s390/scm_block: handle multiple requests in one HW request

Handle up to 8 block layer requests per HW request. These requests
can be processed in parallel on the device leading to better
throughput (and less interrupts). The overhead for additional
requests is small since we don't blindly allocate new aidaws but
try to use what's left of the previous one.

Signed-off-by: Sebastian Ott <sebott@linux.vnet.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>

1 files changed, 31 insertions, 16 deletions

diff --git a/drivers/s390/block/scm_blk_cluster.c b/drivers/s390/block/scm_blk_cluster.c
index 2fd01320b978..09db45296eed 100644
--- a/drivers/s390/block/scm_blk_cluster.c
+++ b/drivers/s390/block/scm_blk_cluster.c
@@ -57,39 +57,52 @@ void scm_request_cluster_init(struct scm_request *scmrq)
 	scmrq->cluster.state = CLUSTER_NONE;
 }
 
-static bool clusters_intersect(struct scm_request *A, struct scm_request *B)
+static bool clusters_intersect(struct request *A, struct request *B)
 {
 	unsigned long firstA, lastA, firstB, lastB;
 
-	firstA = ((u64) blk_rq_pos(A->request) << 9) / CLUSTER_SIZE;
-	lastA = (((u64) blk_rq_pos(A->request) << 9) +
-		    blk_rq_bytes(A->request) - 1) / CLUSTER_SIZE;
+	firstA = ((u64) blk_rq_pos(A) << 9) / CLUSTER_SIZE;
+	lastA = (((u64) blk_rq_pos(A) << 9) +
+		    blk_rq_bytes(A) - 1) / CLUSTER_SIZE;
 
-	firstB = ((u64) blk_rq_pos(B->request) << 9) / CLUSTER_SIZE;
-	lastB = (((u64) blk_rq_pos(B->request) << 9) +
-		    blk_rq_bytes(B->request) - 1) / CLUSTER_SIZE;
+	firstB = ((u64) blk_rq_pos(B) << 9) / CLUSTER_SIZE;
+	lastB = (((u64) blk_rq_pos(B) << 9) +
+		    blk_rq_bytes(B) - 1) / CLUSTER_SIZE;
 
 	return (firstB <= lastA && firstA <= lastB);
 }
 
 bool scm_reserve_cluster(struct scm_request *scmrq)
 {
+	struct request *req = scmrq->request[scmrq->aob->request.msb_count];
 	struct scm_blk_dev *bdev = scmrq->bdev;
 	struct scm_request *iter;
+	int pos, add = 1;
 
 	if (write_cluster_size == 0)
 		return true;
 
 	spin_lock(&bdev->lock);
 	list_for_each_entry(iter, &bdev->cluster_list, cluster.list) {
-		if (clusters_intersect(scmrq, iter) &&
-		    (rq_data_dir(scmrq->request) == WRITE ||
-		     rq_data_dir(iter->request) == WRITE)) {
-			spin_unlock(&bdev->lock);
-			return false;
+		if (iter == scmrq) {
+			/*
+			 * We don't have to use clusters_intersect here, since
+			 * cluster requests are always started separately.
+			 */
+			add = 0;
+			continue;
+		}
+		for (pos = 0; pos <= iter->aob->request.msb_count; pos++) {
+			if (clusters_intersect(req, iter->request[pos]) &&
+			    (rq_data_dir(req) == WRITE ||
+			     rq_data_dir(iter->request[pos]) == WRITE)) {
+				spin_unlock(&bdev->lock);
+				return false;
+			}
 		}
 	}
-	list_add(&scmrq->cluster.list, &bdev->cluster_list);
+	if (add)
+		list_add(&scmrq->cluster.list, &bdev->cluster_list);
 	spin_unlock(&bdev->lock);
 
 	return true;
@@ -118,7 +131,7 @@ static int scm_prepare_cluster_request(struct scm_request *scmrq)
 {
 	struct scm_blk_dev *bdev = scmrq->bdev;
 	struct scm_device *scmdev = bdev->gendisk->private_data;
-	struct request *req = scmrq->request;
+	struct request *req = scmrq->request[0];
 	struct msb *msb = &scmrq->aob->msb[0];
 	struct req_iterator iter;
 	struct aidaw *aidaw;
@@ -183,10 +196,12 @@ static int scm_prepare_cluster_request(struct scm_request *scmrq)
 
 bool scm_need_cluster_request(struct scm_request *scmrq)
 {
-	if (rq_data_dir(scmrq->request) == READ)
+	int pos = scmrq->aob->request.msb_count;
+
+	if (rq_data_dir(scmrq->request[pos]) == READ)
 		return false;
 
-	return blk_rq_bytes(scmrq->request) < CLUSTER_SIZE;
+	return blk_rq_bytes(scmrq->request[pos]) < CLUSTER_SIZE;
 }
 
 /* Called with queue lock held. */