[Bug 207053] fsfreeze deadlock on XFS (the FIFREEZE ioctl and subsequent FITHAW hang indefinitely)

[bugzilla-daemon@bugzilla.kernel.org]

https://bugzilla.kernel.org/show_bug.cgi?id=207053

--- Comment #6 from darrick.wong@oracle.com ---
On Tue, Apr 07, 2020 at 12:37:39PM -0400, Brian Foster wrote:
> On Tue, Apr 07, 2020 at 08:17:38AM -0700, Darrick J. Wong wrote:
> > On Tue, Apr 07, 2020 at 09:18:12AM -0400, Brian Foster wrote:
> > > On Tue, Apr 07, 2020 at 06:41:31AM +0000,
> bugzilla-daemon@bugzilla.kernel.org wrote:
> > > > https://bugzilla.kernel.org/show_bug.cgi?id=207053
> > > > 
> > > > --- Comment #2 from Paul Furtado (paulfurtado91@gmail.com) ---
> > > > Hi Dave,
> > > > 
> > > > Just had another case of this crop up and I was able to get the blocked
> tasks
> > > > output before automation killed the server. Because the log was too
> large to
> > > > attach, I've pasted the output into a github gist here:
> > > >
> https://gist.githubusercontent.com/PaulFurtado/c9bade038b8a5c7ddb53a6e10def058f/raw/ee43926c96c0d6a9ec81a648754c1af599ef0bdd/sysrq_w.log
> > > > 
> > > 
> > > Hm, so it looks like this is stuck between freeze:
> > > 
> > > [377279.630957] fsfreeze        D    0 46819  46337 0x00004084
> > > [377279.634910] Call Trace:
> > > [377279.637594]  ? __schedule+0x292/0x6f0
> > > [377279.640833]  ? xfs_xattr_get+0x51/0x80 [xfs]
> > > [377279.644287]  schedule+0x2f/0xa0
> > > [377279.647286]  schedule_timeout+0x1dd/0x300
> > > [377279.650661]  wait_for_completion+0x126/0x190
> > > [377279.654154]  ? wake_up_q+0x80/0x80
> > > [377279.657277]  ? work_busy+0x80/0x80
> > > [377279.660375]  __flush_work+0x177/0x1b0
> > > [377279.663604]  ? worker_attach_to_pool+0x90/0x90
> > > [377279.667121]  __cancel_work_timer+0x12b/0x1b0
> > > [377279.670571]  ? rcu_sync_enter+0x8b/0xd0
> > > [377279.673864]  xfs_stop_block_reaping+0x15/0x30 [xfs]
> > > [377279.677585]  xfs_fs_freeze+0x15/0x40 [xfs]
> > > [377279.680950]  freeze_super+0xc8/0x190
> > > [377279.684086]  do_vfs_ioctl+0x510/0x630
> > > ...
> > > 
> > > ... and the eofblocks scanner:
> > > 
> > > [377279.422496] Workqueue: xfs-eofblocks/nvme13n1 xfs_eofblocks_worker
> [xfs]
> > > [377279.426971] Call Trace:
> > > [377279.429662]  ? __schedule+0x292/0x6f0
> > > [377279.432839]  schedule+0x2f/0xa0
> > > [377279.435794]  rwsem_down_read_slowpath+0x196/0x530
> > > [377279.439435]  ? kmem_cache_alloc+0x152/0x1f0
> > > [377279.442834]  ? __percpu_down_read+0x49/0x60
> > > [377279.446242]  __percpu_down_read+0x49/0x60
> > > [377279.449586]  __sb_start_write+0x5b/0x60
> > > [377279.452869]  xfs_trans_alloc+0x152/0x160 [xfs]
> > > [377279.456372]  xfs_free_eofblocks+0x12d/0x1f0 [xfs]
> > > [377279.460014]  xfs_inode_free_eofblocks+0x128/0x1a0 [xfs]
> > > [377279.463903]  ? xfs_inode_ag_walk_grab+0x5f/0x90 [xfs]
> > > [377279.467680]  xfs_inode_ag_walk.isra.17+0x1a7/0x410 [xfs]
> > > [377279.471567]  ? __xfs_inode_clear_blocks_tag+0x120/0x120 [xfs]
> > > [377279.475620]  ? kvm_sched_clock_read+0xd/0x20
> > > [377279.479059]  ? sched_clock+0x5/0x10
> > > [377279.482184]  ? __xfs_inode_clear_blocks_tag+0x120/0x120 [xfs]
> > > [377279.486234]  ? radix_tree_gang_lookup_tag+0xa8/0x100
> > > [377279.489974]  ? __xfs_inode_clear_blocks_tag+0x120/0x120 [xfs]
> > > [377279.494041]  xfs_inode_ag_iterator_tag+0x73/0xb0 [xfs]
> > > [377279.497859]  xfs_eofblocks_worker+0x29/0x40 [xfs]
> > > [377279.501484]  process_one_work+0x195/0x380
> > > ...
> > > 
> > > The immediate issue is likely that the eofblocks transaction is not
> > > NOWRITECOUNT (same for the cowblocks scanner, btw), but the problem with
> > > doing that is these helpers are called from other contexts outside of
> > > the background scanners.
> > > 
> > > Perhaps what we need to do here is let these background scanners acquire
> > > a superblock write reference, similar to what Darrick recently added to
> > > scrub..? We'd have to do that from the scanner workqueue task, so it
> > > would probably need to be a trylock so we don't end up in a similar
> > > situation as above. I.e., we'd either get the reference and cause freeze
> > > to wait until it's dropped or bail out if freeze has already stopped the
> > > transaction subsystem. Thoughts?
> > 
> > Hmm, I had a whole gigantic series to refactor all the speculative
> > preallocation gc work into a single thread + radix tree tag; I'll see if
> > that series actually fixed this problem too.
> > 
> > But yes, all background threads that run transactions need to have
> > freezer protection.
> > 
> 
> So something like the following in the meantime, assuming we want a
> backportable fix..? I think this means we could return -EAGAIN from the
> eofblocks ioctl, but afaict if something functionally conflicts with an
> active scan across freeze then perhaps that's preferred.

Apparently I don't have a patch that fixes the speculative gc code.  The
deferred inactivation worker does it, so perhaps I got mixed up. :/

I think a better fix would be to annotate xfs_icache_free_eofblocks and
xfs_icache_free_cowblocks to note that the caller must obtain freeze
protection before calling those functions.  Then we can play whackamole
with the existing callers:

1. xfs_eofblocks_worker and xfs_cowblocks_worker can try to
sb_start_write and just go back to sleep if the fs is frozen.  The
flush_workqueue will then cancel the delayed work and the freeze can
proceed.

2. The buffered write ENOSPC scour-and-retry loops already have freeze
protection because they're file writes, so they don't have to change.

3. XFS_IOC_FREE_EOFBLOCKS can sb_start_write, which means that callers
will sleep on the frozen fs.

--D

> Brian
> 
> diff --git a/fs/xfs/xfs_icache.c b/fs/xfs/xfs_icache.c
> index a7be7a9e5c1a..0f14d58e5bb0 100644
> --- a/fs/xfs/xfs_icache.c
> +++ b/fs/xfs/xfs_icache.c
> @@ -1515,13 +1515,24 @@ __xfs_icache_free_eofblocks(
>                                          void *args),
>       int                     tag)
>  {
> -     int flags = SYNC_TRYLOCK;
> +     int                     flags = SYNC_TRYLOCK;
> +     int                     error;
>  
>       if (eofb && (eofb->eof_flags & XFS_EOF_FLAGS_SYNC))
>               flags = SYNC_WAIT;
>  
> -     return xfs_inode_ag_iterator_tag(mp, execute, flags,
> -                                      eofb, tag);
> +     /*
> +      * freeze waits on background scanner jobs to complete so we cannot
> +      * block on write protection here. Bail if the transaction subsystem is
> +      * already freezing, returning -EAGAIN to notify other callers.
> +      */
> +     if (!sb_start_write_trylock(mp->m_super))
> +             return -EAGAIN;
> +
> +     error = xfs_inode_ag_iterator_tag(mp, execute, flags, eofb, tag);
> +     sb_end_write(mp->m_super);
> +
> +     return error;
>  }
>  
>  int
> 
> > --D
> > 
> > > Brian
> > > 
> > > > 
> > > > Thanks,
> > > > Paul
> > > > 
> > > > -- 
> > > > You are receiving this mail because:
> > > > You are watching the assignee of the bug.
> > > > 
> > > 
> > 
>

-- 
You are receiving this mail because:
You are watching the assignee of the bug.