block: add queue flag for disabling SG merging

If devices are not SG starved, we waste a lot of time potentially
collapsing SG segments. Enough that 1.5% of the CPU time goes
to this, at only 400K IOPS. Add a queue flag, QUEUE_FLAG_NO_SG_MERGE,
which just returns the number of vectors in a bio instead of looping
over all segments and checking for collapsible ones.

Add a BLK_MQ_F_SG_MERGE flag so that drivers can opt-in on the sg
merging, if they so desire.

Signed-off-by: Jens Axboe <axboe@fb.com>

block/blk-merge.c

@@ -13,7 +13,7 @@ static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
 					     struct bio *bio)
 {
 	struct bio_vec bv, bvprv = { NULL };
-	int cluster, high, highprv = 1;
+	int cluster, high, highprv = 1, no_sg_merge;
 	unsigned int seg_size, nr_phys_segs;
 	struct bio *fbio, *bbio;
 	struct bvec_iter iter;
@@ -35,12 +35,21 @@ static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
 	cluster = blk_queue_cluster(q);
 	seg_size = 0;
 	nr_phys_segs = 0;
+	no_sg_merge = test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags);
+	high = 0;
 	for_each_bio(bio) {
 		bio_for_each_segment(bv, bio, iter) {
+			/*
+			 * If SG merging is disabled, each bio vector is
+			 * a segment
+			 */
+			if (no_sg_merge)
+				goto new_segment;
+
 			/*
 			 * the trick here is making sure that a high page is
-			 * never considered part of another segment, since that
+			 * never considered part of another segment, since
-			 * might change with the bounce page.
+			 * that might change with the bounce page.
 			 */
 			high = page_to_pfn(bv.bv_page) > queue_bounce_pfn(q);
 			if (!high && !highprv && cluster) {
@@ -84,11 +93,16 @@ void blk_recalc_rq_segments(struct request *rq)
 
 void blk_recount_segments(struct request_queue *q, struct bio *bio)
 {
-	struct bio *nxt = bio->bi_next;
+	if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags))
+		bio->bi_phys_segments = bio->bi_vcnt;
+	else {
+		struct bio *nxt = bio->bi_next;
+
+		bio->bi_next = NULL;
+		bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
+		bio->bi_next = nxt;
+	}
 
-	bio->bi_next = NULL;
-	bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
-	bio->bi_next = nxt;
 	bio->bi_flags |= (1 << BIO_SEG_VALID);
 }
 EXPORT_SYMBOL(blk_recount_segments);

block/blk-mq.c

@@ -1829,6 +1829,9 @@ struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
 	q->mq_ops = set->ops;
 	q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
 
+	if (!(set->flags & BLK_MQ_F_SG_MERGE))
+		q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
+
 	q->sg_reserved_size = INT_MAX;
 
 	INIT_WORK(&q->requeue_work, blk_mq_requeue_work);

include/linux/blk-mq.h

@@ -129,6 +129,7 @@ enum {
 	BLK_MQ_F_SHOULD_MERGE	= 1 << 0,
 	BLK_MQ_F_SHOULD_SORT	= 1 << 1,
 	BLK_MQ_F_TAG_SHARED	= 1 << 2,
+	BLK_MQ_F_SG_MERGE	= 1 << 3,
 
 	BLK_MQ_S_STOPPED	= 0,
 	BLK_MQ_S_TAG_ACTIVE	= 1,

include/linux/blkdev.h

@@ -510,6 +510,7 @@ struct request_queue {
 #define QUEUE_FLAG_SAME_FORCE  18	/* force complete on same CPU */
 #define QUEUE_FLAG_DEAD        19	/* queue tear-down finished */
 #define QUEUE_FLAG_INIT_DONE   20	/* queue is initialized */
+#define QUEUE_FLAG_NO_SG_MERGE 21	/* don't attempt to merge SG segments*/
 
 #define QUEUE_FLAG_DEFAULT	((1 << QUEUE_FLAG_IO_STAT) |		\
 				 (1 << QUEUE_FLAG_STACKABLE)	|	\