block/rnbd-clt: Dynamically allocate sglist for rnbd_iu

The BMAX_SEGMENT static array for scatterlist is embedded in
rnbd_iu structure to avoid memory allocation in hot IO path.
In many cases, we do need only several sg entries because many IOs
have only several segments.

This patch change rnbd_iu to check the number of segments in the request
and allocate sglist dynamically.

For io path, use sg_alloc_table_chained to allocate sg list faster.

First it makes two sg entries after pdu of request.
The sg_alloc_table_chained uses the pre-allocated sg entries
if the number of segments of the request is less than two.
So it reduces the number of memory allocation.

Signed-off-by: Gioh Kim <gi-oh.kim@cloud.ionos.com>
Signed-off-by: Jack Wang <jinpu.wang@cloud.ionos.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Author: Gioh Kim

drivers/block/rnbd/rnbd-clt.c

@@ -384,6 +384,7 @@ static void rnbd_softirq_done_fn(struct request *rq)
 	struct rnbd_iu *iu;
 
 	iu = blk_mq_rq_to_pdu(rq);
+	sg_free_table_chained(&iu->sgt, RNBD_INLINE_SG_CNT);
 	rnbd_put_permit(sess, iu->permit);
 	blk_mq_end_request(rq, errno_to_blk_status(iu->errno));
 }
@@ -477,7 +478,7 @@ static int send_msg_close(struct rnbd_clt_dev *dev, u32 device_id, bool wait)
 	iu->buf = NULL;
 	iu->dev = dev;
 
-	sg_mark_end(&iu->sglist[0]);
+	sg_alloc_table(&iu->sgt, 1, GFP_KERNEL);
 
 	msg.hdr.type	= cpu_to_le16(RNBD_MSG_CLOSE);
 	msg.device_id	= cpu_to_le32(device_id);
@@ -492,6 +493,7 @@ static int send_msg_close(struct rnbd_clt_dev *dev, u32 device_id, bool wait)
 		err = errno;
 	}
 
+	sg_free_table(&iu->sgt);
 	rnbd_put_iu(sess, iu);
 	return err;
 }
@@ -564,15 +566,16 @@ static int send_msg_open(struct rnbd_clt_dev *dev, bool wait)
 	iu->buf = rsp;
 	iu->dev = dev;
 
-	sg_init_one(iu->sglist, rsp, sizeof(*rsp));
+	sg_alloc_table(&iu->sgt, 1, GFP_KERNEL);
+	sg_init_one(iu->sgt.sgl, rsp, sizeof(*rsp));
 
 	msg.hdr.type	= cpu_to_le16(RNBD_MSG_OPEN);
 	msg.access_mode	= dev->access_mode;
 	strlcpy(msg.dev_name, dev->pathname, sizeof(msg.dev_name));
 
 	WARN_ON(!rnbd_clt_get_dev(dev));
 	err = send_usr_msg(sess->rtrs, READ, iu,
-			   &vec, sizeof(*rsp), iu->sglist, 1,
+			   &vec, sizeof(*rsp), iu->sgt.sgl, 1,
 			   msg_open_conf, &errno, wait);
 	if (err) {
 		rnbd_clt_put_dev(dev);
@@ -582,6 +585,7 @@ static int send_msg_open(struct rnbd_clt_dev *dev, bool wait)
 		err = errno;
 	}
 
+	sg_free_table(&iu->sgt);
 	rnbd_put_iu(sess, iu);
 	return err;
 }
@@ -610,7 +614,8 @@ static int send_msg_sess_info(struct rnbd_clt_session *sess, bool wait)
 	iu->buf = rsp;
 	iu->sess = sess;
 
-	sg_init_one(iu->sglist, rsp, sizeof(*rsp));
+	sg_alloc_table(&iu->sgt, 1, GFP_KERNEL);
+	sg_init_one(iu->sgt.sgl, rsp, sizeof(*rsp));
 
 	msg.hdr.type = cpu_to_le16(RNBD_MSG_SESS_INFO);
 	msg.ver      = RNBD_PROTO_VER_MAJOR;
@@ -626,7 +631,7 @@ static int send_msg_sess_info(struct rnbd_clt_session *sess, bool wait)
 		goto put_iu;
 	}
 	err = send_usr_msg(sess->rtrs, READ, iu,
-			   &vec, sizeof(*rsp), iu->sglist, 1,
+			   &vec, sizeof(*rsp), iu->sgt.sgl, 1,
 			   msg_sess_info_conf, &errno, wait);
 	if (err) {
 		rnbd_clt_put_sess(sess);
@@ -636,7 +641,7 @@ static int send_msg_sess_info(struct rnbd_clt_session *sess, bool wait)
 	} else {
 		err = errno;
 	}
-
+	sg_free_table(&iu->sgt);
 	rnbd_put_iu(sess, iu);
 	return err;
 }
@@ -1016,11 +1021,10 @@ static int rnbd_client_xfer_request(struct rnbd_clt_dev *dev,
 	 * See queue limits.
 	 */
 	if (req_op(rq) != REQ_OP_DISCARD)
-		sg_cnt = blk_rq_map_sg(dev->queue, rq, iu->sglist);
+		sg_cnt = blk_rq_map_sg(dev->queue, rq, iu->sgt.sgl);
 
 	if (sg_cnt == 0)
-		/* Do not forget to mark the end */
-		sg_mark_end(&iu->sglist[0]);
+		sg_mark_end(&iu->sgt.sgl[0]);
 
 	msg.hdr.type	= cpu_to_le16(RNBD_MSG_IO);
 	msg.device_id	= cpu_to_le32(dev->device_id);
@@ -1029,13 +1033,13 @@ static int rnbd_client_xfer_request(struct rnbd_clt_dev *dev,
 		.iov_base = &msg,
 		.iov_len  = sizeof(msg)
 	};
-	size = rnbd_clt_get_sg_size(iu->sglist, sg_cnt);
+	size = rnbd_clt_get_sg_size(iu->sgt.sgl, sg_cnt);
 	req_ops = (struct rtrs_clt_req_ops) {
 		.priv = iu,
 		.conf_fn = msg_io_conf,
 	};
 	err = rtrs_clt_request(rq_data_dir(rq), &req_ops, rtrs, permit,
-			       &vec, 1, size, iu->sglist, sg_cnt);
+			       &vec, 1, size, iu->sgt.sgl, sg_cnt);
 	if (unlikely(err)) {
 		rnbd_clt_err_rl(dev, "RTRS failed to transfer IO, err: %d\n",
 				 err);
@@ -1122,6 +1126,7 @@ static blk_status_t rnbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 	struct rnbd_clt_dev *dev = rq->rq_disk->private_data;
 	struct rnbd_iu *iu = blk_mq_rq_to_pdu(rq);
 	int err;
+	blk_status_t ret = BLK_STS_IOERR;
 
 	if (unlikely(dev->dev_state != DEV_STATE_MAPPED))
 		return BLK_STS_IOERR;
@@ -1133,32 +1138,35 @@ static blk_status_t rnbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 		return BLK_STS_RESOURCE;
 	}
 
+	iu->sgt.sgl = iu->first_sgl;
+	err = sg_alloc_table_chained(&iu->sgt,
+				     /* Even-if the request has no segment,
+				      * sglist must have one entry at least */
+				     blk_rq_nr_phys_segments(rq) ? : 1,
+				     iu->sgt.sgl,
+				     RNBD_INLINE_SG_CNT);
+	if (err) {
+		rnbd_clt_err_rl(dev, "sg_alloc_table_chained ret=%d\n", err);
+		rnbd_clt_dev_kick_mq_queue(dev, hctx, 10/*ms*/);
+		rnbd_put_permit(dev->sess, iu->permit);
+		return BLK_STS_RESOURCE;
+	}
+
 	blk_mq_start_request(rq);
 	err = rnbd_client_xfer_request(dev, rq, iu);
 	if (likely(err == 0))
 		return BLK_STS_OK;
 	if (unlikely(err == -EAGAIN || err == -ENOMEM)) {
 		rnbd_clt_dev_kick_mq_queue(dev, hctx, 10/*ms*/);
-		rnbd_put_permit(dev->sess, iu->permit);
-		return BLK_STS_RESOURCE;
+		ret = BLK_STS_RESOURCE;
 	}
-
+	sg_free_table_chained(&iu->sgt, RNBD_INLINE_SG_CNT);
 	rnbd_put_permit(dev->sess, iu->permit);
-	return BLK_STS_IOERR;
-}
-
-static int rnbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
-			      unsigned int hctx_idx, unsigned int numa_node)
-{
-	struct rnbd_iu *iu = blk_mq_rq_to_pdu(rq);
-
-	sg_init_table(iu->sglist, BMAX_SEGMENTS);
-	return 0;
+	return ret;
 }
 
 static struct blk_mq_ops rnbd_mq_ops = {
 	.queue_rq	= rnbd_queue_rq,
-	.init_request	= rnbd_init_request,
 	.complete	= rnbd_softirq_done_fn,
 };
 
@@ -1172,7 +1180,7 @@ static int setup_mq_tags(struct rnbd_clt_session *sess)
 	tag_set->numa_node		= NUMA_NO_NODE;
 	tag_set->flags		= BLK_MQ_F_SHOULD_MERGE |
 				  BLK_MQ_F_TAG_QUEUE_SHARED;
-	tag_set->cmd_size		= sizeof(struct rnbd_iu);
+	tag_set->cmd_size	= sizeof(struct rnbd_iu) + RNBD_RDMA_SGL_SIZE;
 	tag_set->nr_hw_queues	= num_online_cpus();
 
 	return blk_mq_alloc_tag_set(tag_set);

drivers/block/rnbd/rnbd-clt.h

@@ -44,6 +44,13 @@ struct rnbd_iu_comp {
 	int errno;
 };
 
+#ifdef CONFIG_ARCH_NO_SG_CHAIN
+#define RNBD_INLINE_SG_CNT 0
+#else
+#define RNBD_INLINE_SG_CNT 2
+#endif
+#define RNBD_RDMA_SGL_SIZE (sizeof(struct scatterlist) * RNBD_INLINE_SG_CNT)
+
 struct rnbd_iu {
 	union {
 		struct request *rq; /* for block io */
@@ -56,11 +63,12 @@ struct rnbd_iu {
 		/* use to send msg associated with a sess */
 		struct rnbd_clt_session *sess;
 	};
-	struct scatterlist	sglist[BMAX_SEGMENTS];
+	struct sg_table		sgt;
 	struct work_struct	work;
 	int			errno;
 	struct rnbd_iu_comp	comp;
 	atomic_t		refcount;
+	struct scatterlist	first_sgl[]; /* must be the last one */
 };
 
 struct rnbd_cpu_qlist {