rdma: Add mlx5_dv RDMA provider (daee62a0) · Commits · Public Repositories / spdk

CHANGELOG.md

+7 −0

Original line number	Diff line number	Diff line
		@@ -15,6 +15,13 @@ The software accel engine implemenation has added support for CRC-32C.

		IDXD engine support for CRC-32C has been added.

		### rdma

		A new `rdma` library has been added. It is an abstraction layer over different RDMA providers.
		Two providers are available - verbs (used by default when RDMA is enabled or enabled explicitly
		using --with-rdma=verbs) and mlx5 Direct Verbs aka DV (enabled by --with-rdma=mlx5_dv).
		Using mlx5_dv requires libmlx5 installed on the system.

		## v20.04:

		### configuration

CONFIG

+1 −0

Original line number	Diff line number	Diff line
		@@ -100,6 +100,7 @@ CONFIG_FIO_SOURCE_DIR=/usr/src/fio
		CONFIG_RDMA=n
		CONFIG_RDMA_SEND_WITH_INVAL=n
		CONFIG_RDMA_SET_ACK_TIMEOUT=n
		CONFIG_RDMA_PROV=verbs

		# Enable NVMe Character Devices.
		CONFIG_NVME_CUSE=n

configure

+25 −1

Original line number	Diff line number	Diff line
		@@ -74,7 +74,8 @@ function usage()
		echo " rbd Build Ceph RBD bdev module."
		echo " No path required."
		echo " rdma Build RDMA transport for NVMf target and initiator."
		echo " No path required."
		echo " Accepts optional RDMA provider name. Can be \"verbs\" or \"mlx5_dv\"."
		echo " If no provider specified, \"verbs\" provider is used by default."
		echo " fc Build FC transport for NVMf target."
		echo " If an argument is provided, it is considered a directory containing"
		echo " libufc.a and fc_lld.h. Otherwise the regular system paths will"
		@@ -292,8 +293,13 @@ for i in "$@"; do
		--without-rbd)
		CONFIG[RBD]=n
		;;
		--with-rdma=*)
		CONFIG[RDMA]=y
		CONFIG[RDMA_PROV]=${i#*=}
		;;
		--with-rdma)
		CONFIG[RDMA]=y
		CONFIG[RDMA_PROV]="verbs"
		;;
		--without-rdma)
		CONFIG[RDMA]=n
		@@ -593,6 +599,11 @@ if [[ $sys_name == "FreeBSD" ]]; then
		fi

		if [ "${CONFIG[RDMA]}" = "y" ]; then
		if [[ ! "${CONFIG[RDMA_PROV]}" == "verbs" ]] && [[ ! "${CONFIG[RDMA_PROV]}" == "mlx5_dv" ]]; then
		echo "Invalid RDMA provider specified, must be \"verbs\" or \"mlx5_dv\""
		exit 1
		fi

		if ! echo -e '#include <infiniband/verbs.h>\n#include <rdma/rdma_verbs.h>\n' \
		'int main(void) { return 0; }\n' \
		\| ${BUILD_CMD[@]} -libverbs -lrdmacm - 2>/dev/null; then
		@@ -632,6 +643,19 @@ than or equal to 4.14 will see significantly reduced performance.
		echo "RDMA_OPTION_ID_ACK_TIMEOUT is not supported"
		fi

		if [ "${CONFIG[RDMA_PROV]}" == "mlx5_dv" ]; then
		if ! echo -e '#include <spdk/stdinc.h>\n' \
		'#include <infiniband/mlx5dv.h>\n' \
		'#include <rdma/rdma_cma.h>\n' \
		'int main(void) { return rdma_establish(NULL) \|\| ' \
		'!!IBV_QP_INIT_ATTR_SEND_OPS_FLAGS \|\| !!MLX5_OPCODE_RDMA_WRITE; }\n' \
		\| ${BUILD_CMD[@]} -lmlx5 -I${rootdir}/include -c - 2>/dev/null; then
		echo "mlx5_dv provider is not supported"
		exit 1
		fi
		fi

		echo "Using "${CONFIG[RDMA_PROV]}" RDMA provider"
		fi

		if [[ "${CONFIG[FC]}" = "y" ]]; then

include/spdk_internal/rdma.h

+9 −0

Original line number	Diff line number	Diff line
		@@ -45,6 +45,7 @@ struct spdk_rdma_qp_init_attr {
		struct ibv_srq *srq;
		struct ibv_qp_cap cap;
		struct ibv_pd *pd;
		bool initiator_side;
		};

		struct spdk_rdma_qp {
		@@ -61,6 +62,14 @@ struct spdk_rdma_qp {
		struct spdk_rdma_qp spdk_rdma_qp_create(struct rdma_cm_id cm_id,
		struct spdk_rdma_qp_init_attr *qp_attr);

		/**
		* Complete the connection process, must be called by the active
		* side (NVMEoF initiator) upon receipt RDMA_CM_EVENT_CONNECT_RESPONSE
		* @param spdk_rdma_qp pointer to a qpair
		* @return 0 on success, errno on failure
		*/
		int spdk_rdma_qp_complete_connect(struct spdk_rdma_qp *spdk_rdma_qp);

		/**
		* Destroy RDMA provider specific qpair
		* \param spdk_rdma_qp Pointer to qpair to be destroyed

lib/nvme/nvme_rdma.c

+21 −7

Original line number	Diff line number	Diff line
		@@ -332,13 +332,14 @@ nvme_rdma_qpair_process_cm_event(struct nvme_rdma_qpair *rqpair)
		break;
		case RDMA_CM_EVENT_CONNECT_REQUEST:
		break;
		case RDMA_CM_EVENT_CONNECT_RESPONSE:
		break;
		case RDMA_CM_EVENT_CONNECT_ERROR:
		break;
		case RDMA_CM_EVENT_UNREACHABLE:
		case RDMA_CM_EVENT_REJECTED:
		break;
		case RDMA_CM_EVENT_CONNECT_RESPONSE:
		rc = spdk_rdma_qp_complete_connect(rqpair->rdma_qp);
		/* fall through */
		case RDMA_CM_EVENT_ESTABLISHED:
		accept_data = (struct spdk_nvmf_rdma_accept_private_data *)event->param.conn.private_data;
		if (accept_data == NULL) {
		@@ -443,6 +444,13 @@ nvme_rdma_validate_cm_event(enum rdma_cm_event_type expected_evt_type,
		*/
		if (reaped_evt->event == RDMA_CM_EVENT_REJECTED && reaped_evt->status == 10) {
		rc = -ESTALE;
		} else if (reaped_evt->event == RDMA_CM_EVENT_CONNECT_RESPONSE) {
		/*
		* If we are using a qpair which is not created using rdma cm API
		* then we will receive RDMA_CM_EVENT_CONNECT_RESPONSE instead of
		* RDMA_CM_EVENT_ESTABLISHED.
		*/
		return 0;
		}
		break;
		default:
		@@ -530,6 +538,7 @@ nvme_rdma_qpair_init(struct nvme_rdma_qpair *rqpair)
		attr.cap.max_recv_wr = rqpair->num_entries; /* RECV operations */
		attr.cap.max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, dev_attr.max_sge);
		attr.cap.max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, dev_attr.max_sge);
		attr.initiator_side = true;

		rqpair->rdma_qp = spdk_rdma_qp_create(rqpair->cm_id, &attr);

		@@ -541,7 +550,7 @@ nvme_rdma_qpair_init(struct nvme_rdma_qpair *rqpair)
		rqpair->max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, attr.cap.max_send_sge);
		rqpair->max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, attr.cap.max_recv_sge);

		rctrlr->pd = rqpair->cm_id->qp->pd;
		rctrlr->pd = rqpair->rdma_qp->qp->pd;

		rqpair->cm_id->context = &rqpair->qpair;

		@@ -555,7 +564,7 @@ nvme_rdma_qpair_submit_sends(struct nvme_rdma_qpair *rqpair)
		int rc;

		if (rqpair->sends_to_post.first) {
		rc = ibv_post_send(rqpair->cm_id->qp, rqpair->sends_to_post.first, &bad_send_wr);
		rc = ibv_post_send(rqpair->rdma_qp->qp, rqpair->sends_to_post.first, &bad_send_wr);
		if (spdk_unlikely(rc)) {
		SPDK_ERRLOG("Failed to post WRs on send queue, errno %d (%s), bad_wr %p\n",
		rc, spdk_strerror(rc), bad_send_wr);
		@@ -579,7 +588,7 @@ nvme_rdma_qpair_submit_recvs(struct nvme_rdma_qpair *rqpair)
		int rc;

		if (rqpair->recvs_to_post.first) {
		rc = ibv_post_recv(rqpair->cm_id->qp, rqpair->recvs_to_post.first, &bad_recv_wr);
		rc = ibv_post_recv(rqpair->rdma_qp->qp, rqpair->recvs_to_post.first, &bad_recv_wr);
		if (spdk_unlikely(rc)) {
		SPDK_ERRLOG("Failed to post WRs on receive queue, errno %d (%s), bad_wr %p\n",
		rc, spdk_strerror(rc), bad_recv_wr);
		@@ -933,6 +942,11 @@ nvme_rdma_connect(struct nvme_rdma_qpair *rqpair)
		param.retry_count = ctrlr->opts.transport_retry_count;
		param.rnr_retry_count = 7;

		/* Fields below are ignored by rdma cm if qpair has been
		* created using rdma cm API. */
		param.srq = 0;
		param.qp_num = rqpair->rdma_qp->qp->qp_num;

		ret = rdma_connect(rqpair->cm_id, &param);
		if (ret) {
		SPDK_ERRLOG("nvme rdma connect error\n");
		@@ -944,7 +958,7 @@ nvme_rdma_connect(struct nvme_rdma_qpair *rqpair)
		SPDK_NOTICELOG("Received a stale connection notice during connection.\n");
		return -EAGAIN;
		} else if (ret) {
		SPDK_ERRLOG("RDMA connect error\n");
		SPDK_ERRLOG("RDMA connect error %d\n", ret);
		return -1;
		} else {
		return 0;
		@@ -1033,7 +1047,7 @@ nvme_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2)
		static int
		nvme_rdma_register_mem(struct nvme_rdma_qpair *rqpair)
		{
		struct ibv_pd *pd = rqpair->cm_id->qp->pd;
		struct ibv_pd *pd = rqpair->rdma_qp->qp->pd;
		struct spdk_nvme_rdma_mr_map *mr_map;
		const struct spdk_mem_map_ops nvme_rdma_map_ops = {
		.notify_cb = nvme_rdma_mr_map_notify,