Re: 'ls /mnt/scratch/' freeze(deadlock?) when run xfstest(btrfs/232)

[Filipe Manana <fdmanana@gmail.com>]

On Thu, Apr 22, 2021 at 12:19 AM Qu Wenruo <quwenruo.btrfs@gmx.com> wrote:
>
>
>
> On 2021/4/22 上午12:03, Filipe Manana wrote:
> > On Wed, Apr 21, 2021 at 4:57 PM Wang Yugui <wangyugui@e16-tech.com> wrote:
> >>
> >> Hi,
> >>
> >>> That's the problem, qgroup flushing triggers writeback for an inode
> >>> for which we have a page dirtied and locked.
> >>> This should fix it:  https://pastebin.com/raw/U9GUZiEf
> >>>
> >>> Try it out and I'll write a changelog later.
> >>> Thanks.
> >>
> >> we run xfstest on two server with this patch.
> >> one passed the tests.
> >> but one got a btrfs/232 error.
> >>
> >> btrfs/232 32s ... _check_btrfs_filesystem: filesystem on /dev/nvme0n1p1 is inconsistent
> >> (see /usr/hpc-bio/xfstests/results//btrfs/232.full for details)
> >> ...
> >> [4/7] checking fs roots
> >> root 5 inode 337 errors 400, nbytes wrong
> >> ERROR: errors found in fs roots
> >
> > Ok, that's a different problem caused by something else.
> > It's possible to be due to the recent refactorings for preparation to
> > subpage block size.
>
> This error looks exactly what I have seen during subpage development.
> The subpage bug is caused by incorrect btrfs_invalidatepage() though,
> and not yet merged into misc-next anyway.

Ok, I vaguely remembered you encountering problems with that and
mentioning it somewhere.
I couldn't reproduce it overnight.

>
> I guess it's some error path not clearing extent states correctly, thus
> leaving the inode nbytes accounting wrong.
>
> BTW, the new @in_reclaim_context parameter for start_delalloc_inodes()
> is already in misc-next:

Well, yes, it was added to 5.11 and then backported to the relevant
stable releases.

> commit 3d45f221ce627d13e2e6ef3274f06750c84a6542
> Author: Filipe Manana <fdmanana@suse.com>
> Date:   Wed Dec 2 11:55:58 2020 +0000
>
>      btrfs: fix deadlock when cloning inline extent and low on free
> metadata space
>
> We only need to make btrfs_start_delalloc_snapshot() to accept the new
> parameter and pass in_reclaim_context = true for qgroup.

Correct.
I initially thought about removing the argument and always skip the
flagged inodes, but that would make snapshotting not wait for inline
cloning in progress (which is why I introduced the parameter back
then).
Not that it could cause any problem, just a matter of being consistent
and preserving the behaviour of waiting for any ongoing inline extent
cloning.

>
> Thanks,
> Qu
> >
> > Will try to look into that later.
> >
> > Thanks.
> >
> >> ...
> >>
> >> Best Regards
> >> Wang Yugui (wangyugui@e16-tech.com)
> >> 2021/04/21
> >>
> >
> >



-- 
Filipe David Manana,

“Whether you think you can, or you think you can't — you're right.”