Re: [PATCH] block: fail unaligned bio from submit_bio_noacct()

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

On Thu, Mar 21, 2024 at 11:09:25AM -0600, Jens Axboe wrote:
> On 3/21/24 7:16 AM, Ming Lei wrote:
> > For any bio with data, its start sector and size have to be aligned with
> > the queue's logical block size.
> > 
> > This rule is obvious, but there is still user which may send unaligned
> > bio to block layer, and it is observed that dm-integrity can do that,
> > and cause double free of driver's dma meta buffer.
> > 
> > So failfast unaligned bio from submit_bio_noacct() for avoiding more
> > troubles.
> > 
> > Cc: Mikulas Patocka <mpatocka@redhat.com>
> > Cc: Mike Snitzer <snitzer@kernel.org>
> > Signed-off-by: Ming Lei <ming.lei@redhat.com>
> > ---
> >  block/blk-core.c | 17 +++++++++++++++++
> >  1 file changed, 17 insertions(+)
> > 
> > diff --git a/block/blk-core.c b/block/blk-core.c
> > index a16b5abdbbf5..b1a10187ef74 100644
> > --- a/block/blk-core.c
> > +++ b/block/blk-core.c
> > @@ -729,6 +729,20 @@ void submit_bio_noacct_nocheck(struct bio *bio)
> >  		__submit_bio_noacct(bio);
> >  }
> >  
> > +static bool bio_check_alignment(struct bio *bio, struct request_queue *q)
> > +{
> > +	unsigned int bs = q->limits.logical_block_size;
> > +	unsigned int size = bio->bi_iter.bi_size;
> > +
> > +	if (size & (bs - 1))
> > +		return false;
> > +
> > +	if (size && ((bio->bi_iter.bi_sector << SECTOR_SHIFT) & (bs - 1)))
> > +		return false;
> > +
> > +	return true;
> > +}
> > +
> >  /**
> >   * submit_bio_noacct - re-submit a bio to the block device layer for I/O
> >   * @bio:  The bio describing the location in memory and on the device.
> > @@ -780,6 +794,9 @@ void submit_bio_noacct(struct bio *bio)
> >  		}
> >  	}
> >  
> > +	if (WARN_ON_ONCE(!bio_check_alignment(bio, q)))
> > +		goto end_io;
> > +
> >  	if (!test_bit(QUEUE_FLAG_POLL, &q->queue_flags))
> >  		bio_clear_polled(bio);
> 
> Where is this IO coming from? The normal block level dio has checks. And
> in fact they are expensive... If we add this one, then we should be able
> to kill the block/fops.c checks, no?

I think Most of fs code should send good bio since I didn't trigger it in
xfstests.

But we still have md, dm, bcache and target code which build bio in
their way. The reported issue is from device mapper integrity code.

Yes, all (offset & size) alignment in fops.c shouldn't be needed any more.


Thanks,
Ming