9p: rdma: RDMA Transport Support for 9P
Tom Tucker [Wed, 22 Oct 2008 23:47:39 +0000 (18:47 -0500)]
This patch implements the RDMA transport provider for 9P. It allows
mounts to be performed over iWARP and IB capable network interfaces.

Signed-off-by: Tom Tucker <tom@opengridcomputing.com>
Signed-off-by: Latchesar Ionkov <lionkov@lanl.gov>

include/net/9p/client.h
net/9p/Kconfig
net/9p/Makefile
net/9p/trans_rdma.c [new file with mode: 0644]

index 1f17f3d..9fe45b3 100644 (file)
@@ -206,6 +206,7 @@ int p9_client_wstat(struct p9_fid *fid, struct p9_wstat *wst);
 struct p9_req_t *p9_tag_lookup(struct p9_client *, u16);
 void p9_client_cb(struct p9_client *c, struct p9_req_t *req);
 
+int p9_parse_header(struct p9_fcall *, int32_t *, int8_t *, int16_t *, int);
 int p9stat_read(char *, int, struct p9_wstat *, int);
 void p9stat_free(struct p9_wstat *);
 
index ff34c5a..c42c0c4 100644 (file)
@@ -20,6 +20,12 @@ config NET_9P_VIRTIO
          This builds support for a transports between
          guest partitions and a host partition.
 
+config NET_9P_RDMA
+       depends on NET_9P && INFINIBAND && EXPERIMENTAL
+       tristate "9P RDMA Transport (Experimental)"
+       help
+         This builds support for a RDMA transport.
+
 config NET_9P_DEBUG
        bool "Debug information"
        depends on NET_9P
index 1041b7b..198a640 100644 (file)
@@ -1,5 +1,6 @@
 obj-$(CONFIG_NET_9P) := 9pnet.o
 obj-$(CONFIG_NET_9P_VIRTIO) += 9pnet_virtio.o
+obj-$(CONFIG_NET_9P_RDMA) += 9pnet_rdma.o
 
 9pnet-objs := \
        mod.o \
@@ -11,3 +12,6 @@ obj-$(CONFIG_NET_9P_VIRTIO) += 9pnet_virtio.o
 
 9pnet_virtio-objs := \
        trans_virtio.o \
+
+9pnet_rdma-objs := \
+       trans_rdma.o \
diff --git a/net/9p/trans_rdma.c b/net/9p/trans_rdma.c
new file mode 100644 (file)
index 0000000..8d6cc47
--- /dev/null
@@ -0,0 +1,712 @@
+/*
+ * linux/fs/9p/trans_rdma.c
+ *
+ * RDMA transport layer based on the trans_fd.c implementation.
+ *
+ *  Copyright (C) 2008 by Tom Tucker <tom@opengridcomputing.com>
+ *  Copyright (C) 2006 by Russ Cox <rsc@swtch.com>
+ *  Copyright (C) 2004-2005 by Latchesar Ionkov <lucho@ionkov.net>
+ *  Copyright (C) 2004-2008 by Eric Van Hensbergen <ericvh@gmail.com>
+ *  Copyright (C) 1997-2002 by Ron Minnich <rminnich@sarnoff.com>
+ *
+ *  This program is free software; you can redistribute it and/or modify
+ *  it under the terms of the GNU General Public License version 2
+ *  as published by the Free Software Foundation.
+ *
+ *  This program is distributed in the hope that it will be useful,
+ *  but WITHOUT ANY WARRANTY; without even the implied warranty of
+ *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ *  GNU General Public License for more details.
+ *
+ *  You should have received a copy of the GNU General Public License
+ *  along with this program; if not, write to:
+ *  Free Software Foundation
+ *  51 Franklin Street, Fifth Floor
+ *  Boston, MA  02111-1301  USA
+ *
+ */
+
+#include <linux/in.h>
+#include <linux/module.h>
+#include <linux/net.h>
+#include <linux/ipv6.h>
+#include <linux/kthread.h>
+#include <linux/errno.h>
+#include <linux/kernel.h>
+#include <linux/un.h>
+#include <linux/uaccess.h>
+#include <linux/inet.h>
+#include <linux/idr.h>
+#include <linux/file.h>
+#include <linux/parser.h>
+#include <linux/semaphore.h>
+#include <net/9p/9p.h>
+#include <net/9p/client.h>
+#include <net/9p/transport.h>
+#include <rdma/ib_verbs.h>
+#include <rdma/rdma_cm.h>
+#include <rdma/ib_verbs.h>
+
+#define P9_PORT                        5640
+#define P9_RDMA_SQ_DEPTH       32
+#define P9_RDMA_RQ_DEPTH       32
+#define P9_RDMA_SEND_SGE       4
+#define P9_RDMA_RECV_SGE       4
+#define P9_RDMA_IRD            0
+#define P9_RDMA_ORD            0
+#define P9_RDMA_TIMEOUT                30000           /* 30 seconds */
+#define P9_RDMA_MAXSIZE                (4*4096)        /* Min SGE is 4, so we can
+                                                * safely advertise a maxsize
+                                                * of 64k */
+
+#define P9_RDMA_MAX_SGE (P9_RDMA_MAXSIZE >> PAGE_SHIFT)
+/**
+ * struct p9_trans_rdma - RDMA transport instance
+ *
+ * @state: tracks the transport state machine for connection setup and tear down
+ * @cm_id: The RDMA CM ID
+ * @pd: Protection Domain pointer
+ * @qp: Queue Pair pointer
+ * @cq: Completion Queue pointer
+ * @lkey: The local access only memory region key
+ * @timeout: Number of uSecs to wait for connection management events
+ * @sq_depth: The depth of the Send Queue
+ * @sq_sem: Semaphore for the SQ
+ * @rq_depth: The depth of the Receive Queue.
+ * @addr: The remote peer's address
+ * @req_lock: Protects the active request list
+ * @send_wait: Wait list when the SQ fills up
+ * @cm_done: Completion event for connection management tracking
+ */
+struct p9_trans_rdma {
+       enum {
+               P9_RDMA_INIT,
+               P9_RDMA_ADDR_RESOLVED,
+               P9_RDMA_ROUTE_RESOLVED,
+               P9_RDMA_CONNECTED,
+               P9_RDMA_FLUSHING,
+               P9_RDMA_CLOSING,
+               P9_RDMA_CLOSED,
+       } state;
+       struct rdma_cm_id *cm_id;
+       struct ib_pd *pd;
+       struct ib_qp *qp;
+       struct ib_cq *cq;
+       struct ib_mr *dma_mr;
+       u32 lkey;
+       long timeout;
+       int sq_depth;
+       struct semaphore sq_sem;
+       int rq_depth;
+       atomic_t rq_count;
+       struct sockaddr_in addr;
+       spinlock_t req_lock;
+
+       struct completion cm_done;
+};
+
+/**
+ * p9_rdma_context - Keeps track of in-process WR
+ *
+ * @wc_op: The original WR op for when the CQE completes in error.
+ * @busa: Bus address to unmap when the WR completes
+ * @req: Keeps track of requests (send)
+ * @rc: Keepts track of replies (receive)
+ */
+struct p9_rdma_req;
+struct p9_rdma_context {
+       enum ib_wc_opcode wc_op;
+       dma_addr_t busa;
+       union {
+               struct p9_req_t *req;
+               struct p9_fcall *rc;
+       };
+};
+
+/**
+ * p9_rdma_opts - Collection of mount options
+ * @port: port of connection
+ * @sq_depth: The requested depth of the SQ. This really doesn't need
+ * to be any deeper than the number of threads used in the client
+ * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
+ * @timeout: Time to wait in msecs for CM events
+ */
+struct p9_rdma_opts {
+       short port;
+       int sq_depth;
+       int rq_depth;
+       long timeout;
+};
+
+/*
+ * Option Parsing (code inspired by NFS code)
+ */
+enum {
+       /* Options that take integer arguments */
+       Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout, Opt_err,
+};
+
+static match_table_t tokens = {
+       {Opt_port, "port=%u"},
+       {Opt_sq_depth, "sq=%u"},
+       {Opt_rq_depth, "rq=%u"},
+       {Opt_timeout, "timeout=%u"},
+       {Opt_err, NULL},
+};
+
+/**
+ * parse_options - parse mount options into session structure
+ * @options: options string passed from mount
+ * @opts: transport-specific structure to parse options into
+ *
+ * Returns 0 upon success, -ERRNO upon failure
+ */
+static int parse_opts(char *params, struct p9_rdma_opts *opts)
+{
+       char *p;
+       substring_t args[MAX_OPT_ARGS];
+       int option;
+       char *options;
+       int ret;
+
+       opts->port = P9_PORT;
+       opts->sq_depth = P9_RDMA_SQ_DEPTH;
+       opts->rq_depth = P9_RDMA_RQ_DEPTH;
+       opts->timeout = P9_RDMA_TIMEOUT;
+
+       if (!params)
+               return 0;
+
+       options = kstrdup(params, GFP_KERNEL);
+       if (!options) {
+               P9_DPRINTK(P9_DEBUG_ERROR,
+                          "failed to allocate copy of option string\n");
+               return -ENOMEM;
+       }
+
+       while ((p = strsep(&options, ",")) != NULL) {
+               int token;
+               int r;
+               if (!*p)
+                       continue;
+               token = match_token(p, tokens, args);
+               r = match_int(&args[0], &option);
+               if (r < 0) {
+                       P9_DPRINTK(P9_DEBUG_ERROR,
+                                  "integer field, but no integer?\n");
+                       ret = r;
+                       continue;
+               }
+               switch (token) {
+               case Opt_port:
+                       opts->port = option;
+                       break;
+               case Opt_sq_depth:
+                       opts->sq_depth = option;
+                       break;
+               case Opt_rq_depth:
+                       opts->rq_depth = option;
+                       break;
+               case Opt_timeout:
+                       opts->timeout = option;
+                       break;
+               default:
+                       continue;
+               }
+       }
+       /* RQ must be at least as large as the SQ */
+       opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
+       kfree(options);
+       return 0;
+}
+
+static int
+p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
+{
+       struct p9_client *c = id->context;
+       struct p9_trans_rdma *rdma = c->trans;
+       switch (event->event) {
+       case RDMA_CM_EVENT_ADDR_RESOLVED:
+               BUG_ON(rdma->state != P9_RDMA_INIT);
+               rdma->state = P9_RDMA_ADDR_RESOLVED;
+               break;
+
+       case RDMA_CM_EVENT_ROUTE_RESOLVED:
+               BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
+               rdma->state = P9_RDMA_ROUTE_RESOLVED;
+               break;
+
+       case RDMA_CM_EVENT_ESTABLISHED:
+               BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
+               rdma->state = P9_RDMA_CONNECTED;
+               break;
+
+       case RDMA_CM_EVENT_DISCONNECTED:
+               if (rdma)
+                       rdma->state = P9_RDMA_CLOSED;
+               if (c)
+                       c->status = Disconnected;
+               break;
+
+       case RDMA_CM_EVENT_TIMEWAIT_EXIT:
+               break;
+
+       case RDMA_CM_EVENT_ADDR_CHANGE:
+       case RDMA_CM_EVENT_ROUTE_ERROR:
+       case RDMA_CM_EVENT_DEVICE_REMOVAL:
+       case RDMA_CM_EVENT_MULTICAST_JOIN:
+       case RDMA_CM_EVENT_MULTICAST_ERROR:
+       case RDMA_CM_EVENT_REJECTED:
+       case RDMA_CM_EVENT_CONNECT_REQUEST:
+       case RDMA_CM_EVENT_CONNECT_RESPONSE:
+       case RDMA_CM_EVENT_CONNECT_ERROR:
+       case RDMA_CM_EVENT_ADDR_ERROR:
+       case RDMA_CM_EVENT_UNREACHABLE:
+               c->status = Disconnected;
+               rdma_disconnect(rdma->cm_id);
+               break;
+       default:
+               BUG();
+       }
+       complete(&rdma->cm_done);
+       return 0;
+}
+
+static void
+handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
+           struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
+{
+       struct p9_req_t *req;
+       int err = 0;
+       int16_t tag;
+
+       req = NULL;
+       ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
+                                                        DMA_FROM_DEVICE);
+
+       if (status != IB_WC_SUCCESS)
+               goto err_out;
+
+       err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
+       if (err)
+               goto err_out;
+
+       req = p9_tag_lookup(client, tag);
+       if (!req)
+               goto err_out;
+
+       req->rc = c->rc;
+       p9_client_cb(client, req);
+
+       return;
+
+ err_out:
+       P9_DPRINTK(P9_DEBUG_ERROR, "req %p err %d status %d\n",
+                  req, err, status);
+       rdma->state = P9_RDMA_FLUSHING;
+       client->status = Disconnected;
+       return;
+}
+
+static void
+handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
+           struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
+{
+       ib_dma_unmap_single(rdma->cm_id->device,
+                           c->busa, c->req->tc->size,
+                           DMA_TO_DEVICE);
+}
+
+static void qp_event_handler(struct ib_event *event, void *context)
+{
+       P9_DPRINTK(P9_DEBUG_ERROR, "QP event %d context %p\n", event->event,
+                                                               context);
+}
+
+static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
+{
+       struct p9_client *client = cq_context;
+       struct p9_trans_rdma *rdma = client->trans;
+       int ret;
+       struct ib_wc wc;
+
+       ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
+       while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
+               struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
+
+               switch (c->wc_op) {
+               case IB_WC_RECV:
+                       atomic_dec(&rdma->rq_count);
+                       handle_recv(client, rdma, c, wc.status, wc.byte_len);
+                       break;
+
+               case IB_WC_SEND:
+                       handle_send(client, rdma, c, wc.status, wc.byte_len);
+                       up(&rdma->sq_sem);
+                       break;
+
+               default:
+                       printk(KERN_ERR "9prdma: unexpected completion type, "
+                              "c->wc_op=%d, wc.opcode=%d, status=%d\n",
+                              c->wc_op, wc.opcode, wc.status);
+                       break;
+               }
+               kfree(c);
+       }
+}
+
+static void cq_event_handler(struct ib_event *e, void *v)
+{
+       P9_DPRINTK(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
+}
+
+static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
+{
+       if (!rdma)
+               return;
+
+       if (rdma->dma_mr && !IS_ERR(rdma->dma_mr))
+               ib_dereg_mr(rdma->dma_mr);
+
+       if (rdma->qp && !IS_ERR(rdma->qp))
+               ib_destroy_qp(rdma->qp);
+
+       if (rdma->pd && !IS_ERR(rdma->pd))
+               ib_dealloc_pd(rdma->pd);
+
+       if (rdma->cq && !IS_ERR(rdma->cq))
+               ib_destroy_cq(rdma->cq);
+
+       if (rdma->cm_id && !IS_ERR(rdma->cm_id))
+               rdma_destroy_id(rdma->cm_id);
+
+       kfree(rdma);
+}
+
+static int
+post_recv(struct p9_client *client, struct p9_rdma_context *c)
+{
+       struct p9_trans_rdma *rdma = client->trans;
+       struct ib_recv_wr wr, *bad_wr;
+       struct ib_sge sge;
+
+       c->busa = ib_dma_map_single(rdma->cm_id->device,
+                                   c->rc->sdata, client->msize,
+                                   DMA_FROM_DEVICE);
+       if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
+               goto error;
+
+       sge.addr = c->busa;
+       sge.length = client->msize;
+       sge.lkey = rdma->lkey;
+
+       wr.next = NULL;
+       c->wc_op = IB_WC_RECV;
+       wr.wr_id = (unsigned long) c;
+       wr.sg_list = &sge;
+       wr.num_sge = 1;
+       return ib_post_recv(rdma->qp, &wr, &bad_wr);
+
+ error:
+       P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
+       return -EIO;
+}
+
+static int rdma_request(struct p9_client *client, struct p9_req_t *req)
+{
+       struct p9_trans_rdma *rdma = client->trans;
+       struct ib_send_wr wr, *bad_wr;
+       struct ib_sge sge;
+       int err = 0;
+       unsigned long flags;
+       struct p9_rdma_context *c = NULL;
+       struct p9_rdma_context *rpl_context = NULL;
+
+       /* Allocate an fcall for the reply */
+       rpl_context = kmalloc(sizeof *rpl_context, GFP_KERNEL);
+       if (!rpl_context)
+               goto err_close;
+
+       /*
+        * If the request has a buffer, steal it, otherwise
+        * allocate a new one.  Typically, requests should already
+        * have receive buffers allocated and just swap them around
+        */
+       if (!req->rc) {
+               req->rc = kmalloc(sizeof(struct p9_fcall)+client->msize,
+                                                               GFP_KERNEL);
+               if (req->rc) {
+                       req->rc->sdata = (char *) req->rc +
+                                               sizeof(struct p9_fcall);
+                       req->rc->capacity = client->msize;
+               }
+       }
+       rpl_context->rc = req->rc;
+       if (!rpl_context->rc) {
+               kfree(rpl_context);
+               goto err_close;
+       }
+
+       /*
+        * Post a receive buffer for this request. We need to ensure
+        * there is a reply buffer available for every outstanding
+        * request. A flushed request can result in no reply for an
+        * outstanding request, so we must keep a count to avoid
+        * overflowing the RQ.
+        */
+       if (atomic_inc_return(&rdma->rq_count) <= rdma->rq_depth) {
+               err = post_recv(client, rpl_context);
+               if (err) {
+                       kfree(rpl_context->rc);
+                       kfree(rpl_context);
+                       goto err_close;
+               }
+       } else
+               atomic_dec(&rdma->rq_count);
+
+       /* remove posted receive buffer from request structure */
+       req->rc = NULL;
+
+       /* Post the request */
+       c = kmalloc(sizeof *c, GFP_KERNEL);
+       if (!c)
+               goto err_close;
+       c->req = req;
+
+       c->busa = ib_dma_map_single(rdma->cm_id->device,
+                                   c->req->tc->sdata, c->req->tc->size,
+                                   DMA_TO_DEVICE);
+       if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
+               goto error;
+
+       sge.addr = c->busa;
+       sge.length = c->req->tc->size;
+       sge.lkey = rdma->lkey;
+
+       wr.next = NULL;
+       c->wc_op = IB_WC_SEND;
+       wr.wr_id = (unsigned long) c;
+       wr.opcode = IB_WR_SEND;
+       wr.send_flags = IB_SEND_SIGNALED;
+       wr.sg_list = &sge;
+       wr.num_sge = 1;
+
+       if (down_interruptible(&rdma->sq_sem))
+               goto error;
+
+       return ib_post_send(rdma->qp, &wr, &bad_wr);
+
+ error:
+       P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
+       return -EIO;
+
+ err_close:
+       spin_lock_irqsave(&rdma->req_lock, flags);
+       if (rdma->state < P9_RDMA_CLOSING) {
+               rdma->state = P9_RDMA_CLOSING;
+               spin_unlock_irqrestore(&rdma->req_lock, flags);
+               rdma_disconnect(rdma->cm_id);
+       } else
+               spin_unlock_irqrestore(&rdma->req_lock, flags);
+       return err;
+}
+
+static void rdma_close(struct p9_client *client)
+{
+       struct p9_trans_rdma *rdma;
+
+       if (!client)
+               return;
+
+       rdma = client->trans;
+       if (!rdma)
+               return;
+
+       client->status = Disconnected;
+       rdma_disconnect(rdma->cm_id);
+       rdma_destroy_trans(rdma);
+}
+
+/**
+ * alloc_rdma - Allocate and initialize the rdma transport structure
+ * @msize: MTU
+ * @dotu: Extension attribute
+ * @opts: Mount options structure
+ */
+static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
+{
+       struct p9_trans_rdma *rdma;
+
+       rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
+       if (!rdma)
+               return NULL;
+
+       rdma->sq_depth = opts->sq_depth;
+       rdma->rq_depth = opts->rq_depth;
+       rdma->timeout = opts->timeout;
+       spin_lock_init(&rdma->req_lock);
+       init_completion(&rdma->cm_done);
+       sema_init(&rdma->sq_sem, rdma->sq_depth);
+       atomic_set(&rdma->rq_count, 0);
+
+       return rdma;
+}
+
+/* its not clear to me we can do anything after send has been posted */
+static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
+{
+       return 1;
+}
+
+/**
+ * trans_create_rdma - Transport method for creating atransport instance
+ * @client: client instance
+ * @addr: IP address string
+ * @args: Mount options string
+ */
+static int
+rdma_create_trans(struct p9_client *client, const char *addr, char *args)
+{
+       int err;
+       struct p9_rdma_opts opts;
+       struct p9_trans_rdma *rdma;
+       struct rdma_conn_param conn_param;
+       struct ib_qp_init_attr qp_attr;
+       struct ib_device_attr devattr;
+
+       /* Parse the transport specific mount options */
+       err = parse_opts(args, &opts);
+       if (err < 0)
+               return err;
+
+       /* Create and initialize the RDMA transport structure */
+       rdma = alloc_rdma(&opts);
+       if (!rdma)
+               return -ENOMEM;
+
+       /* Create the RDMA CM ID */
+       rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP);
+       if (IS_ERR(rdma->cm_id))
+               goto error;
+
+       /* Resolve the server's address */
+       rdma->addr.sin_family = AF_INET;
+       rdma->addr.sin_addr.s_addr = in_aton(addr);
+       rdma->addr.sin_port = htons(opts.port);
+       err = rdma_resolve_addr(rdma->cm_id, NULL,
+                               (struct sockaddr *)&rdma->addr,
+                               rdma->timeout);
+       if (err)
+               goto error;
+       err = wait_for_completion_interruptible(&rdma->cm_done);
+       if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
+               goto error;
+
+       /* Resolve the route to the server */
+       err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
+       if (err)
+               goto error;
+       err = wait_for_completion_interruptible(&rdma->cm_done);
+       if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
+               goto error;
+
+       /* Query the device attributes */
+       err = ib_query_device(rdma->cm_id->device, &devattr);
+       if (err)
+               goto error;
+
+       /* Create the Completion Queue */
+       rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
+                               cq_event_handler, client,
+                               opts.sq_depth + opts.rq_depth + 1, 0);
+       if (IS_ERR(rdma->cq))
+               goto error;
+       ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
+
+       /* Create the Protection Domain */
+       rdma->pd = ib_alloc_pd(rdma->cm_id->device);
+       if (IS_ERR(rdma->pd))
+               goto error;
+
+       /* Cache the DMA lkey in the transport */
+       rdma->dma_mr = NULL;
+       if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
+               rdma->lkey = rdma->cm_id->device->local_dma_lkey;
+       else {
+               rdma->dma_mr = ib_get_dma_mr(rdma->pd, IB_ACCESS_LOCAL_WRITE);
+               if (IS_ERR(rdma->dma_mr))
+                       goto error;
+               rdma->lkey = rdma->dma_mr->lkey;
+       }
+
+       /* Create the Queue Pair */
+       memset(&qp_attr, 0, sizeof qp_attr);
+       qp_attr.event_handler = qp_event_handler;
+       qp_attr.qp_context = client;
+       qp_attr.cap.max_send_wr = opts.sq_depth;
+       qp_attr.cap.max_recv_wr = opts.rq_depth;
+       qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
+       qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
+       qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
+       qp_attr.qp_type = IB_QPT_RC;
+       qp_attr.send_cq = rdma->cq;
+       qp_attr.recv_cq = rdma->cq;
+       err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
+       if (err)
+               goto error;
+       rdma->qp = rdma->cm_id->qp;
+
+       /* Request a connection */
+       memset(&conn_param, 0, sizeof(conn_param));
+       conn_param.private_data = NULL;
+       conn_param.private_data_len = 0;
+       conn_param.responder_resources = P9_RDMA_IRD;
+       conn_param.initiator_depth = P9_RDMA_ORD;
+       err = rdma_connect(rdma->cm_id, &conn_param);
+       if (err)
+               goto error;
+       err = wait_for_completion_interruptible(&rdma->cm_done);
+       if (err || (rdma->state != P9_RDMA_CONNECTED))
+               goto error;
+
+       client->trans = rdma;
+       client->status = Connected;
+
+       return 0;
+
+error:
+       rdma_destroy_trans(rdma);
+       return -ENOTCONN;
+}
+
+static struct p9_trans_module p9_rdma_trans = {
+       .name = "rdma",
+       .maxsize = P9_RDMA_MAXSIZE,
+       .def = 0,
+       .owner = THIS_MODULE,
+       .create = rdma_create_trans,
+       .close = rdma_close,
+       .request = rdma_request,
+       .cancel = rdma_cancel,
+};
+
+/**
+ * p9_trans_rdma_init - Register the 9P RDMA transport driver
+ */
+static int __init p9_trans_rdma_init(void)
+{
+       v9fs_register_trans(&p9_rdma_trans);
+       return 0;
+}
+
+static void __exit p9_trans_rdma_exit(void)
+{
+       v9fs_unregister_trans(&p9_rdma_trans);
+}
+
+module_init(p9_trans_rdma_init);
+module_exit(p9_trans_rdma_exit);
+
+MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
+MODULE_DESCRIPTION("RDMA Transport for 9P");
+MODULE_LICENSE("Dual BSD/GPL");