Re: [PATCH v2] btrfs: raid56: data corruption on a device removal

[Ming Lei <ming.lei@redhat.com>]

On Fri, Jan 11, 2019 at 09:08:27AM +0100, Johannes Thumshirn wrote:
> On 10/01/2019 17:49, Johannes Thumshirn wrote:
> > Right, but we're looking for the number of already completed bytes to
> > rewind here, so from bvec.h's docs it is bi_bvec_done.
> > 
> > Dmitriy can you see if this works for you:
> > 
> > diff --git a/fs/btrfs/raid56.c b/fs/btrfs/raid56.c
> > index e74455eb42f9..2d0e2eec5413 100644
> > --- a/fs/btrfs/raid56.c
> > +++ b/fs/btrfs/raid56.c
> > @@ -1350,6 +1350,7 @@ static int find_bio_stripe(struct btrfs_raid_bio
> > *rbio,
> >         struct btrfs_bio_stripe *stripe;
> > 
> >         physical <<= 9;
> > +       physical -= bio->bi_iter.bi_bvec_done;
> 
> OK talked to Hannes about this issue, he says the only way is to save
> the iterator state before submitting the bio.
> 
> So the above is bogus too.
> 
> This also is what Kent said in
> https://marc.info/?l=linux-block&m=153549921116441&w=2

Yeah, .bi_bvec_done can't help, seems you need to take the similar
approach in 7759eb23fd9808a2e ("block: remove bio_rewind_iter()").

In theory, the follow way might work, but still like a hack:

	#define BVEC_ITER_ALL_INIT (struct bvec_iter)                  \
	{                                                              \
	        .bi_sector      = 0,                                   \
	        .bi_size        = UINT_MAX,                            \
	        .bi_idx         = 0,                                   \
	        .bi_bvec_done   = 0,                                   \
	}
	
	old_bi_sector = bio->bi_iter.bi_sector;
	bio->bi_iter = BVEC_ITER_ALL_INIT;
	bio->bi_iter.bi_sector = old_bi_sector;
	bio_advance(bio, where_to_rewind);

This bio can't be the fast-cloned one, which should be satisfied
by fs code.

Also 'where_to_rewind' needs to point to the offset relative to start
of this bio.


Thanks,
Ming