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 #define P9_RDMA_MAX_SGE (P9_RDMA_MAXSIZE >> PAGE_SHIFT)
  63 /**
  64  * struct p9_trans_rdma - RDMA transport instance
  65  *
  66  * @state: tracks the transport state machine for connection setup and tear down
  67  * @cm_id: The RDMA CM ID
  68  * @pd: Protection Domain pointer
  69  * @qp: Queue Pair pointer
  70  * @cq: Completion Queue pointer
  71  * @dm_mr: DMA Memory Region pointer
  72  * @lkey: The local access only memory region key
  73  * @timeout: Number of uSecs to wait for connection management events
  74  * @sq_depth: The depth of the Send Queue
  75  * @sq_sem: Semaphore for the SQ
  76  * @rq_depth: The depth of the Receive Queue.
  77  * @rq_count: Count of requests in the Receive Queue.
  78  * @addr: The remote peer's address
  79  * @req_lock: Protects the active request list
  80  * @cm_done: Completion event for connection management tracking
  81  */
  82 struct p9_trans_rdma {
  83         enum {
  84                 P9_RDMA_INIT,
  85                 P9_RDMA_ADDR_RESOLVED,
  86                 P9_RDMA_ROUTE_RESOLVED,
  87                 P9_RDMA_CONNECTED,
  88                 P9_RDMA_FLUSHING,
  89                 P9_RDMA_CLOSING,
  90                 P9_RDMA_CLOSED,
  91         } state;
  92         struct rdma_cm_id *cm_id;
  93         struct ib_pd *pd;
  94         struct ib_qp *qp;
  95         struct ib_cq *cq;
  96         struct ib_mr *dma_mr;
  97         u32 lkey;
  98         long timeout;
  99         int sq_depth;
 100         struct semaphore sq_sem;
 101         int rq_depth;
 102         atomic_t rq_count;
 103         struct sockaddr_in addr;
 104         spinlock_t req_lock;
 105
 106         struct completion cm_done;
 107 };
 108
 109 /**
 110  * p9_rdma_context - Keeps track of in-process WR
 111  *
 112  * @wc_op: The original WR op for when the CQE completes in error.
 113  * @busa: Bus address to unmap when the WR completes
 114  * @req: Keeps track of requests (send)
 115  * @rc: Keepts track of replies (receive)
 116  */
 117 struct p9_rdma_req;
 118 struct p9_rdma_context {
 119         enum ib_wc_opcode wc_op;
 120         dma_addr_t busa;
 121         union {
 122                 struct p9_req_t *req;
 123                 struct p9_fcall *rc;
 124         };
 125 };
 126
 127 /**
 128  * p9_rdma_opts - Collection of mount options
 129  * @port: port of connection
 130  * @sq_depth: The requested depth of the SQ. This really doesn't need
 131  * to be any deeper than the number of threads used in the client
 132  * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
 133  * @timeout: Time to wait in msecs for CM events
 134  */
 135 struct p9_rdma_opts {
 136         short port;
 137         int sq_depth;
 138         int rq_depth;
 139         long timeout;
 140 };
 141
 142 /*
 143  * Option Parsing (code inspired by NFS code)
 144  */
 145 enum {
 146         /* Options that take integer arguments */
 147         Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout, Opt_err,
 148 };
 149
 150 static match_table_t tokens = {
 151         {Opt_port, "port=%u"},
 152         {Opt_sq_depth, "sq=%u"},
 153         {Opt_rq_depth, "rq=%u"},
 154         {Opt_timeout, "timeout=%u"},
 155         {Opt_err, NULL},
 156 };
 157
 158 /**
 159  * parse_opts - parse mount options into rdma options structure
 160  * @params: options string passed from mount
 161  * @opts: rdma transport-specific structure to parse options into
 162  *
 163  * Returns 0 upon success, -ERRNO upon failure
 164  */
 165 static int parse_opts(char *params, struct p9_rdma_opts *opts)
 166 {
 167         char *p;
 168         substring_t args[MAX_OPT_ARGS];
 169         int option;
 170         char *options, *tmp_options;
 171         int ret;
 172
 173         opts->port = P9_PORT;
 174         opts->sq_depth = P9_RDMA_SQ_DEPTH;
 175         opts->rq_depth = P9_RDMA_RQ_DEPTH;
 176         opts->timeout = P9_RDMA_TIMEOUT;
 177
 178         if (!params)
 179                 return 0;
 180
 181         tmp_options = kstrdup(params, GFP_KERNEL);
 182         if (!tmp_options) {
 183                 P9_DPRINTK(P9_DEBUG_ERROR,
 184                            "failed to allocate copy of option string\n");
 185                 return -ENOMEM;
 186         }
 187         options = tmp_options;
 188
 189         while ((p = strsep(&options, ",")) != NULL) {
 190                 int token;
 191                 int r;
 192                 if (!*p)
 193                         continue;
 194                 token = match_token(p, tokens, args);
 195                 r = match_int(&args[0], &option);
 196                 if (r < 0) {
 197                         P9_DPRINTK(P9_DEBUG_ERROR,
 198                                    "integer field, but no integer?\n");
 199                         ret = r;
 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         req->rc = c->rc;
 301         req->status = REQ_STATUS_RCVD;
 302         p9_client_cb(client, req);
 303
 304         return;
 305
 306  err_out:
 307         P9_DPRINTK(P9_DEBUG_ERROR, "req %p err %d status %d\n",
 308                    req, err, status);
 309         rdma->state = P9_RDMA_FLUSHING;
 310         client->status = Disconnected;
 311 }
 312
 313 static void
 314 handle_send(struct p9_client *client, struct p9_trans_rdma *rdma,
 315             struct p9_rdma_context *c, enum ib_wc_status status, u32 byte_len)
 316 {
 317         ib_dma_unmap_single(rdma->cm_id->device,
 318                             c->busa, c->req->tc->size,
 319                             DMA_TO_DEVICE);
 320 }
 321
 322 static void qp_event_handler(struct ib_event *event, void *context)
 323 {
 324         P9_DPRINTK(P9_DEBUG_ERROR, "QP event %d context %p\n", event->event,
 325                                                                 context);
 326 }
 327
 328 static void cq_comp_handler(struct ib_cq *cq, void *cq_context)
 329 {
 330         struct p9_client *client = cq_context;
 331         struct p9_trans_rdma *rdma = client->trans;
 332         int ret;
 333         struct ib_wc wc;
 334
 335         ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
 336         while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
 337                 struct p9_rdma_context *c = (void *) (unsigned long) wc.wr_id;
 338
 339                 switch (c->wc_op) {
 340                 case IB_WC_RECV:
 341                         atomic_dec(&rdma->rq_count);
 342                         handle_recv(client, rdma, c, wc.status, wc.byte_len);
 343                         break;
 344
 345                 case IB_WC_SEND:
 346                         handle_send(client, rdma, c, wc.status, wc.byte_len);
 347                         up(&rdma->sq_sem);
 348                         break;
 349
 350                 default:
 351                         printk(KERN_ERR "9prdma: unexpected completion type, "
 352                                "c->wc_op=%d, wc.opcode=%d, status=%d\n",
 353                                c->wc_op, wc.opcode, wc.status);
 354                         break;
 355                 }
 356                 kfree(c);
 357         }
 358 }
 359
 360 static void cq_event_handler(struct ib_event *e, void *v)
 361 {
 362         P9_DPRINTK(P9_DEBUG_ERROR, "CQ event %d context %p\n", e->event, v);
 363 }
 364
 365 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
 366 {
 367         if (!rdma)
 368                 return;
 369
 370         if (rdma->dma_mr && !IS_ERR(rdma->dma_mr))
 371                 ib_dereg_mr(rdma->dma_mr);
 372
 373         if (rdma->qp && !IS_ERR(rdma->qp))
 374                 ib_destroy_qp(rdma->qp);
 375
 376         if (rdma->pd && !IS_ERR(rdma->pd))
 377                 ib_dealloc_pd(rdma->pd);
 378
 379         if (rdma->cq && !IS_ERR(rdma->cq))
 380                 ib_destroy_cq(rdma->cq);
 381
 382         if (rdma->cm_id && !IS_ERR(rdma->cm_id))
 383                 rdma_destroy_id(rdma->cm_id);
 384
 385         kfree(rdma);
 386 }
 387
 388 static int
 389 post_recv(struct p9_client *client, struct p9_rdma_context *c)
 390 {
 391         struct p9_trans_rdma *rdma = client->trans;
 392         struct ib_recv_wr wr, *bad_wr;
 393         struct ib_sge sge;
 394
 395         c->busa = ib_dma_map_single(rdma->cm_id->device,
 396                                     c->rc->sdata, client->msize,
 397                                     DMA_FROM_DEVICE);
 398         if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
 399                 goto error;
 400
 401         sge.addr = c->busa;
 402         sge.length = client->msize;
 403         sge.lkey = rdma->lkey;
 404
 405         wr.next = NULL;
 406         c->wc_op = IB_WC_RECV;
 407         wr.wr_id = (unsigned long) c;
 408         wr.sg_list = &sge;
 409         wr.num_sge = 1;
 410         return ib_post_recv(rdma->qp, &wr, &bad_wr);
 411
 412  error:
 413         P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
 414         return -EIO;
 415 }
 416
 417 static int rdma_request(struct p9_client *client, struct p9_req_t *req)
 418 {
 419         struct p9_trans_rdma *rdma = client->trans;
 420         struct ib_send_wr wr, *bad_wr;
 421         struct ib_sge sge;
 422         int err = 0;
 423         unsigned long flags;
 424         struct p9_rdma_context *c = NULL;
 425         struct p9_rdma_context *rpl_context = NULL;
 426
 427         /* Allocate an fcall for the reply */
 428         rpl_context = kmalloc(sizeof *rpl_context, GFP_KERNEL);
 429         if (!rpl_context) {
 430                 err = -ENOMEM;
 431                 goto err_close;
 432         }
 433
 434         /*
 435          * If the request has a buffer, steal it, otherwise
 436          * allocate a new one.  Typically, requests should already
 437          * have receive buffers allocated and just swap them around
 438          */
 439         if (!req->rc) {
 440                 req->rc = kmalloc(sizeof(struct p9_fcall)+client->msize,
 441                                                                 GFP_KERNEL);
 442                 if (req->rc) {
 443                         req->rc->sdata = (char *) req->rc +
 444                                                 sizeof(struct p9_fcall);
 445                         req->rc->capacity = client->msize;
 446                 }
 447         }
 448         rpl_context->rc = req->rc;
 449         if (!rpl_context->rc) {
 450                 err = -ENOMEM;
 451                 goto err_free2;
 452         }
 453
 454         /*
 455          * Post a receive buffer for this request. We need to ensure
 456          * there is a reply buffer available for every outstanding
 457          * request. A flushed request can result in no reply for an
 458          * outstanding request, so we must keep a count to avoid
 459          * overflowing the RQ.
 460          */
 461         if (atomic_inc_return(&rdma->rq_count) <= rdma->rq_depth) {
 462                 err = post_recv(client, rpl_context);
 463                 if (err)
 464                         goto err_free1;
 465         } else
 466                 atomic_dec(&rdma->rq_count);
 467
 468         /* remove posted receive buffer from request structure */
 469         req->rc = NULL;
 470
 471         /* Post the request */
 472         c = kmalloc(sizeof *c, GFP_KERNEL);
 473         if (!c) {
 474                 err = -ENOMEM;
 475                 goto err_free1;
 476         }
 477         c->req = req;
 478
 479         c->busa = ib_dma_map_single(rdma->cm_id->device,
 480                                     c->req->tc->sdata, c->req->tc->size,
 481                                     DMA_TO_DEVICE);
 482         if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
 483                 goto error;
 484
 485         sge.addr = c->busa;
 486         sge.length = c->req->tc->size;
 487         sge.lkey = rdma->lkey;
 488
 489         wr.next = NULL;
 490         c->wc_op = IB_WC_SEND;
 491         wr.wr_id = (unsigned long) c;
 492         wr.opcode = IB_WR_SEND;
 493         wr.send_flags = IB_SEND_SIGNALED;
 494         wr.sg_list = &sge;
 495         wr.num_sge = 1;
 496
 497         if (down_interruptible(&rdma->sq_sem))
 498                 goto error;
 499
 500         return ib_post_send(rdma->qp, &wr, &bad_wr);
 501
 502  error:
 503         kfree(c);
 504         kfree(rpl_context->rc);
 505         kfree(rpl_context);
 506         P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
 507         return -EIO;
 508  err_free1:
 509         kfree(rpl_context->rc);
 510  err_free2:
 511         kfree(rpl_context);
 512  err_close:
 513         spin_lock_irqsave(&rdma->req_lock, flags);
 514         if (rdma->state < P9_RDMA_CLOSING) {
 515                 rdma->state = P9_RDMA_CLOSING;
 516                 spin_unlock_irqrestore(&rdma->req_lock, flags);
 517                 rdma_disconnect(rdma->cm_id);
 518         } else
 519                 spin_unlock_irqrestore(&rdma->req_lock, flags);
 520         return err;
 521 }
 522
 523 static void rdma_close(struct p9_client *client)
 524 {
 525         struct p9_trans_rdma *rdma;
 526
 527         if (!client)
 528                 return;
 529
 530         rdma = client->trans;
 531         if (!rdma)
 532                 return;
 533
 534         client->status = Disconnected;
 535         rdma_disconnect(rdma->cm_id);
 536         rdma_destroy_trans(rdma);
 537 }
 538
 539 /**
 540  * alloc_rdma - Allocate and initialize the rdma transport structure
 541  * @opts: Mount options structure
 542  */
 543 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
 544 {
 545         struct p9_trans_rdma *rdma;
 546
 547         rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
 548         if (!rdma)
 549                 return NULL;
 550
 551         rdma->sq_depth = opts->sq_depth;
 552         rdma->rq_depth = opts->rq_depth;
 553         rdma->timeout = opts->timeout;
 554         spin_lock_init(&rdma->req_lock);
 555         init_completion(&rdma->cm_done);
 556         sema_init(&rdma->sq_sem, rdma->sq_depth);
 557         atomic_set(&rdma->rq_count, 0);
 558
 559         return rdma;
 560 }
 561
 562 /* its not clear to me we can do anything after send has been posted */
 563 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
 564 {
 565         return 1;
 566 }
 567
 568 /**
 569  * trans_create_rdma - Transport method for creating atransport instance
 570  * @client: client instance
 571  * @addr: IP address string
 572  * @args: Mount options string
 573  */
 574 static int
 575 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
 576 {
 577         int err;
 578         struct p9_rdma_opts opts;
 579         struct p9_trans_rdma *rdma;
 580         struct rdma_conn_param conn_param;
 581         struct ib_qp_init_attr qp_attr;
 582         struct ib_device_attr devattr;
 583
 584         /* Parse the transport specific mount options */
 585         err = parse_opts(args, &opts);
 586         if (err < 0)
 587                 return err;
 588
 589         /* Create and initialize the RDMA transport structure */
 590         rdma = alloc_rdma(&opts);
 591         if (!rdma)
 592                 return -ENOMEM;
 593
 594         /* Create the RDMA CM ID */
 595         rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP);
 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");