numam-spdk/lib/nvmf/rdma.c

/*-
 *   BSD LICENSE
 *
 *   Copyright (c) Intel Corporation.
 *   All rights reserved.
 *
 *   Redistribution and use in source and binary forms, with or without
 *   modification, are permitted provided that the following conditions
 *   are met:
 *
 *     * Redistributions of source code must retain the above copyright
 *       notice, this list of conditions and the following disclaimer.
 *     * Redistributions in binary form must reproduce the above copyright
 *       notice, this list of conditions and the following disclaimer in
 *       the documentation and/or other materials provided with the
 *       distribution.
 *     * Neither the name of Intel Corporation nor the names of its
 *       contributors may be used to endorse or promote products derived
 *       from this software without specific prior written permission.
 *
 *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
 *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
 *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
 *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
 *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

#include <arpa/inet.h>
#include <fcntl.h>
#include <infiniband/verbs.h>
#include <rdma/rdma_cma.h>
#include <rdma/rdma_verbs.h>
#include <unistd.h>
#include <stdio.h>
#include <stdint.h>

#include <rte_config.h>
#include <rte_debug.h>
#include <rte_cycles.h>
#include <rte_timer.h>
#include <rte_lcore.h>
#include <rte_malloc.h>

#include "nvmf_internal.h"
#include "request.h"
#include "session.h"
#include "subsystem.h"
#include "transport.h"
#include "spdk/assert.h"
#include "spdk/log.h"
#include "spdk/nvmf_spec.h"
#include "spdk/trace.h"

#define ACCEPT_TIMEOUT (rte_get_timer_hz() >> 10) /* ~1ms */

/*
 RDMA Connection Resouce Defaults
 */
#define NVMF_DEFAULT_TX_SGE		1
#define NVMF_DEFAULT_RX_SGE		2

struct spdk_nvmf_rdma_request {
	struct spdk_nvmf_request	req;

	/* In Capsule data buffer */
	void				*buf;

};

struct spdk_nvmf_rdma_conn {
	struct spdk_nvmf_conn			conn;

	struct rdma_cm_id			*cm_id;
	struct ibv_context			*ctx;
	struct ibv_comp_channel			*comp_channel;
	struct ibv_cq				*cq;
	struct ibv_qp				*qp;

	uint16_t				queue_depth;
	int					outstanding_reqs;

	/* Array of size "queue_depth" containing RDMA requests. */
	struct spdk_nvmf_rdma_request		*reqs;

	/* Array of size "queue_depth" containing 64 byte capsules
	 * used for receive.
	 */
	union nvmf_h2c_msg			*cmds;
	struct ibv_mr				*cmds_mr;

	/* Array of size "queue_depth" containing 16 byte completions
	 * to be sent back to the user.
	 */
	union nvmf_c2h_msg			*cpls;
	struct ibv_mr				*cpls_mr;

	/* Array of size "queue_depth * InCapsuleDataSize" containing
	 * buffers to be used for in capsule data. TODO: Currently, all data
	 * is in capsule.
	 */
	void					*bufs;
	struct ibv_mr				*bufs_mr;

	TAILQ_ENTRY(spdk_nvmf_rdma_conn)	link;
};

/* List of RDMA connections that have not yet received a CONNECT capsule */
static TAILQ_HEAD(, spdk_nvmf_rdma_conn) g_pending_conns = TAILQ_HEAD_INITIALIZER(g_pending_conns);

struct spdk_nvmf_rdma {
	struct rte_timer		acceptor_timer;
	struct rdma_event_channel	*acceptor_event_channel;
	struct rdma_cm_id		*acceptor_listen_id;
};

static struct spdk_nvmf_rdma g_rdma = { };

static inline struct spdk_nvmf_rdma_conn *
get_rdma_conn(struct spdk_nvmf_conn *conn)
{
	return (struct spdk_nvmf_rdma_conn *)((uintptr_t)conn + offsetof(struct spdk_nvmf_rdma_conn, conn));
}

static inline struct spdk_nvmf_rdma_request *
get_rdma_req(struct spdk_nvmf_request *req)
{
	return (struct spdk_nvmf_rdma_request *)((uintptr_t)req + offsetof(struct spdk_nvmf_rdma_request,
			req));
}

static int nvmf_post_rdma_recv(struct spdk_nvmf_request *req);

static void
spdk_nvmf_rdma_conn_destroy(struct spdk_nvmf_rdma_conn *rdma_conn)
{
	if (rdma_conn->cmds_mr) {
		rdma_dereg_mr(rdma_conn->cmds_mr);
	}

	if (rdma_conn->cpls_mr) {
		rdma_dereg_mr(rdma_conn->cpls_mr);
	}

	if (rdma_conn->bufs_mr) {
		rdma_dereg_mr(rdma_conn->bufs_mr);
	}

	if (rdma_conn->cm_id) {
		rdma_destroy_qp(rdma_conn->cm_id);
	}

	if (rdma_conn->cq) {
		ibv_destroy_cq(rdma_conn->cq);
	}

	if (rdma_conn->comp_channel) {
		ibv_destroy_comp_channel(rdma_conn->comp_channel);
	}

	if (rdma_conn->cm_id) {
		rdma_destroy_id(rdma_conn->cm_id);
	}

	/* Free all memory */
	rte_free(rdma_conn->cmds);
	rte_free(rdma_conn->cpls);
	rte_free(rdma_conn->bufs);
	free(rdma_conn->reqs);
	free(rdma_conn);
}

static struct spdk_nvmf_rdma_conn *
spdk_nvmf_rdma_conn_create(struct rdma_cm_id *id, uint16_t queue_depth)
{
	struct spdk_nvmf_rdma_conn	*rdma_conn;
	struct spdk_nvmf_conn		*conn;
	int				rc, i;
	struct ibv_qp_init_attr		attr;
	struct spdk_nvmf_rdma_request	*rdma_req;

	rdma_conn = calloc(1, sizeof(struct spdk_nvmf_rdma_conn));
	if (rdma_conn == NULL) {
		SPDK_ERRLOG("Could not allocate new connection.\n");
		return NULL;
	}

	rdma_conn->queue_depth = queue_depth;
	rdma_conn->ctx = id->verbs;
	rdma_conn->cm_id = id;

	rdma_conn->comp_channel = ibv_create_comp_channel(id->verbs);
	if (!rdma_conn->comp_channel) {
		SPDK_ERRLOG("create completion channel error!\n");
		spdk_nvmf_rdma_conn_destroy(rdma_conn);
		return NULL;
	}

	rc = fcntl(rdma_conn->comp_channel->fd, F_SETFL, O_NONBLOCK);
	if (rc < 0) {
		SPDK_ERRLOG("fcntl to set comp channel to non-blocking failed\n");
		spdk_nvmf_rdma_conn_destroy(rdma_conn);
		return NULL;
	}

	/*
	 * Size the CQ to handle completions for RECV, SEND, and either READ or WRITE.
	 */
	rdma_conn->cq = ibv_create_cq(id->verbs, (queue_depth * 3), rdma_conn, rdma_conn->comp_channel,
				      0);
	if (!rdma_conn->cq) {
		SPDK_ERRLOG("create cq error!\n");
		spdk_nvmf_rdma_conn_destroy(rdma_conn);
		return NULL;
	}

	memset(&attr, 0, sizeof(struct ibv_qp_init_attr));
	attr.qp_type		= IBV_QPT_RC;
	attr.send_cq		= rdma_conn->cq;
	attr.recv_cq		= rdma_conn->cq;
	attr.cap.max_send_wr	= rdma_conn->queue_depth * 2; /* SEND, READ, and WRITE operations */
	attr.cap.max_recv_wr	= rdma_conn->queue_depth; /* RECV operations */
	attr.cap.max_send_sge	= NVMF_DEFAULT_TX_SGE;
	attr.cap.max_recv_sge	= NVMF_DEFAULT_RX_SGE;

	rc = rdma_create_qp(rdma_conn->cm_id, NULL, &attr);
	if (rc) {
		SPDK_ERRLOG("rdma_create_qp failed\n");
		spdk_nvmf_rdma_conn_destroy(rdma_conn);
		return NULL;
	}
	rdma_conn->qp = rdma_conn->cm_id->qp;

	conn = &rdma_conn->conn;
	conn->transport = &spdk_nvmf_transport_rdma;
	id->context = conn;

	SPDK_TRACELOG(SPDK_TRACE_RDMA, "New RDMA Connection: %p\n", conn);

	rdma_conn->reqs = calloc(rdma_conn->queue_depth, sizeof(*rdma_conn->reqs));
	rdma_conn->cmds = rte_calloc("nvmf_rdma_cmd", rdma_conn->queue_depth,
				     sizeof(*rdma_conn->cmds), 0);
	rdma_conn->cpls = rte_calloc("nvmf_rdma_cpl", rdma_conn->queue_depth,
				     sizeof(*rdma_conn->cpls), 0);
	rdma_conn->bufs = rte_calloc("nvmf_rdma_buf", rdma_conn->queue_depth,
				     g_nvmf_tgt.in_capsule_data_size, 0);
	if (!rdma_conn->reqs || !rdma_conn->cmds || !rdma_conn->cpls || !rdma_conn->bufs) {
		SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n");
		spdk_nvmf_rdma_conn_destroy(rdma_conn);
		return NULL;
	}

	rdma_conn->cmds_mr = rdma_reg_msgs(rdma_conn->cm_id, rdma_conn->cmds,
					   queue_depth * sizeof(*rdma_conn->cmds));
	rdma_conn->cpls_mr = rdma_reg_msgs(rdma_conn->cm_id, rdma_conn->cpls,
					   queue_depth * sizeof(*rdma_conn->cpls));
	rdma_conn->bufs_mr = rdma_reg_msgs(rdma_conn->cm_id, rdma_conn->bufs,
					   rdma_conn->queue_depth * g_nvmf_tgt.in_capsule_data_size);
	if (!rdma_conn->cmds_mr || !rdma_conn->cpls_mr || !rdma_conn->bufs_mr) {
		SPDK_ERRLOG("Unable to register required memory for RDMA queue.\n");
		spdk_nvmf_rdma_conn_destroy(rdma_conn);
		return NULL;
	}

	for (i = 0; i < queue_depth; i++) {
		rdma_req = &rdma_conn->reqs[i];

		rdma_req->buf = (void *)((uintptr_t)rdma_conn->bufs + (i * g_nvmf_tgt.in_capsule_data_size));
		rdma_req->req.cmd = &rdma_conn->cmds[i];
		rdma_req->req.rsp = &rdma_conn->cpls[i];
		rdma_req->req.conn = &rdma_conn->conn;

		if (nvmf_post_rdma_recv(&rdma_req->req)) {
			SPDK_ERRLOG("Unable to post capsule for RDMA RECV\n");
			spdk_nvmf_rdma_conn_destroy(rdma_conn);
			return NULL;
		}
	}

	return rdma_conn;
}

static void
nvmf_trace_ibv_sge(struct ibv_sge *sg_list)
{
	if (sg_list) {
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "local addr %p length 0x%x lkey 0x%x\n",
			      (void *)sg_list->addr, sg_list->length, sg_list->lkey);
	}
}

static void
nvmf_ibv_send_wr_init(struct ibv_send_wr *wr,
		      struct spdk_nvmf_request *req,
		      struct ibv_sge *sg_list,
		      enum ibv_wr_opcode opcode,
		      int send_flags)
{
	struct spdk_nvmf_rdma_request *rdma_req = get_rdma_req(req);
	RTE_VERIFY(wr != NULL);
	RTE_VERIFY(sg_list != NULL);

	memset(wr, 0, sizeof(*wr));
	wr->wr_id = (uint64_t)rdma_req;
	wr->next = NULL;
	wr->opcode = opcode;
	wr->send_flags = send_flags;
	wr->sg_list = sg_list;
	wr->num_sge = 1;
}

static void
nvmf_ibv_send_wr_set_rkey(struct ibv_send_wr *wr, struct spdk_nvmf_request *req)
{
	struct spdk_nvme_sgl_descriptor *sgl = &req->cmd->nvme_cmd.dptr.sgl1;

	RTE_VERIFY(sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK);

	wr->wr.rdma.rkey = sgl->keyed.key;
	wr->wr.rdma.remote_addr = sgl->address;

	SPDK_TRACELOG(SPDK_TRACE_RDMA, "rkey %x remote_addr %p\n",
		      wr->wr.rdma.rkey, (void *)wr->wr.rdma.remote_addr);
}

static int
nvmf_post_rdma_read(struct spdk_nvmf_request *req)
{
	struct ibv_send_wr wr, *bad_wr = NULL;
	struct spdk_nvmf_conn *conn = req->conn;
	struct spdk_nvmf_rdma_conn *rdma_conn = get_rdma_conn(conn);
	struct spdk_nvmf_rdma_request *rdma_req = get_rdma_req(req);
	struct ibv_sge sge;
	int rc;

	SPDK_TRACELOG(SPDK_TRACE_RDMA, "RDMA READ POSTED. Request: %p Connection: %p\n", req, conn);

	sge.addr = (uintptr_t)rdma_req->buf;
	sge.lkey = rdma_conn->bufs_mr->lkey;
	sge.length = req->length;
	nvmf_trace_ibv_sge(&sge);

	nvmf_ibv_send_wr_init(&wr, req, &sge, IBV_WR_RDMA_READ, IBV_SEND_SIGNALED);
	nvmf_ibv_send_wr_set_rkey(&wr, req);

	spdk_trace_record(TRACE_RDMA_READ_START, 0, 0, (uintptr_t)req, 0);
	rc = ibv_post_send(rdma_conn->qp, &wr, &bad_wr);
	if (rc) {
		SPDK_ERRLOG("Failure posting rdma read send, rc = 0x%x\n", rc);
	}

	return rc;
}

static int
nvmf_post_rdma_write(struct spdk_nvmf_request *req)
{
	struct ibv_send_wr wr, *bad_wr = NULL;
	struct spdk_nvmf_conn *conn = req->conn;
	struct spdk_nvmf_rdma_conn *rdma_conn = get_rdma_conn(conn);
	struct spdk_nvmf_rdma_request *rdma_req = get_rdma_req(req);
	struct ibv_sge sge;
	int rc;

	SPDK_TRACELOG(SPDK_TRACE_RDMA, "RDMA WRITE POSTED. Request: %p Connection: %p\n", req, conn);

	sge.addr = (uintptr_t)rdma_req->buf;
	sge.lkey = rdma_conn->bufs_mr->lkey;
	sge.length = req->length;
	nvmf_trace_ibv_sge(&sge);

	nvmf_ibv_send_wr_init(&wr, req, &sge, IBV_WR_RDMA_WRITE, 0);
	nvmf_ibv_send_wr_set_rkey(&wr, req);

	spdk_trace_record(TRACE_RDMA_WRITE_START, 0, 0, (uintptr_t)req, 0);
	rc = ibv_post_send(rdma_conn->qp, &wr, &bad_wr);
	if (rc) {
		SPDK_ERRLOG("Failure posting rdma write send, rc = 0x%x\n", rc);
	}

	return rc;
}

static int
nvmf_post_rdma_recv(struct spdk_nvmf_request *req)
{
	struct ibv_recv_wr wr, *bad_wr = NULL;
	struct spdk_nvmf_conn *conn = req->conn;
	struct spdk_nvmf_rdma_conn *rdma_conn = get_rdma_conn(conn);
	struct spdk_nvmf_rdma_request *rdma_req = get_rdma_req(req);
	struct ibv_sge sg_list[2];
	int rc;

	SPDK_TRACELOG(SPDK_TRACE_RDMA, "RDMA RECV POSTED. Request: %p Connection: %p\n", req, conn);

	sg_list[0].addr = (uintptr_t)req->cmd;
	sg_list[0].length = sizeof(*req->cmd);
	sg_list[0].lkey = rdma_conn->cmds_mr->lkey;
	nvmf_trace_ibv_sge(&sg_list[0]);

	sg_list[1].addr = (uintptr_t)rdma_req->buf;
	sg_list[1].length = g_nvmf_tgt.in_capsule_data_size;
	sg_list[1].lkey = rdma_conn->bufs_mr->lkey;
	nvmf_trace_ibv_sge(&sg_list[1]);

	memset(&wr, 0, sizeof(wr));
	wr.wr_id = (uintptr_t)rdma_req;
	wr.next = NULL;
	wr.sg_list = sg_list;
	wr.num_sge = 2;

	rc = ibv_post_recv(rdma_conn->qp, &wr, &bad_wr);
	if (rc) {
		SPDK_ERRLOG("Failure posting rdma recv, rc = 0x%x\n", rc);
	}

	return rc;
}

static int
nvmf_post_rdma_send(struct spdk_nvmf_request *req)
{
	struct ibv_send_wr wr, *bad_wr = NULL;
	struct spdk_nvmf_conn *conn = req->conn;
	struct spdk_nvmf_rdma_conn *rdma_conn = get_rdma_conn(conn);
	struct ibv_sge sge;
	int rc;

	SPDK_TRACELOG(SPDK_TRACE_RDMA, "RDMA SEND POSTED. Request: %p Connection: %p\n", req, conn);

	sge.addr = (uintptr_t)req->rsp;
	sge.length = sizeof(*req->rsp);
	sge.lkey = rdma_conn->cpls_mr->lkey;
	nvmf_trace_ibv_sge(&sge);

	nvmf_ibv_send_wr_init(&wr, req, &sge, IBV_WR_SEND, IBV_SEND_SIGNALED);

	spdk_trace_record(TRACE_NVMF_IO_COMPLETE, 0, 0, (uintptr_t)req, 0);
	rc = ibv_post_send(rdma_conn->qp, &wr, &bad_wr);
	if (rc) {
		SPDK_ERRLOG("Failure posting rdma send for NVMf completion, rc = 0x%x\n", rc);
	}

	return rc;
}

static int
spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req)
{
	struct spdk_nvme_cpl *rsp = &req->rsp->nvme_cpl;
	int ret;

	/* Was the command successful? */
	if (rsp->status.sc == SPDK_NVME_SC_SUCCESS &&
	    req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
		/* Need to transfer data via RDMA Write */
		ret = nvmf_post_rdma_write(req);
		if (ret) {
			SPDK_ERRLOG("Unable to post rdma write tx descriptor\n");
			return -1;
		}
	}

	ret = nvmf_post_rdma_send(req);
	if (ret) {
		SPDK_ERRLOG("Unable to send response capsule\n");
		return -1;
	}

	return 0;
}

static int
spdk_nvmf_rdma_request_release(struct spdk_nvmf_request *req)
{
	struct spdk_nvmf_conn *conn = req->conn;
	struct spdk_nvmf_rdma_conn *rdma_conn = get_rdma_conn(conn);

	if (nvmf_post_rdma_recv(req)) {
		SPDK_ERRLOG("Unable to re-post rx descriptor\n");
		return -1;
	}

	conn->sq_head++;
	if (conn->sq_head == rdma_conn->queue_depth) {
		conn->sq_head = 0;
	}

	return 0;
}

static int
nvmf_rdma_connect(struct rdma_cm_event *event)
{
	struct spdk_nvmf_rdma_conn	*rdma_conn = NULL;
	struct ibv_device_attr		ibdev_attr;
	struct rdma_conn_param		*rdma_param = NULL;
	struct rdma_conn_param		ctrlr_event_data;
	const struct spdk_nvmf_rdma_request_private_data *private_data = NULL;
	struct spdk_nvmf_rdma_accept_private_data accept_data;
	uint16_t			sts = 0;
	uint16_t			queue_depth;
	int 				rc;


	/* Check to make sure we know about this rdma device */
	if (event->id == NULL) {
		SPDK_ERRLOG("connect request: missing cm_id\n");
		goto err0;
	}

	if (event->id->verbs == NULL) {
		SPDK_ERRLOG("connect request: missing cm_id ibv_context\n");
		goto err0;
	}
	SPDK_TRACELOG(SPDK_TRACE_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n",
		      event->id->verbs->device->name, event->id->verbs->device->dev_name);

	/* Figure out the supported queue depth. This is a multi-step process
	 * that takes into account hardware maximums, host provided values,
	 * and our target's internal memory limits */

	SPDK_TRACELOG(SPDK_TRACE_RDMA, "Calculating Queue Depth\n");

	/* Start with the maximum queue depth allowed by the target */
	queue_depth = g_nvmf_tgt.max_queue_depth;
	SPDK_TRACELOG(SPDK_TRACE_RDMA, "Target Max Queue Depth: %d\n", g_nvmf_tgt.max_queue_depth);

	/* Next check the local NIC's hardware limitations */
	rc = ibv_query_device(event->id->verbs, &ibdev_attr);
	if (rc) {
		SPDK_ERRLOG("Failed to query RDMA device attributes\n");
		sts = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
		goto err1;
	}
	SPDK_TRACELOG(SPDK_TRACE_RDMA,
		      "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n",
		      ibdev_attr.max_qp_wr, ibdev_attr.max_qp_rd_atom);
	queue_depth = nvmf_min(queue_depth, ibdev_attr.max_qp_rd_atom);
	queue_depth = nvmf_min(queue_depth, ibdev_attr.max_qp_rd_atom);

	/* Next check the remote NIC's hardware limitations */
	rdma_param = &event->param.conn;
	queue_depth = nvmf_min(queue_depth, rdma_param->initiator_depth);
	SPDK_TRACELOG(SPDK_TRACE_RDMA, "Host NIC Max Read/Write Queue Depth: %d\n",
		      rdma_param->initiator_depth);

	/* Finally check for the host software requested values, which are
	 * optional. */
	if (rdma_param->private_data != NULL &&
	    rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) {
		private_data = rdma_param->private_data;
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize);
		queue_depth = nvmf_min(queue_depth, private_data->hrqsize);
		queue_depth = nvmf_min(queue_depth, private_data->hsqsize);
	}

	SPDK_TRACELOG(SPDK_TRACE_RDMA, "Final Negotiated Queue Depth: %d\n", queue_depth);

	/* Init the NVMf rdma transport connection */
	rdma_conn = spdk_nvmf_rdma_conn_create(event->id, queue_depth);
	if (rdma_conn == NULL) {
		SPDK_ERRLOG("Error on nvmf connection creation\n");
		goto err1;
	}

	/* Add this RDMA connection to the global list until a CONNECT capsule
	 * is received. */
	TAILQ_INSERT_TAIL(&g_pending_conns, rdma_conn, link);

	accept_data.recfmt = 0;
	accept_data.crqsize = rdma_conn->queue_depth;
	ctrlr_event_data = *rdma_param;
	ctrlr_event_data.private_data = &accept_data;
	ctrlr_event_data.private_data_len = sizeof(accept_data);
	if (event->id->ps == RDMA_PS_TCP) {
		ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */
		ctrlr_event_data.initiator_depth = rdma_conn->queue_depth;
	}

	rc = rdma_accept(event->id, &ctrlr_event_data);
	if (rc) {
		SPDK_ERRLOG("Error on rdma_accept\n");
		goto err1;
	}
	SPDK_TRACELOG(SPDK_TRACE_RDMA, "Sent back the accept\n");

	return 0;

err1: {
		struct spdk_nvmf_rdma_reject_private_data rej_data;

		rej_data.status.sc = sts;
		rdma_reject(event->id, &ctrlr_event_data, sizeof(rej_data));
		free(rdma_conn);
	}
err0:
	return -1;
}

static void
spdk_nvmf_handle_disconnect(spdk_event_t event)
{
	struct nvmf_session		*session = spdk_event_get_arg1(event);
	struct spdk_nvmf_conn		*conn = spdk_event_get_arg2(event);

	nvmf_disconnect(session, conn);
}

static int
nvmf_rdma_disconnect(struct rdma_cm_event *evt)
{
	struct spdk_nvmf_conn		*conn;
	struct nvmf_session		*session;
	struct spdk_nvmf_rdma_conn 	*rdma_conn;
	spdk_event_t			event;

	if (evt->id == NULL) {
		SPDK_ERRLOG("disconnect request: missing cm_id\n");
		return -1;
	}

	conn = evt->id->context;
	if (conn == NULL) {
		SPDK_ERRLOG("disconnect request: no active connection\n");
		return -1;
	}
	/* ack the disconnect event before rdma_destroy_id */
	rdma_ack_cm_event(evt);

	rdma_conn = get_rdma_conn(conn);

	session = conn->sess;
	if (session == NULL) {
		/* No session has been established yet. That means the conn
		 * must be in the pending connections list. Remove it. */
		TAILQ_REMOVE(&g_pending_conns, rdma_conn, link);
		spdk_nvmf_rdma_conn_destroy(rdma_conn);
		return 0;
	}

	/* Pass an event to the core that owns this connection */
	event = spdk_event_allocate(session->subsys->poller.lcore,
				    spdk_nvmf_handle_disconnect,
				    session, conn, NULL);
	spdk_event_call(event);

	return 0;
}

#ifdef DEBUG
static const char *CM_EVENT_STR[] = {
	"RDMA_CM_EVENT_ADDR_RESOLVED",
	"RDMA_CM_EVENT_ADDR_ERROR",
	"RDMA_CM_EVENT_ROUTE_RESOLVED",
	"RDMA_CM_EVENT_ROUTE_ERROR",
	"RDMA_CM_EVENT_CONNECT_REQUEST",
	"RDMA_CM_EVENT_CONNECT_RESPONSE",
	"RDMA_CM_EVENT_CONNECT_ERROR",
	"RDMA_CM_EVENT_UNREACHABLE",
	"RDMA_CM_EVENT_REJECTED",
	"RDMA_CM_EVENT_ESTABLISHED",
	"RDMA_CM_EVENT_DISCONNECTED",
	"RDMA_CM_EVENT_DEVICE_REMOVAL",
	"RDMA_CM_EVENT_MULTICAST_JOIN",
	"RDMA_CM_EVENT_MULTICAST_ERROR",
	"RDMA_CM_EVENT_ADDR_CHANGE",
	"RDMA_CM_EVENT_TIMEWAIT_EXIT"
};
#endif /* DEBUG */

static int
spdk_nvmf_request_prep_data(struct spdk_nvmf_request *req,
			    void *in_cap_data, uint32_t in_cap_len,
			    void *bb, uint32_t bb_len)
{
	struct spdk_nvme_cmd *cmd = &req->cmd->nvme_cmd;
	struct spdk_nvme_cpl *rsp = &req->rsp->nvme_cpl;
	enum spdk_nvme_data_transfer xfer;

	req->length = 0;
	req->xfer = SPDK_NVME_DATA_NONE;
	req->data = NULL;

	if (cmd->opc == SPDK_NVME_OPC_FABRIC) {
		xfer = spdk_nvme_opc_get_data_transfer(req->cmd->nvmf_cmd.fctype);
	} else {
		xfer = spdk_nvme_opc_get_data_transfer(cmd->opc);
	}

	if (xfer != SPDK_NVME_DATA_NONE) {
		struct spdk_nvme_sgl_descriptor *sgl = (struct spdk_nvme_sgl_descriptor *)&cmd->dptr.sgl1;

		if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK &&
		    (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS ||
		     sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) {
			if (sgl->keyed.length > bb_len) {
				SPDK_ERRLOG("SGL length 0x%x exceeds BB length 0x%x\n",
					    sgl->keyed.length, bb_len);
				rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
				return -1;
			}

			req->data = bb;
			req->length = sgl->keyed.length;
		} else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK &&
			   sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
			uint64_t offset = sgl->address;
			uint32_t max_len = in_cap_len;

			SPDK_TRACELOG(SPDK_TRACE_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n",
				      offset, sgl->unkeyed.length);

			if (offset > max_len) {
				SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n",
					    offset, max_len);
				rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET;
				return -1;
			}
			max_len -= (uint32_t)offset;

			if (sgl->unkeyed.length > max_len) {
				SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n",
					    sgl->unkeyed.length, max_len);
				rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
				return -1;
			}

			req->data = in_cap_data + offset;
			req->length = sgl->unkeyed.length;
		} else {
			SPDK_ERRLOG("Invalid NVMf I/O Command SGL:  Type 0x%x, Subtype 0x%x\n",
				    sgl->generic.type, sgl->generic.subtype);
			rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID;
			return -1;
		}

		if (req->length == 0) {
			xfer = SPDK_NVME_DATA_NONE;
			req->data = NULL;
		}

		req->xfer = xfer;

		/*
		 * For any I/O that requires data to be
		 * pulled into target BB before processing by
		 * the backend NVMe device
		 */
		if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
			if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK) {
				SPDK_TRACELOG(SPDK_TRACE_NVMF, "Initiating Host to Controller data transfer\n");
				/* Wait for transfer to complete before executing command. */
				return 1;
			}
		}
	}

	if (xfer == SPDK_NVME_DATA_NONE) {
		SPDK_TRACELOG(SPDK_TRACE_NVMF, "No data to transfer\n");
		assert(req->data == NULL);
		assert(req->length == 0);
	} else {
		assert(req->data != NULL);
		assert(req->length != 0);
		SPDK_TRACELOG(SPDK_TRACE_NVMF, "%s data ready\n",
			      xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER ? "Host to Controller" :
			      "Controller to Host");
	}

	return 0;
}

static int spdk_nvmf_rdma_poll(struct spdk_nvmf_conn *conn);

static void
nvmf_rdma_accept(struct rte_timer *timer, void *arg)
{
	struct rdma_cm_event		*event;
	int				rc;
	struct spdk_nvmf_rdma_conn	*rdma_conn, *tmp;

	if (g_rdma.acceptor_event_channel == NULL) {
		return;
	}

	/* Process pending connections for incoming capsules. The only capsule
	 * this should ever find is a CONNECT request. */
	TAILQ_FOREACH_SAFE(rdma_conn, &g_pending_conns, link, tmp) {
		rc = spdk_nvmf_rdma_poll(&rdma_conn->conn);
		if (rc < 0) {
			TAILQ_REMOVE(&g_pending_conns, rdma_conn, link);
			spdk_nvmf_rdma_conn_destroy(rdma_conn);
		} else if (rc > 0) {
			/* At least one request was processed which is assumed to be
			 * a CONNECT. Remove this connection from our list. */
			TAILQ_REMOVE(&g_pending_conns, rdma_conn, link);
		}
	}

	while (1) {
		rc = rdma_get_cm_event(g_rdma.acceptor_event_channel, &event);
		if (rc == 0) {
			SPDK_TRACELOG(SPDK_TRACE_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]);

			switch (event->event) {
			case RDMA_CM_EVENT_CONNECT_REQUEST:
				rc = nvmf_rdma_connect(event);
				if (rc < 0) {
					SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc);
					break;
				}
				break;
			case RDMA_CM_EVENT_ESTABLISHED:
				break;
			case RDMA_CM_EVENT_ADDR_CHANGE:
			case RDMA_CM_EVENT_DISCONNECTED:
			case RDMA_CM_EVENT_DEVICE_REMOVAL:
			case RDMA_CM_EVENT_TIMEWAIT_EXIT:
				rc = nvmf_rdma_disconnect(event);
				if (rc < 0) {
					SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
					break;
				}
				continue;
			default:
				SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event);
				break;
			}

			rdma_ack_cm_event(event);
		} else {
			if (errno != EAGAIN && errno != EWOULDBLOCK) {
				SPDK_ERRLOG("Acceptor Event Error: %s\n", strerror(errno));
			}
			break;
		}
	}
}

static int
spdk_nvmf_rdma_acceptor_start(void)
{
	struct sockaddr_in	addr;
	uint16_t		sin_port;
	int			rc;

	memset(&addr, 0, sizeof(addr));
	addr.sin_family = AF_INET;
	addr.sin_port = g_nvmf_tgt.sin_port;

	/* create an event channel with rdmacm to receive
	   connection oriented requests and notifications */
	g_rdma.acceptor_event_channel = rdma_create_event_channel();
	if (g_rdma.acceptor_event_channel == NULL) {
		SPDK_ERRLOG("rdma_create_event_channel() failed\n");
		return -1;
	}
	rc = fcntl(g_rdma.acceptor_event_channel->fd, F_SETFL, O_NONBLOCK);
	if (rc < 0) {
		SPDK_ERRLOG("fcntl to set fd to non-blocking failed\n");
		goto create_id_error;
	}

	rc = rdma_create_id(g_rdma.acceptor_event_channel, &g_rdma.acceptor_listen_id, NULL, RDMA_PS_TCP);
	if (rc < 0) {
		SPDK_ERRLOG("rdma_create_id() failed\n");
		goto create_id_error;
	}

	rc = rdma_bind_addr(g_rdma.acceptor_listen_id, (struct sockaddr *)&addr);
	if (rc < 0) {
		SPDK_ERRLOG("rdma_bind_addr() failed\n");
		goto listen_error;
	}

	rc = rdma_listen(g_rdma.acceptor_listen_id, 10); /* 10 = backlog */
	if (rc < 0) {
		SPDK_ERRLOG("rdma_listen() failed\n");
		goto listen_error;
	}
	sin_port = ntohs(rdma_get_src_port(g_rdma.acceptor_listen_id));
	SPDK_NOTICELOG("*** NVMf Target Listening on port %d ***\n", sin_port);

	rte_timer_init(&g_rdma.acceptor_timer);
	rte_timer_reset(&g_rdma.acceptor_timer, ACCEPT_TIMEOUT, PERIODICAL,
			rte_lcore_id(), nvmf_rdma_accept, NULL);
	return (rc);

listen_error:
	rdma_destroy_id(g_rdma.acceptor_listen_id);
create_id_error:
	rdma_destroy_event_channel(g_rdma.acceptor_event_channel);
	return -1;
}

static void
spdk_nvmf_rdma_acceptor_stop(void)
{
	SPDK_TRACELOG(SPDK_TRACE_RDMA, "nvmf_acceptor_stop: shutdown\n");
	rte_timer_stop_sync(&g_rdma.acceptor_timer);
}

/*

Initialize with RDMA transport.  Query OFED for device list.

*/
static int
spdk_nvmf_rdma_init(void)
{
	struct ibv_device **dev_list;
	struct ibv_context *ibdev_ctx = NULL;
	struct ibv_device_attr ibdev_attr;
	int num_of_rdma_devices;
	int num_devices_found = 0;
	int i, ret;

	SPDK_NOTICELOG("*** RDMA Transport Init ***\n");

	dev_list = ibv_get_device_list(&num_of_rdma_devices);
	if (!dev_list) {
		SPDK_ERRLOG(" No RDMA verbs devices found\n");
		return -1;
	}
	SPDK_TRACELOG(SPDK_TRACE_RDMA, "%d RDMA verbs device(s) discovered\n", num_of_rdma_devices);

	/* Look through the list of devices for one we support */
	for (i = 0; i < num_of_rdma_devices; i++) {
		SPDK_TRACELOG(SPDK_TRACE_RDMA, " RDMA Device %d:\n", i);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Node type: %d\n", (int)dev_list[i]->node_type);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Transport type: %d\n", (int)dev_list[i]->transport_type);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Name: %s\n", dev_list[i]->name);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Device Name: %s\n", dev_list[i]->dev_name);

		ibdev_ctx = ibv_open_device(dev_list[i]);
		if (!ibdev_ctx) {
			SPDK_ERRLOG(" No rdma context returned for device %d\n", i);
			continue;
		}

		ret = ibv_query_device(ibdev_ctx, &ibdev_attr);
		if (ret) {
			SPDK_ERRLOG(" Failed on query for device %d\n", i);
			ibv_close_device(ibdev_ctx);
			continue;
		}

		/* display device specific attributes */
		SPDK_TRACELOG(SPDK_TRACE_RDMA, " RDMA Device Attributes:\n");
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max MR Size: 0x%llx\n", (long long int)ibdev_attr.max_mr_size);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Page Size Cap: 0x%llx\n",
			      (long long int)ibdev_attr.page_size_cap);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max QPs: 0x%x\n", (int)ibdev_attr.max_qp);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max QP WRs: 0x%x\n", (int)ibdev_attr.max_qp_wr);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max SGE: 0x%x\n", (int)ibdev_attr.max_sge);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max CQs: 0x%x\n", (int)ibdev_attr.max_cq);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max CQE per CQ: 0x%x\n", (int)ibdev_attr.max_cqe);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max MR: 0x%x\n", (int)ibdev_attr.max_mr);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max PD: 0x%x\n", (int)ibdev_attr.max_pd);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max QP RD Atom: 0x%x\n", (int)ibdev_attr.max_qp_rd_atom);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max QP Init RD Atom: 0x%x\n",
			      (int)ibdev_attr.max_qp_init_rd_atom);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max Res RD Atom: 0x%x\n", (int)ibdev_attr.max_res_rd_atom);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max EE: 0x%x\n", (int)ibdev_attr.max_ee);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max SRQ: 0x%x\n", (int)ibdev_attr.max_srq);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max SRQ WR: 0x%x\n", (int)ibdev_attr.max_srq_wr);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max SRQ SGE: 0x%x\n", (int)ibdev_attr.max_srq_sge);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Max PKeys: 0x%x\n", (int)ibdev_attr.max_pkeys);
		SPDK_TRACELOG(SPDK_TRACE_RDMA, "   Phys Port Cnt: %d\n", (int)ibdev_attr.phys_port_cnt);

		num_devices_found++;
	}

	ibv_free_device_list(dev_list);
	SPDK_TRACELOG(SPDK_TRACE_RDMA, "    %d Fabric Intf(s) active\n", num_devices_found);
	return num_devices_found;
}

static int
spdk_nvmf_rdma_fini(void)
{
	/* Nothing to do */
	return 0;
}

static void
spdk_nvmf_rdma_close_conn(struct spdk_nvmf_conn *conn)
{
	struct spdk_nvmf_rdma_conn *rdma_conn = get_rdma_conn(conn);

	return spdk_nvmf_rdma_conn_destroy(rdma_conn);
}

/* Returns the number of times that spdk_nvmf_request_exec was called,
 * or -1 on error.
 */
static int
spdk_nvmf_rdma_poll(struct spdk_nvmf_conn *conn)
{
	struct ibv_wc wc;
	struct spdk_nvmf_rdma_conn *rdma_conn = get_rdma_conn(conn);
	struct spdk_nvmf_rdma_request *rdma_req;
	struct spdk_nvmf_request *req;
	int rc, count;

	count = 0;
	while (true) {
		rc = ibv_poll_cq(rdma_conn->cq, 1, &wc);
		if (rc == 0) // No completions at this time
			break;

		if (rc < 0) {
			SPDK_ERRLOG("Poll CQ error!(%d): %s\n",
				    errno, strerror(errno));
			return -1;
		}

		/* OK, process the single successful cq event */
		if (wc.status) {
			SPDK_TRACELOG(SPDK_TRACE_RDMA, "CQ completion error status %d (%s), exiting handler\n",
				      wc.status, ibv_wc_status_str(wc.status));
			return -1;
		}

		rdma_req = (struct spdk_nvmf_rdma_request *)wc.wr_id;
		if (rdma_req == NULL) {
			SPDK_ERRLOG("Got CQ completion for NULL rdma_req\n");
			return -1;
		}

		req = &rdma_req->req;

		switch (wc.opcode) {
		case IBV_WC_SEND:
			assert(rdma_conn->outstanding_reqs > 0);
			rdma_conn->outstanding_reqs--;
			SPDK_TRACELOG(SPDK_TRACE_RDMA,
				      "RDMA SEND Complete. Request: %p Connection: %p Outstanding I/O: %d\n",
				      req, conn, rdma_conn->outstanding_reqs);
			if (spdk_nvmf_rdma_request_release(req)) {
				return -1;
			}
			break;

		case IBV_WC_RDMA_WRITE:
			/*
			 * Will get this event only if we set IBV_SEND_SIGNALED
			 * flag in rdma_write, to trace rdma write latency
			 */
			SPDK_TRACELOG(SPDK_TRACE_RDMA, "RDMA WRITE Complete. Request: %p Connection: %p\n",
				      req, conn);
			spdk_trace_record(TRACE_RDMA_WRITE_COMPLETE, 0, 0, (uint64_t)req, 0);
			break;

		case IBV_WC_RDMA_READ:
			SPDK_TRACELOG(SPDK_TRACE_RDMA, "RDMA READ Complete. Request: %p Connection: %p\n",
				      req, conn);
			spdk_trace_record(TRACE_RDMA_READ_COMPLETE, 0, 0, (uint64_t)req, 0);
			rc = spdk_nvmf_request_exec(req);
			if (rc) {
				SPDK_ERRLOG("request_exec error %d after RDMA Read completion\n", rc);
				return -1;
			}
			count++;
			break;

		case IBV_WC_RECV:
			if (wc.byte_len < sizeof(struct spdk_nvmf_capsule_cmd)) {
				SPDK_ERRLOG("recv length %u less than capsule header\n", wc.byte_len);
				return -1;
			}

			rdma_conn->outstanding_reqs++;
			SPDK_TRACELOG(SPDK_TRACE_RDMA,
				      "RDMA RECV Complete. Request: %p Connection: %p Outstanding I/O: %d\n",
				      req, conn, rdma_conn->outstanding_reqs);
			spdk_trace_record(TRACE_NVMF_IO_START, 0, 0, (uint64_t)req, 0);

			memset(req->rsp, 0, sizeof(*req->rsp));
			rc = spdk_nvmf_request_prep_data(req,
							 rdma_req->buf,
							 wc.byte_len - sizeof(struct spdk_nvmf_capsule_cmd),
							 rdma_req->buf,
							 g_nvmf_tgt.max_io_size);
			if (rc < 0) {
				SPDK_ERRLOG("prep_data failed\n");
				return spdk_nvmf_request_complete(req);
			} else if (rc == 0) {
				/* Data is immediately available */
				rc = spdk_nvmf_request_exec(req);
				if (rc < 0) {
					SPDK_ERRLOG("Command execution failed\n");
					return -1;
				}
				count++;
			} else {
				/* Start transfer of data from host to target */
				rc = nvmf_post_rdma_read(req);
				if (rc) {
					SPDK_ERRLOG("Unable to transfer data from host to target\n");
					return -1;
				}
			}
			break;

		default:
			SPDK_ERRLOG("Poll cq opcode type unknown!!!!! completion\n");
			return -1;
		}
	}

	return count;
}

static void
nvmf_rdma_discover(struct spdk_nvmf_listen_addr *listen_addr,
		   struct spdk_nvmf_discovery_log_page_entry *entry)
{
	entry->trtype = SPDK_NVMF_TRTYPE_RDMA;
	entry->adrfam = SPDK_NVMF_ADRFAM_IPV4;
	entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_SPECIFIED;

	snprintf(entry->trsvcid, sizeof(entry->trsvcid), "%s", listen_addr->trsvc);
	snprintf(entry->traddr, sizeof(entry->traddr), "%s", listen_addr->traddr);

	entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED;
	entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE;
	entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM;
}

const struct spdk_nvmf_transport spdk_nvmf_transport_rdma = {
	.name = "rdma",
	.transport_init = spdk_nvmf_rdma_init,
	.transport_fini = spdk_nvmf_rdma_fini,
	.transport_start = spdk_nvmf_rdma_acceptor_start,
	.transport_stop = spdk_nvmf_rdma_acceptor_stop,

	.req_complete = spdk_nvmf_rdma_request_complete,
	.req_release = spdk_nvmf_rdma_request_release,

	.conn_fini = spdk_nvmf_rdma_close_conn,
	.conn_poll = spdk_nvmf_rdma_poll,

	.listen_addr_discover = nvmf_rdma_discover,
};

SPDK_LOG_REGISTER_TRACE_FLAG("rdma", SPDK_TRACE_RDMA)