__buffer_migrate_page() vs ll_rw_block()

__buffer_migrate_page() vs ll_rw_block()

[Vincent Whitchurch]

I'm seeing occasional squashfs read failures ("squashfs_read_data failed
to read block") when compaction is run at the same time as reading from
squashfs, with something like the commands below.  The kernel version is
the latest stable/v5.4 kernel, v5.4.123.

 while :; do echo 1 > /proc/sys/vm/compact_memory; done &
 while :; do echo 3 > /proc/sys/vm/drop_caches; find fs/ > /dev/null; done &

On this kernel, squashfs uses ll_rw_block().  The problem is that
ll_rw_block() ignores BHs which it can't get a lock on, but
__buffer_migrate_page() can take the lock on the BHs in order to check
if they can be migrated.  If __buffer_migrate_page() holds the lock at
the same time that ll_rw_block() wants it, the BH is skipped and I/O is
not issued for these blocks, and squashfs ends up seeing
!buffer_uptodate() and erroring out.

On newer kernels, squashfs doesn't use ll_rw_block() anymore, but I
still see other users of that function in other filesystems, and AFAICS
the underlying problem of the race with __buffer_migrate_page() has not
yet been fixed.

I'd be happy to receive any suggestions about the right way to fix this.

Thank you.

Re: __buffer_migrate_page() vs ll_rw_block()

[Jan Kara]

On Tue 01-06-21 15:23:17, Vincent Whitchurch wrote:
> I'm seeing occasional squashfs read failures ("squashfs_read_data failed
> to read block") when compaction is run at the same time as reading from
> squashfs, with something like the commands below.  The kernel version is
> the latest stable/v5.4 kernel, v5.4.123.
> 
>  while :; do echo 1 > /proc/sys/vm/compact_memory; done &
>  while :; do echo 3 > /proc/sys/vm/drop_caches; find fs/ > /dev/null; done &
> 
> On this kernel, squashfs uses ll_rw_block().  The problem is that
> ll_rw_block() ignores BHs which it can't get a lock on, but
> __buffer_migrate_page() can take the lock on the BHs in order to check
> if they can be migrated.  If __buffer_migrate_page() holds the lock at
> the same time that ll_rw_block() wants it, the BH is skipped and I/O is
> not issued for these blocks, and squashfs ends up seeing
> !buffer_uptodate() and erroring out.
> 
> On newer kernels, squashfs doesn't use ll_rw_block() anymore, but I
> still see other users of that function in other filesystems, and AFAICS
> the underlying problem of the race with __buffer_migrate_page() has not
> yet been fixed.
> 
> I'd be happy to receive any suggestions about the right way to fix this.

Well, if you care about squashfs, the best solution would IMO be to
backport 93e72b3c612adc ("squashfs: migrate from ll_rw_block usage to BIO")
and the followup fixes (there are several of them as far as I can see by
Fixes tags) to your kernel - stable tree would be IMHO willing to take the
series if you add your justification that it fixes read failures for you.

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR