net/9p/trans_rdma.c

   1 /*
   2  * linux/fs/9p/trans_rdma.c
   3  *
   4  * RDMA transport layer based on the trans_fd.c implementation.
   5  *
   6  *  Copyright (C) 2008 by Tom Tucker <tom@opengridcomputing.com>
   7  *  Copyright (C) 2006 by Russ Cox <rsc@swtch.com>
   8  *  Copyright (C) 2004-2005 by Latchesar Ionkov <lucho@ionkov.net>
   9  *  Copyright (C) 2004-2008 by Eric Van Hensbergen <ericvh@gmail.com>
  10  *  Copyright (C) 1997-2002 by Ron Minnich <rminnich@sarnoff.com>
  11  *
  12  *  This program is free software; you can redistribute it and/or modify
  13  *  it under the terms of the GNU General Public License version 2
  14  *  as published by the Free Software Foundation.
  15  *
  16  *  This program is distributed in the hope that it will be useful,
  17  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  *  GNU General Public License for more details.
  20  *
  21  *  You should have received a copy of the GNU General Public License
  22  *  along with this program; if not, write to:
  23  *  Free Software Foundation
  24  *  51 Franklin Street, Fifth Floor
  25  *  Boston, MA  02111-1301  USA
  26  *
  27  */
  28
  29 #include <linux/in.h>
  30 #include <linux/module.h>
  31 #include <linux/net.h>
  32 #include <linux/ipv6.h>
  33 #include <linux/kthread.h>
  34 #include <linux/errno.h>
  35 #include <linux/kernel.h>
  36 #include <linux/un.h>
  37 #include <linux/uaccess.h>
  38 #include <linux/inet.h>
  39 #include <linux/idr.h>
  40 #include <linux/file.h>
  41 #include <linux/parser.h>
  42 #include <linux/semaphore.h>
  43 #include <linux/slab.h>
  44 #include <net/9p/9p.h>
  45 #include <net/9p/client.h>
  46 #include <net/9p/transport.h>
  47 #include <rdma/ib_verbs.h>
  48 #include <rdma/rdma_cm.h>
  49
  50 #define P9_PORT                 5640
  51 #define P9_RDMA_SQ_DEPTH        32
  52 #define P9_RDMA_RQ_DEPTH        32
  53 #define P9_RDMA_SEND_SGE        4
  54 #define P9_RDMA_RECV_SGE        4
  55 #define P9_RDMA_IRD             0
  56 #define P9_RDMA_ORD             0
  57 #define P9_RDMA_TIMEOUT         30000           /* 30 seconds */
  58 #define P9_RDMA_MAXSIZE         (4*4096)        /* Min SGE is 4, so we can
  59                                                  * safely advertise a maxsize
  60                                                  * of 64k */
  61
  62 /**
  63  * struct p9_trans_rdma - RDMA transport instance
  64  *
  65  * @state: tracks the transport state machine for connection setup and tear down
  66  * @cm_id: The RDMA CM ID
  67  * @pd: Protection Domain pointer
  68  * @qp: Queue Pair pointer
  69  * @cq: Completion Queue pointer
  70  * @dm_mr: DMA Memory Region pointer
  71  * @lkey: The local access only memory region key
  72  * @timeout: Number of uSecs to wait for connection management events
  73  * @sq_depth: The depth of the Send Queue
  74  * @sq_sem: Semaphore for the SQ
  75  * @rq_depth: The depth of the Receive Queue.
  76  * @rq_count: Count of requests in the Receive Queue.
  77  * @addr: The remote peer's address
  78  * @req_lock: Protects the active request list
  79  * @cm_done: Completion event for connection management tracking
  80  */
  81 struct p9_trans_rdma {
  82         enum {
  83                 P9_RDMA_INIT,
  84                 P9_RDMA_ADDR_RESOLVED,
  85                 P9_RDMA_ROUTE_RESOLVED,
  86                 P9_RDMA_CONNECTED,
  87                 P9_RDMA_FLUSHING,
  88                 P9_RDMA_CLOSING,
  89                 P9_RDMA_CLOSED,
  90         } state;
  91         struct rdma_cm_id *cm_id;
  92         struct ib_pd *pd;
  93         struct ib_qp *qp;
  94         struct ib_cq *cq;
  95         struct ib_mr *dma_mr;
  96         u32 lkey;
  97         long timeout;
  98         int sq_depth;
  99         struct semaphore sq_sem;
 100         int rq_depth;
 101         atomic_t rq_count;
 102         struct sockaddr_in addr;
 103         spinlock_t req_lock;
 104
 105         struct completion cm_done;
 106 };
 107
 108 /**
 109  * p9_rdma_context - Keeps track of in-process WR
 110  *
 111  * @wc_op: The original WR op for when the CQE completes in error.
 112  * @busa: Bus address to unmap when the WR completes
 113  * @req: Keeps track of requests (send)
 114  * @rc: Keepts track of replies (receive)
 115  */
 116 struct p9_rdma_req;
 117 struct p9_rdma_context {
 118         enum ib_wc_opcode wc_op;
 119         dma_addr_t busa;
 120         union {
 121                 struct p9_req_t *req;
 122                 struct p9_fcall *rc;
 123         };
 124 };
 125
 126 /**
 127  * p9_rdma_opts - Collection of mount options
 128  * @port: port of connection
 129  * @sq_depth: The requested depth of the SQ. This really doesn't need
 130  * to be any deeper than the number of threads used in the client
 131  * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
 132  * @timeout: Time to wait in msecs for CM events
 133  */
 134 struct p9_rdma_opts {
 135         short port;
 136         int sq_depth;
 137         int rq_depth;
 138         long timeout;
 139 };
 140
 141 /*
 142  * Option Parsing (code inspired by NFS code)
 143  */
 144 enum {
 145         /* Options that take integer arguments */
 146         Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout, Opt_err,
 147 };
 148
 149 static match_table_t tokens = {
 150         {Opt_port, "port=%u"},
 151         {Opt_sq_depth, "sq=%u"},
 152         {Opt_rq_depth, "rq=%u"},
 153         {Opt_timeout, "timeout=%u"},
 154         {Opt_err, NULL},
 155 };
 156
 157 /**
 158  * parse_opts - parse mount options into rdma options structure
 159  * @params: options string passed from mount
 160  * @opts: rdma transport-specific structure to parse options into
 161  *
 162  * Returns 0 upon success, -ERRNO upon failure
 163  */
 164 static int parse_opts(char *params, struct p9_rdma_opts *opts)
 165 {
 166         char *p;
 167         substring_t args[MAX_OPT_ARGS];
 168         int option;
 169         char *options, *tmp_options;
 170         int ret;
 171
 172         opts->port = P9_PORT;
 173         opts->sq_depth = P9_RDMA_SQ_DEPTH;
 174         opts->rq_depth = P9_RDMA_RQ_DEPTH;
 175         opts->timeout = P9_RDMA_TIMEOUT;
 176
 177         if (!params)
 178                 return 0;
 179
 180         tmp_options = kstrdup(params, GFP_KERNEL);
 181         if (!tmp_options) {
 182                 P9_DPRINTK(P9_DEBUG_ERROR,
 183                            "failed to allocate copy of option string\n");
 184                 return -ENOMEM;
 185         }
 186         options = tmp_options;
 187
 188         while ((p = strsep(&options, ",")) != NULL) {
 189                 int token;
 190                 int r;
 191                 if (!*p)
 192                         continue;
 193                 token = match_token(p, tokens, args);
 194                 r = match_int(&args[0], &option);
 195                 if (r < 0) {
 196                         P9_DPRINTK(P9_DEBUG_ERROR,
 197                                    "integer field, but no integer?\n");
 198                         ret = r;
 199                         continue;
 200                 }
 201                 switch (token) {
 202                 case Opt_port:
 203                         opts->port = option;
 204                         break;
 205                 case Opt_sq_depth:
 206                         opts->sq_depth = option;
 207                         break;
 208                 case Opt_rq_depth:
 209                         opts->rq_depth = option;
 210                         break;
 211                 case Opt_timeout:
 212                         opts->timeout = option;
 213                         break;
 214                 default:
 215                         continue;
 216                 }
 217         }
 218         /* RQ must be at least as large as the SQ */
 219         opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
 220         kfree(tmp_options);
 221         return 0;
 222 }
 223
 224 static int
 225 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
 226 {
 227         struct p9_client *c = id->context;
 228         struct p9_trans_rdma *rdma = c->trans;
 229         switch (event->event) {
 230         case RDMA_CM_EVENT_ADDR_RESOLVED:
 231                 BUG_ON(rdma->state != P9_RDMA_INIT);
 232                 rdma->state = P9_RDMA_ADDR_RESOLVED;
 233                 break;
 234
 235         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 236                 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
 237                 rdma->state = P9_RDMA_ROUTE_RESOLVED;
 238                 break;
 239
 240         case RDMA_CM_EVENT_ESTABLISHED:
 241                 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
 242                 rdma->state = P9_RDMA_CONNECTED;
 243                 break;
 244
 245         case RDMA_CM_EVENT_DISCONNECTED:
 246                 if (rdma)
 247                         rdma->state = P9_RDMA_CLOSED;
 248                 if (c)
 249                         c->status = Disconnected;
 250                 break;
 251
 252         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
 253                 break;
 254
 255         case RDMA_CM_EVENT_ADDR_CHANGE:
 256         case RDMA_CM_EVENT_ROUTE_ERROR:
 257         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 258         case RDMA_CM_EVENT_MULTICAST_JOIN:
 259         case RDMA_CM_EVENT_MULTICAST_ERROR:
 260         case RDMA_CM_EVENT_REJECTED:
 261         case RDMA_CM_EVENT_CONNECT_REQUEST:
 262         case RDMA_CM_EVENT_CONNECT_RESPONSE:
 263         case RDMA_CM_EVENT_CONNECT_ERROR:
 264         case RDMA_CM_EVENT_ADDR_ERROR:
 265         case RDMA_CM_EVENT_UNREACHABLE:
 266                 c->status = Disconnected;
 267                 rdma_disconnect(rdma->cm_id);
 268                 break;
 269         default:
 270                 BUG();
 271         }
 272         complete(&rdma->cm_done);
 273         return 0;
 274 }
 275
 276 static void
 277 handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
 278             struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
 279 {
 280         struct p9_req_t *req;
 281         int err = 0;
 282         int16_t tag;
 283
 284         req = NULL;
 285         ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
 286                                                          DMA_FROM_DEVICE);
 287
 288         if (status != IB_WC_SUCCESS)
 289                 goto err_out;
 290
 291         err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
 292         if (err)
 293                 goto err_out;
 294
 295         req = p9_tag_lookup(client, tag);
 296         if (!req)
 297                 goto err_out;
 298
 299         req->rc = c->rc;
 300         req->status = REQ_STATUS_RCVD;
 301         p9_client_cb(client, req);
 302
 303         return;
 304
 305  err_out:
 306         P9_DPRINTK(P9_DEBUG_ERROR, "req %p err %d status %d\n",
 307                    req, err, status);
 308         rdma->state = P9_RDMA_FLUSHING;
 309         client->status = Disconnected;
 310 }
 311
 312 static void
 313 handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
 314             struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
 315 {
 316         ib_dma_unmap_single(rdma->cm_id->device,
 317                             c->busa, c->req->tc->size,
 318                             DMA_TO_DEVICE);
 319 }
 320
 321 static void qp_event_handler(struct ib_event *event, void *context)
 322 {
 323         P9_DPRINTK(P9_DEBUG_ERROR, "QP event %d context %p\n", event->event,
 324                                                                 context);
 325 }
 326
 327 static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
 328 {
 329         struct p9_client *client = cq_context;
 330         struct p9_trans_rdma *rdma = client->trans;
 331         int ret;
 332         struct ib_wc wc;
 333
 334         ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
 335         while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
 336                 struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
 337
 338                 switch (c->wc_op) {
 339                 case IB_WC_RECV:
 340                         atomic_dec(&rdma->rq_count);
 341                         handle_recv(client, rdma, c, wc.status, wc.byte_len);
 342                         break;
 343
 344                 case IB_WC_SEND:
 345                         handle_send(client, rdma, c, wc.status, wc.byte_len);
 346                         up(&rdma->sq_sem);
 347                         break;
 348
 349                 default:
 350                         printk(KERN_ERR "9prdma: unexpected completion type, "
 351                                "c->wc_op=%d, wc.opcode=%d, status=%d\n",
 352                                c->wc_op, wc.opcode, wc.status);
 353                         break;
 354                 }
 355                 kfree(c);
 356         }
 357 }
 358
 359 static void cq_event_handler(struct ib_event *e, void *v)
 360 {
 361         P9_DPRINTK(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
 362 }
 363
 364 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
 365 {
 366         if (!rdma)
 367                 return;
 368
 369         if (rdma->dma_mr && !IS_ERR(rdma->dma_mr))
 370                 ib_dereg_mr(rdma->dma_mr);
 371
 372         if (rdma->qp && !IS_ERR(rdma->qp))
 373                 ib_destroy_qp(rdma->qp);
 374
 375         if (rdma->pd && !IS_ERR(rdma->pd))
 376                 ib_dealloc_pd(rdma->pd);
 377
 378         if (rdma->cq && !IS_ERR(rdma->cq))
 379                 ib_destroy_cq(rdma->cq);
 380
 381         if (rdma->cm_id && !IS_ERR(rdma->cm_id))
 382                 rdma_destroy_id(rdma->cm_id);
 383
 384         kfree(rdma);
 385 }
 386
 387 static int
 388 post_recv(struct p9_client *client, struct p9_rdma_context *c)
 389 {
 390         struct p9_trans_rdma *rdma = client->trans;
 391         struct ib_recv_wr wr, *bad_wr;
 392         struct ib_sge sge;
 393
 394         c->busa = ib_dma_map_single(rdma->cm_id->device,
 395                                     c->rc->sdata, client->msize,
 396                                     DMA_FROM_DEVICE);
 397         if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
 398                 goto error;
 399
 400         sge.addr = c->busa;
 401         sge.length = client->msize;
 402         sge.lkey = rdma->lkey;
 403
 404         wr.next = NULL;
 405         c->wc_op = IB_WC_RECV;
 406         wr.wr_id = (unsigned long) c;
 407         wr.sg_list = &sge;
 408         wr.num_sge = 1;
 409         return ib_post_recv(rdma->qp, &wr, &bad_wr);
 410
 411  error:
 412         P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
 413         return -EIO;
 414 }
 415
 416 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
 417 {
 418         struct p9_trans_rdma *rdma = client->trans;
 419         struct ib_send_wr wr, *bad_wr;
 420         struct ib_sge sge;
 421         int err = 0;
 422         unsigned long flags;
 423         struct p9_rdma_context *c = NULL;
 424         struct p9_rdma_context *rpl_context = NULL;
 425
 426         /* Allocate an fcall for the reply */
 427         rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
 428         if (!rpl_context) {
 429                 err = -ENOMEM;
 430                 goto err_close;
 431         }
 432
 433         /*
 434          * If the request has a buffer, steal it, otherwise
 435          * allocate a new one.  Typically, requests should already
 436          * have receive buffers allocated and just swap them around
 437          */
 438         if (!req->rc) {
 439                 req->rc = kmalloc(sizeof(struct p9_fcall)+client->msize,
 440                                   GFP_NOFS);
 441                 if (req->rc) {
 442                         req->rc->sdata = (char *) req->rc +
 443                                                 sizeof(struct p9_fcall);
 444                         req->rc->capacity = client->msize;
 445                 }
 446         }
 447         rpl_context->rc = req->rc;
 448         if (!rpl_context->rc) {
 449                 err = -ENOMEM;
 450                 goto err_free2;
 451         }
 452
 453         /*
 454          * Post a receive buffer for this request. We need to ensure
 455          * there is a reply buffer available for every outstanding
 456          * request. A flushed request can result in no reply for an
 457          * outstanding request, so we must keep a count to avoid
 458          * overflowing the RQ.
 459          */
 460         if (atomic_inc_return(&rdma->rq_count) <= rdma->rq_depth) {
 461                 err = post_recv(client, rpl_context);
 462                 if (err)
 463                         goto err_free1;
 464         } else
 465                 atomic_dec(&rdma->rq_count);
 466
 467         /* remove posted receive buffer from request structure */
 468         req->rc = NULL;
 469
 470         /* Post the request */
 471         c = kmalloc(sizeof *c, GFP_NOFS);
 472         if (!c) {
 473                 err = -ENOMEM;
 474                 goto err_free1;
 475         }
 476         c->req = req;
 477
 478         c->busa = ib_dma_map_single(rdma->cm_id->device,
 479                                     c->req->tc->sdata, c->req->tc->size,
 480                                     DMA_TO_DEVICE);
 481         if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
 482                 goto error;
 483
 484         sge.addr = c->busa;
 485         sge.length = c->req->tc->size;
 486         sge.lkey = rdma->lkey;
 487
 488         wr.next = NULL;
 489         c->wc_op = IB_WC_SEND;
 490         wr.wr_id = (unsigned long) c;
 491         wr.opcode = IB_WR_SEND;
 492         wr.send_flags = IB_SEND_SIGNALED;
 493         wr.sg_list = &sge;
 494         wr.num_sge = 1;
 495
 496         if (down_interruptible(&rdma->sq_sem))
 497                 goto error;
 498
 499         return ib_post_send(rdma->qp, &wr, &bad_wr);
 500
 501  error:
 502         kfree(c);
 503         kfree(rpl_context->rc);
 504         kfree(rpl_context);
 505         P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
 506         return -EIO;
 507  err_free1:
 508         kfree(rpl_context->rc);
 509  err_free2:
 510         kfree(rpl_context);
 511  err_close:
 512         spin_lock_irqsave(&rdma->req_lock, flags);
 513         if (rdma->state < P9_RDMA_CLOSING) {
 514                 rdma->state = P9_RDMA_CLOSING;
 515                 spin_unlock_irqrestore(&rdma->req_lock, flags);
 516                 rdma_disconnect(rdma->cm_id);
 517         } else
 518                 spin_unlock_irqrestore(&rdma->req_lock, flags);
 519         return err;
 520 }
 521
 522 static void rdma_close(struct p9_client *client)
 523 {
 524         struct p9_trans_rdma *rdma;
 525
 526         if (!client)
 527                 return;
 528
 529         rdma = client->trans;
 530         if (!rdma)
 531                 return;
 532
 533         client->status = Disconnected;
 534         rdma_disconnect(rdma->cm_id);
 535         rdma_destroy_trans(rdma);
 536 }
 537
 538 /**
 539  * alloc_rdma - Allocate and initialize the rdma transport structure
 540  * @opts: Mount options structure
 541  */
 542 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
 543 {
 544         struct p9_trans_rdma *rdma;
 545
 546         rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
 547         if (!rdma)
 548                 return NULL;
 549
 550         rdma->sq_depth = opts->sq_depth;
 551         rdma->rq_depth = opts->rq_depth;
 552         rdma->timeout = opts->timeout;
 553         spin_lock_init(&rdma->req_lock);
 554         init_completion(&rdma->cm_done);
 555         sema_init(&rdma->sq_sem, rdma->sq_depth);
 556         atomic_set(&rdma->rq_count, 0);
 557
 558         return rdma;
 559 }
 560
 561 /* its not clear to me we can do anything after send has been posted */
 562 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
 563 {
 564         return 1;
 565 }
 566
 567 /**
 568  * trans_create_rdma - Transport method for creating atransport instance
 569  * @client: client instance
 570  * @addr: IP address string
 571  * @args: Mount options string
 572  */
 573 static int
 574 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
 575 {
 576         int err;
 577         struct p9_rdma_opts opts;
 578         struct p9_trans_rdma *rdma;
 579         struct rdma_conn_param conn_param;
 580         struct ib_qp_init_attr qp_attr;
 581         struct ib_device_attr devattr;
 582
 583         /* Parse the transport specific mount options */
 584         err = parse_opts(args, &opts);
 585         if (err < 0)
 586                 return err;
 587
 588         /* Create and initialize the RDMA transport structure */
 589         rdma = alloc_rdma(&opts);
 590         if (!rdma)
 591                 return -ENOMEM;
 592
 593         /* Create the RDMA CM ID */
 594         rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP,
 595                                      IB_QPT_RC);
 596         if (IS_ERR(rdma->cm_id))
 597                 goto error;
 598
 599         /* Associate the client with the transport */
 600         client->trans = rdma;
 601
 602         /* Resolve the server's address */
 603         rdma->addr.sin_family = AF_INET;
 604         rdma->addr.sin_addr.s_addr = in_aton(addr);
 605         rdma->addr.sin_port = htons(opts.port);
 606         err = rdma_resolve_addr(rdma->cm_id, NULL,
 607                                 (struct sockaddr *)&rdma->addr,
 608                                 rdma->timeout);
 609         if (err)
 610                 goto error;
 611         err = wait_for_completion_interruptible(&rdma->cm_done);
 612         if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
 613                 goto error;
 614
 615         /* Resolve the route to the server */
 616         err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
 617         if (err)
 618                 goto error;
 619         err = wait_for_completion_interruptible(&rdma->cm_done);
 620         if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
 621                 goto error;
 622
 623         /* Query the device attributes */
 624         err = ib_query_device(rdma->cm_id->device, &devattr);
 625         if (err)
 626                 goto error;
 627
 628         /* Create the Completion Queue */
 629         rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
 630                                 cq_event_handler, client,
 631                                 opts.sq_depth + opts.rq_depth + 1, 0);
 632         if (IS_ERR(rdma->cq))
 633                 goto error;
 634         ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
 635
 636         /* Create the Protection Domain */
 637         rdma->pd = ib_alloc_pd(rdma->cm_id->device);
 638         if (IS_ERR(rdma->pd))
 639                 goto error;
 640
 641         /* Cache the DMA lkey in the transport */
 642         rdma->dma_mr = NULL;
 643         if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
 644                 rdma->lkey = rdma->cm_id->device->local_dma_lkey;
 645         else {
 646                 rdma->dma_mr = ib_get_dma_mr(rdma->pd, IB_ACCESS_LOCAL_WRITE);
 647                 if (IS_ERR(rdma->dma_mr))
 648                         goto error;
 649                 rdma->lkey = rdma->dma_mr->lkey;
 650         }
 651
 652         /* Create the Queue Pair */
 653         memset(&qp_attr, 0, sizeof qp_attr);
 654         qp_attr.event_handler = qp_event_handler;
 655         qp_attr.qp_context = client;
 656         qp_attr.cap.max_send_wr = opts.sq_depth;
 657         qp_attr.cap.max_recv_wr = opts.rq_depth;
 658         qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
 659         qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
 660         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 661         qp_attr.qp_type = IB_QPT_RC;
 662         qp_attr.send_cq = rdma->cq;
 663         qp_attr.recv_cq = rdma->cq;
 664         err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
 665         if (err)
 666                 goto error;
 667         rdma->qp = rdma->cm_id->qp;
 668
 669         /* Request a connection */
 670         memset(&conn_param, 0, sizeof(conn_param));
 671         conn_param.private_data = NULL;
 672         conn_param.private_data_len = 0;
 673         conn_param.responder_resources = P9_RDMA_IRD;
 674         conn_param.initiator_depth = P9_RDMA_ORD;
 675         err = rdma_connect(rdma->cm_id, &conn_param);
 676         if (err)
 677                 goto error;
 678         err = wait_for_completion_interruptible(&rdma->cm_done);
 679         if (err || (rdma->state != P9_RDMA_CONNECTED))
 680                 goto error;
 681
 682         client->status = Connected;
 683
 684         return 0;
 685
 686 error:
 687         rdma_destroy_trans(rdma);
 688         return -ENOTCONN;
 689 }
 690
 691 static struct p9_trans_module p9_rdma_trans = {
 692         .name = "rdma",
 693         .maxsize = P9_RDMA_MAXSIZE,
 694         .def = 0,
 695         .owner = THIS_MODULE,
 696         .create = rdma_create_trans,
 697         .close = rdma_close,
 698         .request = rdma_request,
 699         .cancel = rdma_cancel,
 700 };
 701
 702 /**
 703  * p9_trans_rdma_init - Register the 9P RDMA transport driver
 704  */
 705 static int __init p9_trans_rdma_init(void)
 706 {
 707         v9fs_register_trans(&p9_rdma_trans);
 708         return 0;
 709 }
 710
 711 static void __exit p9_trans_rdma_exit(void)
 712 {
 713         v9fs_unregister_trans(&p9_rdma_trans);
 714 }
 715
 716 module_init(p9_trans_rdma_init);
 717 module_exit(p9_trans_rdma_exit);
 718
 719 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
 720 MODULE_DESCRIPTION("RDMA Transport for 9P");
 721 MODULE_LICENSE("Dual BSD/GPL");