Re: read corruption with qemu master io_uring engine / linux master / btrfs(?)

[Filipe Manana <fdmanana@kernel.org>]

On Fri, Jul 01, 2022 at 10:25:58AM +0900, Dominique MARTINET wrote:
> Filipe Manana wrote on Thu, Jun 30, 2022 at 04:10:38PM +0100:
> > This may prevent the short reads (not tested yet):
> > 
> > diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
> > index 7a54f964ff37..42fb56ed0021 100644
> > --- a/fs/btrfs/inode.c
> > +++ b/fs/btrfs/inode.c
> > @@ -7684,7 +7684,7 @@ static int btrfs_dio_iomap_begin(struct inode *inode, loff_t start,
> >         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
> >             em->block_start == EXTENT_MAP_INLINE) {
> >                 free_extent_map(em);
> > -               ret = -ENOTBLK;
> > +               ret = (flags & IOMAP_NOWAIT) ? -EAGAIN : -ENOTBLK;
> >                 goto unlock_err;
> >         }
> >  
> > Can you give it a try?
> 
> This appears to do the trick!
> I've also removed the first patch and still cannot see any short reads,
> so this would be enough on its own for my case.

Great.

After my last reply, I retried your reproducer (and qemu image file) on a fs
mounted with -o compress-force=zstd and was able to trigger the short reads
in less than a minute.

Witht that patch I can no longer trigger the short reads too (after running
the reproducer for about 2 hours).

I'll give it some more testing here along with other minor fixes I have for
other scenarios. I'll submit a patchset, with this fix included, on monday.

I'll add your Tested-by tag for this specific patch.

Thanks for all the testing and the reproducer!


> 
> -- 
> Dominique