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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  30
  31 #include <linux/in.h>
  32 #include <linux/module.h>
  33 #include <linux/net.h>
  34 #include <linux/ipv6.h>
  35 #include <linux/kthread.h>
  36 #include <linux/errno.h>
  37 #include <linux/kernel.h>
  38 #include <linux/un.h>
  39 #include <linux/uaccess.h>
  40 #include <linux/inet.h>
  41 #include <linux/idr.h>
  42 #include <linux/file.h>
  43 #include <linux/parser.h>
  44 #include <linux/semaphore.h>
  45 #include <linux/slab.h>
  46 #include <net/9p/9p.h>
  47 #include <net/9p/client.h>
  48 #include <net/9p/transport.h>
  49 #include <rdma/ib_verbs.h>
  50 #include <rdma/rdma_cm.h>
  51
  52 #define P9_PORT                 5640
  53 #define P9_RDMA_SQ_DEPTH        32
  54 #define P9_RDMA_RQ_DEPTH        32
  55 #define P9_RDMA_SEND_SGE        4
  56 #define P9_RDMA_RECV_SGE        4
  57 #define P9_RDMA_IRD             0
  58 #define P9_RDMA_ORD             0
  59 #define P9_RDMA_TIMEOUT         30000           /* 30 seconds */
  60 #define P9_RDMA_MAXSIZE         (1024*1024)     /* 1MB */
  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_sem: Semaphore for the RQ
  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         struct semaphore rq_sem;
 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
 171         opts->port = P9_PORT;
 172         opts->sq_depth = P9_RDMA_SQ_DEPTH;
 173         opts->rq_depth = P9_RDMA_RQ_DEPTH;
 174         opts->timeout = P9_RDMA_TIMEOUT;
 175
 176         if (!params)
 177                 return 0;
 178
 179         tmp_options = kstrdup(params, GFP_KERNEL);
 180         if (!tmp_options) {
 181                 p9_debug(P9_DEBUG_ERROR,
 182                          "failed to allocate copy of option string\n");
 183                 return -ENOMEM;
 184         }
 185         options = tmp_options;
 186
 187         while ((p = strsep(&options, ",")) != NULL) {
 188                 int token;
 189                 int r;
 190                 if (!*p)
 191                         continue;
 192                 token = match_token(p, tokens, args);
 193                 r = match_int(&args[0], &option);
 194                 if (r < 0) {
 195                         p9_debug(P9_DEBUG_ERROR,
 196                                  "integer field, but no integer?\n");
 197                         continue;
 198                 }
 199                 switch (token) {
 200                 case Opt_port:
 201                         opts->port = option;
 202                         break;
 203                 case Opt_sq_depth:
 204                         opts->sq_depth = option;
 205                         break;
 206                 case Opt_rq_depth:
 207                         opts->rq_depth = option;
 208                         break;
 209                 case Opt_timeout:
 210                         opts->timeout = option;
 211                         break;
 212                 default:
 213                         continue;
 214                 }
 215         }
 216         /* RQ must be at least as large as the SQ */
 217         opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
 218         kfree(tmp_options);
 219         return 0;
 220 }
 221
 222 static int
 223 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
 224 {
 225         struct p9_client *c = id->context;
 226         struct p9_trans_rdma *rdma = c->trans;
 227         switch (event->event) {
 228         case RDMA_CM_EVENT_ADDR_RESOLVED:
 229                 BUG_ON(rdma->state != P9_RDMA_INIT);
 230                 rdma->state = P9_RDMA_ADDR_RESOLVED;
 231                 break;
 232
 233         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 234                 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
 235                 rdma->state = P9_RDMA_ROUTE_RESOLVED;
 236                 break;
 237
 238         case RDMA_CM_EVENT_ESTABLISHED:
 239                 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
 240                 rdma->state = P9_RDMA_CONNECTED;
 241                 break;
 242
 243         case RDMA_CM_EVENT_DISCONNECTED:
 244                 if (rdma)
 245                         rdma->state = P9_RDMA_CLOSED;
 246                 if (c)
 247                         c->status = Disconnected;
 248                 break;
 249
 250         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
 251                 break;
 252
 253         case RDMA_CM_EVENT_ADDR_CHANGE:
 254         case RDMA_CM_EVENT_ROUTE_ERROR:
 255         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 256         case RDMA_CM_EVENT_MULTICAST_JOIN:
 257         case RDMA_CM_EVENT_MULTICAST_ERROR:
 258         case RDMA_CM_EVENT_REJECTED:
 259         case RDMA_CM_EVENT_CONNECT_REQUEST:
 260         case RDMA_CM_EVENT_CONNECT_RESPONSE:
 261         case RDMA_CM_EVENT_CONNECT_ERROR:
 262         case RDMA_CM_EVENT_ADDR_ERROR:
 263         case RDMA_CM_EVENT_UNREACHABLE:
 264                 c->status = Disconnected;
 265                 rdma_disconnect(rdma->cm_id);
 266                 break;
 267         default:
 268                 BUG();
 269         }
 270         complete(&rdma->cm_done);
 271         return 0;
 272 }
 273
 274 static void
 275 handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
 276             struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
 277 {
 278         struct p9_req_t *req;
 279         int err = 0;
 280         int16_t tag;
 281
 282         req = NULL;
 283         ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
 284                                                          DMA_FROM_DEVICE);
 285
 286         if (status != IB_WC_SUCCESS)
 287                 goto err_out;
 288
 289         err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
 290         if (err)
 291                 goto err_out;
 292
 293         req = p9_tag_lookup(client, tag);
 294         if (!req)
 295                 goto err_out;
 296
 297         /* Check that we have not yet received a reply for this request.
 298          */
 299         if (unlikely(req->rc)) {
 300                 pr_err("Duplicate reply for request %d", tag);
 301                 goto err_out;
 302         }
 303
 304         req->rc = c->rc;
 305         req->status = REQ_STATUS_RCVD;
 306         p9_client_cb(client, req);
 307
 308         return;
 309
 310  err_out:
 311         p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n", req, err, status);
 312         rdma->state = P9_RDMA_FLUSHING;
 313         client->status = Disconnected;
 314 }
 315
 316 static void
 317 handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
 318             struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
 319 {
 320         ib_dma_unmap_single(rdma->cm_id->device,
 321                             c->busa, c->req->tc->size,
 322                             DMA_TO_DEVICE);
 323 }
 324
 325 static void qp_event_handler(struct ib_event *event, void *context)
 326 {
 327         p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
 328                  event->event, context);
 329 }
 330
 331 static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
 332 {
 333         struct p9_client *client = cq_context;
 334         struct p9_trans_rdma *rdma = client->trans;
 335         int ret;
 336         struct ib_wc wc;
 337
 338         ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
 339         while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
 340                 struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
 341
 342                 switch (c->wc_op) {
 343                 case IB_WC_RECV:
 344                         handle_recv(client, rdma, c, wc.status, wc.byte_len);
 345                         up(&rdma->rq_sem);
 346                         break;
 347
 348                 case IB_WC_SEND:
 349                         handle_send(client, rdma, c, wc.status, wc.byte_len);
 350                         up(&rdma->sq_sem);
 351                         break;
 352
 353                 default:
 354                         pr_err("unexpected completion type, c->wc_op=%d, wc.opcode=%d, status=%d\n",
 355                                c->wc_op, wc.opcode, wc.status);
 356                         break;
 357                 }
 358                 kfree(c);
 359         }
 360 }
 361
 362 static void cq_event_handler(struct ib_event *e, void *v)
 363 {
 364         p9_debug(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
 365 }
 366
 367 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
 368 {
 369         if (!rdma)
 370                 return;
 371
 372         if (rdma->dma_mr && !IS_ERR(rdma->dma_mr))
 373                 ib_dereg_mr(rdma->dma_mr);
 374
 375         if (rdma->qp && !IS_ERR(rdma->qp))
 376                 ib_destroy_qp(rdma->qp);
 377
 378         if (rdma->pd && !IS_ERR(rdma->pd))
 379                 ib_dealloc_pd(rdma->pd);
 380
 381         if (rdma->cq && !IS_ERR(rdma->cq))
 382                 ib_destroy_cq(rdma->cq);
 383
 384         if (rdma->cm_id && !IS_ERR(rdma->cm_id))
 385                 rdma_destroy_id(rdma->cm_id);
 386
 387         kfree(rdma);
 388 }
 389
 390 static int
 391 post_recv(struct p9_client *client, struct p9_rdma_context *c)
 392 {
 393         struct p9_trans_rdma *rdma = client->trans;
 394         struct ib_recv_wr wr, *bad_wr;
 395         struct ib_sge sge;
 396
 397         c->busa = ib_dma_map_single(rdma->cm_id->device,
 398                                     c->rc->sdata, client->msize,
 399                                     DMA_FROM_DEVICE);
 400         if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
 401                 goto error;
 402
 403         sge.addr = c->busa;
 404         sge.length = client->msize;
 405         sge.lkey = rdma->lkey;
 406
 407         wr.next = NULL;
 408         c->wc_op = IB_WC_RECV;
 409         wr.wr_id = (unsigned long) c;
 410         wr.sg_list = &sge;
 411         wr.num_sge = 1;
 412         return ib_post_recv(rdma->qp, &wr, &bad_wr);
 413
 414  error:
 415         p9_debug(P9_DEBUG_ERROR, "EIO\n");
 416         return -EIO;
 417 }
 418
 419 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
 420 {
 421         struct p9_trans_rdma *rdma = client->trans;
 422         struct ib_send_wr wr, *bad_wr;
 423         struct ib_sge sge;
 424         int err = 0;
 425         unsigned long flags;
 426         struct p9_rdma_context *c = NULL;
 427         struct p9_rdma_context *rpl_context = NULL;
 428
 429         /* Allocate an fcall for the reply */
 430         rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
 431         if (!rpl_context) {
 432                 err = -ENOMEM;
 433                 goto recv_error;
 434         }
 435         rpl_context->rc = req->rc;
 436
 437         /*
 438          * Post a receive buffer for this request. We need to ensure
 439          * there is a reply buffer available for every outstanding
 440          * request. A flushed request can result in no reply for an
 441          * outstanding request, so we must keep a count to avoid
 442          * overflowing the RQ.
 443          */
 444         if (down_interruptible(&rdma->rq_sem)) {
 445                 err = -EINTR;
 446                 goto recv_error;
 447         }
 448
 449         err = post_recv(client, rpl_context);
 450         if (err) {
 451                 p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
 452                 goto recv_error;
 453         }
 454
 455         /* remove posted receive buffer from request structure */
 456         req->rc = NULL;
 457
 458         /* Post the request */
 459         c = kmalloc(sizeof *c, GFP_NOFS);
 460         if (!c) {
 461                 err = -ENOMEM;
 462                 goto send_error;
 463         }
 464         c->req = req;
 465
 466         c->busa = ib_dma_map_single(rdma->cm_id->device,
 467                                     c->req->tc->sdata, c->req->tc->size,
 468                                     DMA_TO_DEVICE);
 469         if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
 470                 err = -EIO;
 471                 goto send_error;
 472         }
 473
 474         sge.addr = c->busa;
 475         sge.length = c->req->tc->size;
 476         sge.lkey = rdma->lkey;
 477
 478         wr.next = NULL;
 479         c->wc_op = IB_WC_SEND;
 480         wr.wr_id = (unsigned long) c;
 481         wr.opcode = IB_WR_SEND;
 482         wr.send_flags = IB_SEND_SIGNALED;
 483         wr.sg_list = &sge;
 484         wr.num_sge = 1;
 485
 486         if (down_interruptible(&rdma->sq_sem)) {
 487                 err = -EINTR;
 488                 goto send_error;
 489         }
 490
 491         err = ib_post_send(rdma->qp, &wr, &bad_wr);
 492         if (err)
 493                 goto send_error;
 494
 495         /* Success */
 496         return 0;
 497
 498  /* Handle errors that happened during or while preparing the send: */
 499  send_error:
 500         kfree(c);
 501         p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
 502         return err;
 503
 504  /* Handle errors that happened during or while preparing post_recv(): */
 505  recv_error:
 506         kfree(rpl_context);
 507         spin_lock_irqsave(&rdma->req_lock, flags);
 508         if (rdma->state < P9_RDMA_CLOSING) {
 509                 rdma->state = P9_RDMA_CLOSING;
 510                 spin_unlock_irqrestore(&rdma->req_lock, flags);
 511                 rdma_disconnect(rdma->cm_id);
 512         } else
 513                 spin_unlock_irqrestore(&rdma->req_lock, flags);
 514         return err;
 515 }
 516
 517 static void rdma_close(struct p9_client *client)
 518 {
 519         struct p9_trans_rdma *rdma;
 520
 521         if (!client)
 522                 return;
 523
 524         rdma = client->trans;
 525         if (!rdma)
 526                 return;
 527
 528         client->status = Disconnected;
 529         rdma_disconnect(rdma->cm_id);
 530         rdma_destroy_trans(rdma);
 531 }
 532
 533 /**
 534  * alloc_rdma - Allocate and initialize the rdma transport structure
 535  * @opts: Mount options structure
 536  */
 537 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
 538 {
 539         struct p9_trans_rdma *rdma;
 540
 541         rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
 542         if (!rdma)
 543                 return NULL;
 544
 545         rdma->sq_depth = opts->sq_depth;
 546         rdma->rq_depth = opts->rq_depth;
 547         rdma->timeout = opts->timeout;
 548         spin_lock_init(&rdma->req_lock);
 549         init_completion(&rdma->cm_done);
 550         sema_init(&rdma->sq_sem, rdma->sq_depth);
 551         sema_init(&rdma->rq_sem, rdma->rq_depth);
 552
 553         return rdma;
 554 }
 555
 556 /* its not clear to me we can do anything after send has been posted */
 557 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
 558 {
 559         return 1;
 560 }
 561
 562 /**
 563  * trans_create_rdma - Transport method for creating atransport instance
 564  * @client: client instance
 565  * @addr: IP address string
 566  * @args: Mount options string
 567  */
 568 static int
 569 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
 570 {
 571         int err;
 572         struct p9_rdma_opts opts;
 573         struct p9_trans_rdma *rdma;
 574         struct rdma_conn_param conn_param;
 575         struct ib_qp_init_attr qp_attr;
 576         struct ib_device_attr devattr;
 577
 578         /* Parse the transport specific mount options */
 579         err = parse_opts(args, &opts);
 580         if (err < 0)
 581                 return err;
 582
 583         /* Create and initialize the RDMA transport structure */
 584         rdma = alloc_rdma(&opts);
 585         if (!rdma)
 586                 return -ENOMEM;
 587
 588         /* Create the RDMA CM ID */
 589         rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP,
 590                                      IB_QPT_RC);
 591         if (IS_ERR(rdma->cm_id))
 592                 goto error;
 593
 594         /* Associate the client with the transport */
 595         client->trans = rdma;
 596
 597         /* Resolve the server's address */
 598         rdma->addr.sin_family = AF_INET;
 599         rdma->addr.sin_addr.s_addr = in_aton(addr);
 600         rdma->addr.sin_port = htons(opts.port);
 601         err = rdma_resolve_addr(rdma->cm_id, NULL,
 602                                 (struct sockaddr *)&rdma->addr,
 603                                 rdma->timeout);
 604         if (err)
 605                 goto error;
 606         err = wait_for_completion_interruptible(&rdma->cm_done);
 607         if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
 608                 goto error;
 609
 610         /* Resolve the route to the server */
 611         err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
 612         if (err)
 613                 goto error;
 614         err = wait_for_completion_interruptible(&rdma->cm_done);
 615         if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
 616                 goto error;
 617
 618         /* Query the device attributes */
 619         err = ib_query_device(rdma->cm_id->device, &devattr);
 620         if (err)
 621                 goto error;
 622
 623         /* Create the Completion Queue */
 624         rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
 625                                 cq_event_handler, client,
 626                                 opts.sq_depth + opts.rq_depth + 1, 0);
 627         if (IS_ERR(rdma->cq))
 628                 goto error;
 629         ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
 630
 631         /* Create the Protection Domain */
 632         rdma->pd = ib_alloc_pd(rdma->cm_id->device);
 633         if (IS_ERR(rdma->pd))
 634                 goto error;
 635
 636         /* Cache the DMA lkey in the transport */
 637         rdma->dma_mr = NULL;
 638         if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
 639                 rdma->lkey = rdma->cm_id->device->local_dma_lkey;
 640         else {
 641                 rdma->dma_mr = ib_get_dma_mr(rdma->pd, IB_ACCESS_LOCAL_WRITE);
 642                 if (IS_ERR(rdma->dma_mr))
 643                         goto error;
 644                 rdma->lkey = rdma->dma_mr->lkey;
 645         }
 646
 647         /* Create the Queue Pair */
 648         memset(&qp_attr, 0, sizeof qp_attr);
 649         qp_attr.event_handler = qp_event_handler;
 650         qp_attr.qp_context = client;
 651         qp_attr.cap.max_send_wr = opts.sq_depth;
 652         qp_attr.cap.max_recv_wr = opts.rq_depth;
 653         qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
 654         qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
 655         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 656         qp_attr.qp_type = IB_QPT_RC;
 657         qp_attr.send_cq = rdma->cq;
 658         qp_attr.recv_cq = rdma->cq;
 659         err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
 660         if (err)
 661                 goto error;
 662         rdma->qp = rdma->cm_id->qp;
 663
 664         /* Request a connection */
 665         memset(&conn_param, 0, sizeof(conn_param));
 666         conn_param.private_data = NULL;
 667         conn_param.private_data_len = 0;
 668         conn_param.responder_resources = P9_RDMA_IRD;
 669         conn_param.initiator_depth = P9_RDMA_ORD;
 670         err = rdma_connect(rdma->cm_id, &conn_param);
 671         if (err)
 672                 goto error;
 673         err = wait_for_completion_interruptible(&rdma->cm_done);
 674         if (err || (rdma->state != P9_RDMA_CONNECTED))
 675                 goto error;
 676
 677         client->status = Connected;
 678
 679         return 0;
 680
 681 error:
 682         rdma_destroy_trans(rdma);
 683         return -ENOTCONN;
 684 }
 685
 686 static struct p9_trans_module p9_rdma_trans = {
 687         .name = "rdma",
 688         .maxsize = P9_RDMA_MAXSIZE,
 689         .def = 0,
 690         .owner = THIS_MODULE,
 691         .create = rdma_create_trans,
 692         .close = rdma_close,
 693         .request = rdma_request,
 694         .cancel = rdma_cancel,
 695 };
 696
 697 /**
 698  * p9_trans_rdma_init - Register the 9P RDMA transport driver
 699  */
 700 static int __init p9_trans_rdma_init(void)
 701 {
 702         v9fs_register_trans(&p9_rdma_trans);
 703         return 0;
 704 }
 705
 706 static void __exit p9_trans_rdma_exit(void)
 707 {
 708         v9fs_unregister_trans(&p9_rdma_trans);
 709 }
 710
 711 module_init(p9_trans_rdma_init);
 712 module_exit(p9_trans_rdma_exit);
 713
 714 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
 715 MODULE_DESCRIPTION("RDMA Transport for 9P");
 716 MODULE_LICENSE("Dual BSD/GPL");