block: respect virtual boundary mask in bvecs

With drivers that are settting a virtual boundary constrain, we are
seeing a lot of bio splitting and smaller I/Os being submitted to the
driver.

This happens because the bio gap detection code does not account cases
where PAGE_SIZE - 1 is bigger than queue_virt_boundary() and thus will
split the bio unnecessarily.

Cc: Jan Kara <jack@suse.cz>
Cc: Bart Van Assche <bvanassche@acm.org>
Cc: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Acked-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block/blk-merge.c

@@ -46,7 +46,7 @@ static inline bool bio_will_gap(struct request_queue *q,
 		bio_get_first_bvec(prev_rq->bio, &pb);
 	else
 		bio_get_first_bvec(prev, &pb);
-	if (pb.bv_offset)
+	if (pb.bv_offset & queue_virt_boundary(q))
 		return true;
 
 	/*

block/blk.h

@@ -169,7 +169,7 @@ static inline bool biovec_phys_mergeable(struct request_queue *q,
 static inline bool __bvec_gap_to_prev(struct request_queue *q,
 		struct bio_vec *bprv, unsigned int offset)
 {
-	return offset ||
+	return (offset & queue_virt_boundary(q)) ||
 		((bprv->bv_offset + bprv->bv_len) & queue_virt_boundary(q));
 }