nvmf: add spdk_nvmf_get_optimal_poll_group

This patch is used to do the following work:

1 It is optimized for NVMe/TCP transport. If the qpair's
socket has same NAPI_ID, then the qpair will be handled
by the same polling group.

2. We add a new connection scheduling strategy, named as
ConnectionScheduler in the configuration file. It will be
used to input different scheduler according to the customers'
input.

Signed-off-by: Ziye Yang <ziye.yang@intel.com>
Change-Id: Ifc9246eece0da69bdd39fd63bfdefff18be64132
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/454550
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>

CHANGELOG.md

@@ -25,6 +25,11 @@ Shared receive queue can now be disabled even for NICs that support it using the
 `nvmf_create_transport` RPC method parameter `no_srq`. The actual use of a shared
 receive queue is predicated on hardware support when this flag is not used.
 
+spdk_nvmf_get_optimal_poll_group was added, which is used to return the optimal
+poll group for the qpair. And `ConnectionScheduler` configuration is added into the
+[Nvmf] section in etc/spdk/nvmf.conf.in to demonstrate how to configure the connection
+scheduling strategy among different spdk threads.
+
 ### notify
 
 The function `spdk_notify_get_types()` and `spdk_notify_get_events()` were

etc/spdk/nvmf.conf.in

@@ -65,6 +65,15 @@
   # poll. Units in microseconds.
   AcceptorPollRate 10000
 
+  # Set how the connection is scheduled among multiple threads, current supported string value are
+  # "RoundRobin", "Host", "Transport".
+  # RoundRobin: Schedule the connection with roundrobin manner.
+  # Host: Schedule the connection according to host IP.
+  # Transport: Schedule the connection according to the transport characteristics.
+  #  For example, for  TCP transport, we can schedule the connection according to socket NAPI_ID info.
+  #  The connection which has the same socket NAPI_ID info will be grouped in the same polling group.
+  ConnectionScheduler RoundRobin
+
 # One valid transport type must be set in each [Transport].
 # The first is the case of RDMA transport and the second is the case of TCP transport.
 [Transport]

include/spdk/nvmf.h

@@ -163,6 +163,15 @@ void spdk_nvmf_tgt_accept(struct spdk_nvmf_tgt *tgt, new_qpair_fn cb_fn);
  */
 struct spdk_nvmf_poll_group *spdk_nvmf_poll_group_create(struct spdk_nvmf_tgt *tgt);
 
+/**
+ * Get optimal nvmf poll group for the qpair.
+ *
+ * \param qpair Requested qpair
+ *
+ * \return a poll group on success, or NULL on failure.
+ */
+struct spdk_nvmf_poll_group *spdk_nvmf_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair);
+
 /**
  * Destroy a poll group.
  *

lib/event/subsystems/nvmf/conf.c

@@ -101,16 +101,42 @@ spdk_nvmf_read_config_file_tgt_max_subsystems(struct spdk_conf_section *sp,
 	}
 }
 
-static void
+static int
 spdk_nvmf_read_config_file_tgt_conf(struct spdk_conf_section *sp,
 				    struct spdk_nvmf_tgt_conf *conf)
 {
 	int acceptor_poll_rate;
+	const char *conn_scheduler;
+	int rc = 0;
 
 	acceptor_poll_rate = spdk_conf_section_get_intval(sp, "AcceptorPollRate");
 	if (acceptor_poll_rate >= 0) {
 		conf->acceptor_poll_rate = acceptor_poll_rate;
 	}
+
+	conn_scheduler = spdk_conf_section_get_val(sp, "ConnectionScheduler");
+
+	if (conn_scheduler) {
+		if (strcasecmp(conn_scheduler, "RoundRobin") == 0) {
+			conf->conn_sched = CONNECT_SCHED_ROUND_ROBIN;
+		} else if (strcasecmp(conn_scheduler, "Host") == 0) {
+			conf->conn_sched = CONNECT_SCHED_HOST_IP;
+		} else if (strcasecmp(conn_scheduler, "Transport") == 0) {
+			conf->conn_sched = CONNECT_SCHED_TRANSPORT_OPTIMAL_GROUP;
+		} else {
+			SPDK_ERRLOG("The valid value of ConnectionScheduler should be:\n"
+				    "\t RoundRobin\n"
+				    "\t Host\n"
+				    "\t Transport\n");
+			rc = -1;
+		}
+
+	} else {
+		SPDK_NOTICELOG("The value of ConnectionScheduler is not configured,\n"
+			       "we will use RoundRobin as the default scheduler\n");
+	}
+
+	return rc;
 }
 
 static int
@@ -132,6 +158,7 @@ spdk_nvmf_parse_tgt_conf(void)
 {
 	struct spdk_nvmf_tgt_conf *conf;
 	struct spdk_conf_section *sp;
+	int rc;
 
 	conf = calloc(1, sizeof(*conf));
 	if (!conf) {
@@ -144,7 +171,11 @@ spdk_nvmf_parse_tgt_conf(void)
 
 	sp = spdk_conf_find_section(NULL, "Nvmf");
 	if (sp != NULL) {
-		spdk_nvmf_read_config_file_tgt_conf(sp, conf);
+		rc = spdk_nvmf_read_config_file_tgt_conf(sp, conf);
+		if (rc) {
+			free(conf);
+			return NULL;
+		}
 	}
 
 	return conf;

lib/event/subsystems/nvmf/event_nvmf.h

@@ -48,6 +48,7 @@
 enum spdk_nvmf_connect_sched {
 	CONNECT_SCHED_ROUND_ROBIN = 0,
 	CONNECT_SCHED_HOST_IP,
+	CONNECT_SCHED_TRANSPORT_OPTIMAL_GROUP,
 };
 
 struct spdk_nvmf_tgt_conf {

lib/event/subsystems/nvmf/nvmf_tgt.c

@@ -123,6 +123,30 @@ spdk_nvmf_get_next_pg(void)
 	return pg;
 }
 
+static struct nvmf_tgt_poll_group *
+spdk_nvmf_get_optimal_pg(struct spdk_nvmf_qpair *qpair)
+{
+	struct nvmf_tgt_poll_group *pg, *_pg = NULL;
+	struct spdk_nvmf_poll_group *group = spdk_nvmf_get_optimal_poll_group(qpair);
+
+	if (group == NULL) {
+		_pg = spdk_nvmf_get_next_pg();
+		goto end;
+	}
+
+	TAILQ_FOREACH(pg, &g_poll_groups, link) {
+		if (pg->group == group) {
+			_pg = pg;
+			break;
+		}
+
+	}
+
+end:
+	assert(_pg != NULL);
+	return _pg;
+}
+
 static void
 nvmf_tgt_remove_host_trid(struct spdk_nvmf_qpair *qpair)
 {
@@ -193,6 +217,9 @@ nvmf_tgt_get_pg(struct spdk_nvmf_qpair *qpair)
 			TAILQ_INSERT_TAIL(&g_nvmf_tgt_host_trids, new_trid, link);
 		}
 		break;
+	case CONNECT_SCHED_TRANSPORT_OPTIMAL_GROUP:
+		pg = spdk_nvmf_get_optimal_pg(qpair);
+		break;
 	case CONNECT_SCHED_ROUND_ROBIN:
 	default:
 		pg = spdk_nvmf_get_next_pg();

lib/nvmf/nvmf.c

@@ -855,6 +855,7 @@ spdk_nvmf_poll_group_add_transport(struct spdk_nvmf_poll_group *group,
 		return -1;
 	}
 
+	tgroup->group = group;
 	TAILQ_INSERT_TAIL(&group->tgroups, tgroup, link);
 
 	return 0;
@@ -1211,3 +1212,18 @@ fini:
 		cb_fn(cb_arg, rc);
 	}
 }
+
+
+struct spdk_nvmf_poll_group *
+spdk_nvmf_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair)
+{
+	struct spdk_nvmf_transport_poll_group *tgroup;
+
+	tgroup = spdk_nvmf_transport_get_optimal_poll_group(qpair->transport, qpair);
+
+	if (tgroup == NULL) {
+		return NULL;
+	}
+
+	return tgroup->group;
+}

lib/nvmf/nvmf_internal.h

@@ -109,6 +109,7 @@ struct spdk_nvmf_transport_poll_group {
 	STAILQ_HEAD(, spdk_nvmf_transport_pg_cache_buf)			buf_cache;
 	uint32_t							buf_cache_count;
 	uint32_t							buf_cache_size;
+	struct spdk_nvmf_poll_group					*group;
 	TAILQ_ENTRY(spdk_nvmf_transport_poll_group)			link;
 };