NTFS deadlock on 2.6.18.6

[Sergey Vlasov]

Hello!

I have encountered a deadlock in the NTFS filesystem on a
2.6.18.6-based kernel (x86_64, CONFIG_SMP set, but the machine has
only one CPU (Athlon64 3200+), no PREEMPT).

The kswapd0 and mklocatedb processes were apparently involved in the
deadlock:

kswapd0       D ffff810005dea304     0   163      7           164   162 (L-TLB)
 ffff81003fd6bcc8 0000000000000046 0000000000000020 00000000021aaac0
 000000000000000a ffff81003fe93820 ffff81003fb157a0 0000027894893008
 00000000000308da ffff81003fe93a20 ffff810000000000 ffff810001641670
Call Trace:
 [<ffffffff8025e4cf>] __mutex_lock_slowpath+0x5d/0x98
 [<ffffffff8025e50f>] .text.lock.mutex+0x5/0x14
DWARF2 unwinder stuck at .text.lock.mutex+0x5/0x14
Leftover inexact backtrace:
 [<ffffffff881b58d2>] :ntfs:ntfs_put_inode+0x38/0x7a
 [<ffffffff80229987>] iput+0x3b/0x84
 [<ffffffff881b5a1b>] :ntfs:ntfs_clear_big_inode+0x107/0x121
 [<ffffffff802215f6>] clear_inode+0xc5/0xf6
 [<ffffffff8023362a>] dispose_list+0x56/0xf6
 [<ffffffff8022c47e>] shrink_icache_memory+0x1d4/0x203
 [<ffffffff8023de38>] shrink_slab+0xdc/0x154
 [<ffffffff802545e8>] kswapd+0x320/0x424
 [<ffffffff802921ca>] autoremove_wake_function+0x0/0x2e
 [<ffffffff80292007>] keventd_create_kthread+0x0/0x61
 [<ffffffff802542c8>] kswapd+0x0/0x424
 [<ffffffff80292007>] keventd_create_kthread+0x0/0x61
 [<ffffffff802312d8>] kthread+0xd4/0x107
 [<ffffffff80259d60>] child_rip+0xa/0x12
 [<ffffffff80292007>] keventd_create_kthread+0x0/0x61
 [<ffffffff80231204>] kthread+0x0/0x107
 [<ffffffff80259d56>] child_rip+0x0/0x12

mklocatedb    D ffff810001e0d400     0  4586   4585                     (NOTLB)
 ffff810021fd5c48 0000000000000082 0000000000000000 0000000000000000
 000000000000000a ffff81002728f7e0 ffffffff8048b3c0 000002789a2349cf
 00000000000025c4 ffff81002728f9e0 ffff810000000000 0000000000000000
Call Trace:
 [<ffffffff8024d781>] __wait_on_freeing_inode+0x82/0xa0
 [<ffffffff8024be5e>] find_inode+0x3d/0x6c
 [<ffffffff80256b8c>] ifind+0x34/0x91
 [<ffffffff802cbe81>] iget5_locked+0x6c/0x1a9
 [<ffffffff881b3aea>] :ntfs:ntfs_attr_iget+0x5a/0x5eb
 [<ffffffff881ae308>] :ntfs:ntfs_readdir+0x3f1/0xce1
 [<ffffffff802339d9>] vfs_readdir+0x77/0xa9
 [<ffffffff80237247>] sys_getdents+0x75/0xbd
 [<ffffffff80258ed6>] system_call+0x7e/0x83
DWARF2 unwinder stuck at system_call+0x7e/0x83
Leftover inexact backtrace:

There were some other processes stuck in the D state, but that seems
to be just a result of the above deadlock:

linuxdcpp     D ffff810001e0d400     0  4912   4910          4914  4911 (NOTLB)
 ffff8100283d1bf0 0000000000000082 ffff81000181cf38 ffff8100283d1b98
 0000000000000007 ffff81002ca0c0c0 ffffffff8048b3c0 00000279f4c0a091
 0000000000026b39 ffff81002ca0c2c0 ffffffff00000000 0000000000000000
Call Trace:
 [<ffffffff8025e4cf>] __mutex_lock_slowpath+0x5d/0x98
 [<ffffffff8025e50f>] .text.lock.mutex+0x5/0x14
DWARF2 unwinder stuck at .text.lock.mutex+0x5/0x14
Leftover inexact backtrace:
 [<ffffffff8022c2ea>] shrink_icache_memory+0x40/0x203
 [<ffffffff8023de38>] shrink_slab+0xdc/0x154
 [<ffffffff802b0e1c>] try_to_free_pages+0x179/0x254
 [<ffffffff8020df64>] __alloc_pages+0x1a8/0x2a9
 [<ffffffff80211077>] __do_page_cache_readahead+0x95/0x206
 [<ffffffff802af5a1>] force_page_cache_readahead+0x5f/0x81
 [<ffffffff802b1f06>] sys_madvise+0x2f7/0x3ec
 [<ffffffff80258ed6>] system_call+0x7e/0x83

(apparently waiting for kswapd0 to release iprune_mutex; this path
does not seem to have any FS-related locking, but sys_madvise() has
taken ->mm->mmap_sem for write.)

Other linuxdcpp threads and several ps processes then were stuck
waiting on its ->mm->mmap_sem taken by sys_madvise() above.

So the deadlock seems to be between kswapd0 and mklocatedb.  Note that
vfs_readdir() invoked by mklocatedb has taken i_mutex for the
directory, and kswapd0 is waiting on some i_mutex...

I suspect the following scenario:

 1) kswapd0 runs shrink_icache_memory() (and prune_icache(), which
    apparently was inlined there); prune_icache() notices that some
    attribute inode (probably the index bitmap) for the directory is
    unused, marks that attribute inode with I_FREEING and subsequently
    invokes dispose_list() to free marked inodes.

 2) mklocatedb invokes sys_readdir() on the directory, which grabs
    i_mutex of the directory and proceeds to call the filesystem
    readdir method - ntfs_readdir(), which then finds that it needs
    the bitmap inode and invokes ntfs_attr_iget() to find it.
    ntfs_attr_iget() proceeds down to find_inode(), which notices that
    the inode has I_FREEING set and goes to __wait_on_freeing_inode().

 3) kswapd0 proceeds to call clear_inode() on the attribute inode.
    ntfs_clear_big_inode() calls iput(VFS_I(ni->ext.base_ntfs_ino)) to
    put the base inode (the directory).

 4) iput() calls ntfs_put_inode() for the directory.  At this point
    the directory by chance has exactly two references accounted for
    in ->i_count - one from the file descriptor open by mklocatedb,
    another from the attribute inode (which that iput() is dropping
    now), so ntfs_put_inode() goes to the path which releases
    ni->itype.index.bmp_ino - but it needs to grab ->i_mutex for the
    directory, and that mutex is held by mklocatedb.

 5) Now kswapd0 is waiting for mklocatedb to release ->i_mutex for the
    directory, and mklocatedb is waiting for kswapd0 to finish freeing
    of the attribute inode - deadlock.

Seems that grabbing i_mutex in ntfs_put_inode() is not safe after all
(and lockdep cannot see this deadlock possibility, because one of
waits is __wait_on_freeing_inode - not a standard locking primitive).

-- 
Sergey Vlasov

Attachment: signature.asc
Description: Digital signature