linux/net/9p/trans_rdma.c
<<
>>
Prefs
   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 */
  82struct 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 */
 117struct p9_rdma_req;
 118struct 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 */
 135struct 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 */
 145enum {
 146        /* Options that take integer arguments */
 147        Opt_port, Opt_rq_depth, Opt_sq_depth, Opt_timeout, Opt_err,
 148};
 149
 150static 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 */
 165static 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
 225static int
 226p9_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
 277static void
 278handle_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
 313static void
 314handle_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
 322static 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
 328static 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
 360static 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
 365static 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
 388static int
 389post_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
 417static 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                goto err_close;
 431
 432        /*
 433         * If the request has a buffer, steal it, otherwise
 434         * allocate a new one.  Typically, requests should already
 435         * have receive buffers allocated and just swap them around
 436         */
 437        if (!req->rc) {
 438                req->rc = kmalloc(sizeof(struct p9_fcall)+client->msize,
 439                                                                GFP_KERNEL);
 440                if (req->rc) {
 441                        req->rc->sdata = (char *) req->rc +
 442                                                sizeof(struct p9_fcall);
 443                        req->rc->capacity = client->msize;
 444                }
 445        }
 446        rpl_context->rc = req->rc;
 447        if (!rpl_context->rc) {
 448                kfree(rpl_context);
 449                goto err_close;
 450        }
 451
 452        /*
 453         * Post a receive buffer for this request. We need to ensure
 454         * there is a reply buffer available for every outstanding
 455         * request. A flushed request can result in no reply for an
 456         * outstanding request, so we must keep a count to avoid
 457         * overflowing the RQ.
 458         */
 459        if (atomic_inc_return(&rdma->rq_count) <= rdma->rq_depth) {
 460                err = post_recv(client, rpl_context);
 461                if (err) {
 462                        kfree(rpl_context->rc);
 463                        kfree(rpl_context);
 464                        goto err_close;
 465                }
 466        } else
 467                atomic_dec(&rdma->rq_count);
 468
 469        /* remove posted receive buffer from request structure */
 470        req->rc = NULL;
 471
 472        /* Post the request */
 473        c = kmalloc(sizeof *c, GFP_KERNEL);
 474        if (!c)
 475                goto err_close;
 476        c->req = req;
 477
 478        c->busa = ib_dma_map_single(rdma->cm_id->device,
 479                                    c->req->tc->sdata, c->req->tc->size,
 480                                    DMA_TO_DEVICE);
 481        if (ib_dma_mapping_error(rdma->cm_id->device, c->busa))
 482                goto error;
 483
 484        sge.addr = c->busa;
 485        sge.length = c->req->tc->size;
 486        sge.lkey = rdma->lkey;
 487
 488        wr.next = NULL;
 489        c->wc_op = IB_WC_SEND;
 490        wr.wr_id = (unsigned long) c;
 491        wr.opcode = IB_WR_SEND;
 492        wr.send_flags = IB_SEND_SIGNALED;
 493        wr.sg_list = &sge;
 494        wr.num_sge = 1;
 495
 496        if (down_interruptible(&rdma->sq_sem))
 497                goto error;
 498
 499        return ib_post_send(rdma->qp, &wr, &bad_wr);
 500
 501 error:
 502        P9_DPRINTK(P9_DEBUG_ERROR, "EIO\n");
 503        return -EIO;
 504
 505 err_close:
 506        spin_lock_irqsave(&rdma->req_lock, flags);
 507        if (rdma->state < P9_RDMA_CLOSING) {
 508                rdma->state = P9_RDMA_CLOSING;
 509                spin_unlock_irqrestore(&rdma->req_lock, flags);
 510                rdma_disconnect(rdma->cm_id);
 511        } else
 512                spin_unlock_irqrestore(&rdma->req_lock, flags);
 513        return err;
 514}
 515
 516static void rdma_close(struct p9_client *client)
 517{
 518        struct p9_trans_rdma *rdma;
 519
 520        if (!client)
 521                return;
 522
 523        rdma = client->trans;
 524        if (!rdma)
 525                return;
 526
 527        client->status = Disconnected;
 528        rdma_disconnect(rdma->cm_id);
 529        rdma_destroy_trans(rdma);
 530}
 531
 532/**
 533 * alloc_rdma - Allocate and initialize the rdma transport structure
 534 * @opts: Mount options structure
 535 */
 536static struct p9_trans_rdma *alloc_rdma(struct p9_rdma_opts *opts)
 537{
 538        struct p9_trans_rdma *rdma;
 539
 540        rdma = kzalloc(sizeof(struct p9_trans_rdma), GFP_KERNEL);
 541        if (!rdma)
 542                return NULL;
 543
 544        rdma->sq_depth = opts->sq_depth;
 545        rdma->rq_depth = opts->rq_depth;
 546        rdma->timeout = opts->timeout;
 547        spin_lock_init(&rdma->req_lock);
 548        init_completion(&rdma->cm_done);
 549        sema_init(&rdma->sq_sem, rdma->sq_depth);
 550        atomic_set(&rdma->rq_count, 0);
 551
 552        return rdma;
 553}
 554
 555/* its not clear to me we can do anything after send has been posted */
 556static int rdma_cancel(struct p9_client *client, struct p9_req_t *req)
 557{
 558        return 1;
 559}
 560
 561/**
 562 * trans_create_rdma - Transport method for creating atransport instance
 563 * @client: client instance
 564 * @addr: IP address string
 565 * @args: Mount options string
 566 */
 567static int
 568rdma_create_trans(struct p9_client *client, const char *addr, char *args)
 569{
 570        int err;
 571        struct p9_rdma_opts opts;
 572        struct p9_trans_rdma *rdma;
 573        struct rdma_conn_param conn_param;
 574        struct ib_qp_init_attr qp_attr;
 575        struct ib_device_attr devattr;
 576
 577        /* Parse the transport specific mount options */
 578        err = parse_opts(args, &opts);
 579        if (err < 0)
 580                return err;
 581
 582        /* Create and initialize the RDMA transport structure */
 583        rdma = alloc_rdma(&opts);
 584        if (!rdma)
 585                return -ENOMEM;
 586
 587        /* Create the RDMA CM ID */
 588        rdma->cm_id = rdma_create_id(p9_cm_event_handler, client, RDMA_PS_TCP);
 589        if (IS_ERR(rdma->cm_id))
 590                goto error;
 591
 592        /* Associate the client with the transport */
 593        client->trans = rdma;
 594
 595        /* Resolve the server's address */
 596        rdma->addr.sin_family = AF_INET;
 597        rdma->addr.sin_addr.s_addr = in_aton(addr);
 598        rdma->addr.sin_port = htons(opts.port);
 599        err = rdma_resolve_addr(rdma->cm_id, NULL,
 600                                (struct sockaddr *)&rdma->addr,
 601                                rdma->timeout);
 602        if (err)
 603                goto error;
 604        err = wait_for_completion_interruptible(&rdma->cm_done);
 605        if (err || (rdma->state != P9_RDMA_ADDR_RESOLVED))
 606                goto error;
 607
 608        /* Resolve the route to the server */
 609        err = rdma_resolve_route(rdma->cm_id, rdma->timeout);
 610        if (err)
 611                goto error;
 612        err = wait_for_completion_interruptible(&rdma->cm_done);
 613        if (err || (rdma->state != P9_RDMA_ROUTE_RESOLVED))
 614                goto error;
 615
 616        /* Query the device attributes */
 617        err = ib_query_device(rdma->cm_id->device, &devattr);
 618        if (err)
 619                goto error;
 620
 621        /* Create the Completion Queue */
 622        rdma->cq = ib_create_cq(rdma->cm_id->device, cq_comp_handler,
 623                                cq_event_handler, client,
 624                                opts.sq_depth + opts.rq_depth + 1, 0);
 625        if (IS_ERR(rdma->cq))
 626                goto error;
 627        ib_req_notify_cq(rdma->cq, IB_CQ_NEXT_COMP);
 628
 629        /* Create the Protection Domain */
 630        rdma->pd = ib_alloc_pd(rdma->cm_id->device);
 631        if (IS_ERR(rdma->pd))
 632                goto error;
 633
 634        /* Cache the DMA lkey in the transport */
 635        rdma->dma_mr = NULL;
 636        if (devattr.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
 637                rdma->lkey = rdma->cm_id->device->local_dma_lkey;
 638        else {
 639                rdma->dma_mr = ib_get_dma_mr(rdma->pd, IB_ACCESS_LOCAL_WRITE);
 640                if (IS_ERR(rdma->dma_mr))
 641                        goto error;
 642                rdma->lkey = rdma->dma_mr->lkey;
 643        }
 644
 645        /* Create the Queue Pair */
 646        memset(&qp_attr, 0, sizeof qp_attr);
 647        qp_attr.event_handler = qp_event_handler;
 648        qp_attr.qp_context = client;
 649        qp_attr.cap.max_send_wr = opts.sq_depth;
 650        qp_attr.cap.max_recv_wr = opts.rq_depth;
 651        qp_attr.cap.max_send_sge = P9_RDMA_SEND_SGE;
 652        qp_attr.cap.max_recv_sge = P9_RDMA_RECV_SGE;
 653        qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 654        qp_attr.qp_type = IB_QPT_RC;
 655        qp_attr.send_cq = rdma->cq;
 656        qp_attr.recv_cq = rdma->cq;
 657        err = rdma_create_qp(rdma->cm_id, rdma->pd, &qp_attr);
 658        if (err)
 659                goto error;
 660        rdma->qp = rdma->cm_id->qp;
 661
 662        /* Request a connection */
 663        memset(&conn_param, 0, sizeof(conn_param));
 664        conn_param.private_data = NULL;
 665        conn_param.private_data_len = 0;
 666        conn_param.responder_resources = P9_RDMA_IRD;
 667        conn_param.initiator_depth = P9_RDMA_ORD;
 668        err = rdma_connect(rdma->cm_id, &conn_param);
 669        if (err)
 670                goto error;
 671        err = wait_for_completion_interruptible(&rdma->cm_done);
 672        if (err || (rdma->state != P9_RDMA_CONNECTED))
 673                goto error;
 674
 675        client->status = Connected;
 676
 677        return 0;
 678
 679error:
 680        rdma_destroy_trans(rdma);
 681        return -ENOTCONN;
 682}
 683
 684static struct p9_trans_module p9_rdma_trans = {
 685        .name = "rdma",
 686        .maxsize = P9_RDMA_MAXSIZE,
 687        .def = 0,
 688        .owner = THIS_MODULE,
 689        .create = rdma_create_trans,
 690        .close = rdma_close,
 691        .request = rdma_request,
 692        .cancel = rdma_cancel,
 693};
 694
 695/**
 696 * p9_trans_rdma_init - Register the 9P RDMA transport driver
 697 */
 698static int __init p9_trans_rdma_init(void)
 699{
 700        v9fs_register_trans(&p9_rdma_trans);
 701        return 0;
 702}
 703
 704static void __exit p9_trans_rdma_exit(void)
 705{
 706        v9fs_unregister_trans(&p9_rdma_trans);
 707}
 708
 709module_init(p9_trans_rdma_init);
 710module_exit(p9_trans_rdma_exit);
 711
 712MODULE_AUTHOR("Tom Tucker <tom@opengridcomputing.com>");
 713MODULE_DESCRIPTION("RDMA Transport for 9P");
 714MODULE_LICENSE("Dual BSD/GPL");
 715
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.