From: Arjan van de Ven <arjan@linux.intel.com>
Date: Sun, 30 Jul 2006 10:03:54 +0000 (-0700)
Subject: [PATCH] inotify: fix deadlock found by lockdep
X-Git-Tag: v2.6.18-rc4~105
X-Git-Url: https://err.no/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=5b6509aa8c2f292caea7c0602ec361f920951508;p=linux-2.6

[PATCH] inotify: fix deadlock found by lockdep

This is a real deadlock, a nice complex one:
(warning: long explanation follows so that Andrew can have a complete
patch description)

it's an ABCDA deadlock:

A iprune_mutex
B inode->inotify_mutex
C ih->mutex
D dev->ev_mutex

The AB relationship comes straight from invalidate_inodes()

int invalidate_inodes(struct super_block * sb)
{
        int busy;
        LIST_HEAD(throw_away);

        mutex_lock(&iprune_mutex);
        spin_lock(&inode_lock);
        inotify_unmount_inodes(&sb->s_inodes);

where inotify_umount_inodes() takes the
                mutex_lock(&inode->inotify_mutex);

The BC relationship comes directly from inotify_find_update_watch():
s32 inotify_find_update_watch(struct inotify_handle *ih, struct inode *inode,
                              u32 mask)
{
   ...
        mutex_lock(&inode->inotify_mutex);
        mutex_lock(&ih->mutex);

The CD relationship comes from inotify_rm_wd:
inotify_rm_wd does
        mutex_lock(&inode->inotify_mutex);
        mutex_lock(&ih->mutex)
and then calls inotify_remove_watch_locked() which calls
notify_dev_queue_event() which does
	        mutex_lock(&dev->ev_mutex);

(this strictly is a BCD relationship)

The DA relationship comes from the most interesting part:

  [<ffffffff8022d9f2>] shrink_icache_memory+0x42/0x270
  [<ffffffff80240dc4>] shrink_slab+0x11d/0x1c9
  [<ffffffff802b5104>] try_to_free_pages+0x187/0x244
  [<ffffffff8020efed>] __alloc_pages+0x1cd/0x2e0
  [<ffffffff8025e1f8>] cache_alloc_refill+0x3f8/0x821
  [<ffffffff8020a5e5>] kmem_cache_alloc+0x85/0xcb
  [<ffffffff802db027>] kernel_event+0x2e/0x122
  [<ffffffff8021d61c>] inotify_dev_queue_event+0xcc/0x140

inotify_dev_queue_event schedules a kernel_event which does a
kmem_cache_alloc( , GFP_KERNEL) which may try to shrink slabs, including
the inode cache .. which then takes iprune_mutex.

And voila, there is an AB, a BC, a CD relationship (even a direct BCD),
and also now a DA relationship -> a circular type AB-BA deadlock but
involving 4 locks.

The solution is simple: kernel_event() is NOT allowed to use GFP_KERNEL,
but must use GFP_NOFS to not cause recursion into the VFS.

Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Robert Love <rml@novell.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
---

diff --git a/fs/inotify_user.c b/fs/inotify_user.c
index f2386442ad..017cb0f134 100644
--- a/fs/inotify_user.c
+++ b/fs/inotify_user.c
@@ -187,7 +187,7 @@ static struct inotify_kernel_event * kernel_event(s32 wd, u32 mask, u32 cookie,
 {
 	struct inotify_kernel_event *kevent;
 
-	kevent = kmem_cache_alloc(event_cachep, GFP_KERNEL);
+	kevent = kmem_cache_alloc(event_cachep, GFP_NOFS);
 	if (unlikely(!kevent))
 		return NULL;