Re: circular locking dependency in dm

[Kiyoshi Ueda <k-ueda@ct.jp.nec.com>]

Hi Christof, Mike and all,

This is just an information for all dm developers.
Please see below.

On 09/24/2009 05:11 PM +0900, Christof Schmitt wrote:
> While testing some zfcp patches with the current linux-2.6 git
> checkout i got this:
> 
> =======================================================
> [ INFO: possible circular locking dependency detected ]
> 2.6.31-08358-g94a8d5c-dirty #2
> -------------------------------------------------------
> multipathd/2409 is trying to acquire lock:
>  (_event_lock){+.+.+.}, at: [<00000000002e5efc>] dm_table_event+0x50/0x80
> 
> but task is already holding lock:
>  (_hash_lock){++++..}, at: [<00000000002e9d84>] dev_remove+0x2c/0xf0
> 
> which lock already depends on the new lock.
> 
> 
> the existing dependency chain (in reverse order) is:
> 
> -> #1 (_hash_lock){++++..}:
>        [<0000000000086d92>] __lock_acquire+0xe2e/0x19cc
>        [<00000000000879a0>] lock_acquire+0x70/0x98
>        [<00000000004eb486>] down_read+0x5e/0x9c
>        [<00000000002e81ba>] dm_copy_name_and_uuid+0x3e/0xec
>        [<00000000002e131a>] dm_send_uevents+0x9a/0x180
>        [<00000000002e1eee>] event_callback+0xae/0xf8
>        [<00000000002e5f12>] dm_table_event+0x66/0x80
>        [<000000000006b0a4>] worker_thread+0x24c/0x308
>        [<0000000000071302>] kthread+0x9a/0xa4
>        [<000000000001c5de>] kernel_thread_starter+0x6/0xc
>        [<000000000001c5d8>] kernel_thread_starter+0x0/0xc
> 
> -> #0 (_event_lock){+.+.+.}:
>        [<00000000000877e2>] __lock_acquire+0x187e/0x19cc
>        [<00000000000879a0>] lock_acquire+0x70/0x98
>        [<00000000004ea7a0>] mutex_lock_nested+0x78/0x3f0
>        [<00000000002e5efc>] dm_table_event+0x50/0x80
>        [<00000000002e9d1a>] __hash_remove+0xaa/0xe8
>        [<00000000002e9db6>] dev_remove+0x5e/0xf0
>        [<00000000002ea5e4>] ctl_ioctl+0x20c/0x288
>        [<00000000002ea68a>] dm_ctl_ioctl+0x2a/0x38
>        [<00000000000fc6cc>] vfs_ioctl+0x4c/0xdc
>        [<00000000000fcbdc>] do_vfs_ioctl+0x3a8/0x5c8
>        [<00000000000fce4c>] SyS_ioctl+0x50/0x8c
>        [<0000000000029cf6>] sysc_noemu+0x10/0x16
>        [<0000020000264286>] 0x20000264286
> 
> other info that might help us debug this:
> 
> 1 lock held by multipathd/2409:
>  #0:  (_hash_lock){++++..}, at: [<00000000002e9d84>] dev_remove+0x2c/0xf0
> 
> stack backtrace:
> CPU: 0 Not tainted 2.6.31-08358-g94a8d5c-dirty #2
> Process multipathd (pid: 2409, task: 000000002f6c6038, ksp: 000000002c8c7aa8)
>        0000000000000000 000000002c8c7998 0000000000000002 0000000000000000 
>        000000002c8c7a38 000000002c8c79b0 000000002c8c79b0 00000000004e8582 
>        0000000000000000 0000000000000001 0000000000000000 0000000000000000 
>        000000000000000d 0000000000000000 000000002c8c7a00 000000000000000e 
>        00000000004faf90 00000000000179e0 000000002c8c7998 000000002c8c79e0 
> Call Trace:
> ([<00000000000178e2>] show_trace+0xee/0x144)
>  [<00000000000858ca>] print_circular_bug+0x112/0x124
>  [<00000000000877e2>] __lock_acquire+0x187e/0x19cc
>  [<00000000000879a0>] lock_acquire+0x70/0x98
>  [<00000000004ea7a0>] mutex_lock_nested+0x78/0x3f0
>  [<00000000002e5efc>] dm_table_event+0x50/0x80
>  [<00000000002e9d1a>] __hash_remove+0xaa/0xe8
>  [<00000000002e9db6>] dev_remove+0x5e/0xf0
>  [<00000000002ea5e4>] ctl_ioctl+0x20c/0x288
>  [<00000000002ea68a>] dm_ctl_ioctl+0x2a/0x38
>  [<00000000000fc6cc>] vfs_ioctl+0x4c/0xdc
>  [<00000000000fcbdc>] do_vfs_ioctl+0x3a8/0x5c8
>  [<00000000000fce4c>] SyS_ioctl+0x50/0x8c
>  [<0000000000029cf6>] sysc_noemu+0x10/0x16
>  [<0000020000264286>] 0x20000264286
> INFO: lockdep is turned off.

I think this is not a regression in the recent kernel and this is
a known AB-BA deadlock issue which was introduced by this commit:
---------------------------------------------------------------------
    commit 7a8c3d3b92883798e4ead21dd48c16db0ec0ff6f
    Author: Mike Anderson <andmike@linux.vnet.ibm.com>
    Date:   Fri Oct 19 22:48:01 2007 +0100

        dm: uevent generate events

        This patch adds support for the dm_path_event dm_send_event
        functions which create and send udev events.
---------------------------------------------------------------------


Other similar deadlock possibilities which are caused by the patch
were pointed out here:
    http://marc.info/?l=dm-devel&m=123130584210651&w=2

But all problems above haven't been fixed yet.

Thanks,
Kiyoshi Ueda