trans_rdma.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * RDMA transport layer based on the trans_fd.c implementation.
   4  *
   5  *  Copyright (C) 2008 by Tom Tucker <tom@opengridcomputing.com>
   6  *  Copyright (C) 2006 by Russ Cox <rsc@swtch.com>
   7  *  Copyright (C) 2004-2005 by Latchesar Ionkov <lucho@ionkov.net>
   8  *  Copyright (C) 2004-2008 by Eric Van Hensbergen <ericvh@gmail.com>
   9  *  Copyright (C) 1997-2002 by Ron Minnich <rminnich@sarnoff.com>
  10  */
  11
  12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  13
  14 #include <linux/in.h>
  15 #include <linux/module.h>
  16 #include <linux/net.h>
  17 #include <linux/ipv6.h>
  18 #include <linux/kthread.h>
  19 #include <linux/errno.h>
  20 #include <linux/kernel.h>
  21 #include <linux/un.h>
  22 #include <linux/uaccess.h>
  23 #include <linux/inet.h>
  24 #include <linux/idr.h>
  25 #include <linux/file.h>
  26 #include <linux/parser.h>
  27 #include <linux/semaphore.h>
  28 #include <linux/slab.h>
  29 #include <linux/seq_file.h>
  30 #include <net/9p/9p.h>
  31 #include <net/9p/client.h>
  32 #include <net/9p/transport.h>
  33 #include <rdma/ib_verbs.h>
  34 #include <rdma/rdma_cm.h>
  35
  36 #define P9_PORT                 5640
  37 #define P9_RDMA_SQ_DEPTH        32
  38 #define P9_RDMA_RQ_DEPTH        32
  39 #define P9_RDMA_SEND_SGE        4
  40 #define P9_RDMA_RECV_SGE        4
  41 #define P9_RDMA_IRD             0
  42 #define P9_RDMA_ORD             0
  43 #define P9_RDMA_TIMEOUT         30000           /* 30 seconds */
  44 #define P9_RDMA_MAXSIZE         (1024*1024)     /* 1MB */
  45
  46 /**
  47  * struct p9_trans_rdma - RDMA transport instance
  48  *
  49  * @state: tracks the transport state machine for connection setup and tear down
  50  * @cm_id: The RDMA CM ID
  51  * @pd: Protection Domain pointer
  52  * @qp: Queue Pair pointer
  53  * @cq: Completion Queue pointer
  54  * @timeout: Number of uSecs to wait for connection management events
  55  * @privport: Whether a privileged port may be used
  56  * @port: The port to use
  57  * @sq_depth: The depth of the Send Queue
  58  * @sq_sem: Semaphore for the SQ
  59  * @rq_depth: The depth of the Receive Queue.
  60  * @rq_sem: Semaphore for the RQ
  61  * @excess_rc : Amount of posted Receive Contexts without a pending request.
  62  *              See rdma_request()
  63  * @addr: The remote peer's address
  64  * @req_lock: Protects the active request list
  65  * @cm_done: Completion event for connection management tracking
  66  */
  67 struct p9_trans_rdma {
  68         enum {
  69                 P9_RDMA_INIT,
  70                 P9_RDMA_ADDR_RESOLVED,
  71                 P9_RDMA_ROUTE_RESOLVED,
  72                 P9_RDMA_CONNECTED,
  73                 P9_RDMA_FLUSHING,
  74                 P9_RDMA_CLOSING,
  75                 P9_RDMA_CLOSED,
  76         } state;
  77         struct rdma_cm_id *cm_id;
  78         struct ib_pd *pd;
  79         struct ib_qp *qp;
  80         struct ib_cq *cq;
  81         long timeout;
  82         bool privport;
  83         u16 port;
  84         int sq_depth;
  85         struct semaphore sq_sem;
  86         int rq_depth;
  87         struct semaphore rq_sem;
  88         atomic_t excess_rc;
  89         struct sockaddr_in addr;
  90         spinlock_t req_lock;
  91
  92         struct completion cm_done;
  93 };
  94
  95 struct p9_rdma_req;
  96
  97 /**
  98  * struct p9_rdma_context - Keeps track of in-process WR
  99  *
 100  * @cqe: completion queue entry
 101  * @busa: Bus address to unmap when the WR completes
 102  * @req: Keeps track of requests (send)
 103  * @rc: Keepts track of replies (receive)
 104  */
 105 struct p9_rdma_context {
 106         struct ib_cqe cqe;
 107         dma_addr_t busa;
 108         union {
 109                 struct p9_req_t *req;
 110                 struct p9_fcall rc;
 111         };
 112 };
 113
 114 /**
 115  * struct p9_rdma_opts - Collection of mount options
 116  * @port: port of connection
 117  * @privport: Whether a privileged port may be used
 118  * @sq_depth: The requested depth of the SQ. This really doesn't need
 119  * to be any deeper than the number of threads used in the client
 120  * @rq_depth: The depth of the RQ. Should be greater than or equal to SQ depth
 121  * @timeout: Time to wait in msecs for CM events
 122  */
 123 struct p9_rdma_opts {
 124         short port;
 125         bool privport;
 126         int sq_depth;
 127         int rq_depth;
 128         long timeout;
 129 };
 130
 131 /*
 132  * Option Parsing (code inspired by NFS code)
 133  */
 134 enum {
 135         /* Options that take integer arguments */
 136         Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout,
 137         /* Options that take no argument */
 138         Opt_privport,
 139         Opt_err,
 140 };
 141
 142 static match_table_t tokens = {
 143         {Opt_port, "port=%u"},
 144         {Opt_sq_depth, "sq=%u"},
 145         {Opt_rq_depth, "rq=%u"},
 146         {Opt_timeout, "timeout=%u"},
 147         {Opt_privport, "privport"},
 148         {Opt_err, NULL},
 149 };
 150
 151 static int p9_rdma_show_options(struct seq_file *m, struct p9_client *clnt)
 152 {
 153         struct p9_trans_rdma *rdma = clnt->trans;
 154
 155         if (rdma->port != P9_PORT)
 156                 seq_printf(m, ",port=%u", rdma->port);
 157         if (rdma->sq_depth != P9_RDMA_SQ_DEPTH)
 158                 seq_printf(m, ",sq=%u", rdma->sq_depth);
 159         if (rdma->rq_depth != P9_RDMA_RQ_DEPTH)
 160                 seq_printf(m, ",rq=%u", rdma->rq_depth);
 161         if (rdma->timeout != P9_RDMA_TIMEOUT)
 162                 seq_printf(m, ",timeout=%lu", rdma->timeout);
 163         if (rdma->privport)
 164                 seq_puts(m, ",privport");
 165         return 0;
 166 }
 167
 168 /**
 169  * parse_opts - parse mount options into rdma options structure
 170  * @params: options string passed from mount
 171  * @opts: rdma transport-specific structure to parse options into
 172  *
 173  * Returns 0 upon success, -ERRNO upon failure
 174  */
 175 static int parse_opts(char *params, struct p9_rdma_opts *opts)
 176 {
 177         char *p;
 178         substring_t args[MAX_OPT_ARGS];
 179         int option;
 180         char *options, *tmp_options;
 181
 182         opts->port = P9_PORT;
 183         opts->sq_depth = P9_RDMA_SQ_DEPTH;
 184         opts->rq_depth = P9_RDMA_RQ_DEPTH;
 185         opts->timeout = P9_RDMA_TIMEOUT;
 186         opts->privport = false;
 187
 188         if (!params)
 189                 return 0;
 190
 191         tmp_options = kstrdup(params, GFP_KERNEL);
 192         if (!tmp_options) {
 193                 p9_debug(P9_DEBUG_ERROR,
 194                          "failed to allocate copy of option string\n");
 195                 return -ENOMEM;
 196         }
 197         options = tmp_options;
 198
 199         while ((p = strsep(&options, ",")) != NULL) {
 200                 int token;
 201                 int r;
 202                 if (!*p)
 203                         continue;
 204                 token = match_token(p, tokens, args);
 205                 if ((token != Opt_err) && (token != Opt_privport)) {
 206                         r = match_int(&args[0], &option);
 207                         if (r < 0) {
 208                                 p9_debug(P9_DEBUG_ERROR,
 209                                          "integer field, but no integer?\n");
 210                                 continue;
 211                         }
 212                 }
 213                 switch (token) {
 214                 case Opt_port:
 215                         opts->port = option;
 216                         break;
 217                 case Opt_sq_depth:
 218                         opts->sq_depth = option;
 219                         break;
 220                 case Opt_rq_depth:
 221                         opts->rq_depth = option;
 222                         break;
 223                 case Opt_timeout:
 224                         opts->timeout = option;
 225                         break;
 226                 case Opt_privport:
 227                         opts->privport = true;
 228                         break;
 229                 default:
 230                         continue;
 231                 }
 232         }
 233         /* RQ must be at least as large as the SQ */
 234         opts->rq_depth = max(opts->rq_depth, opts->sq_depth);
 235         kfree(tmp_options);
 236         return 0;
 237 }
 238
 239 static int
 240 p9_cm_event_handler(struct rdma_cm_id *id, struct rdma_cm_event *event)
 241 {
 242         struct p9_client *c = id->context;
 243         struct p9_trans_rdma *rdma = c->trans;
 244         switch (event->event) {
 245         case RDMA_CM_EVENT_ADDR_RESOLVED:
 246                 BUG_ON(rdma->state != P9_RDMA_INIT);
 247                 rdma->state = P9_RDMA_ADDR_RESOLVED;
 248                 break;
 249
 250         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 251                 BUG_ON(rdma->state != P9_RDMA_ADDR_RESOLVED);
 252                 rdma->state = P9_RDMA_ROUTE_RESOLVED;
 253                 break;
 254
 255         case RDMA_CM_EVENT_ESTABLISHED:
 256                 BUG_ON(rdma->state != P9_RDMA_ROUTE_RESOLVED);
 257                 rdma->state = P9_RDMA_CONNECTED;
 258                 break;
 259
 260         case RDMA_CM_EVENT_DISCONNECTED:
 261                 if (rdma)
 262                         rdma->state = P9_RDMA_CLOSED;
 263                 c->status = Disconnected;
 264                 break;
 265
 266         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
 267                 break;
 268
 269         case RDMA_CM_EVENT_ADDR_CHANGE:
 270         case RDMA_CM_EVENT_ROUTE_ERROR:
 271         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 272         case RDMA_CM_EVENT_MULTICAST_JOIN:
 273         case RDMA_CM_EVENT_MULTICAST_ERROR:
 274         case RDMA_CM_EVENT_REJECTED:
 275         case RDMA_CM_EVENT_CONNECT_REQUEST:
 276         case RDMA_CM_EVENT_CONNECT_RESPONSE:
 277         case RDMA_CM_EVENT_CONNECT_ERROR:
 278         case RDMA_CM_EVENT_ADDR_ERROR:
 279         case RDMA_CM_EVENT_UNREACHABLE:
 280                 c->status = Disconnected;
 281                 rdma_disconnect(rdma->cm_id);
 282                 break;
 283         default:
 284                 BUG();
 285         }
 286         complete(&rdma->cm_done);
 287         return 0;
 288 }
 289
 290 static void
 291 recv_done(struct ib_cq *cq, struct ib_wc *wc)
 292 {
 293         struct p9_client *client = cq->cq_context;
 294         struct p9_trans_rdma *rdma = client->trans;
 295         struct p9_rdma_context *c =
 296                 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
 297         struct p9_req_t *req;
 298         int err = 0;
 299         int16_t tag;
 300
 301         req = NULL;
 302         ib_dma_unmap_single(rdma->cm_id->device, c->busa, client->msize,
 303                                                          DMA_FROM_DEVICE);
 304
 305         if (wc->status != IB_WC_SUCCESS)
 306                 goto err_out;
 307
 308         c->rc.size = wc->byte_len;
 309         err = p9_parse_header(&c->rc, NULL, NULL, &tag, 1);
 310         if (err)
 311                 goto err_out;
 312
 313         req = p9_tag_lookup(client, tag);
 314         if (!req)
 315                 goto err_out;
 316
 317         /* Check that we have not yet received a reply for this request.
 318          */
 319         if (unlikely(req->rc.sdata)) {
 320                 pr_err("Duplicate reply for request %d", tag);
 321                 goto err_out;
 322         }
 323
 324         req->rc.size = c->rc.size;
 325         req->rc.sdata = c->rc.sdata;
 326         p9_client_cb(client, req, REQ_STATUS_RCVD);
 327
 328  out:
 329         up(&rdma->rq_sem);
 330         kfree(c);
 331         return;
 332
 333  err_out:
 334         p9_debug(P9_DEBUG_ERROR, "req %p err %d status %d\n",
 335                         req, err, wc->status);
 336         rdma->state = P9_RDMA_FLUSHING;
 337         client->status = Disconnected;
 338         goto out;
 339 }
 340
 341 static void
 342 send_done(struct ib_cq *cq, struct ib_wc *wc)
 343 {
 344         struct p9_client *client = cq->cq_context;
 345         struct p9_trans_rdma *rdma = client->trans;
 346         struct p9_rdma_context *c =
 347                 container_of(wc->wr_cqe, struct p9_rdma_context, cqe);
 348
 349         ib_dma_unmap_single(rdma->cm_id->device,
 350                             c->busa, c->req->tc.size,
 351                             DMA_TO_DEVICE);
 352         up(&rdma->sq_sem);
 353         p9_req_put(client, c->req);
 354         kfree(c);
 355 }
 356
 357 static void qp_event_handler(struct ib_event *event, void *context)
 358 {
 359         p9_debug(P9_DEBUG_ERROR, "QP event %d context %p\n",
 360                  event->event, context);
 361 }
 362
 363 static void rdma_destroy_trans(struct p9_trans_rdma *rdma)
 364 {
 365         if (!rdma)
 366                 return;
 367
 368         if (rdma->qp && !IS_ERR(rdma->qp))
 369                 ib_destroy_qp(rdma->qp);
 370
 371         if (rdma->pd && !IS_ERR(rdma->pd))
 372                 ib_dealloc_pd(rdma->pd);
 373
 374         if (rdma->cq && !IS_ERR(rdma->cq))
 375                 ib_free_cq(rdma->cq);
 376
 377         if (rdma->cm_id && !IS_ERR(rdma->cm_id))
 378                 rdma_destroy_id(rdma->cm_id);
 379
 380         kfree(rdma);
 381 }
 382
 383 static int
 384 post_recv(struct p9_client *client, struct p9_rdma_context *c)
 385 {
 386         struct p9_trans_rdma *rdma = client->trans;
 387         struct ib_recv_wr wr;
 388         struct ib_sge sge;
 389         int ret;
 390
 391         c->busa = ib_dma_map_single(rdma->cm_id->device,
 392                                     c->rc.sdata, client->msize,
 393                                     DMA_FROM_DEVICE);
 394         if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
 395                 goto error;
 396
 397         c->cqe.done = recv_done;
 398
 399         sge.addr = c->busa;
 400         sge.length = client->msize;
 401         sge.lkey = rdma->pd->local_dma_lkey;
 402
 403         wr.next = NULL;
 404         wr.wr_cqe = &c->cqe;
 405         wr.sg_list = &sge;
 406         wr.num_sge = 1;
 407
 408         ret = ib_post_recv(rdma->qp, &wr, NULL);
 409         if (ret)
 410                 ib_dma_unmap_single(rdma->cm_id->device, c->busa,
 411                                     client->msize, DMA_FROM_DEVICE);
 412         return ret;
 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;
 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         /* When an error occurs between posting the recv and the send,
 430          * there will be a receive context posted without a pending request.
 431          * Since there is no way to "un-post" it, we remember it and skip
 432          * post_recv() for the next request.
 433          * So here,
 434          * see if we are this `next request' and need to absorb an excess rc.
 435          * If yes, then drop and free our own, and do not recv_post().
 436          **/
 437         if (unlikely(atomic_read(&rdma->excess_rc) > 0)) {
 438                 if ((atomic_sub_return(1, &rdma->excess_rc) >= 0)) {
 439                         /* Got one! */
 440                         p9_fcall_fini(&req->rc);
 441                         req->rc.sdata = NULL;
 442                         goto dont_need_post_recv;
 443                 } else {
 444                         /* We raced and lost. */
 445                         atomic_inc(&rdma->excess_rc);
 446                 }
 447         }
 448
 449         /* Allocate an fcall for the reply */
 450         rpl_context = kmalloc(sizeof *rpl_context, GFP_NOFS);
 451         if (!rpl_context) {
 452                 err = -ENOMEM;
 453                 goto recv_error;
 454         }
 455         rpl_context->rc.sdata = req->rc.sdata;
 456
 457         /*
 458          * Post a receive buffer for this request. We need to ensure
 459          * there is a reply buffer available for every outstanding
 460          * request. A flushed request can result in no reply for an
 461          * outstanding request, so we must keep a count to avoid
 462          * overflowing the RQ.
 463          */
 464         if (down_interruptible(&rdma->rq_sem)) {
 465                 err = -EINTR;
 466                 goto recv_error;
 467         }
 468
 469         err = post_recv(client, rpl_context);
 470         if (err) {
 471                 p9_debug(P9_DEBUG_ERROR, "POST RECV failed: %d\n", err);
 472                 goto recv_error;
 473         }
 474         /* remove posted receive buffer from request structure */
 475         req->rc.sdata = NULL;
 476
 477 dont_need_post_recv:
 478         /* Post the request */
 479         c = kmalloc(sizeof *c, GFP_NOFS);
 480         if (!c) {
 481                 err = -ENOMEM;
 482                 goto send_error;
 483         }
 484         c->req = req;
 485
 486         c->busa = ib_dma_map_single(rdma->cm_id->device,
 487                                     c->req->tc.sdata, c->req->tc.size,
 488                                     DMA_TO_DEVICE);
 489         if (ib_dma_mapping_error(rdma->cm_id->device, c->busa)) {
 490                 err = -EIO;
 491                 goto send_error;
 492         }
 493
 494         c->cqe.done = send_done;
 495
 496         sge.addr = c->busa;
 497         sge.length = c->req->tc.size;
 498         sge.lkey = rdma->pd->local_dma_lkey;
 499
 500         wr.next = NULL;
 501         wr.wr_cqe = &c->cqe;
 502         wr.opcode = IB_WR_SEND;
 503         wr.send_flags = IB_SEND_SIGNALED;
 504         wr.sg_list = &sge;
 505         wr.num_sge = 1;
 506
 507         if (down_interruptible(&rdma->sq_sem)) {
 508                 err = -EINTR;
 509                 goto dma_unmap;
 510         }
 511
 512         /* Mark request as `sent' *before* we actually send it,
 513          * because doing if after could erase the REQ_STATUS_RCVD
 514          * status in case of a very fast reply.
 515          */
 516         WRITE_ONCE(req->status, REQ_STATUS_SENT);
 517         err = ib_post_send(rdma->qp, &wr, NULL);
 518         if (err)
 519                 goto dma_unmap;
 520
 521         /* Success */
 522         return 0;
 523
 524 dma_unmap:
 525         ib_dma_unmap_single(rdma->cm_id->device, c->busa,
 526                             c->req->tc.size, DMA_TO_DEVICE);
 527  /* Handle errors that happened during or while preparing the send: */
 528  send_error:
 529         WRITE_ONCE(req->status, REQ_STATUS_ERROR);
 530         kfree(c);
 531         p9_debug(P9_DEBUG_ERROR, "Error %d in rdma_request()\n", err);
 532
 533         /* Ach.
 534          *  We did recv_post(), but not send. We have one recv_post in excess.
 535          */
 536         atomic_inc(&rdma->excess_rc);
 537         return err;
 538
 539  /* Handle errors that happened during or while preparing post_recv(): */
 540  recv_error:
 541         kfree(rpl_context);
 542         spin_lock_irqsave(&rdma->req_lock, flags);
 543         if (err != -EINTR && rdma->state < P9_RDMA_CLOSING) {
 544                 rdma->state = P9_RDMA_CLOSING;
 545                 spin_unlock_irqrestore(&rdma->req_lock, flags);
 546                 rdma_disconnect(rdma->cm_id);
 547         } else
 548                 spin_unlock_irqrestore(&rdma->req_lock, flags);
 549         return err;
 550 }
 551
 552 static void rdma_close(struct p9_client *client)
 553 {
 554         struct p9_trans_rdma *rdma;
 555
 556         if (!client)
 557                 return;
 558
 559         rdma = client->trans;
 560         if (!rdma)
 561                 return;
 562
 563         client->status = Disconnected;
 564         rdma_disconnect(rdma->cm_id);
 565         rdma_destroy_trans(rdma);
 566 }
 567
 568 /**
 569  * alloc_rdma - Allocate and initialize the rdma transport structure
 570  * @opts: Mount options structure
 571  */
 572 static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
 573 {
 574         struct p9_trans_rdma *rdma;
 575
 576         rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
 577         if (!rdma)
 578                 return NULL;
 579
 580         rdma->port = opts->port;
 581         rdma->privport = opts->privport;
 582         rdma->sq_depth = opts->sq_depth;
 583         rdma->rq_depth = opts->rq_depth;
 584         rdma->timeout = opts->timeout;
 585         spin_lock_init(&rdma->req_lock);
 586         init_completion(&rdma->cm_done);
 587         sema_init(&rdma->sq_sem, rdma->sq_depth);
 588         sema_init(&rdma->rq_sem, rdma->rq_depth);
 589         atomic_set(&rdma->excess_rc, 0);
 590
 591         return rdma;
 592 }
 593
 594 static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
 595 {
 596         /* Nothing to do here.
 597          * We will take care of it (if we have to) in rdma_cancelled()
 598          */
 599         return 1;
 600 }
 601
 602 /* A request has been fully flushed without a reply.
 603  * That means we have posted one buffer in excess.
 604  */
 605 static int rdma_cancelled(struct p9_client *client, struct p9_req_t *req)
 606 {
 607         struct p9_trans_rdma *rdma = client->trans;
 608         atomic_inc(&rdma->excess_rc);
 609         return 0;
 610 }
 611
 612 static int p9_rdma_bind_privport(struct p9_trans_rdma *rdma)
 613 {
 614         struct sockaddr_in cl = {
 615                 .sin_family = AF_INET,
 616                 .sin_addr.s_addr = htonl(INADDR_ANY),
 617         };
 618         int port, err = -EINVAL;
 619
 620         for (port = P9_DEF_MAX_RESVPORT; port >= P9_DEF_MIN_RESVPORT; port--) {
 621                 cl.sin_port = htons((ushort)port);
 622                 err = rdma_bind_addr(rdma->cm_id, (struct sockaddr *)&cl);
 623                 if (err != -EADDRINUSE)
 624                         break;
 625         }
 626         return err;
 627 }
 628
 629 /**
 630  * rdma_create_trans - Transport method for creating a transport instance
 631  * @client: client instance
 632  * @addr: IP address string
 633  * @args: Mount options string
 634  */
 635 static int
 636 rdma_create_trans(struct p9_client *client, const char *addr, char *args)
 637 {
 638         int err;
 639         struct p9_rdma_opts opts;
 640         struct p9_trans_rdma *rdma;
 641         struct rdma_conn_param conn_param;
 642         struct ib_qp_init_attr qp_attr;
 643
 644         if (addr == NULL)
 645                 return -EINVAL;
 646
 647         /* Parse the transport specific mount options */
 648         err = parse_opts(args, &opts);
 649         if (err < 0)
 650                 return err;
 651
 652         /* Create and initialize the RDMA transport structure */
 653         rdma = alloc_rdma(&opts);
 654         if (!rdma)
 655                 return -ENOMEM;
 656
 657         /* Create the RDMA CM ID */
 658         rdma->cm_id = rdma_create_id(&init_net, p9_cm_event_handler, client,
 659                                      RDMA_PS_TCP, IB_QPT_RC);
 660         if (IS_ERR(rdma->cm_id))
 661                 goto error;
 662
 663         /* Associate the client with the transport */
 664         client->trans = rdma;
 665
 666         /* Bind to a privileged port if we need to */
 667         if (opts.privport) {
 668                 err = p9_rdma_bind_privport(rdma);
 669                 if (err < 0) {
 670                         pr_err("%s (%d): problem binding to privport: %d\n",
 671                                __func__, task_pid_nr(current), -err);
 672                         goto error;
 673                 }
 674         }
 675
 676         /* Resolve the server's address */
 677         rdma->addr.sin_family = AF_INET;
 678         rdma->addr.sin_addr.s_addr = in_aton(addr);
 679         rdma->addr.sin_port = htons(opts.port);
 680         err = rdma_resolve_addr(rdma->cm_id, NULL,
 681                                 (struct sockaddr *)&rdma->addr,
 682                                 rdma->timeout);
 683         if (err)
 684                 goto error;
 685         err = wait_for_completion_interruptible(&rdma->cm_done);
 686         if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
 687                 goto error;
 688
 689         /* Resolve the route to the server */
 690         err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
 691         if (err)
 692                 goto error;
 693         err = wait_for_completion_interruptible(&rdma->cm_done);
 694         if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
 695                 goto error;
 696
 697         /* Create the Completion Queue */
 698         rdma->cq = ib_alloc_cq_any(rdma->cm_id->device, client,
 699                                    opts.sq_depth + opts.rq_depth + 1,
 700                                    IB_POLL_SOFTIRQ);
 701         if (IS_ERR(rdma->cq))
 702                 goto error;
 703
 704         /* Create the Protection Domain */
 705         rdma->pd = ib_alloc_pd(rdma->cm_id->device, 0);
 706         if (IS_ERR(rdma->pd))
 707                 goto error;
 708
 709         /* Create the Queue Pair */
 710         memset(&qp_attr, 0, sizeof qp_attr);
 711         qp_attr.event_handler = qp_event_handler;
 712         qp_attr.qp_context = client;
 713         qp_attr.cap.max_send_wr = opts.sq_depth;
 714         qp_attr.cap.max_recv_wr = opts.rq_depth;
 715         qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
 716         qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
 717         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 718         qp_attr.qp_type = IB_QPT_RC;
 719         qp_attr.send_cq = rdma->cq;
 720         qp_attr.recv_cq = rdma->cq;
 721         err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
 722         if (err)
 723                 goto error;
 724         rdma->qp = rdma->cm_id->qp;
 725
 726         /* Request a connection */
 727         memset(&conn_param, 0, sizeof(conn_param));
 728         conn_param.private_data = NULL;
 729         conn_param.private_data_len = 0;
 730         conn_param.responder_resources = P9_RDMA_IRD;
 731         conn_param.initiator_depth = P9_RDMA_ORD;
 732         err = rdma_connect(rdma->cm_id, &conn_param);
 733         if (err)
 734                 goto error;
 735         err = wait_for_completion_interruptible(&rdma->cm_done);
 736         if (err || (rdma->state != P9_RDMA_CONNECTED))
 737                 goto error;
 738
 739         client->status = Connected;
 740
 741         return 0;
 742
 743 error:
 744         rdma_destroy_trans(rdma);
 745         return -ENOTCONN;
 746 }
 747
 748 static struct p9_trans_module p9_rdma_trans = {
 749         .name = "rdma",
 750         .maxsize = P9_RDMA_MAXSIZE,
 751         .pooled_rbuffers = true,
 752         .def = 0,
 753         .owner = THIS_MODULE,
 754         .create = rdma_create_trans,
 755         .close = rdma_close,
 756         .request = rdma_request,
 757         .cancel = rdma_cancel,
 758         .cancelled = rdma_cancelled,
 759         .show_options = p9_rdma_show_options,
 760 };
 761
 762 /**
 763  * p9_trans_rdma_init - Register the 9P RDMA transport driver
 764  */
 765 static int __init p9_trans_rdma_init(void)
 766 {
 767         v9fs_register_trans(&p9_rdma_trans);
 768         return 0;
 769 }
 770
 771 static void __exit p9_trans_rdma_exit(void)
 772 {
 773         v9fs_unregister_trans(&p9_rdma_trans);
 774 }
 775
 776 module_init(p9_trans_rdma_init);
 777 module_exit(p9_trans_rdma_exit);
 778 MODULE_ALIAS_9P("rdma");
 779
 780 MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
 781 MODULE_DESCRIPTION("RDMA Transport for 9P");
 782 MODULE_LICENSE("Dual BSD/GPL");