net/9p/trans_rdma.c

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