENT_ON_CHILD set will trigger all watches and cause
spurious events.
This case was overlooked in commit e42e2773.
#define _GNU_SOURCE
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <signal.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <string.h>
static void create_event(int s, siginfo_t* si, void* p)
{
printf("create\n");
}
static void delete_event(int s, siginfo_t* si, void* p)
{
printf("delete\n");
}
int main (void) {
struct sigaction action;
char *tmpdir, *file;
int fd1, fd2;
sigemptyset (&action.sa_mask);
action.sa_flags = SA_SIGINFO;
action.sa_sigaction = create_event;
sigaction (SIGRTMIN + 0, &action, NULL);
action.sa_sigaction = delete_event;
sigaction (SIGRTMIN + 1, &action, NULL);
# define TMPDIR "/tmp/test.XXXXXX"
tmpdir = malloc(strlen(TMPDIR) + 1);
strcpy(tmpdir, TMPDIR);
mkdtemp(tmpdir);
# define TMPFILE "/file"
file = malloc(strlen(tmpdir) + strlen(TMPFILE) + 1);
sprintf(file, "%s/%s", tmpdir, TMPFILE);
fd1 = open (tmpdir, O_RDONLY);
fcntl(fd1, F_SETSIG, SIGRTMIN);
fcntl(fd1, F_NOTIFY, DN_MULTISHOT | DN_CREATE);
fd2 = open (tmpdir, O_RDONLY);
fcntl(fd2, F_SETSIG, SIGRTMIN + 1);
fcntl(fd2, F_NOTIFY, DN_MULTISHOT | DN_DELETE);
if (fork()) {
/* This triggers a create event */
creat(file, 0600);
/* This triggers a create and delete event (!) */
unlink(file);
} else {
sleep(1);
rmdir(tmpdir);
}
return 0;
}
Signed-off-by: Andreas Gruenbacher <agruen@suse.de>
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
If we do rename a dir entry, like this:
rename("/tmp/ino7UrgoJ.rename1", "/tmp/ino7UrgoJ.rename2")
rename("/tmp/ino7UrgoJ.rename2", "/tmp/ino7UrgoJ")
The duplicate events should be coalesced into a single event. But those two
events do not be coalesced into a single event, due to some bad check in
event_compare(). It can not match the two NULL inodes as the same event.
Signed-off-by: Wei Yongjun <yjwei@cn.fujitsu.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
fsnotify_add_mark is supposed to add a mark to the g_list and i_list and to
set the group and inode for the mark. fsnotify_destroy_mark_by_entry uses
the fact that ->group != NULL to know if this group should be destroyed or
if it's already been done.
But fsnotify_add_mark sets the group and inode before it actually adds the
mark to the i_list and g_list. This can result in a race in inotify, it
requires 3 threads.
sys_inotify_add_watch("file") sys_inotify_add_watch("file") sys_inotify_rm_watch([a])
inotify_update_watch()
inotify_new_watch()
inotify_add_to_idr()
^--- returns wd = [a]
inotfiy_update_watch()
inotify_new_watch()
inotify_add_to_idr()
fsnotify_add_mark()
^--- returns wd = [b]
returns to userspace;
inotify_idr_find([a])
^--- gives us the pointer from task 1
fsnotify_add_mark()
^--- this is going to set the mark->group and mark->inode fields, but will
return -EEXIST because of the race with [b].
fsnotify_destroy_mark()
^--- since ->group != NULL we call back
into inotify_freeing_mark() which calls
inotify_remove_from_idr([a])
since fsnotify_add_mark() failed we call:
inotify_remove_from_idr([a]) <------WHOOPS it's not in the idr, this could
have been any entry added later!
The fix is to make sure we don't set mark->group until we are sure the mark is
on the inode and fsnotify_add_mark will return success.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
| |
Seperating the addition and update of marks in inotify resulted in a
regression in that inotify never gets events. The inotify group mask is
always 0. This mask should be updated any time a new mark is added.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
| |
0db501bd0610ee0c0 introduced a regresion in that it now sends a nul
terminator but the length accounting when checking for space or
reporting to userspace did not take this into account. This corrects
all of the rounding logic.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
|
|
| |
When an event has no pathname, there's no need to pad it with a null byte and
therefore generate an inotify_event sized block of zeros. This fixes a
regression introduced by commit 0db501bd0610ee0c0aca84d927f90bcccd09e2bd where
my system wouldn't finish booting because some process was being confused by
this.
Signed-off-by: Brian Rogers <brian@xyzw.org>
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Before the rewrite copy_event_to_user always wrote a terqminating '\0'
byte to user space after the filename. Since the rewrite that
terminating byte was skipped if your filename is exactly a multiple of
event_size. Ouch!
So add one byte to name_size before we round up and use clear_user to
set userspace to zero like /dev/zero does instead of copying the
strange nul_inotify_event. I can't quite convince myself len_to_zero
will never exceed 16 and even if it doesn't clear_user should be more
efficient and a more accurate reflection of what the code is trying to
do.
Signed-off-by: Eric W. Biederman <ebiederm@aristanetworks.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
| |
The are races around the idr storage of inotify watches. It's possible
that a watch could be found from sys_inotify_rm_watch() in the idr, but it
could be removed from the idr before that code does it's removal. Move the
locking and the refcnt'ing so that these have to happen atomically.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
| |
If an inotify watch is left in the idr when an fsnotify group is destroyed
this will lead to a BUG. This is not a dangerous situation and really
indicates a programming bug and leak of memory. This patch changes it to
use a WARN and a printk rather than killing people's boxes.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
| |
There is nothing known wrong with the inotify watch addition/modification
but this patch seperates the two code paths to make them each easy to
verify as correct.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The inotify_add_watch man page specifies that inotify_add_watch() will
return a non-negative integer. However, historically the inotify
watches started at 1, not at 0.
Turns out that the inotifywait program provided by the inotify-tools
package doesn't properly handle a 0 watch descriptor. In 7e790dd5 we
changed from starting at 1 to starting at 0. This patch starts at 1,
just like in previous kernels, but also just like in previous kernels
it's possible for it to wrap back to 0. This preserves the kernel
functionality exactly like it was before the patch (neither method broke
the spec)
Signed-off-by: Eric Paris <eparis@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
|
|
|
|
|
|
|
|
|
| |
In f44aebcc the tail drop logic of events with no file backing
(q_overflow and in_ignored) was reversed so IN_IGNORED events would
never be tail dropped. This now means that Q_OVERFLOW events are NOT
tail dropped. The fix is to not tail drop IN_IGNORED, but to tail drop
Q_OVERFLOW.
Signed-off-by: Eric Paris <eparis@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
|
|
|
|
|
|
|
|
|
|
| |
inotify decides if private data it passed to get added to an event was
used by checking list_empty(). But it's possible that the event may
have been dequeued and the private event removed so it would look empty.
The fix is to use the return code from fsnotify_add_notify_event rather
than looking at the list.
Signed-off-by: Eric Paris <eparis@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
inotify can have a watchs removed under filesystem reclaim.
=================================
[ INFO: inconsistent lock state ]
2.6.31-rc2 #16
---------------------------------
inconsistent {IN-RECLAIM_FS-W} -> {RECLAIM_FS-ON-W} usage.
khubd/217 [HC0[0]:SC0[0]:HE1:SE1] takes:
(iprune_mutex){+.+.?.}, at: [<c10ba899>] invalidate_inodes+0x20/0xe3
{IN-RECLAIM_FS-W} state was registered at:
[<c10536ab>] __lock_acquire+0x2c9/0xac4
[<c1053f45>] lock_acquire+0x9f/0xc2
[<c1308872>] __mutex_lock_common+0x2d/0x323
[<c1308c00>] mutex_lock_nested+0x2e/0x36
[<c10ba6ff>] shrink_icache_memory+0x38/0x1b2
[<c108bfb6>] shrink_slab+0xe2/0x13c
[<c108c3e1>] kswapd+0x3d1/0x55d
[<c10449b5>] kthread+0x66/0x6b
[<c1003fdf>] kernel_thread_helper+0x7/0x10
[<ffffffff>] 0xffffffff
Two things are needed to fix this. First we need a method to tell
fsnotify_create_event() to use GFP_NOFS and second we need to stop using
one global IN_IGNORED event and allocate them one at a time. This solves
current issues with multiple IN_IGNORED on a queue having tail drop
problems and simplifies the allocations since we don't have to worry about
two tasks opperating on the IGNORED event concurrently.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
|
|
|
| |
fsnotify drops new events when they are the same as the tail event on the
queue to be sent to userspace. The problem is that if the event comes with
a path we forget to break out of the switch statement and fall into the
code path which matches on events that do not have any type of file backed
information (things like IN_UNMOUNT and IN_Q_OVERFLOW). The problem is
that this code thinks all such events should be dropped. Fix is to add a
break.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
inotify drops events if the last event on the queue is the same as the
current event. But it does 2 things wrong. First it is comparing old->inode
with new->inode. But after an event if put on the queue the ->inode is no
longer allowed to be used. It's possible between the last event and this new
event the inode could be reused and we would falsely match the inode's memory
address between two differing events.
The second problem is that when a file is removed fsnotify is passed the
negative dentry for the removed object rather than the postive dentry from
immediately before the removal. This mean the (broken) inotify tail drop code
was matching the NULL ->inode of differing events.
The fix is to check the file name which is stored with events when doing the
tail drop instead of wrongly checking the address of the stored ->inode.
Reported-by: Scott James Remnant <scott@ubuntu.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
|
|
| |
fsnotify doens't give the user anything. If someone chooses inotify or
dnotify it should build fsnotify, if they don't select one it shouldn't be
built. This patch changes fsnotify to be a def_bool=n and makes everything
else select it. Also fixes the issue people complained about on lwn where
gdm hung because they didn't have inotify and they didn't get the inotify
build option.....
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
|
|
| |
inotify_update_watch could leave things in a horrid state on a number of
error paths. We could try to remove idr entries that didn't exist, we
could send an IN_IGNORED to userspace for watches that don't exist, and a
bit of other stupidity. Clean these up by doing the idr addition before we
put the mark on the inode since we can clean that up on error and getting
off the inode's mark list is hard.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
inotify_add_watch had a couple of problems. The biggest being that if
inotify_add_watch was called on the same inode twice (to update or change the
event mask) a refence was taken on the original inode mark by
fsnotify_find_mark_entry but was not being dropped at the end of the
inotify_add_watch call. Thus if inotify_rm_watch was called although the mark
was removed from the inode, the refcnt wouldn't hit zero and we would leak
memory.
Reported-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
| |
The inotify rewrite forgot to drop the inotify watch use cound when a watch
was removed. This means that a single inotify fd can only ever register a
maximum of /proc/sys/fs/max_user_watches even if some of those had been
freed.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
| |
The per-user inotify_devs value is incremented each time a new file is
allocated, but never decremented. This led to inotify_init failing after a
limited number of calls.
Signed-off-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
inotify_destroy_mark_entry could get called twice for the same mark since it
is called directly in inotify_rm_watch and when the mark is being destroyed for
another reason. As an example assume that the file being watched was just
deleted so inotify_destroy_mark_entry would get called from the path
fsnotify_inoderemove() -> fsnotify_destroy_marks_by_inode() ->
fsnotify_destroy_mark_entry() -> inotify_destroy_mark_entry(). If this
happened at the same time as userspace tried to remove a watch via
inotify_rm_watch we could attempt to remove the mark from the idr twice and
could thus double dec the ref cnt and potentially could be in a use after
free/double free situation. The fix is to have inotify_rm_watch use the
generic recursive safe fsnotify_destroy_mark_by_entry() so we are sure the
inotify_destroy_mark_entry() function can only be called one.
This patch also renames the function to inotify_ingored_remove_idr() so it is
clear what is actually going on in the function.
Hopefully this fixes:
[ 20.342058] idr_remove called for id=20 which is not allocated.
[ 20.348000] Pid: 1860, comm: udevd Not tainted 2.6.30-tip #1077
[ 20.353933] Call Trace:
[ 20.356410] [<ffffffff811a82b7>] idr_remove+0x115/0x18f
[ 20.361737] [<ffffffff8134259d>] ? _spin_lock+0x6d/0x75
[ 20.367061] [<ffffffff8111640a>] ? inotify_destroy_mark_entry+0xa3/0xcf
[ 20.373771] [<ffffffff8111641e>] inotify_destroy_mark_entry+0xb7/0xcf
[ 20.380306] [<ffffffff81115913>] inotify_freeing_mark+0xe/0x10
[ 20.386238] [<ffffffff8111410d>] fsnotify_destroy_mark_by_entry+0x143/0x170
[ 20.393293] [<ffffffff811163a3>] inotify_destroy_mark_entry+0x3c/0xcf
[ 20.399829] [<ffffffff811164d1>] sys_inotify_rm_watch+0x9b/0xc6
[ 20.405850] [<ffffffff8100bcdb>] system_call_fastpath+0x16/0x1b
Reported-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Eric Paris <eparis@redhat.com>
Tested-by: Peter Ziljlstra <peterz@infradead.org>
|
|
|
|
|
|
|
|
| |
Most fsnotify listeners (all but inotify) do not care about marks being
freed. Allow groups to set freeing_mark to null and do not call any
function if it is set that way.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
| |
inotify and dnotify will both indicate that they want any event which came
from a child inode. The fix is to mask off FS_EVENT_ON_CHILD when deciding
if inotify or dnotify is interested in a given event.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
| |
entry->lock is needed to make sure entry->mask does not change while
manipulating it. In dnotify_should_send_event() we don't care if we get an
old or a new mask value out of this entry so there is no point it taking
the lock.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
| |
dnotify_should send event assigned a bool using ?true:false when computing
a bit operation. This is poitless and the bool type does this for us.
Signed-off-by: Eric Paris <eparis@redhat.com>
|
|
|
|
|
|
|
|
|
|
|
|
| |
Reimplement inotify_user using fsnotify. This should be feature for feature
exactly the same as the original inotify_user. This does not make any changes
to the in kernel inotify feature used by audit. Those patches (and the eventual
removal of in kernel inotify) will come after the new inotify_user proves to be
working correctly.
Signed-off-by: Eric Paris <eparis@redhat.com>
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
|
|
|
|
|
|
|
|
|
|
| |
When an fs is unmounted with an fsnotify mark entry attached to one of its
inodes we need to destroy that mark entry and we also (like inotify) send
an unmount event.
Signed-off-by: Eric Paris <eparis@redhat.com>
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch pins any inodes with an fsnotify mark in core. The idea is that
as soon as the mark is removed from the inode->fsnotify_mark_entries list
the inode will be iput. In reality is doesn't quite work exactly this way.
The igrab will happen when the mark is added to an inode, but the iput will
happen when the inode pointer is NULL'd inside the mark.
It's possible that 2 racing things will try to remove the mark from
different directions. One may try to remove the mark because of an
explicit request and one might try to remove it because the inode was
deleted. It's possible that the removal because of inode deletion will
remove the mark from the inode's list, but the removal by explicit request
will actually set entry->inode == NULL; and call the iput. This is safe.
Signed-off-by: Eric Paris <eparis@redhat.com>
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
|
|
|
|
|
|
|
|
|
|
| |
inotify needs per group information attached to events. This patch allows
groups to attach private information and implements a callback so that
information can be freed when an event is being destroyed.
Signed-off-by: Eric Paris <eparis@redhat.com>
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
|
|
|
|
|
|
|
|
|
|
|
| |
As part of the standard inotify events it includes a correlation cookie
between two dentry move operations. This patch includes the same behaviour
in fsnotify events. It is needed so that inotify userspace can be
implemented on top of fsnotify.
Signed-off-by: Eric Paris <eparis@redhat.com>
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
|
|
|
|
|
|
|
|
|
|
| |
When inotify wants to send events to a directory about a child it includes
the name of the original file. This patch collects that filename and makes
it available for notification.
Signed-off-by: Eric Paris <eparis@redhat.com>
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
|
|
|
|
|
|
|
|
|
|
|
| |
inotify needs to do asyc notification in which event information is stored
on a queue until the listener is ready to receive it. This patch
implements a generic notification queue for inotify (and later fanotify) to
store events to be sent at a later time.
Signed-off-by: Eric Paris <eparis@redhat.com>
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
|
|
|
|
|
|
|
|
| |
Reimplement dnotify using fsnotify.
Signed-off-by: Eric Paris <eparis@redhat.com>
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
|
|
|
|
|
|
|
|
|
|
|
| |
inotify and dnotify both use a similar parent notification mechanism. We
add a generic parent notification mechanism to fsnotify for both of these
to use. This new machanism also adds the dentry flag optimization which
exists for inotify to dnotify.
Signed-off-by: Eric Paris <eparis@redhat.com>
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch creates a way for fsnotify groups to attach marks to inodes.
These marks have little meaning to the generic fsnotify infrastructure
and thus their meaning should be interpreted by the group that attached
them to the inode's list.
dnotify and inotify will make use of these markings to indicate which
inodes are of interest to their respective groups. But this implementation
has the useful property that in the future other listeners could actually
use the marks for the exact opposite reason, aka to indicate which inodes
it had NO interest in.
Signed-off-by: Eric Paris <eparis@redhat.com>
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
fsnotify is a backend for filesystem notification. fsnotify does
not provide any userspace interface but does provide the basis
needed for other notification schemes such as dnotify. fsnotify
can be extended to be the backend for inotify or the upcoming
fanotify. fsnotify provides a mechanism for "groups" to register for
some set of filesystem events and to then deliver those events to
those groups for processing.
fsnotify has a number of benefits, the first being actually shrinking the size
of an inode. Before fsnotify to support both dnotify and inotify an inode had
unsigned long i_dnotify_mask; /* Directory notify events */
struct dnotify_struct *i_dnotify; /* for directory notifications */
struct list_head inotify_watches; /* watches on this inode */
struct mutex inotify_mutex; /* protects the watches list
But with fsnotify this same functionallity (and more) is done with just
__u32 i_fsnotify_mask; /* all events for this inode */
struct hlist_head i_fsnotify_mark_entries; /* marks on this inode */
That's right, inotify, dnotify, and fanotify all in 64 bits. We used that
much space just in inotify_watches alone, before this patch set.
fsnotify object lifetime and locking is MUCH better than what we have today.
inotify locking is incredibly complex. See 8f7b0ba1c8539 as an example of
what's been busted since inception. inotify needs to know internal semantics
of superblock destruction and unmounting to function. The inode pinning and
vfs contortions are horrible.
no fsnotify implementers do allocation under locks. This means things like
f04b30de3 which (due to an overabundance of caution) changes GFP_KERNEL to
GFP_NOFS can be reverted. There are no longer any allocation rules when using
or implementing your own fsnotify listener.
fsnotify paves the way for fanotify. In brief fanotify is a notification
mechanism that delivers the lisener both an 'event' and an open file descriptor
to the object in question. This means that fanotify is pathname agnostic.
Some on lkml may not care for the original companies or users that pushed for
TALPA, but fanotify was designed with flexibility and input for other users in
mind. The readahead group expressed interest in fanotify as it could be used
to profile disk access on boot without breaking the audit system. The desktop
search groups have also expressed interest in fanotify as it solves a number
of the race conditions and problems present with managing inotify when more
than a limited number of specific files are of interest. fanotify can provide
for a userspace access control system which makes it a clean interface for AV
vendors to hook without trying to do binary patching on the syscall table,
LSM, and everywhere else they do their things today. With this patch series
fanotify can be implemented in less than 1200 lines of easy to review code.
Almost all of which is the socket based user interface.
This patch series builds fsnotify to the point that it can implement
dnotify and inotify_user. Patches exist and will be sent soon after
acceptance to finish the in kernel inotify conversion (audit) and implement
fanotify.
Signed-off-by: Eric Paris <eparis@redhat.com>
Acked-by: Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@lst.de>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
There is what we believe to be a false positive reported by lockdep.
inotify_inode_queue_event() => take inotify_mutex => kernel_event() =>
kmalloc() => SLOB => alloc_pages_node() => page reclaim => slab reclaim =>
dcache reclaim => inotify_inode_is_dead => take inotify_mutex => deadlock
The plan is to fix this via lockdep annotation, but that is proving to be
quite involved.
The patch flips the allocation over to GFP_NFS to shut the warning up, for
the 2.6.30 release.
Hopefully we will fix this for real in 2.6.31. I'll queue a patch in -mm
to switch it back to GFP_KERNEL so we don't forget.
=================================
[ INFO: inconsistent lock state ]
2.6.30-rc2-next-20090417 #203
---------------------------------
inconsistent {RECLAIM_FS-ON-W} -> {IN-RECLAIM_FS-W} usage.
kswapd0/380 [HC0[0]:SC0[0]:HE1:SE1] takes:
(&inode->inotify_mutex){+.+.?.}, at: [<ffffffff8112f1b5>] inotify_inode_is_dead+0x35/0xb0
{RECLAIM_FS-ON-W} state was registered at:
[<ffffffff81079188>] mark_held_locks+0x68/0x90
[<ffffffff810792a5>] lockdep_trace_alloc+0xf5/0x100
[<ffffffff810f5261>] __kmalloc_node+0x31/0x1e0
[<ffffffff81130652>] kernel_event+0xe2/0x190
[<ffffffff81130826>] inotify_dev_queue_event+0x126/0x230
[<ffffffff8112f096>] inotify_inode_queue_event+0xc6/0x110
[<ffffffff8110444d>] vfs_create+0xcd/0x140
[<ffffffff8110825d>] do_filp_open+0x88d/0xa20
[<ffffffff810f6b68>] do_sys_open+0x98/0x140
[<ffffffff810f6c50>] sys_open+0x20/0x30
[<ffffffff8100c272>] system_call_fastpath+0x16/0x1b
[<ffffffffffffffff>] 0xffffffffffffffff
irq event stamp: 690455
hardirqs last enabled at (690455): [<ffffffff81564fe4>] _spin_unlock_irqrestore+0x44/0x80
hardirqs last disabled at (690454): [<ffffffff81565372>] _spin_lock_irqsave+0x32/0xa0
softirqs last enabled at (690178): [<ffffffff81052282>] __do_softirq+0x202/0x220
softirqs last disabled at (690157): [<ffffffff8100d50c>] call_softirq+0x1c/0x50
other info that might help us debug this:
2 locks held by kswapd0/380:
#0: (shrinker_rwsem){++++..}, at: [<ffffffff810d0bd7>] shrink_slab+0x37/0x180
#1: (&type->s_umount_key#17){++++..}, at: [<ffffffff8110cfbf>] shrink_dcache_memory+0x11f/0x1e0
stack backtrace:
Pid: 380, comm: kswapd0 Not tainted 2.6.30-rc2-next-20090417 #203
Call Trace:
[<ffffffff810789ef>] print_usage_bug+0x19f/0x200
[<ffffffff81018bff>] ? save_stack_trace+0x2f/0x50
[<ffffffff81078f0b>] mark_lock+0x4bb/0x6d0
[<ffffffff810799e0>] ? check_usage_forwards+0x0/0xc0
[<ffffffff8107b142>] __lock_acquire+0xc62/0x1ae0
[<ffffffff810f478c>] ? slob_free+0x10c/0x370
[<ffffffff8107c0a1>] lock_acquire+0xe1/0x120
[<ffffffff8112f1b5>] ? inotify_inode_is_dead+0x35/0xb0
[<ffffffff81562d43>] mutex_lock_nested+0x63/0x420
[<ffffffff8112f1b5>] ? inotify_inode_is_dead+0x35/0xb0
[<ffffffff8112f1b5>] ? inotify_inode_is_dead+0x35/0xb0
[<ffffffff81012fe9>] ? sched_clock+0x9/0x10
[<ffffffff81077165>] ? lock_release_holdtime+0x35/0x1c0
[<ffffffff8112f1b5>] inotify_inode_is_dead+0x35/0xb0
[<ffffffff8110c9dc>] dentry_iput+0xbc/0xe0
[<ffffffff8110cb23>] d_kill+0x33/0x60
[<ffffffff8110ce23>] __shrink_dcache_sb+0x2d3/0x350
[<ffffffff8110cffa>] shrink_dcache_memory+0x15a/0x1e0
[<ffffffff810d0cc5>] shrink_slab+0x125/0x180
[<ffffffff810d1540>] kswapd+0x560/0x7a0
[<ffffffff810ce160>] ? isolate_pages_global+0x0/0x2c0
[<ffffffff81065a30>] ? autoremove_wake_function+0x0/0x40
[<ffffffff8107953d>] ? trace_hardirqs_on+0xd/0x10
[<ffffffff810d0fe0>] ? kswapd+0x0/0x7a0
[<ffffffff8106555b>] kthread+0x5b/0xa0
[<ffffffff8100d40a>] child_rip+0xa/0x20
[<ffffffff8100cdd0>] ? restore_args+0x0/0x30
[<ffffffff81065500>] ? kthread+0x0/0xa0
[<ffffffff8100d400>] ? child_rip+0x0/0x20
[eparis@redhat.com: fix audit too]
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Matt Mackall <mpm@selenic.com>
Cc: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-off-by: Eric Paris <eparis@redhat.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
To be on the safe side, it should be less fragile to exclude I_NEW inodes
from inode list scans by default (unless there is an important reason to
have them).
Normally they will get excluded (eg. by zero refcount or writecount etc),
however it is a bit fragile for list walkers to know exactly what parts of
the inode state is set up and valid to test when in I_NEW. So along these
lines, move I_NEW checks upward as well (sometimes taking I_FREEING etc
checks with them too -- this shouldn't be a problem should it?)
Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Jan Kara <jack@suse.cz>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Enhanced lockdep coverage of __GFP_NOFS turned up this new lockdep
assert:
[ 1093.677775]
[ 1093.677781] =================================
[ 1093.680031] [ INFO: inconsistent lock state ]
[ 1093.680031] 2.6.29-rc5-tip-01504-gb49eca1-dirty #1
[ 1093.680031] ---------------------------------
[ 1093.680031] inconsistent {RECLAIM_FS-ON-W} -> {IN-RECLAIM_FS-W} usage.
[ 1093.680031] kswapd0/308 [HC0[0]:SC0[0]:HE1:SE1] takes:
[ 1093.680031] (&inode->inotify_mutex){+.+.?.}, at: [<c0205942>] inotify_inode_is_dead+0x20/0x80
[ 1093.680031] {RECLAIM_FS-ON-W} state was registered at:
[ 1093.680031] [<c01696b9>] mark_held_locks+0x43/0x5b
[ 1093.680031] [<c016baa4>] lockdep_trace_alloc+0x6c/0x6e
[ 1093.680031] [<c01cf8b0>] kmem_cache_alloc+0x20/0x150
[ 1093.680031] [<c040d0ec>] idr_pre_get+0x27/0x6c
[ 1093.680031] [<c02056e3>] inotify_handle_get_wd+0x25/0xad
[ 1093.680031] [<c0205f43>] inotify_add_watch+0x7a/0x129
[ 1093.680031] [<c020679e>] sys_inotify_add_watch+0x20f/0x250
[ 1093.680031] [<c010389e>] sysenter_do_call+0x12/0x35
[ 1093.680031] [<ffffffff>] 0xffffffff
[ 1093.680031] irq event stamp: 60417
[ 1093.680031] hardirqs last enabled at (60417): [<c018d5f5>] call_rcu+0x53/0x59
[ 1093.680031] hardirqs last disabled at (60416): [<c018d5b9>] call_rcu+0x17/0x59
[ 1093.680031] softirqs last enabled at (59656): [<c0146229>] __do_softirq+0x157/0x16b
[ 1093.680031] softirqs last disabled at (59651): [<c0106293>] do_softirq+0x74/0x15d
[ 1093.680031]
[ 1093.680031] other info that might help us debug this:
[ 1093.680031] 2 locks held by kswapd0/308:
[ 1093.680031] #0: (shrinker_rwsem){++++..}, at: [<c01b0502>] shrink_slab+0x36/0x189
[ 1093.680031] #1: (&type->s_umount_key#4){+++++.}, at: [<c01e6d77>] shrink_dcache_memory+0x110/0x1fb
[ 1093.680031]
[ 1093.680031] stack backtrace:
[ 1093.680031] Pid: 308, comm: kswapd0 Not tainted 2.6.29-rc5-tip-01504-gb49eca1-dirty #1
[ 1093.680031] Call Trace:
[ 1093.680031] [<c016947a>] valid_state+0x12a/0x13d
[ 1093.680031] [<c016954e>] mark_lock+0xc1/0x1e9
[ 1093.680031] [<c016a5b4>] ? check_usage_forwards+0x0/0x3f
[ 1093.680031] [<c016ab74>] __lock_acquire+0x2c6/0xac8
[ 1093.680031] [<c01688d9>] ? register_lock_class+0x17/0x228
[ 1093.680031] [<c016b3d3>] lock_acquire+0x5d/0x7a
[ 1093.680031] [<c0205942>] ? inotify_inode_is_dead+0x20/0x80
[ 1093.680031] [<c08824c4>] __mutex_lock_common+0x3a/0x4cb
[ 1093.680031] [<c0205942>] ? inotify_inode_is_dead+0x20/0x80
[ 1093.680031] [<c08829ed>] mutex_lock_nested+0x2e/0x36
[ 1093.680031] [<c0205942>] ? inotify_inode_is_dead+0x20/0x80
[ 1093.680031] [<c0205942>] inotify_inode_is_dead+0x20/0x80
[ 1093.680031] [<c01e6672>] dentry_iput+0x90/0xc2
[ 1093.680031] [<c01e67a3>] d_kill+0x21/0x45
[ 1093.680031] [<c01e6a46>] __shrink_dcache_sb+0x27f/0x355
[ 1093.680031] [<c01e6dc5>] shrink_dcache_memory+0x15e/0x1fb
[ 1093.680031] [<c01b05ed>] shrink_slab+0x121/0x189
[ 1093.680031] [<c01b0d12>] kswapd+0x39f/0x561
[ 1093.680031] [<c01ae499>] ? isolate_pages_global+0x0/0x233
[ 1093.680031] [<c0157eae>] ? autoremove_wake_function+0x0/0x43
[ 1093.680031] [<c01b0973>] ? kswapd+0x0/0x561
[ 1093.680031] [<c0157daf>] kthread+0x41/0x82
[ 1093.680031] [<c0157d6e>] ? kthread+0x0/0x82
[ 1093.680031] [<c01043ab>] kernel_thread_helper+0x7/0x10
inotify_handle_get_wd() does idr_pre_get() which does a
kmem_cache_alloc() without __GFP_FS - and is hence deadlockable under
extreme MM pressure.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: MinChan Kim <minchan.kim@gmail.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
If userspace supplies an invalid pointer to a read() of an inotify
instance, the inotify device's event list mutex is unlocked twice.
This causes an unbalance which effectively leaves the data structure
unprotected, and we can trigger oopses by accessing the inotify
instance from different tasks concurrently.
The best fix (contributed largely by Linus) is a total rewrite
of the function in question:
On Thu, Jan 22, 2009 at 7:05 AM, Linus Torvalds wrote:
> The thing to notice is that:
>
> - locking is done in just one place, and there is no question about it
> not having an unlock.
>
> - that whole double-while(1)-loop thing is gone.
>
> - use multiple functions to make nesting and error handling sane
>
> - do error testing after doing the things you always need to do, ie do
> this:
>
> mutex_lock(..)
> ret = function_call();
> mutex_unlock(..)
>
> .. test ret here ..
>
> instead of doing conditional exits with unlocking or freeing.
>
> So if the code is written in this way, it may still be buggy, but at least
> it's not buggy because of subtle "forgot to unlock" or "forgot to free"
> issues.
>
> This _always_ unlocks if it locked, and it always frees if it got a
> non-error kevent.
Cc: John McCutchan <ttb@tentacle.dhs.org>
Cc: Robert Love <rlove@google.com>
Cc: <stable@kernel.org>
Signed-off-by: Vegard Nossum <vegard.nossum@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
|
|
| |
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
|