[RFC PATCH 00/11] xfs: introduce the free inode btree

[Brian Foster]

On 09/06/2013 05:35 PM, Dave Chinner wrote:
> On Thu, Sep 05, 2013 at 05:17:10PM -0400, Michael L. Semon wrote:
> ....
>> [  814.376620] XFS (sdb4): Mounting Filesystem
>> [  815.050778] XFS (sdb4): Ending clean mount
>> [  823.169368] 
>> [  823.170932] ======================================================
>> [  823.172146] [ INFO: possible circular locking dependency detected ]
>> [  823.172146] 3.11.0+ #5 Not tainted
>> [  823.172146] -------------------------------------------------------
>> [  823.172146] dirstress/5276 is trying to acquire lock:
>> [  823.172146]  (sb_internal){.+.+.+}, at: [<c11c5e60>] xfs_trans_alloc+0x1f/0x35
>> [  823.172146] 
>> [  823.172146] but task is already holding lock:
>> [  823.172146]  (&(&ip->i_lock)->mr_lock){+++++.}, at: [<c1206cfb>] xfs_ilock+0x100/0x1f1
>> [  823.172146] 
>> [  823.172146] which lock already depends on the new lock.
>> [  823.172146] 
>> [  823.172146] 
>> [  823.172146] the existing dependency chain (in reverse order) is:
>> [  823.172146] 
>> [  823.172146] -> #1 (&(&ip->i_lock)->mr_lock){+++++.}:
>> [  823.172146]        [<c1070a11>] __lock_acquire+0x345/0xa11
>> [  823.172146]        [<c1071c45>] lock_acquire+0x88/0x17e
>> [  823.172146]        [<c14bff98>] _raw_spin_lock+0x47/0x74
>> [  823.172146]        [<c1116247>] __mark_inode_dirty+0x171/0x38c
>> [  823.172146]        [<c111acab>] __set_page_dirty+0x5f/0x95
>> [  823.172146]        [<c111b93e>] mark_buffer_dirty+0x58/0x12b
>> [  823.172146]        [<c111baff>] __block_commit_write.isra.17+0x64/0x82
>> [  823.172146]        [<c111c197>] block_write_end+0x2b/0x53
>> [  823.172146]        [<c111c201>] generic_write_end+0x42/0x9a
>> [  823.172146]        [<c11a42d5>] xfs_vm_write_end+0x60/0xbe
>> [  823.172146]        [<c10bd47a>] generic_file_buffered_write+0x140/0x20f
>> [  823.172146]        [<c11b2347>] xfs_file_buffered_aio_write+0x10b/0x205
>> [  823.172146]        [<c11b24ee>] xfs_file_aio_write+0xad/0xec
>> [  823.172146]        [<c10f0c5f>] do_sync_write+0x60/0x87
>> [  823.172146]        [<c10f0e0f>] vfs_write+0x9c/0x189
>> [  823.172146]        [<c10f0fc6>] SyS_write+0x49/0x81
>> [  823.172146]        [<c14c14bb>] sysenter_do_call+0x12/0x32
>> [  823.172146] 
>> [  823.172146] -> #0 (sb_internal){.+.+.+}:
>> [  823.172146]        [<c106e972>] validate_chain.isra.35+0xfc7/0xff4
>> [  823.172146]        [<c1070a11>] __lock_acquire+0x345/0xa11
>> [  823.172146]        [<c1071c45>] lock_acquire+0x88/0x17e
>> [  823.172146]        [<c10f36eb>] __sb_start_write+0xad/0x177
>> [  823.172146]        [<c11c5e60>] xfs_trans_alloc+0x1f/0x35
>> [  823.172146]        [<c120a823>] xfs_inactive+0x129/0x4a3
>> [  823.172146]        [<c11c280d>] xfs_fs_evict_inode+0x6c/0x114
>> [  823.172146]        [<c1106678>] evict+0x8e/0x15d
>> [  823.172146]        [<c1107126>] iput+0xc4/0x138
>> [  823.172146]        [<c1103504>] dput+0x1b2/0x257
>> [  823.172146]        [<c10f1a30>] __fput+0x140/0x1eb
>> [  823.172146]        [<c10f1b0f>] ____fput+0xd/0xf
>> [  823.172146]        [<c1048477>] task_work_run+0x67/0x90
>> [  823.172146]        [<c1001ea5>] do_notify_resume+0x61/0x63
>> [  823.172146]        [<c14c0cfa>] work_notifysig+0x1f/0x25
>> [  823.172146] 
>> [  823.172146] other info that might help us debug this:
>> [  823.172146] 
>> [  823.172146]  Possible unsafe locking scenario:
>> [  823.172146] 
>> [  823.172146]        CPU0                    CPU1
>> [  823.172146]        ----                    ----
>> [  823.172146]   lock(&(&ip->i_lock)->mr_lock);
>> [  823.172146]                                lock(sb_internal);
>> [  823.172146]                                lock(&(&ip->i_lock)->mr_lock);
>> [  823.172146]   lock(sb_internal);
> 
> Ah, now there's something I missed in all the xfs_inactive
> transaction rework - you can't call
> xfs_trans_alloc()/xfs-trans_reserve with the XFS_ILOCK_??? held.
> It's not the freeze locks you really have to worry about deadlocking
> if you do, it's deadlocking against log space that is much more
> likely.
> 
> i.e. if you hold the ILOCK, the AIL can't get it to flush the inode
> to disk. That means if the inode you hold locked is pinning the tail
> of the log and there is no logspace for the transaction you are
> about to run, xfs_trans_reserve() will block forever waiting for the
> inode to be flushed and the tail of the log to move forward. This
> will end up blocking all further reservations and hence deadlock the
> filesystem...
> 
> Brian, if you rewrite xfs_inactive in the way that I suggested, this
> problem goes away ;)
> 
> Thanks for reporting this, Michael.
> 

Oh, very interesting scenario. Thanks again for catching this, Michael,
and for the analysis, Dave. I'll get this cleaned up in the next
revision as well.

Brian

> Cheers,
> 
> Dave.
>