Linux-RDMA Archive on lore.kernel.org
 help / color / Atom feed
* [PATCHv4 for-next 08/19] block/rnbd-clt: Replace {NO_WAIT,WAIT} with RTRS_PERMIT_{WAIT,NOWAIT}
       [not found] <20210414122402.203388-1-gi-oh.kim@ionos.com>
@ 2021-04-14 12:23 ` Gioh Kim
  2021-04-14 12:23 ` [PATCHv4 for-next 13/19] block/rnbd-clt: Support polling mode for IO latency optimization Gioh Kim
                   ` (2 subsequent siblings)
  3 siblings, 0 replies; 10+ messages in thread
From: Gioh Kim @ 2021-04-14 12:23 UTC (permalink / raw)
  To: linux-block
  Cc: axboe, hch, sagi, bvanassche, haris.iqbal, jinpu.wang, Gioh Kim,
	Leon Romanovsky, linux-rdma, Guoqing Jiang, Gioh Kim,
	Chaitanya Kulkarni, Jason Gunthorpe

From: Gioh Kim <gi-oh.kim@cloud.ionos.com>

They are defined with the same value and similar meaning, let's remove
one of them, then we can remove {WAIT,NOWAIT}.

Also change the type of 'wait' from 'int' to 'enum wait_type' to make
it clear.

Cc: Leon Romanovsky <leonro@nvidia.com>
Cc: linux-rdma@vger.kernel.org
Signed-off-by: Guoqing Jiang <guoqing.jiang@ionos.com>
Reviewed-by: Md Haris Iqbal <haris.iqbal@ionos.com>
Signed-off-by: Gioh Kim <gi-oh.kim@ionos.com>
Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Acked-by: Jason Gunthorpe <jgg@nvidia.com>
---
 drivers/block/rnbd/rnbd-clt.c          | 42 +++++++++++---------------
 drivers/infiniband/ulp/rtrs/rtrs-clt.c |  4 +--
 drivers/infiniband/ulp/rtrs/rtrs.h     |  6 ++--
 3 files changed, 22 insertions(+), 30 deletions(-)

diff --git a/drivers/block/rnbd/rnbd-clt.c b/drivers/block/rnbd/rnbd-clt.c
index 4e687ec88721..652b41cc4492 100644
--- a/drivers/block/rnbd/rnbd-clt.c
+++ b/drivers/block/rnbd/rnbd-clt.c
@@ -312,13 +312,11 @@ static void rnbd_rerun_all_if_idle(struct rnbd_clt_session *sess)
 
 static struct rtrs_permit *rnbd_get_permit(struct rnbd_clt_session *sess,
 					     enum rtrs_clt_con_type con_type,
-					     int wait)
+					     enum wait_type wait)
 {
 	struct rtrs_permit *permit;
 
-	permit = rtrs_clt_get_permit(sess->rtrs, con_type,
-				      wait ? RTRS_PERMIT_WAIT :
-				      RTRS_PERMIT_NOWAIT);
+	permit = rtrs_clt_get_permit(sess->rtrs, con_type, wait);
 	if (likely(permit))
 		/* We have a subtle rare case here, when all permits can be
 		 * consumed before busy counter increased.  This is safe,
@@ -344,7 +342,7 @@ static void rnbd_put_permit(struct rnbd_clt_session *sess,
 
 static struct rnbd_iu *rnbd_get_iu(struct rnbd_clt_session *sess,
 				     enum rtrs_clt_con_type con_type,
-				     int wait)
+				     enum wait_type wait)
 {
 	struct rnbd_iu *iu;
 	struct rtrs_permit *permit;
@@ -354,9 +352,7 @@ static struct rnbd_iu *rnbd_get_iu(struct rnbd_clt_session *sess,
 		return NULL;
 	}
 
-	permit = rnbd_get_permit(sess, con_type,
-				  wait ? RTRS_PERMIT_WAIT :
-				  RTRS_PERMIT_NOWAIT);
+	permit = rnbd_get_permit(sess, con_type, wait);
 	if (unlikely(!permit)) {
 		kfree(iu);
 		return NULL;
@@ -435,16 +431,11 @@ static void msg_conf(void *priv, int errno)
 	schedule_work(&iu->work);
 }
 
-enum wait_type {
-	NO_WAIT = 0,
-	WAIT    = 1
-};
-
 static int send_usr_msg(struct rtrs_clt *rtrs, int dir,
 			struct rnbd_iu *iu, struct kvec *vec,
 			size_t len, struct scatterlist *sg, unsigned int sg_len,
 			void (*conf)(struct work_struct *work),
-			int *errno, enum wait_type wait)
+			int *errno, int wait)
 {
 	int err;
 	struct rtrs_clt_req_ops req_ops;
@@ -476,7 +467,8 @@ static void msg_close_conf(struct work_struct *work)
 	rnbd_clt_put_dev(dev);
 }
 
-static int send_msg_close(struct rnbd_clt_dev *dev, u32 device_id, bool wait)
+static int send_msg_close(struct rnbd_clt_dev *dev, u32 device_id,
+			  enum wait_type wait)
 {
 	struct rnbd_clt_session *sess = dev->sess;
 	struct rnbd_msg_close msg;
@@ -530,7 +522,7 @@ static void msg_open_conf(struct work_struct *work)
 			 * If server thinks its fine, but we fail to process
 			 * then be nice and send a close to server.
 			 */
-			(void)send_msg_close(dev, device_id, NO_WAIT);
+			send_msg_close(dev, device_id, RTRS_PERMIT_NOWAIT);
 		}
 	}
 	kfree(rsp);
@@ -554,7 +546,7 @@ static void msg_sess_info_conf(struct work_struct *work)
 	rnbd_clt_put_sess(sess);
 }
 
-static int send_msg_open(struct rnbd_clt_dev *dev, bool wait)
+static int send_msg_open(struct rnbd_clt_dev *dev, enum wait_type wait)
 {
 	struct rnbd_clt_session *sess = dev->sess;
 	struct rnbd_msg_open_rsp *rsp;
@@ -601,7 +593,7 @@ static int send_msg_open(struct rnbd_clt_dev *dev, bool wait)
 	return err;
 }
 
-static int send_msg_sess_info(struct rnbd_clt_session *sess, bool wait)
+static int send_msg_sess_info(struct rnbd_clt_session *sess, enum wait_type wait)
 {
 	struct rnbd_msg_sess_info_rsp *rsp;
 	struct rnbd_msg_sess_info msg;
@@ -687,7 +679,7 @@ static void remap_devs(struct rnbd_clt_session *sess)
 	 * be asynchronous.
 	 */
 
-	err = send_msg_sess_info(sess, NO_WAIT);
+	err = send_msg_sess_info(sess, RTRS_PERMIT_NOWAIT);
 	if (err) {
 		pr_err("send_msg_sess_info(\"%s\"): %d\n", sess->sessname, err);
 		return;
@@ -711,7 +703,7 @@ static void remap_devs(struct rnbd_clt_session *sess)
 			continue;
 
 		rnbd_clt_info(dev, "session reconnected, remapping device\n");
-		err = send_msg_open(dev, NO_WAIT);
+		err = send_msg_open(dev, RTRS_PERMIT_NOWAIT);
 		if (err) {
 			rnbd_clt_err(dev, "send_msg_open(): %d\n", err);
 			break;
@@ -1242,7 +1234,7 @@ find_and_get_or_create_sess(const char *sessname,
 	if (err)
 		goto close_rtrs;
 
-	err = send_msg_sess_info(sess, WAIT);
+	err = send_msg_sess_info(sess, RTRS_PERMIT_WAIT);
 	if (err)
 		goto close_rtrs;
 
@@ -1525,7 +1517,7 @@ struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
 		ret = -EEXIST;
 		goto put_dev;
 	}
-	ret = send_msg_open(dev, WAIT);
+	ret = send_msg_open(dev, RTRS_PERMIT_WAIT);
 	if (ret) {
 		rnbd_clt_err(dev,
 			      "map_device: failed, can't open remote device, err: %d\n",
@@ -1559,7 +1551,7 @@ struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
 	return dev;
 
 send_close:
-	send_msg_close(dev, dev->device_id, WAIT);
+	send_msg_close(dev, dev->device_id, RTRS_PERMIT_WAIT);
 del_dev:
 	delete_dev(dev);
 put_dev:
@@ -1619,7 +1611,7 @@ int rnbd_clt_unmap_device(struct rnbd_clt_dev *dev, bool force,
 	destroy_sysfs(dev, sysfs_self);
 	destroy_gen_disk(dev);
 	if (was_mapped && sess->rtrs)
-		send_msg_close(dev, dev->device_id, WAIT);
+		send_msg_close(dev, dev->device_id, RTRS_PERMIT_WAIT);
 
 	rnbd_clt_info(dev, "Device is unmapped\n");
 
@@ -1653,7 +1645,7 @@ int rnbd_clt_remap_device(struct rnbd_clt_dev *dev)
 	mutex_unlock(&dev->lock);
 	if (!err) {
 		rnbd_clt_info(dev, "Remapping device.\n");
-		err = send_msg_open(dev, WAIT);
+		err = send_msg_open(dev, RTRS_PERMIT_WAIT);
 		if (err)
 			rnbd_clt_err(dev, "remap_device: %d\n", err);
 	}
diff --git a/drivers/infiniband/ulp/rtrs/rtrs-clt.c b/drivers/infiniband/ulp/rtrs/rtrs-clt.c
index 0a08b4b742a3..7efd49bdc78c 100644
--- a/drivers/infiniband/ulp/rtrs/rtrs-clt.c
+++ b/drivers/infiniband/ulp/rtrs/rtrs-clt.c
@@ -103,11 +103,11 @@ static inline void __rtrs_put_permit(struct rtrs_clt *clt,
  *    up earlier.
  *
  * Context:
- *    Can sleep if @wait == RTRS_TAG_WAIT
+ *    Can sleep if @wait == RTRS_PERMIT_WAIT
  */
 struct rtrs_permit *rtrs_clt_get_permit(struct rtrs_clt *clt,
 					  enum rtrs_clt_con_type con_type,
-					  int can_wait)
+					  enum wait_type can_wait)
 {
 	struct rtrs_permit *permit;
 	DEFINE_WAIT(wait);
diff --git a/drivers/infiniband/ulp/rtrs/rtrs.h b/drivers/infiniband/ulp/rtrs/rtrs.h
index 8738e90e715a..2db1b5eb3ab0 100644
--- a/drivers/infiniband/ulp/rtrs/rtrs.h
+++ b/drivers/infiniband/ulp/rtrs/rtrs.h
@@ -63,9 +63,9 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
 
 void rtrs_clt_close(struct rtrs_clt *sess);
 
-enum {
+enum wait_type {
 	RTRS_PERMIT_NOWAIT = 0,
-	RTRS_PERMIT_WAIT   = 1,
+	RTRS_PERMIT_WAIT   = 1
 };
 
 /**
@@ -81,7 +81,7 @@ enum rtrs_clt_con_type {
 
 struct rtrs_permit *rtrs_clt_get_permit(struct rtrs_clt *sess,
 				    enum rtrs_clt_con_type con_type,
-				    int wait);
+				    enum wait_type wait);
 
 void rtrs_clt_put_permit(struct rtrs_clt *sess, struct rtrs_permit *permit);
 
-- 
2.25.1


^ permalink raw reply	[flat|nested] 10+ messages in thread

* [PATCHv4 for-next 13/19] block/rnbd-clt: Support polling mode for IO latency optimization
       [not found] <20210414122402.203388-1-gi-oh.kim@ionos.com>
  2021-04-14 12:23 ` [PATCHv4 for-next 08/19] block/rnbd-clt: Replace {NO_WAIT,WAIT} with RTRS_PERMIT_{WAIT,NOWAIT} Gioh Kim
@ 2021-04-14 12:23 ` Gioh Kim
  2021-04-18  8:36   ` Leon Romanovsky
  2021-04-14 12:23 ` [PATCHv4 for-next 15/19] block/rnbd-srv: Remove unused arguments of rnbd_srv_rdma_ev Gioh Kim
  2021-04-14 12:24 ` [PATCHv4 for-next 17/19] block/rnbd-clt: Remove max_segment_size Gioh Kim
  3 siblings, 1 reply; 10+ messages in thread
From: Gioh Kim @ 2021-04-14 12:23 UTC (permalink / raw)
  To: linux-block
  Cc: axboe, hch, sagi, bvanassche, haris.iqbal, jinpu.wang, Gioh Kim,
	Leon Romanovsky, linux-rdma, Gioh Kim, Jason Gunthorpe

From: Gioh Kim <gi-oh.kim@cloud.ionos.com>

RNBD can make double-queues for irq-mode and poll-mode.
For example, on 4-CPU system 8 request-queues are created,
4 for irq-mode and 4 for poll-mode.
If the IO has HIPRI flag, the block-layer will call .poll function
of RNBD. Then IO is sent to the poll-mode queue.
Add optional nr_poll_queues argument for map_devices interface.

To support polling of RNBD, RTRS client creates connections
for both of irq-mode and direct-poll-mode.

For example, on 4-CPU system it could've create 5 connections:
con[0] => user message (softirq cq)
con[1:4] => softirq cq

After this patch, it can create 9 connections:
con[0] => user message (softirq cq)
con[1:4] => softirq cq
con[5:8] => DIRECT-POLL cq

Cc: Leon Romanovsky <leonro@nvidia.com>
Cc: linux-rdma@vger.kernel.org
Signed-off-by: Gioh Kim <gi-oh.kim@ionos.com>
Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
Acked-by: Jason Gunthorpe <jgg@nvidia.com>
---
 drivers/block/rnbd/rnbd-clt-sysfs.c    | 56 +++++++++++++----
 drivers/block/rnbd/rnbd-clt.c          | 85 +++++++++++++++++++++++---
 drivers/block/rnbd/rnbd-clt.h          |  5 +-
 drivers/infiniband/ulp/rtrs/rtrs-clt.c | 62 +++++++++++++++----
 drivers/infiniband/ulp/rtrs/rtrs-pri.h |  1 +
 drivers/infiniband/ulp/rtrs/rtrs.h     |  3 +-
 6 files changed, 178 insertions(+), 34 deletions(-)

diff --git a/drivers/block/rnbd/rnbd-clt-sysfs.c b/drivers/block/rnbd/rnbd-clt-sysfs.c
index 49015f428e67..bd111ebceb75 100644
--- a/drivers/block/rnbd/rnbd-clt-sysfs.c
+++ b/drivers/block/rnbd/rnbd-clt-sysfs.c
@@ -34,6 +34,7 @@ enum {
 	RNBD_OPT_DEV_PATH	= 1 << 2,
 	RNBD_OPT_ACCESS_MODE	= 1 << 3,
 	RNBD_OPT_SESSNAME	= 1 << 6,
+	RNBD_OPT_NR_POLL_QUEUES	= 1 << 7,
 };
 
 static const unsigned int rnbd_opt_mandatory[] = {
@@ -42,12 +43,13 @@ static const unsigned int rnbd_opt_mandatory[] = {
 };
 
 static const match_table_t rnbd_opt_tokens = {
-	{RNBD_OPT_PATH,		"path=%s"	},
-	{RNBD_OPT_DEV_PATH,	"device_path=%s"},
-	{RNBD_OPT_DEST_PORT,	"dest_port=%d"  },
-	{RNBD_OPT_ACCESS_MODE,	"access_mode=%s"},
-	{RNBD_OPT_SESSNAME,	"sessname=%s"	},
-	{RNBD_OPT_ERR,		NULL		},
+	{RNBD_OPT_PATH,			"path=%s"		},
+	{RNBD_OPT_DEV_PATH,		"device_path=%s"	},
+	{RNBD_OPT_DEST_PORT,		"dest_port=%d"		},
+	{RNBD_OPT_ACCESS_MODE,		"access_mode=%s"	},
+	{RNBD_OPT_SESSNAME,		"sessname=%s"		},
+	{RNBD_OPT_NR_POLL_QUEUES,	"nr_poll_queues=%d"	},
+	{RNBD_OPT_ERR,			NULL			},
 };
 
 struct rnbd_map_options {
@@ -57,6 +59,7 @@ struct rnbd_map_options {
 	char *pathname;
 	u16 *dest_port;
 	enum rnbd_access_mode *access_mode;
+	u32 *nr_poll_queues;
 };
 
 static int rnbd_clt_parse_map_options(const char *buf, size_t max_path_cnt,
@@ -68,7 +71,7 @@ static int rnbd_clt_parse_map_options(const char *buf, size_t max_path_cnt,
 	int opt_mask = 0;
 	int token;
 	int ret = -EINVAL;
-	int i, dest_port;
+	int i, dest_port, nr_poll_queues;
 	int p_cnt = 0;
 
 	options = kstrdup(buf, GFP_KERNEL);
@@ -178,6 +181,19 @@ static int rnbd_clt_parse_map_options(const char *buf, size_t max_path_cnt,
 			kfree(p);
 			break;
 
+		case RNBD_OPT_NR_POLL_QUEUES:
+			if (match_int(args, &nr_poll_queues) || nr_poll_queues < -1 ||
+			    nr_poll_queues > (int)nr_cpu_ids) {
+				pr_err("bad nr_poll_queues parameter '%d'\n",
+				       nr_poll_queues);
+				ret = -EINVAL;
+				goto out;
+			}
+			if (nr_poll_queues == -1)
+				nr_poll_queues = nr_cpu_ids;
+			*opt->nr_poll_queues = nr_poll_queues;
+			break;
+
 		default:
 			pr_err("map_device: Unknown parameter or missing value '%s'\n",
 			       p);
@@ -227,6 +243,20 @@ static ssize_t state_show(struct kobject *kobj,
 
 static struct kobj_attribute rnbd_clt_state_attr = __ATTR_RO(state);
 
+static ssize_t nr_poll_queues_show(struct kobject *kobj,
+				   struct kobj_attribute *attr, char *page)
+{
+	struct rnbd_clt_dev *dev;
+
+	dev = container_of(kobj, struct rnbd_clt_dev, kobj);
+
+	return snprintf(page, PAGE_SIZE, "%d\n",
+			dev->nr_poll_queues);
+}
+
+static struct kobj_attribute rnbd_clt_nr_poll_queues =
+	__ATTR_RO(nr_poll_queues);
+
 static ssize_t mapping_path_show(struct kobject *kobj,
 				 struct kobj_attribute *attr, char *page)
 {
@@ -421,6 +451,7 @@ static struct attribute *rnbd_dev_attrs[] = {
 	&rnbd_clt_state_attr.attr,
 	&rnbd_clt_session_attr.attr,
 	&rnbd_clt_access_mode.attr,
+	&rnbd_clt_nr_poll_queues.attr,
 	NULL,
 };
 
@@ -469,7 +500,7 @@ static ssize_t rnbd_clt_map_device_show(struct kobject *kobj,
 					 char *page)
 {
 	return scnprintf(page, PAGE_SIZE,
-			 "Usage: echo \"[dest_port=server port number] sessname=<name of the rtrs session> path=<[srcaddr@]dstaddr> [path=<[srcaddr@]dstaddr>] device_path=<full path on remote side> [access_mode=<ro|rw|migration>]\" > %s\n\naddr ::= [ ip:<ipv4> | ip:<ipv6> | gid:<gid> ]\n",
+			 "Usage: echo \"[dest_port=server port number] sessname=<name of the rtrs session> path=<[srcaddr@]dstaddr> [path=<[srcaddr@]dstaddr>] device_path=<full path on remote side> [access_mode=<ro|rw|migration>] [nr_poll_queues=<number of queues>]\" > %s\n\naddr ::= [ ip:<ipv4> | ip:<ipv6> | gid:<gid> ]\n",
 			 attr->attr.name);
 }
 
@@ -541,6 +572,7 @@ static ssize_t rnbd_clt_map_device_store(struct kobject *kobj,
 	char sessname[NAME_MAX];
 	enum rnbd_access_mode access_mode = RNBD_ACCESS_RW;
 	u16 port_nr = RTRS_PORT;
+	u32 nr_poll_queues = 0;
 
 	struct sockaddr_storage *addrs;
 	struct rtrs_addr paths[6];
@@ -552,6 +584,7 @@ static ssize_t rnbd_clt_map_device_store(struct kobject *kobj,
 	opt.pathname = pathname;
 	opt.dest_port = &port_nr;
 	opt.access_mode = &access_mode;
+	opt.nr_poll_queues = &nr_poll_queues;
 	addrs = kcalloc(ARRAY_SIZE(paths) * 2, sizeof(*addrs), GFP_KERNEL);
 	if (!addrs)
 		return -ENOMEM;
@@ -565,12 +598,13 @@ static ssize_t rnbd_clt_map_device_store(struct kobject *kobj,
 	if (ret)
 		goto out;
 
-	pr_info("Mapping device %s on session %s, (access_mode: %s)\n",
+	pr_info("Mapping device %s on session %s, (access_mode: %s, nr_poll_queues: %d)\n",
 		pathname, sessname,
-		rnbd_access_mode_str(access_mode));
+		rnbd_access_mode_str(access_mode),
+		nr_poll_queues);
 
 	dev = rnbd_clt_map_device(sessname, paths, path_cnt, port_nr, pathname,
-				  access_mode);
+				  access_mode, nr_poll_queues);
 	if (IS_ERR(dev)) {
 		ret = PTR_ERR(dev);
 		goto out;
diff --git a/drivers/block/rnbd/rnbd-clt.c b/drivers/block/rnbd/rnbd-clt.c
index 9b44aac680d5..63719ec04d58 100644
--- a/drivers/block/rnbd/rnbd-clt.c
+++ b/drivers/block/rnbd/rnbd-clt.c
@@ -1165,9 +1165,54 @@ static blk_status_t rnbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 	return ret;
 }
 
+static int rnbd_rdma_poll(struct blk_mq_hw_ctx *hctx)
+{
+	struct rnbd_queue *q = hctx->driver_data;
+	struct rnbd_clt_dev *dev = q->dev;
+	int cnt;
+
+	cnt = rtrs_clt_rdma_cq_direct(dev->sess->rtrs, hctx->queue_num);
+	return cnt;
+}
+
+static int rnbd_rdma_map_queues(struct blk_mq_tag_set *set)
+{
+	struct rnbd_clt_session *sess = set->driver_data;
+
+	/* shared read/write queues */
+	set->map[HCTX_TYPE_DEFAULT].nr_queues = num_online_cpus();
+	set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
+	set->map[HCTX_TYPE_READ].nr_queues = num_online_cpus();
+	set->map[HCTX_TYPE_READ].queue_offset = 0;
+	blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
+	blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
+
+	if (sess->nr_poll_queues) {
+		/* dedicated queue for poll */
+		set->map[HCTX_TYPE_POLL].nr_queues = sess->nr_poll_queues;
+		set->map[HCTX_TYPE_POLL].queue_offset = set->map[HCTX_TYPE_READ].queue_offset +
+			set->map[HCTX_TYPE_READ].nr_queues;
+		blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
+		pr_info("[session=%s] mapped %d/%d/%d default/read/poll queues.\n",
+			sess->sessname,
+			set->map[HCTX_TYPE_DEFAULT].nr_queues,
+			set->map[HCTX_TYPE_READ].nr_queues,
+			set->map[HCTX_TYPE_POLL].nr_queues);
+	} else {
+		pr_info("[session=%s] mapped %d/%d default/read queues.\n",
+			sess->sessname,
+			set->map[HCTX_TYPE_DEFAULT].nr_queues,
+			set->map[HCTX_TYPE_READ].nr_queues);
+	}
+
+	return 0;
+}
+
 static struct blk_mq_ops rnbd_mq_ops = {
 	.queue_rq	= rnbd_queue_rq,
 	.complete	= rnbd_softirq_done_fn,
+	.map_queues     = rnbd_rdma_map_queues,
+	.poll           = rnbd_rdma_poll,
 };
 
 static int setup_mq_tags(struct rnbd_clt_session *sess)
@@ -1181,7 +1226,15 @@ static int setup_mq_tags(struct rnbd_clt_session *sess)
 	tag_set->flags		= BLK_MQ_F_SHOULD_MERGE |
 				  BLK_MQ_F_TAG_QUEUE_SHARED;
 	tag_set->cmd_size	= sizeof(struct rnbd_iu) + RNBD_RDMA_SGL_SIZE;
-	tag_set->nr_hw_queues	= num_online_cpus();
+
+	/* for HCTX_TYPE_DEFAULT, HCTX_TYPE_READ, HCTX_TYPE_POLL */
+	tag_set->nr_maps        = sess->nr_poll_queues ? HCTX_MAX_TYPES : 2;
+	/*
+	 * HCTX_TYPE_DEFAULT and HCTX_TYPE_READ share one set of queues
+	 * others are for HCTX_TYPE_POLL
+	 */
+	tag_set->nr_hw_queues	= num_online_cpus() + sess->nr_poll_queues;
+	tag_set->driver_data    = sess;
 
 	return blk_mq_alloc_tag_set(tag_set);
 }
@@ -1189,7 +1242,7 @@ static int setup_mq_tags(struct rnbd_clt_session *sess)
 static struct rnbd_clt_session *
 find_and_get_or_create_sess(const char *sessname,
 			    const struct rtrs_addr *paths,
-			    size_t path_cnt, u16 port_nr)
+			    size_t path_cnt, u16 port_nr, u32 nr_poll_queues)
 {
 	struct rnbd_clt_session *sess;
 	struct rtrs_attrs attrs;
@@ -1198,6 +1251,17 @@ find_and_get_or_create_sess(const char *sessname,
 	struct rtrs_clt_ops rtrs_ops;
 
 	sess = find_or_create_sess(sessname, &first);
+	if (sess == ERR_PTR(-ENOMEM))
+		return ERR_PTR(-ENOMEM);
+	else if ((nr_poll_queues && !first) ||  (!nr_poll_queues && sess->nr_poll_queues)) {
+		/*
+		 * A device MUST have its own session to use the polling-mode.
+		 * It must fail to map new device with the same session.
+		 */
+		err = -EINVAL;
+		goto put_sess;
+	}
+
 	if (!first)
 		return sess;
 
@@ -1219,7 +1283,7 @@ find_and_get_or_create_sess(const char *sessname,
 				   0, /* Do not use pdu of rtrs */
 				   RECONNECT_DELAY, BMAX_SEGMENTS,
 				   BLK_MAX_SEGMENT_SIZE,
-				   MAX_RECONNECTS);
+				   MAX_RECONNECTS, nr_poll_queues);
 	if (IS_ERR(sess->rtrs)) {
 		err = PTR_ERR(sess->rtrs);
 		goto wake_up_and_put;
@@ -1227,6 +1291,7 @@ find_and_get_or_create_sess(const char *sessname,
 	rtrs_clt_query(sess->rtrs, &attrs);
 	sess->max_io_size = attrs.max_io_size;
 	sess->queue_depth = attrs.queue_depth;
+	sess->nr_poll_queues = nr_poll_queues;
 
 	err = setup_mq_tags(sess);
 	if (err)
@@ -1370,7 +1435,8 @@ static int rnbd_client_setup_device(struct rnbd_clt_dev *dev)
 
 static struct rnbd_clt_dev *init_dev(struct rnbd_clt_session *sess,
 				      enum rnbd_access_mode access_mode,
-				      const char *pathname)
+				      const char *pathname,
+				      u32 nr_poll_queues)
 {
 	struct rnbd_clt_dev *dev;
 	int ret;
@@ -1379,7 +1445,8 @@ static struct rnbd_clt_dev *init_dev(struct rnbd_clt_session *sess,
 	if (!dev)
 		return ERR_PTR(-ENOMEM);
 
-	dev->hw_queues = kcalloc(nr_cpu_ids, sizeof(*dev->hw_queues),
+	dev->hw_queues = kcalloc(nr_cpu_ids /* softirq */ + nr_poll_queues /* poll */,
+				 sizeof(*dev->hw_queues),
 				 GFP_KERNEL);
 	if (!dev->hw_queues) {
 		ret = -ENOMEM;
@@ -1405,6 +1472,7 @@ static struct rnbd_clt_dev *init_dev(struct rnbd_clt_session *sess,
 	dev->clt_device_id	= ret;
 	dev->sess		= sess;
 	dev->access_mode	= access_mode;
+	dev->nr_poll_queues	= nr_poll_queues;
 	mutex_init(&dev->lock);
 	refcount_set(&dev->refcount, 1);
 	dev->dev_state = DEV_STATE_INIT;
@@ -1491,7 +1559,8 @@ struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
 					   struct rtrs_addr *paths,
 					   size_t path_cnt, u16 port_nr,
 					   const char *pathname,
-					   enum rnbd_access_mode access_mode)
+					   enum rnbd_access_mode access_mode,
+					   u32 nr_poll_queues)
 {
 	struct rnbd_clt_session *sess;
 	struct rnbd_clt_dev *dev;
@@ -1500,11 +1569,11 @@ struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
 	if (unlikely(exists_devpath(pathname, sessname)))
 		return ERR_PTR(-EEXIST);
 
-	sess = find_and_get_or_create_sess(sessname, paths, path_cnt, port_nr);
+	sess = find_and_get_or_create_sess(sessname, paths, path_cnt, port_nr, nr_poll_queues);
 	if (IS_ERR(sess))
 		return ERR_CAST(sess);
 
-	dev = init_dev(sess, access_mode, pathname);
+	dev = init_dev(sess, access_mode, pathname, nr_poll_queues);
 	if (IS_ERR(dev)) {
 		pr_err("map_device: failed to map device '%s' from session %s, can't initialize device, err: %ld\n",
 		       pathname, sess->sessname, PTR_ERR(dev));
diff --git a/drivers/block/rnbd/rnbd-clt.h b/drivers/block/rnbd/rnbd-clt.h
index 714d426b449b..451e7383738f 100644
--- a/drivers/block/rnbd/rnbd-clt.h
+++ b/drivers/block/rnbd/rnbd-clt.h
@@ -90,6 +90,7 @@ struct rnbd_clt_session {
 	int			queue_depth;
 	u32			max_io_size;
 	struct blk_mq_tag_set	tag_set;
+	u32			nr_poll_queues;
 	struct mutex		lock; /* protects state and devs_list */
 	struct list_head        devs_list; /* list of struct rnbd_clt_dev */
 	refcount_t		refcount;
@@ -118,6 +119,7 @@ struct rnbd_clt_dev {
 	enum rnbd_clt_dev_state	dev_state;
 	char			*pathname;
 	enum rnbd_access_mode	access_mode;
+	u32			nr_poll_queues;
 	bool			read_only;
 	bool			rotational;
 	bool			wc;
@@ -147,7 +149,8 @@ struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
 					   struct rtrs_addr *paths,
 					   size_t path_cnt, u16 port_nr,
 					   const char *pathname,
-					   enum rnbd_access_mode access_mode);
+					   enum rnbd_access_mode access_mode,
+					   u32 nr_poll_queues);
 int rnbd_clt_unmap_device(struct rnbd_clt_dev *dev, bool force,
 			   const struct attribute *sysfs_self);
 
diff --git a/drivers/infiniband/ulp/rtrs/rtrs-clt.c b/drivers/infiniband/ulp/rtrs/rtrs-clt.c
index 7efd49bdc78c..467d135a82cf 100644
--- a/drivers/infiniband/ulp/rtrs/rtrs-clt.c
+++ b/drivers/infiniband/ulp/rtrs/rtrs-clt.c
@@ -174,7 +174,7 @@ struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_sess *sess,
 	int id = 0;
 
 	if (likely(permit->con_type == RTRS_IO_CON))
-		id = (permit->cpu_id % (sess->s.con_num - 1)) + 1;
+		id = (permit->cpu_id % (sess->s.irq_con_num - 1)) + 1;
 
 	return to_clt_con(sess->s.con[id]);
 }
@@ -1400,23 +1400,29 @@ static void rtrs_clt_close_work(struct work_struct *work);
 static struct rtrs_clt_sess *alloc_sess(struct rtrs_clt *clt,
 					 const struct rtrs_addr *path,
 					 size_t con_num, u16 max_segments,
-					 size_t max_segment_size)
+					 size_t max_segment_size, u32 nr_poll_queues)
 {
 	struct rtrs_clt_sess *sess;
 	int err = -ENOMEM;
 	int cpu;
+	size_t total_con;
 
 	sess = kzalloc(sizeof(*sess), GFP_KERNEL);
 	if (!sess)
 		goto err;
 
-	/* Extra connection for user messages */
-	con_num += 1;
-
-	sess->s.con = kcalloc(con_num, sizeof(*sess->s.con), GFP_KERNEL);
+	/*
+	 * irqmode and poll
+	 * +1: Extra connection for user messages
+	 */
+	total_con = con_num + nr_poll_queues + 1;
+	sess->s.con = kcalloc(total_con, sizeof(*sess->s.con), GFP_KERNEL);
 	if (!sess->s.con)
 		goto err_free_sess;
 
+	sess->s.con_num = total_con;
+	sess->s.irq_con_num = con_num + 1;
+
 	sess->stats = kzalloc(sizeof(*sess->stats), GFP_KERNEL);
 	if (!sess->stats)
 		goto err_free_con;
@@ -1435,7 +1441,6 @@ static struct rtrs_clt_sess *alloc_sess(struct rtrs_clt *clt,
 		memcpy(&sess->s.src_addr, path->src,
 		       rdma_addr_size((struct sockaddr *)path->src));
 	strlcpy(sess->s.sessname, clt->sessname, sizeof(sess->s.sessname));
-	sess->s.con_num = con_num;
 	sess->clt = clt;
 	sess->max_pages_per_mr = max_segments * max_segment_size >> 12;
 	init_waitqueue_head(&sess->state_wq);
@@ -1576,9 +1581,14 @@ static int create_con_cq_qp(struct rtrs_clt_con *con)
 	}
 	cq_size = max_send_wr + max_recv_wr;
 	cq_vector = con->cpu % sess->s.dev->ib_dev->num_comp_vectors;
-	err = rtrs_cq_qp_create(&sess->s, &con->c, sess->max_send_sge,
-				 cq_vector, cq_size, max_send_wr,
-				 max_recv_wr, IB_POLL_SOFTIRQ);
+	if (con->c.cid >= sess->s.irq_con_num)
+		err = rtrs_cq_qp_create(&sess->s, &con->c, sess->max_send_sge,
+					cq_vector, cq_size, max_send_wr,
+					max_recv_wr, IB_POLL_DIRECT);
+	else
+		err = rtrs_cq_qp_create(&sess->s, &con->c, sess->max_send_sge,
+					cq_vector, cq_size, max_send_wr,
+					max_recv_wr, IB_POLL_SOFTIRQ);
 	/*
 	 * In case of error we do not bother to clean previous allocations,
 	 * since destroy_con_cq_qp() must be called.
@@ -2631,6 +2641,7 @@ static void free_clt(struct rtrs_clt *clt)
  * @max_segment_size: Max. size of one segment
  * @max_reconnect_attempts: Number of times to reconnect on error before giving
  *			    up, 0 for * disabled, -1 for forever
+ * @nr_poll_queues: number of polling mode connection using IB_POLL_DIRECT flag
  *
  * Starts session establishment with the rtrs_server. The function can block
  * up to ~2000ms before it returns.
@@ -2644,7 +2655,7 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
 				 size_t pdu_sz, u8 reconnect_delay_sec,
 				 u16 max_segments,
 				 size_t max_segment_size,
-				 s16 max_reconnect_attempts)
+				 s16 max_reconnect_attempts, u32 nr_poll_queues)
 {
 	struct rtrs_clt_sess *sess, *tmp;
 	struct rtrs_clt *clt;
@@ -2662,7 +2673,7 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
 		struct rtrs_clt_sess *sess;
 
 		sess = alloc_sess(clt, &paths[i], nr_cpu_ids,
-				  max_segments, max_segment_size);
+				  max_segments, max_segment_size, nr_poll_queues);
 		if (IS_ERR(sess)) {
 			err = PTR_ERR(sess);
 			goto close_all_sess;
@@ -2887,6 +2898,31 @@ int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
 }
 EXPORT_SYMBOL(rtrs_clt_request);
 
+int rtrs_clt_rdma_cq_direct(struct rtrs_clt *clt, unsigned int index)
+{
+	int cnt;
+	struct rtrs_con *con;
+	struct rtrs_clt_sess *sess;
+	struct path_it it;
+
+	rcu_read_lock();
+	for (path_it_init(&it, clt);
+	     (sess = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
+		if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED))
+			continue;
+
+		con = sess->s.con[index + 1];
+		cnt = ib_process_cq_direct(con->cq, -1);
+		if (likely(cnt))
+			break;
+	}
+	path_it_deinit(&it);
+	rcu_read_unlock();
+
+	return cnt;
+}
+EXPORT_SYMBOL(rtrs_clt_rdma_cq_direct);
+
 /**
  * rtrs_clt_query() - queries RTRS session attributes
  *@clt: session pointer
@@ -2916,7 +2952,7 @@ int rtrs_clt_create_path_from_sysfs(struct rtrs_clt *clt,
 	int err;
 
 	sess = alloc_sess(clt, addr, nr_cpu_ids, clt->max_segments,
-			  clt->max_segment_size);
+			  clt->max_segment_size, 0);
 	if (IS_ERR(sess))
 		return PTR_ERR(sess);
 
diff --git a/drivers/infiniband/ulp/rtrs/rtrs-pri.h b/drivers/infiniband/ulp/rtrs/rtrs-pri.h
index 8caad0a2322b..00eb45053339 100644
--- a/drivers/infiniband/ulp/rtrs/rtrs-pri.h
+++ b/drivers/infiniband/ulp/rtrs/rtrs-pri.h
@@ -101,6 +101,7 @@ struct rtrs_sess {
 	uuid_t			uuid;
 	struct rtrs_con	**con;
 	unsigned int		con_num;
+	unsigned int		irq_con_num;
 	unsigned int		recon_cnt;
 	struct rtrs_ib_dev	*dev;
 	int			dev_ref;
diff --git a/drivers/infiniband/ulp/rtrs/rtrs.h b/drivers/infiniband/ulp/rtrs/rtrs.h
index 2db1b5eb3ab0..f891fbe7abe6 100644
--- a/drivers/infiniband/ulp/rtrs/rtrs.h
+++ b/drivers/infiniband/ulp/rtrs/rtrs.h
@@ -59,7 +59,7 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
 				 size_t pdu_sz, u8 reconnect_delay_sec,
 				 u16 max_segments,
 				 size_t max_segment_size,
-				 s16 max_reconnect_attempts);
+				 s16 max_reconnect_attempts, u32 nr_poll_queues);
 
 void rtrs_clt_close(struct rtrs_clt *sess);
 
@@ -103,6 +103,7 @@ int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
 		     struct rtrs_clt *sess, struct rtrs_permit *permit,
 		     const struct kvec *vec, size_t nr, size_t len,
 		     struct scatterlist *sg, unsigned int sg_cnt);
+int rtrs_clt_rdma_cq_direct(struct rtrs_clt *clt, unsigned int index);
 
 /**
  * rtrs_attrs - RTRS session attributes
-- 
2.25.1


^ permalink raw reply	[flat|nested] 10+ messages in thread

* [PATCHv4 for-next 15/19] block/rnbd-srv: Remove unused arguments of rnbd_srv_rdma_ev
       [not found] <20210414122402.203388-1-gi-oh.kim@ionos.com>
  2021-04-14 12:23 ` [PATCHv4 for-next 08/19] block/rnbd-clt: Replace {NO_WAIT,WAIT} with RTRS_PERMIT_{WAIT,NOWAIT} Gioh Kim
  2021-04-14 12:23 ` [PATCHv4 for-next 13/19] block/rnbd-clt: Support polling mode for IO latency optimization Gioh Kim
@ 2021-04-14 12:23 ` Gioh Kim
  2021-04-14 12:24 ` [PATCHv4 for-next 17/19] block/rnbd-clt: Remove max_segment_size Gioh Kim
  3 siblings, 0 replies; 10+ messages in thread
From: Gioh Kim @ 2021-04-14 12:23 UTC (permalink / raw)
  To: linux-block
  Cc: axboe, hch, sagi, bvanassche, haris.iqbal, jinpu.wang, Gioh Kim,
	Leon Romanovsky, linux-rdma, Aleksei Marov, Gioh Kim,
	Chaitanya Kulkarni, Jason Gunthorpe

From: Gioh Kim <gi-oh.kim@cloud.ionos.com>

struct rtrs_srv is not used when handling rnbd_srv_rdma_ev messages, so
cleaned up
rdma_ev function pointer in rtrs_srv_ops also is changed.

Cc: Leon Romanovsky <leonro@nvidia.com>
Cc: linux-rdma@vger.kernel.org
Signed-off-by: Aleksei Marov <aleksei.marov@ionos.com>
Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
Signed-off-by: Gioh Kim <gi-oh.kim@ionos.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Acked-by: Jason Gunthorpe <jgg@nvidia.com>
---
 drivers/block/rnbd/rnbd-srv.c          | 39 ++++++++++----------------
 drivers/infiniband/ulp/rtrs/rtrs-srv.c |  4 +--
 drivers/infiniband/ulp/rtrs/rtrs.h     |  3 +-
 3 files changed, 18 insertions(+), 28 deletions(-)

diff --git a/drivers/block/rnbd/rnbd-srv.c b/drivers/block/rnbd/rnbd-srv.c
index a9bb414f7442..abacd9ef10d6 100644
--- a/drivers/block/rnbd/rnbd-srv.c
+++ b/drivers/block/rnbd/rnbd-srv.c
@@ -114,8 +114,7 @@ rnbd_get_sess_dev(int dev_id, struct rnbd_srv_session *srv_sess)
 	return sess_dev;
 }
 
-static int process_rdma(struct rtrs_srv *sess,
-			struct rnbd_srv_session *srv_sess,
+static int process_rdma(struct rnbd_srv_session *srv_sess,
 			struct rtrs_srv_op *id, void *data, u32 datalen,
 			const void *usr, size_t usrlen)
 {
@@ -344,8 +343,7 @@ void rnbd_srv_sess_dev_force_close(struct rnbd_srv_sess_dev *sess_dev,
 	mutex_unlock(&sess->lock);
 }
 
-static int process_msg_close(struct rtrs_srv *rtrs,
-			     struct rnbd_srv_session *srv_sess,
+static int process_msg_close(struct rnbd_srv_session *srv_sess,
 			     void *data, size_t datalen, const void *usr,
 			     size_t usrlen)
 {
@@ -364,20 +362,18 @@ static int process_msg_close(struct rtrs_srv *rtrs,
 	return 0;
 }
 
-static int process_msg_open(struct rtrs_srv *rtrs,
-			    struct rnbd_srv_session *srv_sess,
+static int process_msg_open(struct rnbd_srv_session *srv_sess,
 			    const void *msg, size_t len,
 			    void *data, size_t datalen);
 
-static int process_msg_sess_info(struct rtrs_srv *rtrs,
-				 struct rnbd_srv_session *srv_sess,
+static int process_msg_sess_info(struct rnbd_srv_session *srv_sess,
 				 const void *msg, size_t len,
 				 void *data, size_t datalen);
 
-static int rnbd_srv_rdma_ev(struct rtrs_srv *rtrs, void *priv,
-			     struct rtrs_srv_op *id, int dir,
-			     void *data, size_t datalen, const void *usr,
-			     size_t usrlen)
+static int rnbd_srv_rdma_ev(void *priv,
+			    struct rtrs_srv_op *id, int dir,
+			    void *data, size_t datalen, const void *usr,
+			    size_t usrlen)
 {
 	struct rnbd_srv_session *srv_sess = priv;
 	const struct rnbd_msg_hdr *hdr = usr;
@@ -391,19 +387,16 @@ static int rnbd_srv_rdma_ev(struct rtrs_srv *rtrs, void *priv,
 
 	switch (type) {
 	case RNBD_MSG_IO:
-		return process_rdma(rtrs, srv_sess, id, data, datalen, usr,
-				    usrlen);
+		return process_rdma(srv_sess, id, data, datalen, usr, usrlen);
 	case RNBD_MSG_CLOSE:
-		ret = process_msg_close(rtrs, srv_sess, data, datalen,
-					usr, usrlen);
+		ret = process_msg_close(srv_sess, data, datalen, usr, usrlen);
 		break;
 	case RNBD_MSG_OPEN:
-		ret = process_msg_open(rtrs, srv_sess, usr, usrlen,
-				       data, datalen);
+		ret = process_msg_open(srv_sess, usr, usrlen, data, datalen);
 		break;
 	case RNBD_MSG_SESS_INFO:
-		ret = process_msg_sess_info(rtrs, srv_sess, usr, usrlen,
-					    data, datalen);
+		ret = process_msg_sess_info(srv_sess, usr, usrlen, data,
+					    datalen);
 		break;
 	default:
 		pr_warn("Received unexpected message type %d with dir %d from session %s\n",
@@ -656,8 +649,7 @@ static char *rnbd_srv_get_full_path(struct rnbd_srv_session *srv_sess,
 	return full_path;
 }
 
-static int process_msg_sess_info(struct rtrs_srv *rtrs,
-				 struct rnbd_srv_session *srv_sess,
+static int process_msg_sess_info(struct rnbd_srv_session *srv_sess,
 				 const void *msg, size_t len,
 				 void *data, size_t datalen)
 {
@@ -698,8 +690,7 @@ find_srv_sess_dev(struct rnbd_srv_session *srv_sess, const char *dev_name)
 	return NULL;
 }
 
-static int process_msg_open(struct rtrs_srv *rtrs,
-			    struct rnbd_srv_session *srv_sess,
+static int process_msg_open(struct rnbd_srv_session *srv_sess,
 			    const void *msg, size_t len,
 			    void *data, size_t datalen)
 {
diff --git a/drivers/infiniband/ulp/rtrs/rtrs-srv.c b/drivers/infiniband/ulp/rtrs/rtrs-srv.c
index d071809e3ed2..f7aa2a7e7442 100644
--- a/drivers/infiniband/ulp/rtrs/rtrs-srv.c
+++ b/drivers/infiniband/ulp/rtrs/rtrs-srv.c
@@ -998,7 +998,7 @@ static void process_read(struct rtrs_srv_con *con,
 	usr_len = le16_to_cpu(msg->usr_len);
 	data_len = off - usr_len;
 	data = page_address(srv->chunks[buf_id]);
-	ret = ctx->ops.rdma_ev(srv, srv->priv, id, READ, data, data_len,
+	ret = ctx->ops.rdma_ev(srv->priv, id, READ, data, data_len,
 			   data + data_len, usr_len);
 
 	if (unlikely(ret)) {
@@ -1051,7 +1051,7 @@ static void process_write(struct rtrs_srv_con *con,
 	usr_len = le16_to_cpu(req->usr_len);
 	data_len = off - usr_len;
 	data = page_address(srv->chunks[buf_id]);
-	ret = ctx->ops.rdma_ev(srv, srv->priv, id, WRITE, data, data_len,
+	ret = ctx->ops.rdma_ev(srv->priv, id, WRITE, data, data_len,
 			   data + data_len, usr_len);
 	if (unlikely(ret)) {
 		rtrs_err_rl(s,
diff --git a/drivers/infiniband/ulp/rtrs/rtrs.h b/drivers/infiniband/ulp/rtrs/rtrs.h
index f891fbe7abe6..b0f56ffeff88 100644
--- a/drivers/infiniband/ulp/rtrs/rtrs.h
+++ b/drivers/infiniband/ulp/rtrs/rtrs.h
@@ -139,7 +139,6 @@ struct rtrs_srv_ops {
 	 *			message for the data transfer will be sent to
 	 *			the client.
 
-	 *	@sess:		Session
 	 *	@priv:		Private data set by rtrs_srv_set_sess_priv()
 	 *	@id:		internal RTRS operation id
 	 *	@dir:		READ/WRITE
@@ -153,7 +152,7 @@ struct rtrs_srv_ops {
 	 *	@usr:		The extra user message sent by the client (%vec)
 	 *	@usrlen:	Size of the user message
 	 */
-	int (*rdma_ev)(struct rtrs_srv *sess, void *priv,
+	int (*rdma_ev)(void *priv,
 		       struct rtrs_srv_op *id, int dir,
 		       void *data, size_t datalen, const void *usr,
 		       size_t usrlen);
-- 
2.25.1


^ permalink raw reply	[flat|nested] 10+ messages in thread

* [PATCHv4 for-next 17/19] block/rnbd-clt: Remove max_segment_size
       [not found] <20210414122402.203388-1-gi-oh.kim@ionos.com>
                   ` (2 preceding siblings ...)
  2021-04-14 12:23 ` [PATCHv4 for-next 15/19] block/rnbd-srv: Remove unused arguments of rnbd_srv_rdma_ev Gioh Kim
@ 2021-04-14 12:24 ` Gioh Kim
  3 siblings, 0 replies; 10+ messages in thread
From: Gioh Kim @ 2021-04-14 12:24 UTC (permalink / raw)
  To: linux-block
  Cc: axboe, hch, sagi, bvanassche, haris.iqbal, jinpu.wang, Jack Wang,
	Leon Romanovsky, linux-rdma, Gioh Kim, Jason Gunthorpe

From: Jack Wang <jinpu.wang@cloud.ionos.com>

We always map with SZ_4K, so do not need max_segment_size.

Cc: Leon Romanovsky <leonro@nvidia.com>
Cc: linux-rdma@vger.kernel.org
Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
Reviewed-by: Md Haris Iqbal <haris.iqbal@ionos.com>
Signed-off-by: Gioh Kim <gi-oh.kim@ionos.com>
Acked-by: Jason Gunthorpe <jgg@nvidia.com>
---
 drivers/block/rnbd/rnbd-clt.c          |  1 -
 drivers/infiniband/ulp/rtrs/rtrs-clt.c | 15 +++++----------
 drivers/infiniband/ulp/rtrs/rtrs-clt.h |  1 -
 drivers/infiniband/ulp/rtrs/rtrs.h     |  1 -
 4 files changed, 5 insertions(+), 13 deletions(-)

diff --git a/drivers/block/rnbd/rnbd-clt.c b/drivers/block/rnbd/rnbd-clt.c
index 1fe010ed6f69..7446660eb7f2 100644
--- a/drivers/block/rnbd/rnbd-clt.c
+++ b/drivers/block/rnbd/rnbd-clt.c
@@ -1289,7 +1289,6 @@ find_and_get_or_create_sess(const char *sessname,
 				   paths, path_cnt, port_nr,
 				   0, /* Do not use pdu of rtrs */
 				   RECONNECT_DELAY, BMAX_SEGMENTS,
-				   BLK_MAX_SEGMENT_SIZE,
 				   MAX_RECONNECTS, nr_poll_queues);
 	if (IS_ERR(sess->rtrs)) {
 		err = PTR_ERR(sess->rtrs);
diff --git a/drivers/infiniband/ulp/rtrs/rtrs-clt.c b/drivers/infiniband/ulp/rtrs/rtrs-clt.c
index 467d135a82cf..1603e0c399e8 100644
--- a/drivers/infiniband/ulp/rtrs/rtrs-clt.c
+++ b/drivers/infiniband/ulp/rtrs/rtrs-clt.c
@@ -1400,7 +1400,7 @@ static void rtrs_clt_close_work(struct work_struct *work);
 static struct rtrs_clt_sess *alloc_sess(struct rtrs_clt *clt,
 					 const struct rtrs_addr *path,
 					 size_t con_num, u16 max_segments,
-					 size_t max_segment_size, u32 nr_poll_queues)
+					 u32 nr_poll_queues)
 {
 	struct rtrs_clt_sess *sess;
 	int err = -ENOMEM;
@@ -1442,7 +1442,7 @@ static struct rtrs_clt_sess *alloc_sess(struct rtrs_clt *clt,
 		       rdma_addr_size((struct sockaddr *)path->src));
 	strlcpy(sess->s.sessname, clt->sessname, sizeof(sess->s.sessname));
 	sess->clt = clt;
-	sess->max_pages_per_mr = max_segments * max_segment_size >> 12;
+	sess->max_pages_per_mr = max_segments;
 	init_waitqueue_head(&sess->state_wq);
 	sess->state = RTRS_CLT_CONNECTING;
 	atomic_set(&sess->connected_cnt, 0);
@@ -2538,7 +2538,6 @@ static struct rtrs_clt *alloc_clt(const char *sessname, size_t paths_num,
 				  void	(*link_ev)(void *priv,
 						   enum rtrs_clt_link_ev ev),
 				  unsigned int max_segments,
-				  size_t max_segment_size,
 				  unsigned int reconnect_delay_sec,
 				  unsigned int max_reconnect_attempts)
 {
@@ -2568,7 +2567,6 @@ static struct rtrs_clt *alloc_clt(const char *sessname, size_t paths_num,
 	clt->port = port;
 	clt->pdu_sz = pdu_sz;
 	clt->max_segments = max_segments;
-	clt->max_segment_size = max_segment_size;
 	clt->reconnect_delay_sec = reconnect_delay_sec;
 	clt->max_reconnect_attempts = max_reconnect_attempts;
 	clt->priv = priv;
@@ -2638,7 +2636,6 @@ static void free_clt(struct rtrs_clt *clt)
  * @pdu_sz: Size of extra payload which can be accessed after permit allocation.
  * @reconnect_delay_sec: time between reconnect tries
  * @max_segments: Max. number of segments per IO request
- * @max_segment_size: Max. size of one segment
  * @max_reconnect_attempts: Number of times to reconnect on error before giving
  *			    up, 0 for * disabled, -1 for forever
  * @nr_poll_queues: number of polling mode connection using IB_POLL_DIRECT flag
@@ -2654,7 +2651,6 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
 				 size_t paths_num, u16 port,
 				 size_t pdu_sz, u8 reconnect_delay_sec,
 				 u16 max_segments,
-				 size_t max_segment_size,
 				 s16 max_reconnect_attempts, u32 nr_poll_queues)
 {
 	struct rtrs_clt_sess *sess, *tmp;
@@ -2663,7 +2659,7 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
 
 	clt = alloc_clt(sessname, paths_num, port, pdu_sz, ops->priv,
 			ops->link_ev,
-			max_segments, max_segment_size, reconnect_delay_sec,
+			max_segments, reconnect_delay_sec,
 			max_reconnect_attempts);
 	if (IS_ERR(clt)) {
 		err = PTR_ERR(clt);
@@ -2673,7 +2669,7 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
 		struct rtrs_clt_sess *sess;
 
 		sess = alloc_sess(clt, &paths[i], nr_cpu_ids,
-				  max_segments, max_segment_size, nr_poll_queues);
+				  max_segments, nr_poll_queues);
 		if (IS_ERR(sess)) {
 			err = PTR_ERR(sess);
 			goto close_all_sess;
@@ -2951,8 +2947,7 @@ int rtrs_clt_create_path_from_sysfs(struct rtrs_clt *clt,
 	struct rtrs_clt_sess *sess;
 	int err;
 
-	sess = alloc_sess(clt, addr, nr_cpu_ids, clt->max_segments,
-			  clt->max_segment_size, 0);
+	sess = alloc_sess(clt, addr, nr_cpu_ids, clt->max_segments, 0);
 	if (IS_ERR(sess))
 		return PTR_ERR(sess);
 
diff --git a/drivers/infiniband/ulp/rtrs/rtrs-clt.h b/drivers/infiniband/ulp/rtrs/rtrs-clt.h
index 692bc83e1f09..98ba5d0a48b8 100644
--- a/drivers/infiniband/ulp/rtrs/rtrs-clt.h
+++ b/drivers/infiniband/ulp/rtrs/rtrs-clt.h
@@ -166,7 +166,6 @@ struct rtrs_clt {
 	unsigned int		max_reconnect_attempts;
 	unsigned int		reconnect_delay_sec;
 	unsigned int		max_segments;
-	size_t			max_segment_size;
 	void			*permits;
 	unsigned long		*permits_map;
 	size_t			queue_depth;
diff --git a/drivers/infiniband/ulp/rtrs/rtrs.h b/drivers/infiniband/ulp/rtrs/rtrs.h
index b0f56ffeff88..bebaa94c4728 100644
--- a/drivers/infiniband/ulp/rtrs/rtrs.h
+++ b/drivers/infiniband/ulp/rtrs/rtrs.h
@@ -58,7 +58,6 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
 				 size_t path_cnt, u16 port,
 				 size_t pdu_sz, u8 reconnect_delay_sec,
 				 u16 max_segments,
-				 size_t max_segment_size,
 				 s16 max_reconnect_attempts, u32 nr_poll_queues);
 
 void rtrs_clt_close(struct rtrs_clt *sess);
-- 
2.25.1


^ permalink raw reply	[flat|nested] 10+ messages in thread

* Re: [PATCHv4 for-next 13/19] block/rnbd-clt: Support polling mode for IO latency optimization
  2021-04-14 12:23 ` [PATCHv4 for-next 13/19] block/rnbd-clt: Support polling mode for IO latency optimization Gioh Kim
@ 2021-04-18  8:36   ` Leon Romanovsky
  2021-04-19  5:12     ` Gioh Kim
  0 siblings, 1 reply; 10+ messages in thread
From: Leon Romanovsky @ 2021-04-18  8:36 UTC (permalink / raw)
  To: Gioh Kim
  Cc: linux-block, axboe, hch, sagi, bvanassche, haris.iqbal,
	jinpu.wang, Gioh Kim, linux-rdma, Jason Gunthorpe

On Wed, Apr 14, 2021 at 02:23:56PM +0200, Gioh Kim wrote:
> From: Gioh Kim <gi-oh.kim@cloud.ionos.com>
> 
> RNBD can make double-queues for irq-mode and poll-mode.
> For example, on 4-CPU system 8 request-queues are created,
> 4 for irq-mode and 4 for poll-mode.
> If the IO has HIPRI flag, the block-layer will call .poll function
> of RNBD. Then IO is sent to the poll-mode queue.
> Add optional nr_poll_queues argument for map_devices interface.
> 
> To support polling of RNBD, RTRS client creates connections
> for both of irq-mode and direct-poll-mode.
> 
> For example, on 4-CPU system it could've create 5 connections:
> con[0] => user message (softirq cq)
> con[1:4] => softirq cq
> 
> After this patch, it can create 9 connections:
> con[0] => user message (softirq cq)
> con[1:4] => softirq cq
> con[5:8] => DIRECT-POLL cq
> 
> Cc: Leon Romanovsky <leonro@nvidia.com>
> Cc: linux-rdma@vger.kernel.org
> Signed-off-by: Gioh Kim <gi-oh.kim@ionos.com>
> Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
> Acked-by: Jason Gunthorpe <jgg@nvidia.com>
> ---
>  drivers/block/rnbd/rnbd-clt-sysfs.c    | 56 +++++++++++++----
>  drivers/block/rnbd/rnbd-clt.c          | 85 +++++++++++++++++++++++---
>  drivers/block/rnbd/rnbd-clt.h          |  5 +-
>  drivers/infiniband/ulp/rtrs/rtrs-clt.c | 62 +++++++++++++++----
>  drivers/infiniband/ulp/rtrs/rtrs-pri.h |  1 +
>  drivers/infiniband/ulp/rtrs/rtrs.h     |  3 +-
>  6 files changed, 178 insertions(+), 34 deletions(-)
> 
> diff --git a/drivers/block/rnbd/rnbd-clt-sysfs.c b/drivers/block/rnbd/rnbd-clt-sysfs.c
> index 49015f428e67..bd111ebceb75 100644
> --- a/drivers/block/rnbd/rnbd-clt-sysfs.c
> +++ b/drivers/block/rnbd/rnbd-clt-sysfs.c
> @@ -34,6 +34,7 @@ enum {
>  	RNBD_OPT_DEV_PATH	= 1 << 2,
>  	RNBD_OPT_ACCESS_MODE	= 1 << 3,
>  	RNBD_OPT_SESSNAME	= 1 << 6,
> +	RNBD_OPT_NR_POLL_QUEUES	= 1 << 7,
>  };
>  
>  static const unsigned int rnbd_opt_mandatory[] = {
> @@ -42,12 +43,13 @@ static const unsigned int rnbd_opt_mandatory[] = {
>  };
>  
>  static const match_table_t rnbd_opt_tokens = {
> -	{RNBD_OPT_PATH,		"path=%s"	},
> -	{RNBD_OPT_DEV_PATH,	"device_path=%s"},
> -	{RNBD_OPT_DEST_PORT,	"dest_port=%d"  },
> -	{RNBD_OPT_ACCESS_MODE,	"access_mode=%s"},
> -	{RNBD_OPT_SESSNAME,	"sessname=%s"	},
> -	{RNBD_OPT_ERR,		NULL		},
> +	{RNBD_OPT_PATH,			"path=%s"		},
> +	{RNBD_OPT_DEV_PATH,		"device_path=%s"	},
> +	{RNBD_OPT_DEST_PORT,		"dest_port=%d"		},
> +	{RNBD_OPT_ACCESS_MODE,		"access_mode=%s"	},
> +	{RNBD_OPT_SESSNAME,		"sessname=%s"		},
> +	{RNBD_OPT_NR_POLL_QUEUES,	"nr_poll_queues=%d"	},
> +	{RNBD_OPT_ERR,			NULL			},
>  };
>  
>  struct rnbd_map_options {
> @@ -57,6 +59,7 @@ struct rnbd_map_options {
>  	char *pathname;
>  	u16 *dest_port;
>  	enum rnbd_access_mode *access_mode;
> +	u32 *nr_poll_queues;
>  };
>  
>  static int rnbd_clt_parse_map_options(const char *buf, size_t max_path_cnt,
> @@ -68,7 +71,7 @@ static int rnbd_clt_parse_map_options(const char *buf, size_t max_path_cnt,
>  	int opt_mask = 0;
>  	int token;
>  	int ret = -EINVAL;
> -	int i, dest_port;
> +	int i, dest_port, nr_poll_queues;
>  	int p_cnt = 0;
>  
>  	options = kstrdup(buf, GFP_KERNEL);
> @@ -178,6 +181,19 @@ static int rnbd_clt_parse_map_options(const char *buf, size_t max_path_cnt,
>  			kfree(p);
>  			break;
>  
> +		case RNBD_OPT_NR_POLL_QUEUES:
> +			if (match_int(args, &nr_poll_queues) || nr_poll_queues < -1 ||
> +			    nr_poll_queues > (int)nr_cpu_ids) {
> +				pr_err("bad nr_poll_queues parameter '%d'\n",
> +				       nr_poll_queues);
> +				ret = -EINVAL;
> +				goto out;
> +			}
> +			if (nr_poll_queues == -1)
> +				nr_poll_queues = nr_cpu_ids;
> +			*opt->nr_poll_queues = nr_poll_queues;
> +			break;
> +
>  		default:
>  			pr_err("map_device: Unknown parameter or missing value '%s'\n",
>  			       p);
> @@ -227,6 +243,20 @@ static ssize_t state_show(struct kobject *kobj,
>  
>  static struct kobj_attribute rnbd_clt_state_attr = __ATTR_RO(state);
>  
> +static ssize_t nr_poll_queues_show(struct kobject *kobj,
> +				   struct kobj_attribute *attr, char *page)
> +{
> +	struct rnbd_clt_dev *dev;
> +
> +	dev = container_of(kobj, struct rnbd_clt_dev, kobj);
> +
> +	return snprintf(page, PAGE_SIZE, "%d\n",
> +			dev->nr_poll_queues);
> +}

Didn't Greg ask you to use sysfs_emit() here?

> +
> +static struct kobj_attribute rnbd_clt_nr_poll_queues =
> +	__ATTR_RO(nr_poll_queues);
> +
>  static ssize_t mapping_path_show(struct kobject *kobj,
>  				 struct kobj_attribute *attr, char *page)
>  {
> @@ -421,6 +451,7 @@ static struct attribute *rnbd_dev_attrs[] = {
>  	&rnbd_clt_state_attr.attr,
>  	&rnbd_clt_session_attr.attr,
>  	&rnbd_clt_access_mode.attr,
> +	&rnbd_clt_nr_poll_queues.attr,
>  	NULL,
>  };
>  
> @@ -469,7 +500,7 @@ static ssize_t rnbd_clt_map_device_show(struct kobject *kobj,
>  					 char *page)
>  {
>  	return scnprintf(page, PAGE_SIZE,
> -			 "Usage: echo \"[dest_port=server port number] sessname=<name of the rtrs session> path=<[srcaddr@]dstaddr> [path=<[srcaddr@]dstaddr>] device_path=<full path on remote side> [access_mode=<ro|rw|migration>]\" > %s\n\naddr ::= [ ip:<ipv4> | ip:<ipv6> | gid:<gid> ]\n",
> +			 "Usage: echo \"[dest_port=server port number] sessname=<name of the rtrs session> path=<[srcaddr@]dstaddr> [path=<[srcaddr@]dstaddr>] device_path=<full path on remote side> [access_mode=<ro|rw|migration>] [nr_poll_queues=<number of queues>]\" > %s\n\naddr ::= [ ip:<ipv4> | ip:<ipv6> | gid:<gid> ]\n",
>  			 attr->attr.name);
>  }
>  
> @@ -541,6 +572,7 @@ static ssize_t rnbd_clt_map_device_store(struct kobject *kobj,
>  	char sessname[NAME_MAX];
>  	enum rnbd_access_mode access_mode = RNBD_ACCESS_RW;
>  	u16 port_nr = RTRS_PORT;
> +	u32 nr_poll_queues = 0;
>  
>  	struct sockaddr_storage *addrs;
>  	struct rtrs_addr paths[6];
> @@ -552,6 +584,7 @@ static ssize_t rnbd_clt_map_device_store(struct kobject *kobj,
>  	opt.pathname = pathname;
>  	opt.dest_port = &port_nr;
>  	opt.access_mode = &access_mode;
> +	opt.nr_poll_queues = &nr_poll_queues;
>  	addrs = kcalloc(ARRAY_SIZE(paths) * 2, sizeof(*addrs), GFP_KERNEL);
>  	if (!addrs)
>  		return -ENOMEM;
> @@ -565,12 +598,13 @@ static ssize_t rnbd_clt_map_device_store(struct kobject *kobj,
>  	if (ret)
>  		goto out;
>  
> -	pr_info("Mapping device %s on session %s, (access_mode: %s)\n",
> +	pr_info("Mapping device %s on session %s, (access_mode: %s, nr_poll_queues: %d)\n",
>  		pathname, sessname,
> -		rnbd_access_mode_str(access_mode));
> +		rnbd_access_mode_str(access_mode),
> +		nr_poll_queues);
>  
>  	dev = rnbd_clt_map_device(sessname, paths, path_cnt, port_nr, pathname,
> -				  access_mode);
> +				  access_mode, nr_poll_queues);
>  	if (IS_ERR(dev)) {
>  		ret = PTR_ERR(dev);
>  		goto out;
> diff --git a/drivers/block/rnbd/rnbd-clt.c b/drivers/block/rnbd/rnbd-clt.c
> index 9b44aac680d5..63719ec04d58 100644
> --- a/drivers/block/rnbd/rnbd-clt.c
> +++ b/drivers/block/rnbd/rnbd-clt.c
> @@ -1165,9 +1165,54 @@ static blk_status_t rnbd_queue_rq(struct blk_mq_hw_ctx *hctx,
>  	return ret;
>  }
>  
> +static int rnbd_rdma_poll(struct blk_mq_hw_ctx *hctx)
> +{
> +	struct rnbd_queue *q = hctx->driver_data;
> +	struct rnbd_clt_dev *dev = q->dev;
> +	int cnt;
> +
> +	cnt = rtrs_clt_rdma_cq_direct(dev->sess->rtrs, hctx->queue_num);
> +	return cnt;
> +}
> +
> +static int rnbd_rdma_map_queues(struct blk_mq_tag_set *set)
> +{
> +	struct rnbd_clt_session *sess = set->driver_data;
> +
> +	/* shared read/write queues */
> +	set->map[HCTX_TYPE_DEFAULT].nr_queues = num_online_cpus();
> +	set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
> +	set->map[HCTX_TYPE_READ].nr_queues = num_online_cpus();
> +	set->map[HCTX_TYPE_READ].queue_offset = 0;
> +	blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
> +	blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
> +
> +	if (sess->nr_poll_queues) {
> +		/* dedicated queue for poll */
> +		set->map[HCTX_TYPE_POLL].nr_queues = sess->nr_poll_queues;
> +		set->map[HCTX_TYPE_POLL].queue_offset = set->map[HCTX_TYPE_READ].queue_offset +
> +			set->map[HCTX_TYPE_READ].nr_queues;
> +		blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
> +		pr_info("[session=%s] mapped %d/%d/%d default/read/poll queues.\n",
> +			sess->sessname,
> +			set->map[HCTX_TYPE_DEFAULT].nr_queues,
> +			set->map[HCTX_TYPE_READ].nr_queues,
> +			set->map[HCTX_TYPE_POLL].nr_queues);
> +	} else {
> +		pr_info("[session=%s] mapped %d/%d default/read queues.\n",
> +			sess->sessname,
> +			set->map[HCTX_TYPE_DEFAULT].nr_queues,
> +			set->map[HCTX_TYPE_READ].nr_queues);
> +	}
> +
> +	return 0;
> +}
> +
>  static struct blk_mq_ops rnbd_mq_ops = {
>  	.queue_rq	= rnbd_queue_rq,
>  	.complete	= rnbd_softirq_done_fn,
> +	.map_queues     = rnbd_rdma_map_queues,
> +	.poll           = rnbd_rdma_poll,
>  };
>  
>  static int setup_mq_tags(struct rnbd_clt_session *sess)
> @@ -1181,7 +1226,15 @@ static int setup_mq_tags(struct rnbd_clt_session *sess)
>  	tag_set->flags		= BLK_MQ_F_SHOULD_MERGE |
>  				  BLK_MQ_F_TAG_QUEUE_SHARED;
>  	tag_set->cmd_size	= sizeof(struct rnbd_iu) + RNBD_RDMA_SGL_SIZE;
> -	tag_set->nr_hw_queues	= num_online_cpus();
> +
> +	/* for HCTX_TYPE_DEFAULT, HCTX_TYPE_READ, HCTX_TYPE_POLL */
> +	tag_set->nr_maps        = sess->nr_poll_queues ? HCTX_MAX_TYPES : 2;
> +	/*
> +	 * HCTX_TYPE_DEFAULT and HCTX_TYPE_READ share one set of queues
> +	 * others are for HCTX_TYPE_POLL
> +	 */
> +	tag_set->nr_hw_queues	= num_online_cpus() + sess->nr_poll_queues;
> +	tag_set->driver_data    = sess;
>  
>  	return blk_mq_alloc_tag_set(tag_set);
>  }
> @@ -1189,7 +1242,7 @@ static int setup_mq_tags(struct rnbd_clt_session *sess)
>  static struct rnbd_clt_session *
>  find_and_get_or_create_sess(const char *sessname,
>  			    const struct rtrs_addr *paths,
> -			    size_t path_cnt, u16 port_nr)
> +			    size_t path_cnt, u16 port_nr, u32 nr_poll_queues)
>  {
>  	struct rnbd_clt_session *sess;
>  	struct rtrs_attrs attrs;
> @@ -1198,6 +1251,17 @@ find_and_get_or_create_sess(const char *sessname,
>  	struct rtrs_clt_ops rtrs_ops;
>  
>  	sess = find_or_create_sess(sessname, &first);
> +	if (sess == ERR_PTR(-ENOMEM))
> +		return ERR_PTR(-ENOMEM);
> +	else if ((nr_poll_queues && !first) ||  (!nr_poll_queues && sess->nr_poll_queues)) {
> +		/*
> +		 * A device MUST have its own session to use the polling-mode.
> +		 * It must fail to map new device with the same session.
> +		 */
> +		err = -EINVAL;
> +		goto put_sess;
> +	}
> +
>  	if (!first)
>  		return sess;
>  
> @@ -1219,7 +1283,7 @@ find_and_get_or_create_sess(const char *sessname,
>  				   0, /* Do not use pdu of rtrs */
>  				   RECONNECT_DELAY, BMAX_SEGMENTS,
>  				   BLK_MAX_SEGMENT_SIZE,
> -				   MAX_RECONNECTS);
> +				   MAX_RECONNECTS, nr_poll_queues);
>  	if (IS_ERR(sess->rtrs)) {
>  		err = PTR_ERR(sess->rtrs);
>  		goto wake_up_and_put;
> @@ -1227,6 +1291,7 @@ find_and_get_or_create_sess(const char *sessname,
>  	rtrs_clt_query(sess->rtrs, &attrs);
>  	sess->max_io_size = attrs.max_io_size;
>  	sess->queue_depth = attrs.queue_depth;
> +	sess->nr_poll_queues = nr_poll_queues;
>  
>  	err = setup_mq_tags(sess);
>  	if (err)
> @@ -1370,7 +1435,8 @@ static int rnbd_client_setup_device(struct rnbd_clt_dev *dev)
>  
>  static struct rnbd_clt_dev *init_dev(struct rnbd_clt_session *sess,
>  				      enum rnbd_access_mode access_mode,
> -				      const char *pathname)
> +				      const char *pathname,
> +				      u32 nr_poll_queues)
>  {
>  	struct rnbd_clt_dev *dev;
>  	int ret;
> @@ -1379,7 +1445,8 @@ static struct rnbd_clt_dev *init_dev(struct rnbd_clt_session *sess,
>  	if (!dev)
>  		return ERR_PTR(-ENOMEM);
>  
> -	dev->hw_queues = kcalloc(nr_cpu_ids, sizeof(*dev->hw_queues),
> +	dev->hw_queues = kcalloc(nr_cpu_ids /* softirq */ + nr_poll_queues /* poll */,

Please don't add comments in the middle of function call.

> +				 sizeof(*dev->hw_queues),
>  				 GFP_KERNEL);
>  	if (!dev->hw_queues) {
>  		ret = -ENOMEM;
> @@ -1405,6 +1472,7 @@ static struct rnbd_clt_dev *init_dev(struct rnbd_clt_session *sess,
>  	dev->clt_device_id	= ret;
>  	dev->sess		= sess;
>  	dev->access_mode	= access_mode;
> +	dev->nr_poll_queues	= nr_poll_queues;
>  	mutex_init(&dev->lock);
>  	refcount_set(&dev->refcount, 1);
>  	dev->dev_state = DEV_STATE_INIT;
> @@ -1491,7 +1559,8 @@ struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
>  					   struct rtrs_addr *paths,
>  					   size_t path_cnt, u16 port_nr,
>  					   const char *pathname,
> -					   enum rnbd_access_mode access_mode)
> +					   enum rnbd_access_mode access_mode,
> +					   u32 nr_poll_queues)
>  {
>  	struct rnbd_clt_session *sess;
>  	struct rnbd_clt_dev *dev;
> @@ -1500,11 +1569,11 @@ struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
>  	if (unlikely(exists_devpath(pathname, sessname)))
>  		return ERR_PTR(-EEXIST);
>  
> -	sess = find_and_get_or_create_sess(sessname, paths, path_cnt, port_nr);
> +	sess = find_and_get_or_create_sess(sessname, paths, path_cnt, port_nr, nr_poll_queues);
>  	if (IS_ERR(sess))
>  		return ERR_CAST(sess);
>  
> -	dev = init_dev(sess, access_mode, pathname);
> +	dev = init_dev(sess, access_mode, pathname, nr_poll_queues);
>  	if (IS_ERR(dev)) {
>  		pr_err("map_device: failed to map device '%s' from session %s, can't initialize device, err: %ld\n",
>  		       pathname, sess->sessname, PTR_ERR(dev));
> diff --git a/drivers/block/rnbd/rnbd-clt.h b/drivers/block/rnbd/rnbd-clt.h
> index 714d426b449b..451e7383738f 100644
> --- a/drivers/block/rnbd/rnbd-clt.h
> +++ b/drivers/block/rnbd/rnbd-clt.h
> @@ -90,6 +90,7 @@ struct rnbd_clt_session {
>  	int			queue_depth;
>  	u32			max_io_size;
>  	struct blk_mq_tag_set	tag_set;
> +	u32			nr_poll_queues;
>  	struct mutex		lock; /* protects state and devs_list */
>  	struct list_head        devs_list; /* list of struct rnbd_clt_dev */
>  	refcount_t		refcount;
> @@ -118,6 +119,7 @@ struct rnbd_clt_dev {
>  	enum rnbd_clt_dev_state	dev_state;
>  	char			*pathname;
>  	enum rnbd_access_mode	access_mode;
> +	u32			nr_poll_queues;
>  	bool			read_only;
>  	bool			rotational;
>  	bool			wc;
> @@ -147,7 +149,8 @@ struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
>  					   struct rtrs_addr *paths,
>  					   size_t path_cnt, u16 port_nr,
>  					   const char *pathname,
> -					   enum rnbd_access_mode access_mode);
> +					   enum rnbd_access_mode access_mode,
> +					   u32 nr_poll_queues);
>  int rnbd_clt_unmap_device(struct rnbd_clt_dev *dev, bool force,
>  			   const struct attribute *sysfs_self);
>  
> diff --git a/drivers/infiniband/ulp/rtrs/rtrs-clt.c b/drivers/infiniband/ulp/rtrs/rtrs-clt.c
> index 7efd49bdc78c..467d135a82cf 100644
> --- a/drivers/infiniband/ulp/rtrs/rtrs-clt.c
> +++ b/drivers/infiniband/ulp/rtrs/rtrs-clt.c
> @@ -174,7 +174,7 @@ struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_sess *sess,
>  	int id = 0;
>  
>  	if (likely(permit->con_type == RTRS_IO_CON))
> -		id = (permit->cpu_id % (sess->s.con_num - 1)) + 1;
> +		id = (permit->cpu_id % (sess->s.irq_con_num - 1)) + 1;
>  
>  	return to_clt_con(sess->s.con[id]);
>  }
> @@ -1400,23 +1400,29 @@ static void rtrs_clt_close_work(struct work_struct *work);
>  static struct rtrs_clt_sess *alloc_sess(struct rtrs_clt *clt,
>  					 const struct rtrs_addr *path,
>  					 size_t con_num, u16 max_segments,
> -					 size_t max_segment_size)
> +					 size_t max_segment_size, u32 nr_poll_queues)
>  {
>  	struct rtrs_clt_sess *sess;
>  	int err = -ENOMEM;
>  	int cpu;
> +	size_t total_con;
>  
>  	sess = kzalloc(sizeof(*sess), GFP_KERNEL);
>  	if (!sess)
>  		goto err;
>  
> -	/* Extra connection for user messages */
> -	con_num += 1;
> -
> -	sess->s.con = kcalloc(con_num, sizeof(*sess->s.con), GFP_KERNEL);
> +	/*
> +	 * irqmode and poll
> +	 * +1: Extra connection for user messages
> +	 */
> +	total_con = con_num + nr_poll_queues + 1;
> +	sess->s.con = kcalloc(total_con, sizeof(*sess->s.con), GFP_KERNEL);
>  	if (!sess->s.con)
>  		goto err_free_sess;
>  
> +	sess->s.con_num = total_con;
> +	sess->s.irq_con_num = con_num + 1;
> +
>  	sess->stats = kzalloc(sizeof(*sess->stats), GFP_KERNEL);
>  	if (!sess->stats)
>  		goto err_free_con;
> @@ -1435,7 +1441,6 @@ static struct rtrs_clt_sess *alloc_sess(struct rtrs_clt *clt,
>  		memcpy(&sess->s.src_addr, path->src,
>  		       rdma_addr_size((struct sockaddr *)path->src));
>  	strlcpy(sess->s.sessname, clt->sessname, sizeof(sess->s.sessname));
> -	sess->s.con_num = con_num;
>  	sess->clt = clt;
>  	sess->max_pages_per_mr = max_segments * max_segment_size >> 12;
>  	init_waitqueue_head(&sess->state_wq);
> @@ -1576,9 +1581,14 @@ static int create_con_cq_qp(struct rtrs_clt_con *con)
>  	}
>  	cq_size = max_send_wr + max_recv_wr;
>  	cq_vector = con->cpu % sess->s.dev->ib_dev->num_comp_vectors;
> -	err = rtrs_cq_qp_create(&sess->s, &con->c, sess->max_send_sge,
> -				 cq_vector, cq_size, max_send_wr,
> -				 max_recv_wr, IB_POLL_SOFTIRQ);
> +	if (con->c.cid >= sess->s.irq_con_num)
> +		err = rtrs_cq_qp_create(&sess->s, &con->c, sess->max_send_sge,
> +					cq_vector, cq_size, max_send_wr,
> +					max_recv_wr, IB_POLL_DIRECT);
> +	else
> +		err = rtrs_cq_qp_create(&sess->s, &con->c, sess->max_send_sge,
> +					cq_vector, cq_size, max_send_wr,
> +					max_recv_wr, IB_POLL_SOFTIRQ);
>  	/*
>  	 * In case of error we do not bother to clean previous allocations,
>  	 * since destroy_con_cq_qp() must be called.
> @@ -2631,6 +2641,7 @@ static void free_clt(struct rtrs_clt *clt)
>   * @max_segment_size: Max. size of one segment
>   * @max_reconnect_attempts: Number of times to reconnect on error before giving
>   *			    up, 0 for * disabled, -1 for forever
> + * @nr_poll_queues: number of polling mode connection using IB_POLL_DIRECT flag
>   *
>   * Starts session establishment with the rtrs_server. The function can block
>   * up to ~2000ms before it returns.
> @@ -2644,7 +2655,7 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
>  				 size_t pdu_sz, u8 reconnect_delay_sec,
>  				 u16 max_segments,
>  				 size_t max_segment_size,
> -				 s16 max_reconnect_attempts)
> +				 s16 max_reconnect_attempts, u32 nr_poll_queues)
>  {
>  	struct rtrs_clt_sess *sess, *tmp;
>  	struct rtrs_clt *clt;
> @@ -2662,7 +2673,7 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
>  		struct rtrs_clt_sess *sess;
>  
>  		sess = alloc_sess(clt, &paths[i], nr_cpu_ids,
> -				  max_segments, max_segment_size);
> +				  max_segments, max_segment_size, nr_poll_queues);
>  		if (IS_ERR(sess)) {
>  			err = PTR_ERR(sess);
>  			goto close_all_sess;
> @@ -2887,6 +2898,31 @@ int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
>  }
>  EXPORT_SYMBOL(rtrs_clt_request);
>  
> +int rtrs_clt_rdma_cq_direct(struct rtrs_clt *clt, unsigned int index)
> +{
> +	int cnt;
> +	struct rtrs_con *con;
> +	struct rtrs_clt_sess *sess;
> +	struct path_it it;
> +
> +	rcu_read_lock();
> +	for (path_it_init(&it, clt);
> +	     (sess = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
> +		if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED))

We talked about useless likely/unlikely in your workloads.

> +			continue;
> +
> +		con = sess->s.con[index + 1];
> +		cnt = ib_process_cq_direct(con->cq, -1);
> +		if (likely(cnt))
> +			break;
> +	}
> +	path_it_deinit(&it);
> +	rcu_read_unlock();
> +
> +	return cnt;
> +}
> +EXPORT_SYMBOL(rtrs_clt_rdma_cq_direct);
> +
>  /**
>   * rtrs_clt_query() - queries RTRS session attributes
>   *@clt: session pointer
> @@ -2916,7 +2952,7 @@ int rtrs_clt_create_path_from_sysfs(struct rtrs_clt *clt,
>  	int err;
>  
>  	sess = alloc_sess(clt, addr, nr_cpu_ids, clt->max_segments,
> -			  clt->max_segment_size);
> +			  clt->max_segment_size, 0);
>  	if (IS_ERR(sess))
>  		return PTR_ERR(sess);
>  
> diff --git a/drivers/infiniband/ulp/rtrs/rtrs-pri.h b/drivers/infiniband/ulp/rtrs/rtrs-pri.h
> index 8caad0a2322b..00eb45053339 100644
> --- a/drivers/infiniband/ulp/rtrs/rtrs-pri.h
> +++ b/drivers/infiniband/ulp/rtrs/rtrs-pri.h
> @@ -101,6 +101,7 @@ struct rtrs_sess {
>  	uuid_t			uuid;
>  	struct rtrs_con	**con;
>  	unsigned int		con_num;
> +	unsigned int		irq_con_num;
>  	unsigned int		recon_cnt;
>  	struct rtrs_ib_dev	*dev;
>  	int			dev_ref;
> diff --git a/drivers/infiniband/ulp/rtrs/rtrs.h b/drivers/infiniband/ulp/rtrs/rtrs.h
> index 2db1b5eb3ab0..f891fbe7abe6 100644
> --- a/drivers/infiniband/ulp/rtrs/rtrs.h
> +++ b/drivers/infiniband/ulp/rtrs/rtrs.h
> @@ -59,7 +59,7 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
>  				 size_t pdu_sz, u8 reconnect_delay_sec,
>  				 u16 max_segments,
>  				 size_t max_segment_size,
> -				 s16 max_reconnect_attempts);
> +				 s16 max_reconnect_attempts, u32 nr_poll_queues);
>  
>  void rtrs_clt_close(struct rtrs_clt *sess);
>  
> @@ -103,6 +103,7 @@ int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
>  		     struct rtrs_clt *sess, struct rtrs_permit *permit,
>  		     const struct kvec *vec, size_t nr, size_t len,
>  		     struct scatterlist *sg, unsigned int sg_cnt);
> +int rtrs_clt_rdma_cq_direct(struct rtrs_clt *clt, unsigned int index);
>  
>  /**
>   * rtrs_attrs - RTRS session attributes
> -- 
> 2.25.1
> 

^ permalink raw reply	[flat|nested] 10+ messages in thread

* Re: [PATCHv4 for-next 13/19] block/rnbd-clt: Support polling mode for IO latency optimization
  2021-04-18  8:36   ` Leon Romanovsky
@ 2021-04-19  5:12     ` Gioh Kim
  2021-04-19  5:20       ` Leon Romanovsky
  0 siblings, 1 reply; 10+ messages in thread
From: Gioh Kim @ 2021-04-19  5:12 UTC (permalink / raw)
  To: Leon Romanovsky
  Cc: linux-block, Jens Axboe, hch, sagi, Bart Van Assche, Haris Iqbal,
	Jinpu Wang, Gioh Kim, linux-rdma, Jason Gunthorpe

On Sun, Apr 18, 2021 at 10:36 AM Leon Romanovsky <leon@kernel.org> wrote:
>
> On Wed, Apr 14, 2021 at 02:23:56PM +0200, Gioh Kim wrote:
> > From: Gioh Kim <gi-oh.kim@cloud.ionos.com>
> >
> > RNBD can make double-queues for irq-mode and poll-mode.
> > For example, on 4-CPU system 8 request-queues are created,
> > 4 for irq-mode and 4 for poll-mode.
> > If the IO has HIPRI flag, the block-layer will call .poll function
> > of RNBD. Then IO is sent to the poll-mode queue.
> > Add optional nr_poll_queues argument for map_devices interface.
> >
> > To support polling of RNBD, RTRS client creates connections
> > for both of irq-mode and direct-poll-mode.
> >
> > For example, on 4-CPU system it could've create 5 connections:
> > con[0] => user message (softirq cq)
> > con[1:4] => softirq cq
> >
> > After this patch, it can create 9 connections:
> > con[0] => user message (softirq cq)
> > con[1:4] => softirq cq
> > con[5:8] => DIRECT-POLL cq
> >
> > Cc: Leon Romanovsky <leonro@nvidia.com>
> > Cc: linux-rdma@vger.kernel.org
> > Signed-off-by: Gioh Kim <gi-oh.kim@ionos.com>
> > Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
> > Acked-by: Jason Gunthorpe <jgg@nvidia.com>
> > ---
> >  drivers/block/rnbd/rnbd-clt-sysfs.c    | 56 +++++++++++++----
> >  drivers/block/rnbd/rnbd-clt.c          | 85 +++++++++++++++++++++++---
> >  drivers/block/rnbd/rnbd-clt.h          |  5 +-
> >  drivers/infiniband/ulp/rtrs/rtrs-clt.c | 62 +++++++++++++++----
> >  drivers/infiniband/ulp/rtrs/rtrs-pri.h |  1 +
> >  drivers/infiniband/ulp/rtrs/rtrs.h     |  3 +-
> >  6 files changed, 178 insertions(+), 34 deletions(-)
> >
> > diff --git a/drivers/block/rnbd/rnbd-clt-sysfs.c b/drivers/block/rnbd/rnbd-clt-sysfs.c
> > index 49015f428e67..bd111ebceb75 100644
> > --- a/drivers/block/rnbd/rnbd-clt-sysfs.c
> > +++ b/drivers/block/rnbd/rnbd-clt-sysfs.c
> > @@ -34,6 +34,7 @@ enum {
> >       RNBD_OPT_DEV_PATH       = 1 << 2,
> >       RNBD_OPT_ACCESS_MODE    = 1 << 3,
> >       RNBD_OPT_SESSNAME       = 1 << 6,
> > +     RNBD_OPT_NR_POLL_QUEUES = 1 << 7,
> >  };
> >
> >  static const unsigned int rnbd_opt_mandatory[] = {
> > @@ -42,12 +43,13 @@ static const unsigned int rnbd_opt_mandatory[] = {
> >  };
> >
> >  static const match_table_t rnbd_opt_tokens = {
> > -     {RNBD_OPT_PATH,         "path=%s"       },
> > -     {RNBD_OPT_DEV_PATH,     "device_path=%s"},
> > -     {RNBD_OPT_DEST_PORT,    "dest_port=%d"  },
> > -     {RNBD_OPT_ACCESS_MODE,  "access_mode=%s"},
> > -     {RNBD_OPT_SESSNAME,     "sessname=%s"   },
> > -     {RNBD_OPT_ERR,          NULL            },
> > +     {RNBD_OPT_PATH,                 "path=%s"               },
> > +     {RNBD_OPT_DEV_PATH,             "device_path=%s"        },
> > +     {RNBD_OPT_DEST_PORT,            "dest_port=%d"          },
> > +     {RNBD_OPT_ACCESS_MODE,          "access_mode=%s"        },
> > +     {RNBD_OPT_SESSNAME,             "sessname=%s"           },
> > +     {RNBD_OPT_NR_POLL_QUEUES,       "nr_poll_queues=%d"     },
> > +     {RNBD_OPT_ERR,                  NULL                    },
> >  };
> >
> >  struct rnbd_map_options {
> > @@ -57,6 +59,7 @@ struct rnbd_map_options {
> >       char *pathname;
> >       u16 *dest_port;
> >       enum rnbd_access_mode *access_mode;
> > +     u32 *nr_poll_queues;
> >  };
> >
> >  static int rnbd_clt_parse_map_options(const char *buf, size_t max_path_cnt,
> > @@ -68,7 +71,7 @@ static int rnbd_clt_parse_map_options(const char *buf, size_t max_path_cnt,
> >       int opt_mask = 0;
> >       int token;
> >       int ret = -EINVAL;
> > -     int i, dest_port;
> > +     int i, dest_port, nr_poll_queues;
> >       int p_cnt = 0;
> >
> >       options = kstrdup(buf, GFP_KERNEL);
> > @@ -178,6 +181,19 @@ static int rnbd_clt_parse_map_options(const char *buf, size_t max_path_cnt,
> >                       kfree(p);
> >                       break;
> >
> > +             case RNBD_OPT_NR_POLL_QUEUES:
> > +                     if (match_int(args, &nr_poll_queues) || nr_poll_queues < -1 ||
> > +                         nr_poll_queues > (int)nr_cpu_ids) {
> > +                             pr_err("bad nr_poll_queues parameter '%d'\n",
> > +                                    nr_poll_queues);
> > +                             ret = -EINVAL;
> > +                             goto out;
> > +                     }
> > +                     if (nr_poll_queues == -1)
> > +                             nr_poll_queues = nr_cpu_ids;
> > +                     *opt->nr_poll_queues = nr_poll_queues;
> > +                     break;
> > +
> >               default:
> >                       pr_err("map_device: Unknown parameter or missing value '%s'\n",
> >                              p);
> > @@ -227,6 +243,20 @@ static ssize_t state_show(struct kobject *kobj,
> >
> >  static struct kobj_attribute rnbd_clt_state_attr = __ATTR_RO(state);
> >
> > +static ssize_t nr_poll_queues_show(struct kobject *kobj,
> > +                                struct kobj_attribute *attr, char *page)
> > +{
> > +     struct rnbd_clt_dev *dev;
> > +
> > +     dev = container_of(kobj, struct rnbd_clt_dev, kobj);
> > +
> > +     return snprintf(page, PAGE_SIZE, "%d\n",
> > +                     dev->nr_poll_queues);
> > +}
>
> Didn't Greg ask you to use sysfs_emit() here?

Right, I missed it.
I will fix it for next round.


>
> > +
> > +static struct kobj_attribute rnbd_clt_nr_poll_queues =
> > +     __ATTR_RO(nr_poll_queues);
> > +
> >  static ssize_t mapping_path_show(struct kobject *kobj,
> >                                struct kobj_attribute *attr, char *page)
> >  {
> > @@ -421,6 +451,7 @@ static struct attribute *rnbd_dev_attrs[] = {
> >       &rnbd_clt_state_attr.attr,
> >       &rnbd_clt_session_attr.attr,
> >       &rnbd_clt_access_mode.attr,
> > +     &rnbd_clt_nr_poll_queues.attr,
> >       NULL,
> >  };
> >
> > @@ -469,7 +500,7 @@ static ssize_t rnbd_clt_map_device_show(struct kobject *kobj,
> >                                        char *page)
> >  {
> >       return scnprintf(page, PAGE_SIZE,
> > -                      "Usage: echo \"[dest_port=server port number] sessname=<name of the rtrs session> path=<[srcaddr@]dstaddr> [path=<[srcaddr@]dstaddr>] device_path=<full path on remote side> [access_mode=<ro|rw|migration>]\" > %s\n\naddr ::= [ ip:<ipv4> | ip:<ipv6> | gid:<gid> ]\n",
> > +                      "Usage: echo \"[dest_port=server port number] sessname=<name of the rtrs session> path=<[srcaddr@]dstaddr> [path=<[srcaddr@]dstaddr>] device_path=<full path on remote side> [access_mode=<ro|rw|migration>] [nr_poll_queues=<number of queues>]\" > %s\n\naddr ::= [ ip:<ipv4> | ip:<ipv6> | gid:<gid> ]\n",
> >                        attr->attr.name);
> >  }
> >
> > @@ -541,6 +572,7 @@ static ssize_t rnbd_clt_map_device_store(struct kobject *kobj,
> >       char sessname[NAME_MAX];
> >       enum rnbd_access_mode access_mode = RNBD_ACCESS_RW;
> >       u16 port_nr = RTRS_PORT;
> > +     u32 nr_poll_queues = 0;
> >
> >       struct sockaddr_storage *addrs;
> >       struct rtrs_addr paths[6];
> > @@ -552,6 +584,7 @@ static ssize_t rnbd_clt_map_device_store(struct kobject *kobj,
> >       opt.pathname = pathname;
> >       opt.dest_port = &port_nr;
> >       opt.access_mode = &access_mode;
> > +     opt.nr_poll_queues = &nr_poll_queues;
> >       addrs = kcalloc(ARRAY_SIZE(paths) * 2, sizeof(*addrs), GFP_KERNEL);
> >       if (!addrs)
> >               return -ENOMEM;
> > @@ -565,12 +598,13 @@ static ssize_t rnbd_clt_map_device_store(struct kobject *kobj,
> >       if (ret)
> >               goto out;
> >
> > -     pr_info("Mapping device %s on session %s, (access_mode: %s)\n",
> > +     pr_info("Mapping device %s on session %s, (access_mode: %s, nr_poll_queues: %d)\n",
> >               pathname, sessname,
> > -             rnbd_access_mode_str(access_mode));
> > +             rnbd_access_mode_str(access_mode),
> > +             nr_poll_queues);
> >
> >       dev = rnbd_clt_map_device(sessname, paths, path_cnt, port_nr, pathname,
> > -                               access_mode);
> > +                               access_mode, nr_poll_queues);
> >       if (IS_ERR(dev)) {
> >               ret = PTR_ERR(dev);
> >               goto out;
> > diff --git a/drivers/block/rnbd/rnbd-clt.c b/drivers/block/rnbd/rnbd-clt.c
> > index 9b44aac680d5..63719ec04d58 100644
> > --- a/drivers/block/rnbd/rnbd-clt.c
> > +++ b/drivers/block/rnbd/rnbd-clt.c
> > @@ -1165,9 +1165,54 @@ static blk_status_t rnbd_queue_rq(struct blk_mq_hw_ctx *hctx,
> >       return ret;
> >  }
> >
> > +static int rnbd_rdma_poll(struct blk_mq_hw_ctx *hctx)
> > +{
> > +     struct rnbd_queue *q = hctx->driver_data;
> > +     struct rnbd_clt_dev *dev = q->dev;
> > +     int cnt;
> > +
> > +     cnt = rtrs_clt_rdma_cq_direct(dev->sess->rtrs, hctx->queue_num);
> > +     return cnt;
> > +}
> > +
> > +static int rnbd_rdma_map_queues(struct blk_mq_tag_set *set)
> > +{
> > +     struct rnbd_clt_session *sess = set->driver_data;
> > +
> > +     /* shared read/write queues */
> > +     set->map[HCTX_TYPE_DEFAULT].nr_queues = num_online_cpus();
> > +     set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
> > +     set->map[HCTX_TYPE_READ].nr_queues = num_online_cpus();
> > +     set->map[HCTX_TYPE_READ].queue_offset = 0;
> > +     blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
> > +     blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
> > +
> > +     if (sess->nr_poll_queues) {
> > +             /* dedicated queue for poll */
> > +             set->map[HCTX_TYPE_POLL].nr_queues = sess->nr_poll_queues;
> > +             set->map[HCTX_TYPE_POLL].queue_offset = set->map[HCTX_TYPE_READ].queue_offset +
> > +                     set->map[HCTX_TYPE_READ].nr_queues;
> > +             blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
> > +             pr_info("[session=%s] mapped %d/%d/%d default/read/poll queues.\n",
> > +                     sess->sessname,
> > +                     set->map[HCTX_TYPE_DEFAULT].nr_queues,
> > +                     set->map[HCTX_TYPE_READ].nr_queues,
> > +                     set->map[HCTX_TYPE_POLL].nr_queues);
> > +     } else {
> > +             pr_info("[session=%s] mapped %d/%d default/read queues.\n",
> > +                     sess->sessname,
> > +                     set->map[HCTX_TYPE_DEFAULT].nr_queues,
> > +                     set->map[HCTX_TYPE_READ].nr_queues);
> > +     }
> > +
> > +     return 0;
> > +}
> > +
> >  static struct blk_mq_ops rnbd_mq_ops = {
> >       .queue_rq       = rnbd_queue_rq,
> >       .complete       = rnbd_softirq_done_fn,
> > +     .map_queues     = rnbd_rdma_map_queues,
> > +     .poll           = rnbd_rdma_poll,
> >  };
> >
> >  static int setup_mq_tags(struct rnbd_clt_session *sess)
> > @@ -1181,7 +1226,15 @@ static int setup_mq_tags(struct rnbd_clt_session *sess)
> >       tag_set->flags          = BLK_MQ_F_SHOULD_MERGE |
> >                                 BLK_MQ_F_TAG_QUEUE_SHARED;
> >       tag_set->cmd_size       = sizeof(struct rnbd_iu) + RNBD_RDMA_SGL_SIZE;
> > -     tag_set->nr_hw_queues   = num_online_cpus();
> > +
> > +     /* for HCTX_TYPE_DEFAULT, HCTX_TYPE_READ, HCTX_TYPE_POLL */
> > +     tag_set->nr_maps        = sess->nr_poll_queues ? HCTX_MAX_TYPES : 2;
> > +     /*
> > +      * HCTX_TYPE_DEFAULT and HCTX_TYPE_READ share one set of queues
> > +      * others are for HCTX_TYPE_POLL
> > +      */
> > +     tag_set->nr_hw_queues   = num_online_cpus() + sess->nr_poll_queues;
> > +     tag_set->driver_data    = sess;
> >
> >       return blk_mq_alloc_tag_set(tag_set);
> >  }
> > @@ -1189,7 +1242,7 @@ static int setup_mq_tags(struct rnbd_clt_session *sess)
> >  static struct rnbd_clt_session *
> >  find_and_get_or_create_sess(const char *sessname,
> >                           const struct rtrs_addr *paths,
> > -                         size_t path_cnt, u16 port_nr)
> > +                         size_t path_cnt, u16 port_nr, u32 nr_poll_queues)
> >  {
> >       struct rnbd_clt_session *sess;
> >       struct rtrs_attrs attrs;
> > @@ -1198,6 +1251,17 @@ find_and_get_or_create_sess(const char *sessname,
> >       struct rtrs_clt_ops rtrs_ops;
> >
> >       sess = find_or_create_sess(sessname, &first);
> > +     if (sess == ERR_PTR(-ENOMEM))
> > +             return ERR_PTR(-ENOMEM);
> > +     else if ((nr_poll_queues && !first) ||  (!nr_poll_queues && sess->nr_poll_queues)) {
> > +             /*
> > +              * A device MUST have its own session to use the polling-mode.
> > +              * It must fail to map new device with the same session.
> > +              */
> > +             err = -EINVAL;
> > +             goto put_sess;
> > +     }
> > +
> >       if (!first)
> >               return sess;
> >
> > @@ -1219,7 +1283,7 @@ find_and_get_or_create_sess(const char *sessname,
> >                                  0, /* Do not use pdu of rtrs */
> >                                  RECONNECT_DELAY, BMAX_SEGMENTS,
> >                                  BLK_MAX_SEGMENT_SIZE,
> > -                                MAX_RECONNECTS);
> > +                                MAX_RECONNECTS, nr_poll_queues);
> >       if (IS_ERR(sess->rtrs)) {
> >               err = PTR_ERR(sess->rtrs);
> >               goto wake_up_and_put;
> > @@ -1227,6 +1291,7 @@ find_and_get_or_create_sess(const char *sessname,
> >       rtrs_clt_query(sess->rtrs, &attrs);
> >       sess->max_io_size = attrs.max_io_size;
> >       sess->queue_depth = attrs.queue_depth;
> > +     sess->nr_poll_queues = nr_poll_queues;
> >
> >       err = setup_mq_tags(sess);
> >       if (err)
> > @@ -1370,7 +1435,8 @@ static int rnbd_client_setup_device(struct rnbd_clt_dev *dev)
> >
> >  static struct rnbd_clt_dev *init_dev(struct rnbd_clt_session *sess,
> >                                     enum rnbd_access_mode access_mode,
> > -                                   const char *pathname)
> > +                                   const char *pathname,
> > +                                   u32 nr_poll_queues)
> >  {
> >       struct rnbd_clt_dev *dev;
> >       int ret;
> > @@ -1379,7 +1445,8 @@ static struct rnbd_clt_dev *init_dev(struct rnbd_clt_session *sess,
> >       if (!dev)
> >               return ERR_PTR(-ENOMEM);
> >
> > -     dev->hw_queues = kcalloc(nr_cpu_ids, sizeof(*dev->hw_queues),
> > +     dev->hw_queues = kcalloc(nr_cpu_ids /* softirq */ + nr_poll_queues /* poll */,
>
> Please don't add comments in the middle of function call.

Ok, I will fix it for next round.


>
> > +                              sizeof(*dev->hw_queues),
> >                                GFP_KERNEL);
> >       if (!dev->hw_queues) {
> >               ret = -ENOMEM;
> > @@ -1405,6 +1472,7 @@ static struct rnbd_clt_dev *init_dev(struct rnbd_clt_session *sess,
> >       dev->clt_device_id      = ret;
> >       dev->sess               = sess;
> >       dev->access_mode        = access_mode;
> > +     dev->nr_poll_queues     = nr_poll_queues;
> >       mutex_init(&dev->lock);
> >       refcount_set(&dev->refcount, 1);
> >       dev->dev_state = DEV_STATE_INIT;
> > @@ -1491,7 +1559,8 @@ struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
> >                                          struct rtrs_addr *paths,
> >                                          size_t path_cnt, u16 port_nr,
> >                                          const char *pathname,
> > -                                        enum rnbd_access_mode access_mode)
> > +                                        enum rnbd_access_mode access_mode,
> > +                                        u32 nr_poll_queues)
> >  {
> >       struct rnbd_clt_session *sess;
> >       struct rnbd_clt_dev *dev;
> > @@ -1500,11 +1569,11 @@ struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
> >       if (unlikely(exists_devpath(pathname, sessname)))
> >               return ERR_PTR(-EEXIST);
> >
> > -     sess = find_and_get_or_create_sess(sessname, paths, path_cnt, port_nr);
> > +     sess = find_and_get_or_create_sess(sessname, paths, path_cnt, port_nr, nr_poll_queues);
> >       if (IS_ERR(sess))
> >               return ERR_CAST(sess);
> >
> > -     dev = init_dev(sess, access_mode, pathname);
> > +     dev = init_dev(sess, access_mode, pathname, nr_poll_queues);
> >       if (IS_ERR(dev)) {
> >               pr_err("map_device: failed to map device '%s' from session %s, can't initialize device, err: %ld\n",
> >                      pathname, sess->sessname, PTR_ERR(dev));
> > diff --git a/drivers/block/rnbd/rnbd-clt.h b/drivers/block/rnbd/rnbd-clt.h
> > index 714d426b449b..451e7383738f 100644
> > --- a/drivers/block/rnbd/rnbd-clt.h
> > +++ b/drivers/block/rnbd/rnbd-clt.h
> > @@ -90,6 +90,7 @@ struct rnbd_clt_session {
> >       int                     queue_depth;
> >       u32                     max_io_size;
> >       struct blk_mq_tag_set   tag_set;
> > +     u32                     nr_poll_queues;
> >       struct mutex            lock; /* protects state and devs_list */
> >       struct list_head        devs_list; /* list of struct rnbd_clt_dev */
> >       refcount_t              refcount;
> > @@ -118,6 +119,7 @@ struct rnbd_clt_dev {
> >       enum rnbd_clt_dev_state dev_state;
> >       char                    *pathname;
> >       enum rnbd_access_mode   access_mode;
> > +     u32                     nr_poll_queues;
> >       bool                    read_only;
> >       bool                    rotational;
> >       bool                    wc;
> > @@ -147,7 +149,8 @@ struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
> >                                          struct rtrs_addr *paths,
> >                                          size_t path_cnt, u16 port_nr,
> >                                          const char *pathname,
> > -                                        enum rnbd_access_mode access_mode);
> > +                                        enum rnbd_access_mode access_mode,
> > +                                        u32 nr_poll_queues);
> >  int rnbd_clt_unmap_device(struct rnbd_clt_dev *dev, bool force,
> >                          const struct attribute *sysfs_self);
> >
> > diff --git a/drivers/infiniband/ulp/rtrs/rtrs-clt.c b/drivers/infiniband/ulp/rtrs/rtrs-clt.c
> > index 7efd49bdc78c..467d135a82cf 100644
> > --- a/drivers/infiniband/ulp/rtrs/rtrs-clt.c
> > +++ b/drivers/infiniband/ulp/rtrs/rtrs-clt.c
> > @@ -174,7 +174,7 @@ struct rtrs_clt_con *rtrs_permit_to_clt_con(struct rtrs_clt_sess *sess,
> >       int id = 0;
> >
> >       if (likely(permit->con_type == RTRS_IO_CON))
> > -             id = (permit->cpu_id % (sess->s.con_num - 1)) + 1;
> > +             id = (permit->cpu_id % (sess->s.irq_con_num - 1)) + 1;
> >
> >       return to_clt_con(sess->s.con[id]);
> >  }
> > @@ -1400,23 +1400,29 @@ static void rtrs_clt_close_work(struct work_struct *work);
> >  static struct rtrs_clt_sess *alloc_sess(struct rtrs_clt *clt,
> >                                        const struct rtrs_addr *path,
> >                                        size_t con_num, u16 max_segments,
> > -                                      size_t max_segment_size)
> > +                                      size_t max_segment_size, u32 nr_poll_queues)
> >  {
> >       struct rtrs_clt_sess *sess;
> >       int err = -ENOMEM;
> >       int cpu;
> > +     size_t total_con;
> >
> >       sess = kzalloc(sizeof(*sess), GFP_KERNEL);
> >       if (!sess)
> >               goto err;
> >
> > -     /* Extra connection for user messages */
> > -     con_num += 1;
> > -
> > -     sess->s.con = kcalloc(con_num, sizeof(*sess->s.con), GFP_KERNEL);
> > +     /*
> > +      * irqmode and poll
> > +      * +1: Extra connection for user messages
> > +      */
> > +     total_con = con_num + nr_poll_queues + 1;
> > +     sess->s.con = kcalloc(total_con, sizeof(*sess->s.con), GFP_KERNEL);
> >       if (!sess->s.con)
> >               goto err_free_sess;
> >
> > +     sess->s.con_num = total_con;
> > +     sess->s.irq_con_num = con_num + 1;
> > +
> >       sess->stats = kzalloc(sizeof(*sess->stats), GFP_KERNEL);
> >       if (!sess->stats)
> >               goto err_free_con;
> > @@ -1435,7 +1441,6 @@ static struct rtrs_clt_sess *alloc_sess(struct rtrs_clt *clt,
> >               memcpy(&sess->s.src_addr, path->src,
> >                      rdma_addr_size((struct sockaddr *)path->src));
> >       strlcpy(sess->s.sessname, clt->sessname, sizeof(sess->s.sessname));
> > -     sess->s.con_num = con_num;
> >       sess->clt = clt;
> >       sess->max_pages_per_mr = max_segments * max_segment_size >> 12;
> >       init_waitqueue_head(&sess->state_wq);
> > @@ -1576,9 +1581,14 @@ static int create_con_cq_qp(struct rtrs_clt_con *con)
> >       }
> >       cq_size = max_send_wr + max_recv_wr;
> >       cq_vector = con->cpu % sess->s.dev->ib_dev->num_comp_vectors;
> > -     err = rtrs_cq_qp_create(&sess->s, &con->c, sess->max_send_sge,
> > -                              cq_vector, cq_size, max_send_wr,
> > -                              max_recv_wr, IB_POLL_SOFTIRQ);
> > +     if (con->c.cid >= sess->s.irq_con_num)
> > +             err = rtrs_cq_qp_create(&sess->s, &con->c, sess->max_send_sge,
> > +                                     cq_vector, cq_size, max_send_wr,
> > +                                     max_recv_wr, IB_POLL_DIRECT);
> > +     else
> > +             err = rtrs_cq_qp_create(&sess->s, &con->c, sess->max_send_sge,
> > +                                     cq_vector, cq_size, max_send_wr,
> > +                                     max_recv_wr, IB_POLL_SOFTIRQ);
> >       /*
> >        * In case of error we do not bother to clean previous allocations,
> >        * since destroy_con_cq_qp() must be called.
> > @@ -2631,6 +2641,7 @@ static void free_clt(struct rtrs_clt *clt)
> >   * @max_segment_size: Max. size of one segment
> >   * @max_reconnect_attempts: Number of times to reconnect on error before giving
> >   *                       up, 0 for * disabled, -1 for forever
> > + * @nr_poll_queues: number of polling mode connection using IB_POLL_DIRECT flag
> >   *
> >   * Starts session establishment with the rtrs_server. The function can block
> >   * up to ~2000ms before it returns.
> > @@ -2644,7 +2655,7 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
> >                                size_t pdu_sz, u8 reconnect_delay_sec,
> >                                u16 max_segments,
> >                                size_t max_segment_size,
> > -                              s16 max_reconnect_attempts)
> > +                              s16 max_reconnect_attempts, u32 nr_poll_queues)
> >  {
> >       struct rtrs_clt_sess *sess, *tmp;
> >       struct rtrs_clt *clt;
> > @@ -2662,7 +2673,7 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
> >               struct rtrs_clt_sess *sess;
> >
> >               sess = alloc_sess(clt, &paths[i], nr_cpu_ids,
> > -                               max_segments, max_segment_size);
> > +                               max_segments, max_segment_size, nr_poll_queues);
> >               if (IS_ERR(sess)) {
> >                       err = PTR_ERR(sess);
> >                       goto close_all_sess;
> > @@ -2887,6 +2898,31 @@ int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
> >  }
> >  EXPORT_SYMBOL(rtrs_clt_request);
> >
> > +int rtrs_clt_rdma_cq_direct(struct rtrs_clt *clt, unsigned int index)
> > +{
> > +     int cnt;
> > +     struct rtrs_con *con;
> > +     struct rtrs_clt_sess *sess;
> > +     struct path_it it;
> > +
> > +     rcu_read_lock();
> > +     for (path_it_init(&it, clt);
> > +          (sess = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
> > +             if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED))
>
> We talked about useless likely/unlikely in your workloads.

Right, I've made a patch to remove all likely/unlikely
and will send with the next patch set.

I thought it could be better for review to keep the patches
in the patch set. So if this set is applied, I will send a small patch set
to remove likely/unlikely and do some cleanup.

>
> > +                     continue;
> > +
> > +             con = sess->s.con[index + 1];
> > +             cnt = ib_process_cq_direct(con->cq, -1);
> > +             if (likely(cnt))
> > +                     break;
> > +     }
> > +     path_it_deinit(&it);
> > +     rcu_read_unlock();
> > +
> > +     return cnt;
> > +}
> > +EXPORT_SYMBOL(rtrs_clt_rdma_cq_direct);
> > +
> >  /**
> >   * rtrs_clt_query() - queries RTRS session attributes
> >   *@clt: session pointer
> > @@ -2916,7 +2952,7 @@ int rtrs_clt_create_path_from_sysfs(struct rtrs_clt *clt,
> >       int err;
> >
> >       sess = alloc_sess(clt, addr, nr_cpu_ids, clt->max_segments,
> > -                       clt->max_segment_size);
> > +                       clt->max_segment_size, 0);
> >       if (IS_ERR(sess))
> >               return PTR_ERR(sess);
> >
> > diff --git a/drivers/infiniband/ulp/rtrs/rtrs-pri.h b/drivers/infiniband/ulp/rtrs/rtrs-pri.h
> > index 8caad0a2322b..00eb45053339 100644
> > --- a/drivers/infiniband/ulp/rtrs/rtrs-pri.h
> > +++ b/drivers/infiniband/ulp/rtrs/rtrs-pri.h
> > @@ -101,6 +101,7 @@ struct rtrs_sess {
> >       uuid_t                  uuid;
> >       struct rtrs_con **con;
> >       unsigned int            con_num;
> > +     unsigned int            irq_con_num;
> >       unsigned int            recon_cnt;
> >       struct rtrs_ib_dev      *dev;
> >       int                     dev_ref;
> > diff --git a/drivers/infiniband/ulp/rtrs/rtrs.h b/drivers/infiniband/ulp/rtrs/rtrs.h
> > index 2db1b5eb3ab0..f891fbe7abe6 100644
> > --- a/drivers/infiniband/ulp/rtrs/rtrs.h
> > +++ b/drivers/infiniband/ulp/rtrs/rtrs.h
> > @@ -59,7 +59,7 @@ struct rtrs_clt *rtrs_clt_open(struct rtrs_clt_ops *ops,
> >                                size_t pdu_sz, u8 reconnect_delay_sec,
> >                                u16 max_segments,
> >                                size_t max_segment_size,
> > -                              s16 max_reconnect_attempts);
> > +                              s16 max_reconnect_attempts, u32 nr_poll_queues);
> >
> >  void rtrs_clt_close(struct rtrs_clt *sess);
> >
> > @@ -103,6 +103,7 @@ int rtrs_clt_request(int dir, struct rtrs_clt_req_ops *ops,
> >                    struct rtrs_clt *sess, struct rtrs_permit *permit,
> >                    const struct kvec *vec, size_t nr, size_t len,
> >                    struct scatterlist *sg, unsigned int sg_cnt);
> > +int rtrs_clt_rdma_cq_direct(struct rtrs_clt *clt, unsigned int index);
> >
> >  /**
> >   * rtrs_attrs - RTRS session attributes
> > --
> > 2.25.1
> >

Thank you for the review.
I will send V5 soon.

^ permalink raw reply	[flat|nested] 10+ messages in thread

* Re: [PATCHv4 for-next 13/19] block/rnbd-clt: Support polling mode for IO latency optimization
  2021-04-19  5:12     ` Gioh Kim
@ 2021-04-19  5:20       ` Leon Romanovsky
  2021-04-19  5:51         ` Gioh Kim
  0 siblings, 1 reply; 10+ messages in thread
From: Leon Romanovsky @ 2021-04-19  5:20 UTC (permalink / raw)
  To: Gioh Kim
  Cc: linux-block, Jens Axboe, hch, sagi, Bart Van Assche, Haris Iqbal,
	Jinpu Wang, Gioh Kim, linux-rdma, Jason Gunthorpe

On Mon, Apr 19, 2021 at 07:12:09AM +0200, Gioh Kim wrote:
> On Sun, Apr 18, 2021 at 10:36 AM Leon Romanovsky <leon@kernel.org> wrote:
> >
> > On Wed, Apr 14, 2021 at 02:23:56PM +0200, Gioh Kim wrote:
> > > From: Gioh Kim <gi-oh.kim@cloud.ionos.com>
> > >
> > > RNBD can make double-queues for irq-mode and poll-mode.
> > > For example, on 4-CPU system 8 request-queues are created,
> > > 4 for irq-mode and 4 for poll-mode.
> > > If the IO has HIPRI flag, the block-layer will call .poll function
> > > of RNBD. Then IO is sent to the poll-mode queue.
> > > Add optional nr_poll_queues argument for map_devices interface.
> > >
> > > To support polling of RNBD, RTRS client creates connections
> > > for both of irq-mode and direct-poll-mode.
> > >
> > > For example, on 4-CPU system it could've create 5 connections:
> > > con[0] => user message (softirq cq)
> > > con[1:4] => softirq cq
> > >
> > > After this patch, it can create 9 connections:
> > > con[0] => user message (softirq cq)
> > > con[1:4] => softirq cq
> > > con[5:8] => DIRECT-POLL cq

<...>

> > > +int rtrs_clt_rdma_cq_direct(struct rtrs_clt *clt, unsigned int index)
> > > +{
> > > +     int cnt;
> > > +     struct rtrs_con *con;
> > > +     struct rtrs_clt_sess *sess;
> > > +     struct path_it it;
> > > +
> > > +     rcu_read_lock();
> > > +     for (path_it_init(&it, clt);
> > > +          (sess = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
> > > +             if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED))
> >
> > We talked about useless likely/unlikely in your workloads.
> 
> Right, I've made a patch to remove all likely/unlikely
> and will send with the next patch set.

This specific line is "brand new". We don't add code that will be
removed in next patch.

Thanks

^ permalink raw reply	[flat|nested] 10+ messages in thread

* Re: [PATCHv4 for-next 13/19] block/rnbd-clt: Support polling mode for IO latency optimization
  2021-04-19  5:20       ` Leon Romanovsky
@ 2021-04-19  5:51         ` Gioh Kim
  2021-04-19  6:09           ` Leon Romanovsky
  0 siblings, 1 reply; 10+ messages in thread
From: Gioh Kim @ 2021-04-19  5:51 UTC (permalink / raw)
  To: Leon Romanovsky
  Cc: linux-block, Jens Axboe, hch, sagi, Bart Van Assche, Haris Iqbal,
	Jinpu Wang, Gioh Kim, linux-rdma, Jason Gunthorpe

On Mon, Apr 19, 2021 at 7:46 AM Leon Romanovsky <leon@kernel.org> wrote:
>
> On Mon, Apr 19, 2021 at 07:12:09AM +0200, Gioh Kim wrote:
> > On Sun, Apr 18, 2021 at 10:36 AM Leon Romanovsky <leon@kernel.org> wrote:
> > >
> > > On Wed, Apr 14, 2021 at 02:23:56PM +0200, Gioh Kim wrote:
> > > > From: Gioh Kim <gi-oh.kim@cloud.ionos.com>
> > > >
> > > > RNBD can make double-queues for irq-mode and poll-mode.
> > > > For example, on 4-CPU system 8 request-queues are created,
> > > > 4 for irq-mode and 4 for poll-mode.
> > > > If the IO has HIPRI flag, the block-layer will call .poll function
> > > > of RNBD. Then IO is sent to the poll-mode queue.
> > > > Add optional nr_poll_queues argument for map_devices interface.
> > > >
> > > > To support polling of RNBD, RTRS client creates connections
> > > > for both of irq-mode and direct-poll-mode.
> > > >
> > > > For example, on 4-CPU system it could've create 5 connections:
> > > > con[0] => user message (softirq cq)
> > > > con[1:4] => softirq cq
> > > >
> > > > After this patch, it can create 9 connections:
> > > > con[0] => user message (softirq cq)
> > > > con[1:4] => softirq cq
> > > > con[5:8] => DIRECT-POLL cq
>
> <...>

I am sorry that I don't understand exactly.
Do I need to change them to "con<5..8>"?


>
> > > > +int rtrs_clt_rdma_cq_direct(struct rtrs_clt *clt, unsigned int index)
> > > > +{
> > > > +     int cnt;
> > > > +     struct rtrs_con *con;
> > > > +     struct rtrs_clt_sess *sess;
> > > > +     struct path_it it;
> > > > +
> > > > +     rcu_read_lock();
> > > > +     for (path_it_init(&it, clt);
> > > > +          (sess = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
> > > > +             if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED))
> > >
> > > We talked about useless likely/unlikely in your workloads.
> >
> > Right, I've made a patch to remove all likely/unlikely
> > and will send with the next patch set.
>
> This specific line is "brand new". We don't add code that will be
> removed in next patch.

Ah, ok. So you mean,
1. remove unlikely from that line
2. send a patch to remove all likely/unlikely for next round

Am I right?

^ permalink raw reply	[flat|nested] 10+ messages in thread

* Re: [PATCHv4 for-next 13/19] block/rnbd-clt: Support polling mode for IO latency optimization
  2021-04-19  5:51         ` Gioh Kim
@ 2021-04-19  6:09           ` Leon Romanovsky
  2021-04-19  6:15             ` Gioh Kim
  0 siblings, 1 reply; 10+ messages in thread
From: Leon Romanovsky @ 2021-04-19  6:09 UTC (permalink / raw)
  To: Gioh Kim
  Cc: linux-block, Jens Axboe, hch, sagi, Bart Van Assche, Haris Iqbal,
	Jinpu Wang, Gioh Kim, linux-rdma, Jason Gunthorpe

On Mon, Apr 19, 2021 at 07:51:34AM +0200, Gioh Kim wrote:
> On Mon, Apr 19, 2021 at 7:46 AM Leon Romanovsky <leon@kernel.org> wrote:
> >
> > On Mon, Apr 19, 2021 at 07:12:09AM +0200, Gioh Kim wrote:
> > > On Sun, Apr 18, 2021 at 10:36 AM Leon Romanovsky <leon@kernel.org> wrote:
> > > >
> > > > On Wed, Apr 14, 2021 at 02:23:56PM +0200, Gioh Kim wrote:
> > > > > From: Gioh Kim <gi-oh.kim@cloud.ionos.com>
> > > > >
> > > > > RNBD can make double-queues for irq-mode and poll-mode.
> > > > > For example, on 4-CPU system 8 request-queues are created,
> > > > > 4 for irq-mode and 4 for poll-mode.
> > > > > If the IO has HIPRI flag, the block-layer will call .poll function
> > > > > of RNBD. Then IO is sent to the poll-mode queue.
> > > > > Add optional nr_poll_queues argument for map_devices interface.
> > > > >
> > > > > To support polling of RNBD, RTRS client creates connections
> > > > > for both of irq-mode and direct-poll-mode.
> > > > >
> > > > > For example, on 4-CPU system it could've create 5 connections:
> > > > > con[0] => user message (softirq cq)
> > > > > con[1:4] => softirq cq
> > > > >
> > > > > After this patch, it can create 9 connections:
> > > > > con[0] => user message (softirq cq)
> > > > > con[1:4] => softirq cq
> > > > > con[5:8] => DIRECT-POLL cq
> >
> > <...>
> 
> I am sorry that I don't understand exactly.
> Do I need to change them to "con<5..8>"?

No, I just removed not relevant text and replaced it with <...> in
automatic way :).

> 
> 
> >
> > > > > +int rtrs_clt_rdma_cq_direct(struct rtrs_clt *clt, unsigned int index)
> > > > > +{
> > > > > +     int cnt;
> > > > > +     struct rtrs_con *con;
> > > > > +     struct rtrs_clt_sess *sess;
> > > > > +     struct path_it it;
> > > > > +
> > > > > +     rcu_read_lock();
> > > > > +     for (path_it_init(&it, clt);
> > > > > +          (sess = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
> > > > > +             if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED))
> > > >
> > > > We talked about useless likely/unlikely in your workloads.
> > >
> > > Right, I've made a patch to remove all likely/unlikely
> > > and will send with the next patch set.
> >
> > This specific line is "brand new". We don't add code that will be
> > removed in next patch.
> 
> Ah, ok. So you mean,
> 1. remove unlikely from that line
> 2. send a patch to remove all likely/unlikely for next round
> 
> Am I right?

Right

^ permalink raw reply	[flat|nested] 10+ messages in thread

* Re: [PATCHv4 for-next 13/19] block/rnbd-clt: Support polling mode for IO latency optimization
  2021-04-19  6:09           ` Leon Romanovsky
@ 2021-04-19  6:15             ` Gioh Kim
  0 siblings, 0 replies; 10+ messages in thread
From: Gioh Kim @ 2021-04-19  6:15 UTC (permalink / raw)
  To: Leon Romanovsky
  Cc: linux-block, Jens Axboe, hch, sagi, Bart Van Assche, Haris Iqbal,
	Jinpu Wang, Gioh Kim, linux-rdma, Jason Gunthorpe

On Mon, Apr 19, 2021 at 8:09 AM Leon Romanovsky <leon@kernel.org> wrote:
>
> On Mon, Apr 19, 2021 at 07:51:34AM +0200, Gioh Kim wrote:
> > On Mon, Apr 19, 2021 at 7:46 AM Leon Romanovsky <leon@kernel.org> wrote:
> > >
> > > On Mon, Apr 19, 2021 at 07:12:09AM +0200, Gioh Kim wrote:
> > > > On Sun, Apr 18, 2021 at 10:36 AM Leon Romanovsky <leon@kernel.org> wrote:
> > > > >
> > > > > On Wed, Apr 14, 2021 at 02:23:56PM +0200, Gioh Kim wrote:
> > > > > > From: Gioh Kim <gi-oh.kim@cloud.ionos.com>
> > > > > >
> > > > > > RNBD can make double-queues for irq-mode and poll-mode.
> > > > > > For example, on 4-CPU system 8 request-queues are created,
> > > > > > 4 for irq-mode and 4 for poll-mode.
> > > > > > If the IO has HIPRI flag, the block-layer will call .poll function
> > > > > > of RNBD. Then IO is sent to the poll-mode queue.
> > > > > > Add optional nr_poll_queues argument for map_devices interface.
> > > > > >
> > > > > > To support polling of RNBD, RTRS client creates connections
> > > > > > for both of irq-mode and direct-poll-mode.
> > > > > >
> > > > > > For example, on 4-CPU system it could've create 5 connections:
> > > > > > con[0] => user message (softirq cq)
> > > > > > con[1:4] => softirq cq
> > > > > >
> > > > > > After this patch, it can create 9 connections:
> > > > > > con[0] => user message (softirq cq)
> > > > > > con[1:4] => softirq cq
> > > > > > con[5:8] => DIRECT-POLL cq
> > >
> > > <...>
> >
> > I am sorry that I don't understand exactly.
> > Do I need to change them to "con<5..8>"?
>
> No, I just removed not relevant text and replaced it with <...> in
> automatic way :).

Oh ;-)

>
> >
> >
> > >
> > > > > > +int rtrs_clt_rdma_cq_direct(struct rtrs_clt *clt, unsigned int index)
> > > > > > +{
> > > > > > +     int cnt;
> > > > > > +     struct rtrs_con *con;
> > > > > > +     struct rtrs_clt_sess *sess;
> > > > > > +     struct path_it it;
> > > > > > +
> > > > > > +     rcu_read_lock();
> > > > > > +     for (path_it_init(&it, clt);
> > > > > > +          (sess = it.next_path(&it)) && it.i < it.clt->paths_num; it.i++) {
> > > > > > +             if (unlikely(READ_ONCE(sess->state) != RTRS_CLT_CONNECTED))
> > > > >
> > > > > We talked about useless likely/unlikely in your workloads.
> > > >
> > > > Right, I've made a patch to remove all likely/unlikely
> > > > and will send with the next patch set.
> > >
> > > This specific line is "brand new". We don't add code that will be
> > > removed in next patch.
> >
> > Ah, ok. So you mean,
> > 1. remove unlikely from that line
> > 2. send a patch to remove all likely/unlikely for next round
> >
> > Am I right?
>
> Right

Thank you very much.
I will send V5 soon.

^ permalink raw reply	[flat|nested] 10+ messages in thread

end of thread, back to index

Thread overview: 10+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
     [not found] <20210414122402.203388-1-gi-oh.kim@ionos.com>
2021-04-14 12:23 ` [PATCHv4 for-next 08/19] block/rnbd-clt: Replace {NO_WAIT,WAIT} with RTRS_PERMIT_{WAIT,NOWAIT} Gioh Kim
2021-04-14 12:23 ` [PATCHv4 for-next 13/19] block/rnbd-clt: Support polling mode for IO latency optimization Gioh Kim
2021-04-18  8:36   ` Leon Romanovsky
2021-04-19  5:12     ` Gioh Kim
2021-04-19  5:20       ` Leon Romanovsky
2021-04-19  5:51         ` Gioh Kim
2021-04-19  6:09           ` Leon Romanovsky
2021-04-19  6:15             ` Gioh Kim
2021-04-14 12:23 ` [PATCHv4 for-next 15/19] block/rnbd-srv: Remove unused arguments of rnbd_srv_rdma_ev Gioh Kim
2021-04-14 12:24 ` [PATCHv4 for-next 17/19] block/rnbd-clt: Remove max_segment_size Gioh Kim

Linux-RDMA Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/linux-rdma/0 linux-rdma/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 linux-rdma linux-rdma/ https://lore.kernel.org/linux-rdma \
		linux-rdma@vger.kernel.org
	public-inbox-index linux-rdma

Example config snippet for mirrors

Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.kernel.vger.linux-rdma


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git