Konstanin Khlebnikov reports that a dangerous race between umount and
shmem_writepage can be reproduced by this script:

  for i in {1..300} ; do
	mkdir $i
	while true ; do
		mount -t tmpfs none $i
		dd if=/dev/zero of=$i/test bs=1M count=$(($RANDOM % 100))
		umount $i
	done &
  done

on a 6xCPU node with 8Gb RAM: kernel very unstable after this accident. =)

Kernel log:

  VFS: Busy inodes after unmount of tmpfs.
                 Self-destruct in 5 seconds.  Have a nice day...

  WARNING: at lib/list_debug.c:53 __list_del_entry+0x8d/0x98()
  list_del corruption. prev->next should be ffff880222fdaac8, but was (null)
  Pid: 11222, comm: mount.tmpfs Not tainted 2.6.39-rc2+ #4
  Call Trace:
   warn_slowpath_common+0x80/0x98
   warn_slowpath_fmt+0x41/0x43
   __list_del_entry+0x8d/0x98
   evict+0x50/0x113
   iput+0x138/0x141
  ...
  BUG: unable to handle kernel paging request at ffffffffffffffff
  IP: shmem_free_blocks+0x18/0x4c
  Pid: 10422, comm: dd Tainted: G        W   2.6.39-rc2+ #4
  Call Trace:
   shmem_recalc_inode+0x61/0x66
   shmem_writepage+0xba/0x1dc
   pageout+0x13c/0x24c
   shrink_page_list+0x28e/0x4be
   shrink_inactive_list+0x21f/0x382
  ...

shmem_writepage() calls igrab() on the inode for the page which came from
page reclaim, to add it later into shmem_swaplist for swapoff operation.

This igrab() can race with super-block deactivating process:

  shrink_inactive_list()          deactivate_super()
  pageout()                       tmpfs_fs_type->kill_sb()
  shmem_writepage()               kill_litter_super()
                                  generic_shutdown_super()
                                   evict_inodes()
   igrab()
                                    atomic_read(&inode->i_count)
                                     skip-inode
   iput()
                                   if (!list_empty(&sb->s_inodes))
                                          printk("VFS: Busy inodes after...

This igrap-iput pair was added in commit 1b1b32f2 "tmpfs: fix
shmem_swaplist races" based on incorrect assumptions: igrab() protects the
inode from concurrent eviction by deletion, but it does nothing to protect
it from concurrent unmounting, which goes ahead despite the raised
i_count.

So this use of igrab() was wrong all along, but the race made much worse
in 2.6.37 when commit 63997e98

 "split invalidate_inodes()" replaced
two attempts at invalidate_inodes() by a single evict_inodes().

Konstantin posted a plausible patch, raising sb->s_active too: I'm unsure
whether it was correct or not; but burnt once by igrab(), I am sure that
we don't want to rely more deeply upon externals here.

Fix it by adding the inode to shmem_swaplist earlier, while the page lock
on page in page cache still secures the inode against eviction, without
artifically raising i_count.  It was originally added later because
shmem_unuse_inode() is liable to remove an inode from the list while it's
unswapped; but we can guard against that by taking spinlock before
dropping mutex.

Reported-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Signed-off-by: Hugh Dickins <hughd@google.com>
Tested-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>