Commit db5c3ce3 authored by Xiaodong Liu's avatar Xiaodong Liu Committed by Jim Harris
Browse files

nvmf/rdma: dynamically enlarge CQ size 

Assigned CQ size when creating CQ may run over due to
heavy workload with too many qpairs. Enlarge it dynamically
can prevent IBV_EVENT_CQ_ERR caused by CQ's runover.
This patch fixes issue #498:
https://github.com/spdk/spdk/issues/498



Change-Id: I6c2d7194d4147d812d49d4fe787fcba5c6bbede9
Signed-off-by: default avatarXiaodong Liu <xiaodong.liu@intel.com>
Reviewed-on: https://review.gerrithub.io/c/440853


Reviewed-by: default avatarShuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: default avatarSeth Howell <seth.howell5141@gmail.com>
Reviewed-by: default avatarBen Walker <benjamin.walker@intel.com>
Reviewed-by: default avatarJim Harris <james.r.harris@intel.com>
Tested-by: default avatarSPDK CI Jenkins <sys_sgci@intel.com>
Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
parent 0656ba94

lib/nvmf/rdma.c

+46 −3
Original line number Diff line number Diff line
@@ -59,7 +59,8 @@ 
#define NVMF_DEFAULT_RX_SGE		2



/* The RDMA completion queue size */

#define NVMF_RDMA_CQ_SIZE	4096

#define DEFAULT_NVMF_RDMA_CQ_SIZE	4096

#define MAX_WR_PER_QP(queue_depth)	(queue_depth * 3 + 2)



enum spdk_nvmf_rdma_request_state {

	/* The request is not currently in use */
@@ -329,6 +330,8 @@ struct spdk_nvmf_rdma_poller { 
	struct spdk_nvmf_rdma_device		*device;

	struct spdk_nvmf_rdma_poll_group	*group;



	int					num_cqe;

	int					required_num_wr;

	struct ibv_cq				*cq;



	TAILQ_HEAD(, spdk_nvmf_rdma_qpair)	qpairs;
@@ -659,6 +662,10 @@ spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair) 
	if (rqpair->cm_id) {

		rdma_destroy_qp(rqpair->cm_id);

		rdma_destroy_id(rqpair->cm_id);



		if (rqpair->poller) {

			rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth);

		}

	}



	if (rqpair->mgmt_channel) {
@@ -679,7 +686,8 @@ spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 
{

	struct spdk_nvmf_rdma_transport *rtransport;

	struct spdk_nvmf_rdma_qpair	*rqpair;

	int				rc, i;

	struct spdk_nvmf_rdma_poller	*rpoller;

	int				rc, i, num_cqe, required_num_wr;;

	struct spdk_nvmf_rdma_recv	*rdma_recv;

	struct spdk_nvmf_rdma_request	*rdma_req;

	struct spdk_nvmf_transport	*transport;
@@ -703,6 +711,38 @@ spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 
	ibv_init_attr.cap.max_send_sge	= spdk_min(device->attr.max_sge, NVMF_DEFAULT_TX_SGE);

	ibv_init_attr.cap.max_recv_sge	= spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE);



	/* Enlarge CQ size dynamically */

	rpoller = rqpair->poller;

	required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth);

	num_cqe = rpoller->num_cqe;

	if (num_cqe < required_num_wr) {

		num_cqe = spdk_max(num_cqe * 2, required_num_wr);

		num_cqe = spdk_min(num_cqe, device->attr.max_cqe);

	}



	if (rpoller->num_cqe != num_cqe) {

		if (required_num_wr > device->attr.max_cqe) {

			SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n",

				    required_num_wr, device->attr.max_cqe);

			rdma_destroy_id(rqpair->cm_id);

			rqpair->cm_id = NULL;

			spdk_nvmf_rdma_qpair_destroy(rqpair);

			return -1;

		}



		SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe);

		rc = ibv_resize_cq(rpoller->cq, num_cqe);

		if (rc) {

			SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno));

			rdma_destroy_id(rqpair->cm_id);

			rqpair->cm_id = NULL;

			spdk_nvmf_rdma_qpair_destroy(rqpair);

			return -1;

		}



		rpoller->num_cqe = num_cqe;

	}



	rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &ibv_init_attr);

	if (rc) {

		SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno));
@@ -712,6 +752,8 @@ spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair) 
		return -1;

	}



	rpoller->required_num_wr = required_num_wr;



	rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge);

	rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge);

	spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0);
@@ -2423,12 +2465,13 @@ spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport) 


		TAILQ_INIT(&poller->qpairs);



		poller->cq = ibv_create_cq(device->context, NVMF_RDMA_CQ_SIZE, poller, NULL, 0);

		poller->cq = ibv_create_cq(device->context, DEFAULT_NVMF_RDMA_CQ_SIZE, poller, NULL, 0);

		if (!poller->cq) {

			SPDK_ERRLOG("Unable to create completion queue\n");

			free(poller);

			goto err_exit;

		}

		poller->num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE;



		TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link);

	}