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  * @excess_rc : Amount of posted Receive Contexts without a pending request.
  78  *              See rdma_request()
  79  * @addr: The remote peer's address
  80  * @req_lock: Protects the active request list
  81  * @cm_done: Completion event for connection management tracking
  82  */
  83 struct p9_trans_rdma {
  84         enum {
  85                 P9_RDMA_INIT,
  86                 P9_RDMA_ADDR_RESOLVED,
  87                 P9_RDMA_ROUTE_RESOLVED,
  88                 P9_RDMA_CONNECTED,
  89                 P9_RDMA_FLUSHING,
  90                 P9_RDMA_CLOSING,
  91                 P9_RDMA_CLOSED,
  92         } state;
  93         struct rdma_cm_id *cm_id;
  94         struct ib_pd *pd;
  95         struct ib_qp *qp;
  96         struct ib_cq *cq;
  97         struct ib_mr *dma_mr;
  98         u32 lkey;
  99         long timeout;
 100         int sq_depth;
 101         struct semaphore sq_sem;
 102         int rq_depth;
 103         struct semaphore rq_sem;
 104         atomic_t excess_rc;
 105         struct sockaddr_in addr;
 106         spinlock_t req_lock;
 107
 108         struct completion cm_done;
 109 };
 110
 111 /**
 112  * p9_rdma_context - Keeps track of in-process WR
 113  *
 114  * @wc_op: The original WR op for when the CQE completes in error.
 115  * @busa: Bus address to unmap when the WR completes
 116  * @req: Keeps track of requests (send)
 117  * @rc: Keepts track of replies (receive)
 118  */
 119 struct p9_rdma_req;
 120 struct p9_rdma_context {
 121         enum ib_wc_opcode wc_op;
 122         dma_addr_t busa;
 123         union {
 124                 struct p9_req_t *req;
 125                 struct p9_fcall *rc;
 126         };
 127 };
 128
 129 /**
 130  * p9_rdma_opts - Collection of mount options
 131  * @port: port of connection
 132  * @sq_depth: The requested depth of the SQ. This really doesn't need
 133  * to be any deeper than the number of threads used in the client
 134  * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
 135  * @timeout: Time to wait in msecs for CM events
 136  */
 137 struct p9_rdma_opts {
 138         short port;
 139         int sq_depth;
 140         int rq_depth;
 141         long timeout;
 142 };
 143
 144 /*
 145  * Option Parsing (code inspired by NFS code)
 146  */
 147 enum {
 148         /* Options that take integer arguments */
 149         Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout, Opt_err,
 150 };
 151
 152 static match_table_t tokens = {
 153         {Opt_port, "port=%u"},
 154         {Opt_sq_depth, "sq=%u"},
 155         {Opt_rq_depth, "rq=%u"},
 156         {Opt_timeout, "timeout=%u"},
 157         {Opt_err, NULL},
 158 };
 159
 160 /**
 161  * parse_opts - parse mount options into rdma options structure
 162  * @params: options string passed from mount
 163  * @opts: rdma transport-specific structure to parse options into
 164  *
 165  * Returns 0 upon success, -ERRNO upon failure
 166  */
 167 static int parse_opts(char *params, struct p9_rdma_opts *opts)
 168 {
 169         char *p;
 170         substring_t args[MAX_OPT_ARGS];
 171         int option;
 172         char *options, *tmp_options;
 173
 174         opts->port = P9_PORT;
 175         opts->sq_depth = P9_RDMA_SQ_DEPTH;
 176         opts->rq_depth = P9_RDMA_RQ_DEPTH;
 177         opts->timeout = P9_RDMA_TIMEOUT;
 178
 179         if (!params)
 180                 return 0;
 181
 182         tmp_options = kstrdup(params, GFP_KERNEL);
 183         if (!tmp_options) {
 184                 p9_debug(P9_DEBUG_ERROR,
 185                          "failed to allocate copy of option string\n");
 186                 return -ENOMEM;
 187         }
 188         options = tmp_options;
 189
 190         while ((p = strsep(&options, ",")) != NULL) {
 191                 int token;
 192                 int r;
 193                 if (!*p)
 194                         continue;
 195                 token = match_token(p, tokens, args);
 196                 r = match_int(&args[0], &option);
 197                 if (r < 0) {
 198                         p9_debug(P9_DEBUG_ERROR,
 199                                  "integer field, but no integer?\n");
 200                         continue;
 201                 }
 202                 switch (token) {
 203                 case Opt_port:
 204                         opts->port = option;
 205                         break;
 206                 case Opt_sq_depth:
 207                         opts->sq_depth = option;
 208                         break;
 209                 case Opt_rq_depth:
 210                         opts->rq_depth = option;
 211                         break;
 212                 case Opt_timeout:
 213                         opts->timeout = option;
 214                         break;
 215                 default:
 216                         continue;
 217                 }
 218         }
 219         /* RQ must be at least as large as the SQ */
 220         opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
 221         kfree(tmp_options);
 222         return 0;
 223 }
 224
 225 static int
 226 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
 227 {
 228         struct p9_client *c = id->context;
 229         struct p9_trans_rdma *rdma = c->trans;
 230         switch (event->event) {
 231         case RDMA_CM_EVENT_ADDR_RESOLVED:
 232                 BUG_ON(rdma->state != P9_RDMA_INIT);
 233                 rdma->state = P9_RDMA_ADDR_RESOLVED;
 234                 break;
 235
 236         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 237                 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
 238                 rdma->state = P9_RDMA_ROUTE_RESOLVED;
 239                 break;
 240
 241         case RDMA_CM_EVENT_ESTABLISHED:
 242                 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
 243                 rdma->state = P9_RDMA_CONNECTED;
 244                 break;
 245
 246         case RDMA_CM_EVENT_DISCONNECTED:
 247                 if (rdma)
 248                         rdma->state = P9_RDMA_CLOSED;
 249                 if (c)
 250                         c->status = Disconnected;
 251                 break;
 252
 253         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
 254                 break;
 255
 256         case RDMA_CM_EVENT_ADDR_CHANGE:
 257         case RDMA_CM_EVENT_ROUTE_ERROR:
 258         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 259         case RDMA_CM_EVENT_MULTICAST_JOIN:
 260         case RDMA_CM_EVENT_MULTICAST_ERROR:
 261         case RDMA_CM_EVENT_REJECTED:
 262         case RDMA_CM_EVENT_CONNECT_REQUEST:
 263         case RDMA_CM_EVENT_CONNECT_RESPONSE:
 264         case RDMA_CM_EVENT_CONNECT_ERROR:
 265         case RDMA_CM_EVENT_ADDR_ERROR:
 266         case RDMA_CM_EVENT_UNREACHABLE:
 267                 c->status = Disconnected;
 268                 rdma_disconnect(rdma->cm_id);
 269                 break;
 270         default:
 271                 BUG();
 272         }
 273         complete(&rdma->cm_done);
 274         return 0;
 275 }
 276
 277 static void
 278 handle_recv(struct p9_client *client, struct p9_trans_rdma *rdma,
 279             struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
 280 {
 281         struct p9_req_t *req;
 282         int err = 0;
 283         int16_t tag;
 284
 285         req = NULL;
 286         ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
 287                                                          DMA_FROM_DEVICE);
 288
 289         if (status != IB_WC_SUCCESS)
 290                 goto err_out;
 291
 292         err = p9_parse_header(c->rc, NULL, NULL, &tag, 1);
 293         if (err)
 294                 goto err_out;
 295
 296         req = p9_tag_lookup(client, tag);
 297         if (!req)
 298                 goto err_out;
 299
 300         /* Check that we have not yet received a reply for this request.
 301          */
 302         if (unlikely(req->rc)) {
 303                 pr_err("Duplicate reply for request %d", tag);
 304                 goto err_out;
 305         }
 306
 307         req->rc = c->rc;
 308         p9_client_cb(client, req, REQ_STATUS_RCVD);
 309
 310         return;
 311
 312  err_out:
 313         p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n", req, err, status);
 314         rdma->state = P9_RDMA_FLUSHING;
 315         client->status = Disconnected;
 316 }
 317
 318 static void
 319 handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
 320             struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
 321 {
 322         ib_dma_unmap_single(rdma->cm_id->device,
 323                             c->busa, c->req->tc->size,
 324                             DMA_TO_DEVICE);
 325 }
 326
 327 static void qp_event_handler(struct ib_event *event, void *context)
 328 {
 329         p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
 330                  event->event, context);
 331 }
 332
 333 static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
 334 {
 335         struct p9_client *client = cq_context;
 336         struct p9_trans_rdma *rdma = client->trans;
 337         int ret;
 338         struct ib_wc wc;
 339
 340         ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
 341         while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
 342                 struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
 343
 344                 switch (c->wc_op) {
 345                 case IB_WC_RECV:
 346                         handle_recv(client, rdma, c, wc.status, wc.byte_len);
 347                         up(&rdma->rq_sem);
 348                         break;
 349
 350                 case IB_WC_SEND:
 351                         handle_send(client, rdma, c, wc.status, wc.byte_len);
 352                         up(&rdma->sq_sem);
 353                         break;
 354
 355                 default:
 356                         pr_err("unexpected completion type, c->wc_op=%d, wc.opcode=%d, status=%d\n",
 357                                c->wc_op, wc.opcode, wc.status);
 358                         break;
 359                 }
 360                 kfree(c);
 361         }
 362 }
 363
 364 static void cq_event_handler(struct ib_event *e, void *v)
 365 {
 366         p9_debug(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
 367 }
 368
 369 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
 370 {
 371         if (!rdma)
 372                 return;
 373
 374         if (rdma->dma_mr && !IS_ERR(rdma->dma_mr))
 375                 ib_dereg_mr(rdma->dma_mr);
 376
 377         if (rdma->qp && !IS_ERR(rdma->qp))
 378                 ib_destroy_qp(rdma->qp);
 379
 380         if (rdma->pd && !IS_ERR(rdma->pd))
 381                 ib_dealloc_pd(rdma->pd);
 382
 383         if (rdma->cq && !IS_ERR(rdma->cq))
 384                 ib_destroy_cq(rdma->cq);
 385
 386         if (rdma->cm_id && !IS_ERR(rdma->cm_id))
 387                 rdma_destroy_id(rdma->cm_id);
 388
 389         kfree(rdma);
 390 }
 391
 392 static int
 393 post_recv(struct p9_client *client, struct p9_rdma_context *c)
 394 {
 395         struct p9_trans_rdma *rdma = client->trans;
 396         struct ib_recv_wr wr, *bad_wr;
 397         struct ib_sge sge;
 398
 399         c->busa = ib_dma_map_single(rdma->cm_id->device,
 400                                     c->rc->sdata, client->msize,
 401                                     DMA_FROM_DEVICE);
 402         if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
 403                 goto error;
 404
 405         sge.addr = c->busa;
 406         sge.length = client->msize;
 407         sge.lkey = rdma->lkey;
 408
 409         wr.next = NULL;
 410         c->wc_op = IB_WC_RECV;
 411         wr.wr_id = (unsigned long) c;
 412         wr.sg_list = &sge;
 413         wr.num_sge = 1;
 414         return ib_post_recv(rdma->qp, &wr, &bad_wr);
 415
 416  error:
 417         p9_debug(P9_DEBUG_ERROR, "EIO\n");
 418         return -EIO;
 419 }
 420
 421 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
 422 {
 423         struct p9_trans_rdma *rdma = client->trans;
 424         struct ib_send_wr wr, *bad_wr;
 425         struct ib_sge sge;
 426         int err = 0;
 427         unsigned long flags;
 428         struct p9_rdma_context *c = NULL;
 429         struct p9_rdma_context *rpl_context = NULL;
 430
 431         /* When an error occurs between posting the recv and the send,
 432          * there will be a receive context posted without a pending request.
 433          * Since there is no way to "un-post" it, we remember it and skip
 434          * post_recv() for the next request.
 435          * So here,
 436          * see if we are this `next request' and need to absorb an excess rc.
 437          * If yes, then drop and free our own, and do not recv_post().
 438          **/
 439         if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
 440                 if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
 441                         /* Got one ! */
 442                         kfree(req->rc);
 443                         req->rc = NULL;
 444                         goto dont_need_post_recv;
 445                 } else {
 446                         /* We raced and lost. */
 447                         atomic_inc(&rdma->excess_rc);
 448                 }
 449         }
 450
 451         /* Allocate an fcall for the reply */
 452         rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
 453         if (!rpl_context) {
 454                 err = -ENOMEM;
 455                 goto recv_error;
 456         }
 457         rpl_context->rc = req->rc;
 458
 459         /*
 460          * Post a receive buffer for this request. We need to ensure
 461          * there is a reply buffer available for every outstanding
 462          * request. A flushed request can result in no reply for an
 463          * outstanding request, so we must keep a count to avoid
 464          * overflowing the RQ.
 465          */
 466         if (down_interruptible(&rdma->rq_sem)) {
 467                 err = -EINTR;
 468                 goto recv_error;
 469         }
 470
 471         err = post_recv(client, rpl_context);
 472         if (err) {
 473                 p9_debug(P9_DEBUG_FCALL, "POST RECV failed\n");
 474                 goto recv_error;
 475         }
 476         /* remove posted receive buffer from request structure */
 477         req->rc = NULL;
 478
 479 dont_need_post_recv:
 480         /* Post the request */
 481         c = kmalloc(sizeof *c, GFP_NOFS);
 482         if (!c) {
 483                 err = -ENOMEM;
 484                 goto send_error;
 485         }
 486         c->req = req;
 487
 488         c->busa = ib_dma_map_single(rdma->cm_id->device,
 489                                     c->req->tc->sdata, c->req->tc->size,
 490                                     DMA_TO_DEVICE);
 491         if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
 492                 err = -EIO;
 493                 goto send_error;
 494         }
 495
 496         sge.addr = c->busa;
 497         sge.length = c->req->tc->size;
 498         sge.lkey = rdma->lkey;
 499
 500         wr.next = NULL;
 501         c->wc_op = IB_WC_SEND;
 502         wr.wr_id = (unsigned long) c;
 503         wr.opcode = IB_WR_SEND;
 504         wr.send_flags = IB_SEND_SIGNALED;
 505         wr.sg_list = &sge;
 506         wr.num_sge = 1;
 507
 508         if (down_interruptible(&rdma->sq_sem)) {
 509                 err = -EINTR;
 510                 goto send_error;
 511         }
 512
 513         err = ib_post_send(rdma->qp, &wr, &bad_wr);
 514         if (err)
 515                 goto send_error;
 516
 517         /* Success */
 518         return 0;
 519
 520  /* Handle errors that happened during or while preparing the send: */
 521  send_error:
 522         kfree(c);
 523         p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
 524
 525         /* Ach.
 526          *  We did recv_post(), but not send. We have one recv_post in excess.
 527          */
 528         atomic_inc(&rdma->excess_rc);
 529         return err;
 530
 531  /* Handle errors that happened during or while preparing post_recv(): */
 532  recv_error:
 533         kfree(rpl_context);
 534         spin_lock_irqsave(&rdma->req_lock, flags);
 535         if (rdma->state < P9_RDMA_CLOSING) {
 536                 rdma->state = P9_RDMA_CLOSING;
 537                 spin_unlock_irqrestore(&rdma->req_lock, flags);
 538                 rdma_disconnect(rdma->cm_id);
 539         } else
 540                 spin_unlock_irqrestore(&rdma->req_lock, flags);
 541         return err;
 542 }
 543
 544 static void rdma_close(struct p9_client *client)
 545 {
 546         struct p9_trans_rdma *rdma;
 547
 548         if (!client)
 549                 return;
 550
 551         rdma = client->trans;
 552         if (!rdma)
 553                 return;
 554
 555         client->status = Disconnected;
 556         rdma_disconnect(rdma->cm_id);
 557         rdma_destroy_trans(rdma);
 558 }
 559
 560 /**
 561  * alloc_rdma - Allocate and initialize the rdma transport structure
 562  * @opts: Mount options structure
 563  */
 564 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
 565 {
 566         struct p9_trans_rdma *rdma;
 567
 568         rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
 569         if (!rdma)
 570                 return NULL;
 571
 572         rdma->sq_depth = opts->sq_depth;
 573         rdma->rq_depth = opts->rq_depth;
 574         rdma->timeout = opts->timeout;
 575         spin_lock_init(&rdma->req_lock);
 576         init_completion(&rdma->cm_done);
 577         sema_init(&rdma->sq_sem, rdma->sq_depth);
 578         sema_init(&rdma->rq_sem, rdma->rq_depth);
 579         atomic_set(&rdma->excess_rc, 0);
 580
 581         return rdma;
 582 }
 583
 584 /* its not clear to me we can do anything after send has been posted */
 585 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
 586 {
 587         return 1;
 588 }
 589
 590 /**
 591  * trans_create_rdma - Transport method for creating atransport instance
 592  * @client: client instance
 593  * @addr: IP address string
 594  * @args: Mount options string
 595  */
 596 static int
 597 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
 598 {
 599         int err;
 600         struct p9_rdma_opts opts;
 601         struct p9_trans_rdma *rdma;
 602         struct rdma_conn_param conn_param;
 603         struct ib_qp_init_attr qp_attr;
 604         struct ib_device_attr devattr;
 605
 606         /* Parse the transport specific mount options */
 607         err = parse_opts(args, &opts);
 608         if (err < 0)
 609                 return err;
 610
 611         /* Create and initialize the RDMA transport structure */
 612         rdma = alloc_rdma(&opts);
 613         if (!rdma)
 614                 return -ENOMEM;
 615
 616         /* Create the RDMA CM ID */
 617         rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP,
 618                                      IB_QPT_RC);
 619         if (IS_ERR(rdma->cm_id))
 620                 goto error;
 621
 622         /* Associate the client with the transport */
 623         client->trans = rdma;
 624
 625         /* Resolve the server's address */
 626         rdma->addr.sin_family = AF_INET;
 627         rdma->addr.sin_addr.s_addr = in_aton(addr);
 628         rdma->addr.sin_port = htons(opts.port);
 629         err = rdma_resolve_addr(rdma->cm_id, NULL,
 630                                 (struct sockaddr *)&rdma->addr,
 631                                 rdma->timeout);
 632         if (err)
 633                 goto error;
 634         err = wait_for_completion_interruptible(&rdma->cm_done);
 635         if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
 636                 goto error;
 637
 638         /* Resolve the route to the server */
 639         err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
 640         if (err)
 641                 goto error;
 642         err = wait_for_completion_interruptible(&rdma->cm_done);
 643         if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
 644                 goto error;
 645
 646         /* Query the device attributes */
 647         err = ib_query_device(rdma->cm_id->device, &devattr);
 648         if (err)
 649                 goto error;
 650
 651         /* Create the Completion Queue */
 652         rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
 653                                 cq_event_handler, client,
 654                                 opts.sq_depth + opts.rq_depth + 1, 0);
 655         if (IS_ERR(rdma->cq))
 656                 goto error;
 657         ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
 658
 659         /* Create the Protection Domain */
 660         rdma->pd = ib_alloc_pd(rdma->cm_id->device);
 661         if (IS_ERR(rdma->pd))
 662                 goto error;
 663
 664         /* Cache the DMA lkey in the transport */
 665         rdma->dma_mr = NULL;
 666         if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
 667                 rdma->lkey = rdma->cm_id->device->local_dma_lkey;
 668         else {
 669                 rdma->dma_mr = ib_get_dma_mr(rdma->pd, IB_ACCESS_LOCAL_WRITE);
 670                 if (IS_ERR(rdma->dma_mr))
 671                         goto error;
 672                 rdma->lkey = rdma->dma_mr->lkey;
 673         }
 674
 675         /* Create the Queue Pair */
 676         memset(&qp_attr, 0, sizeof qp_attr);
 677         qp_attr.event_handler = qp_event_handler;
 678         qp_attr.qp_context = client;
 679         qp_attr.cap.max_send_wr = opts.sq_depth;
 680         qp_attr.cap.max_recv_wr = opts.rq_depth;
 681         qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
 682         qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
 683         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 684         qp_attr.qp_type = IB_QPT_RC;
 685         qp_attr.send_cq = rdma->cq;
 686         qp_attr.recv_cq = rdma->cq;
 687         err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
 688         if (err)
 689                 goto error;
 690         rdma->qp = rdma->cm_id->qp;
 691
 692         /* Request a connection */
 693         memset(&conn_param, 0, sizeof(conn_param));
 694         conn_param.private_data = NULL;
 695         conn_param.private_data_len = 0;
 696         conn_param.responder_resources = P9_RDMA_IRD;
 697         conn_param.initiator_depth = P9_RDMA_ORD;
 698         err = rdma_connect(rdma->cm_id, &conn_param);
 699         if (err)
 700                 goto error;
 701         err = wait_for_completion_interruptible(&rdma->cm_done);
 702         if (err || (rdma->state != P9_RDMA_CONNECTED))
 703                 goto error;
 704
 705         client->status = Connected;
 706
 707         return 0;
 708
 709 error:
 710         rdma_destroy_trans(rdma);
 711         return -ENOTCONN;
 712 }
 713
 714 static struct p9_trans_module p9_rdma_trans = {
 715         .name = "rdma",
 716         .maxsize = P9_RDMA_MAXSIZE,
 717         .def = 0,
 718         .owner = THIS_MODULE,
 719         .create = rdma_create_trans,
 720         .close = rdma_close,
 721         .request = rdma_request,
 722         .cancel = rdma_cancel,
 723 };
 724
 725 /**
 726  * p9_trans_rdma_init - Register the 9P RDMA transport driver
 727  */
 728 static int __init p9_trans_rdma_init(void)
 729 {
 730         v9fs_register_trans(&p9_rdma_trans);
 731         return 0;
 732 }
 733
 734 static void __exit p9_trans_rdma_exit(void)
 735 {
 736         v9fs_unregister_trans(&p9_rdma_trans);
 737 }
 738
 739 module_init(p9_trans_rdma_init);
 740 module_exit(p9_trans_rdma_exit);
 741
 742 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
 743 MODULE_DESCRIPTION("RDMA Transport for 9P");
 744 MODULE_LICENSE("Dual BSD/GPL");