nvmf/rdma: Add NVMf RDMA transport pending statistics

This patch adds statistics for pending state in NVMf RDMA subsytem
which may help to detect lack of resources and adjust configuration
correctly.

Signed-off-by: Evgeniy Kochetov <evgeniik@mellanox.com>
Change-Id: I9560d931c0dfb469659be42e13b8302c52912420
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/452300
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Paul Luse <paul.e.luse@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>

doc/jsonrpc.md

@@ -4177,16 +4177,23 @@ Example response:
         "transports": [
           {
             "trtype": "RDMA",
+            "pending_data_buffer": 0,
             "devices": [
               {
                 "name": "mlx5_1",
                 "polls": 1536729,
-                "completions": 0
+                "completions": 0,
+                "pending_free_request": 0,
+                "pending_rdma_read": 0,
+                "pending_rdma_write": 0
               },
               {
                 "name": "mlx5_0",
                 "polls": 1536729,
-                "completions": 18667357
+                "completions": 18667357,
+                "pending_free_request": 0,
+                "pending_rdma_read": 337602,
+                "pending_rdma_write": 0
               }
             ]
           }

include/spdk/nvmf.h

@@ -89,12 +89,16 @@ struct spdk_nvmf_rdma_device_stat {
 	const char *name;
 	uint64_t polls;
 	uint64_t completions;
+	uint64_t pending_free_request;
+	uint64_t pending_rdma_read;
+	uint64_t pending_rdma_write;
 };
 
 struct spdk_nvmf_transport_poll_group_stat {
 	spdk_nvme_transport_type_t trtype;
 	union {
 		struct {
+			uint64_t pending_data_buffer;
 			uint64_t num_devices;
 			struct spdk_nvmf_rdma_device_stat *devices;
 		} rdma;

lib/event/subsystems/nvmf/nvmf_rpc.c

@@ -1619,12 +1619,19 @@ write_nvmf_transport_stats(struct spdk_json_write_ctx *w,
 				     spdk_nvme_transport_id_trtype_str(stat->trtype));
 	switch (stat->trtype) {
 	case SPDK_NVME_TRANSPORT_RDMA:
+		spdk_json_write_named_uint64(w, "pending_data_buffer", stat->rdma.pending_data_buffer);
 		spdk_json_write_named_array_begin(w, "devices");
 		for (i = 0; i < stat->rdma.num_devices; ++i) {
 			spdk_json_write_object_begin(w);
 			spdk_json_write_named_string(w, "name", stat->rdma.devices[i].name);
 			spdk_json_write_named_uint64(w, "polls", stat->rdma.devices[i].polls);
 			spdk_json_write_named_uint64(w, "completions", stat->rdma.devices[i].completions);
+			spdk_json_write_named_uint64(w, "pending_free_request",
+						     stat->rdma.devices[i].pending_free_request);
+			spdk_json_write_named_uint64(w, "pending_rdma_read",
+						     stat->rdma.devices[i].pending_rdma_read);
+			spdk_json_write_named_uint64(w, "pending_rdma_write",
+						     stat->rdma.devices[i].pending_rdma_write);
 			spdk_json_write_object_end(w);
 		}
 		spdk_json_write_array_end(w);

lib/nvmf/rdma.c

@@ -412,6 +412,9 @@ struct spdk_nvmf_rdma_qpair {
 struct spdk_nvmf_rdma_poller_stat {
 	uint64_t				completions;
 	uint64_t				polls;
+	uint64_t				pending_free_request;
+	uint64_t				pending_rdma_read;
+	uint64_t				pending_rdma_write;
 };
 
 struct spdk_nvmf_rdma_poller {
@@ -440,6 +443,10 @@ struct spdk_nvmf_rdma_poller {
 	TAILQ_ENTRY(spdk_nvmf_rdma_poller)	link;
 };
 
+struct spdk_nvmf_rdma_poll_group_stat {
+	uint64_t				pending_data_buffer;
+};
+
 struct spdk_nvmf_rdma_poll_group {
 	struct spdk_nvmf_transport_poll_group	group;
 
@@ -447,6 +454,8 @@ struct spdk_nvmf_rdma_poll_group {
 	STAILQ_HEAD(, spdk_nvmf_rdma_request)	pending_data_buf_queue;
 
 	TAILQ_HEAD(, spdk_nvmf_rdma_poller)	pollers;
+
+	struct spdk_nvmf_rdma_poll_group_stat	stat;
 };
 
 /* Assuming rdma_cm uses just one protection domain per ibv_context. */
@@ -1950,6 +1959,7 @@ spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
 
 			if (!rdma_req->req.data) {
 				/* No buffers available. */
+				rgroup->stat.pending_data_buffer++;
 				break;
 			}
 
@@ -1977,6 +1987,7 @@ spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
 			if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth
 			    || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) {
 				/* We can only have so many WRs outstanding. we have to wait until some finish. */
+				rqpair->poller->stat.pending_rdma_read++;
 				break;
 			}
 
@@ -2031,6 +2042,7 @@ spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
 			    rqpair->max_send_depth) {
 				/* We can only have so many WRs outstanding. we have to wait until some finish.
 				 * +1 since each request has an additional wr in the resp. */
+				rqpair->poller->stat.pending_rdma_write++;
 				break;
 			}
 
@@ -2616,6 +2628,9 @@ spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport
 			break;
 		}
 	}
+	if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) {
+		rqpair->poller->stat.pending_free_request++;
+	}
 }
 
 static void
@@ -3695,6 +3710,7 @@ spdk_nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt,
 				return -ENOMEM;
 			}
 
+			(*stat)->rdma.pending_data_buffer = rgroup->stat.pending_data_buffer;
 			(*stat)->rdma.num_devices = num_devices;
 			num_devices = 0;
 			TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
@@ -3702,6 +3718,9 @@ spdk_nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt,
 				device_stat->name = ibv_get_device_name(rpoller->device->context->device);
 				device_stat->polls = rpoller->stat.polls;
 				device_stat->completions = rpoller->stat.completions;
+				device_stat->pending_free_request = rpoller->stat.pending_free_request;
+				device_stat->pending_rdma_read = rpoller->stat.pending_rdma_read;
+				device_stat->pending_rdma_write = rpoller->stat.pending_rdma_write;
 			}
 			return 0;
 		}