[Top] [All Lists]

Re: Possible deadlock when deleting from realtime section

To: Denny Priebe <denny.priebe@xxxxxxxxxxxxxx>
Subject: Re: Possible deadlock when deleting from realtime section
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Wed, 27 Oct 2010 09:58:39 +1100
Cc: xfs@xxxxxxxxxxx
In-reply-to: <AANLkTi=Zq3mh=0Q-g6oi-OqYFRsENiTjLcHfvPNsOkGa@xxxxxxxxxxxxxx>
References: <AANLkTi=Zq3mh=0Q-g6oi-OqYFRsENiTjLcHfvPNsOkGa@xxxxxxxxxxxxxx>
User-agent: Mutt/1.5.20 (2009-06-14)
On Mon, Oct 25, 2010 at 03:59:22PM +0000, Denny Priebe wrote:
> Hi,
> I'm experiencing a deadlock situation when deleting directories placed in
> the realtime section. This is reproducable with kernel versions and
> 2.6.36-rc8. I haven't tried final 2.6.36 yet. The same setup is working
> perfectly without using the realtime section. The file system has been
> created with
>  mkfs.xfs -f -l logdev=/dev/sdb1 -r rtdev=/dev/sdb3,extsize=256k /dev/sdb2

FYI, using an external log on the same device as the main device will be
slower than using the default internal log....

> and mounted with
>  mount -t xfs -o logdev=/dev/sdb1,rtdev=/dev/sdb3,sunit=512,swidth=2048 \
>  /dev/sdb2 /there
> This is where rm blocks:
> SysRq : Show Blocked State
>  task                        PC stack   pid father
> rm            D 0000000000000005     0  1705   1658 0x00000080
> ffff88018c40d858 0000000000000086 ffff880100000000 ffff88018c40c010
> ffff88018c40dfd8 00000000000148c0 ffff88018b1c2e20 ffff88018b1c31d8
> ffff88018b1c31d0 00000000000148c0 00000000000148c0 ffff88018c40dfd8
> Call Trace:
> [<ffffffff81437739>] rwsem_down_failed_common+0xd3/0x105
> [<ffffffff8143777e>] rwsem_down_write_failed+0x13/0x15
> [<ffffffff81202ad3>] call_rwsem_down_write_failed+0x13/0x20
> [<ffffffff81436e3d>] ? down_write+0x40/0x44
> [<ffffffffa033be9b>] xfs_ilock+0x4a/0x9a [xfs]
> [<ffffffffa033c413>] xfs_iget+0x34c/0x5cd [xfs]
> [<ffffffffa0351fa9>] xfs_trans_iget+0x1b/0x56 [xfs]
> [<ffffffffa0313035>] xfs_rtfree_extent+0x37/0xdc [xfs]
> [<ffffffffa033fce2>] ? xfs_iext_remove+0xc0/0xd2 [xfs]
> [<ffffffffa0324ffe>] ? xfs_bmap_del_extent+0x2e8/0x93d [xfs]
> [<ffffffffa0324e87>] xfs_bmap_del_extent+0x171/0x93d [xfs]
> [<ffffffffa03506eb>] ? xfs_trans_commit_iclog+0x2ba/0x2d3 [xfs]
> [<ffffffffa0325dc3>] xfs_bunmapi+0x770/0xa28 [xfs]
> [<ffffffffa033e44e>] xfs_itruncate_finish+0x185/0x2b8 [xfs]
> [<ffffffffa0354437>] xfs_inactive+0x1c8/0x3d5 [xfs]
> [<ffffffffa035f38c>] xfs_fs_evict_inode+0xd5/0xdd [xfs]
> [<ffffffff8111282b>] evict+0x22/0x92
> [<ffffffff81112c60>] iput+0x1bc/0x225
> [<ffffffff8110b079>] do_unlinkat+0x103/0x156
> [<ffffffff811086c5>] ? path_put+0x1d/0x22
> [<ffffffff81093e2c>] ? audit_syscall_entry+0x119/0x145
> [<ffffffff8110b205>] sys_unlinkat+0x24/0x26
> [<ffffffff81009ac2>] system_call_fastpath+0x16/0x1b

So it is blocked trying to lock the allocation bitmap inode.
Hmmm, I suspect that XFS_ITRUNC_MAX_EXTENTS is the start of
the problem here.

i.e. what I think might be the problem is that xfs_bunmapi() is
trying to free two extents in the one transaction, and what we see
above is the second extent being freed via xfs_rtfree_extent().
The bitmap inode won't be unlocked until the transaction commits,
so the second call to xfs_trans_iget() in the same transaction will
hang like this.

Hmmm. Looks like we broke recursive inode locking in
xfs_trans_iget() in commit aa72a5cf00001d0b952c7c755be404b9118ceb2e
("xfs: simplify xfs_trans_iget"). The changelog says:

    A quick audit of the callers of xfs_trans_iget shows that no
    caller really relies on this behaviour fortunately - xfs_ialloc
    allows this inode from disk so it must not be there before, and
    all the RT allocator routines only every add each RT bitmap
    inode once.

What we didn't take into account is multiple RT allocator calls in
the same transaction context. Let me think about the best way to fix


Dave Chinner

<Prev in Thread] Current Thread [Next in Thread>