SAFE public projects git trees. - safe/jmp/linux-2.6/blob - net/sunrpc/xprtrdma/svc_rdma_transport.c

   1 /*
   2  * Copyright (c) 2005-2007 Network Appliance, Inc. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the BSD-type
   8  * license below:
   9  *
  10  * Redistribution and use in source and binary forms, with or without
  11  * modification, are permitted provided that the following conditions
  12  * are met:
  13  *
  14  *      Redistributions of source code must retain the above copyright
  15  *      notice, this list of conditions and the following disclaimer.
  16  *
  17  *      Redistributions in binary form must reproduce the above
  18  *      copyright notice, this list of conditions and the following
  19  *      disclaimer in the documentation and/or other materials provided
  20  *      with the distribution.
  21  *
  22  *      Neither the name of the Network Appliance, Inc. nor the names of
  23  *      its contributors may be used to endorse or promote products
  24  *      derived from this software without specific prior written
  25  *      permission.
  26  *
  27  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  28  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  29  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  30  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  31  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  32  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  33  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  34  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  35  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  36  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  37  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  38  *
  39  * Author: Tom Tucker <tom@opengridcomputing.com>
  40  */
  41
  42 #include <linux/sunrpc/svc_xprt.h>
  43 #include <linux/sunrpc/debug.h>
  44 #include <linux/sunrpc/rpc_rdma.h>
  45 #include <linux/spinlock.h>
  46 #include <rdma/ib_verbs.h>
  47 #include <rdma/rdma_cm.h>
  48 #include <linux/sunrpc/svc_rdma.h>
  49
  50 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
  51
  52 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
  53                                         struct sockaddr *sa, int salen,
  54                                         int flags);
  55 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt);
  56 static void svc_rdma_release_rqst(struct svc_rqst *);
  57 static void dto_tasklet_func(unsigned long data);
  58 static void svc_rdma_detach(struct svc_xprt *xprt);
  59 static void svc_rdma_free(struct svc_xprt *xprt);
  60 static int svc_rdma_has_wspace(struct svc_xprt *xprt);
  61 static void rq_cq_reap(struct svcxprt_rdma *xprt);
  62 static void sq_cq_reap(struct svcxprt_rdma *xprt);
  63
  64 DECLARE_TASKLET(dto_tasklet, dto_tasklet_func, 0UL);
  65 static DEFINE_SPINLOCK(dto_lock);
  66 static LIST_HEAD(dto_xprt_q);
  67
  68 static struct svc_xprt_ops svc_rdma_ops = {
  69         .xpo_create = svc_rdma_create,
  70         .xpo_recvfrom = svc_rdma_recvfrom,
  71         .xpo_sendto = svc_rdma_sendto,
  72         .xpo_release_rqst = svc_rdma_release_rqst,
  73         .xpo_detach = svc_rdma_detach,
  74         .xpo_free = svc_rdma_free,
  75         .xpo_prep_reply_hdr = svc_rdma_prep_reply_hdr,
  76         .xpo_has_wspace = svc_rdma_has_wspace,
  77         .xpo_accept = svc_rdma_accept,
  78 };
  79
  80 struct svc_xprt_class svc_rdma_class = {
  81         .xcl_name = "rdma",
  82         .xcl_owner = THIS_MODULE,
  83         .xcl_ops = &svc_rdma_ops,
  84         .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
  85 };
  86
  87 static int rdma_bump_context_cache(struct svcxprt_rdma *xprt)
  88 {
  89         int target;
  90         int at_least_one = 0;
  91         struct svc_rdma_op_ctxt *ctxt;
  92
  93         target = min(xprt->sc_ctxt_cnt + xprt->sc_ctxt_bump,
  94                      xprt->sc_ctxt_max);
  95
  96         spin_lock_bh(&xprt->sc_ctxt_lock);
  97         while (xprt->sc_ctxt_cnt < target) {
  98                 xprt->sc_ctxt_cnt++;
  99                 spin_unlock_bh(&xprt->sc_ctxt_lock);
 100
 101                 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
 102
 103                 spin_lock_bh(&xprt->sc_ctxt_lock);
 104                 if (ctxt) {
 105                         at_least_one = 1;
 106                         ctxt->next = xprt->sc_ctxt_head;
 107                         xprt->sc_ctxt_head = ctxt;
 108                 } else {
 109                         /* kmalloc failed...give up for now */
 110                         xprt->sc_ctxt_cnt--;
 111                         break;
 112                 }
 113         }
 114         spin_unlock_bh(&xprt->sc_ctxt_lock);
 115         dprintk("svcrdma: sc_ctxt_max=%d, sc_ctxt_cnt=%d\n",
 116                 xprt->sc_ctxt_max, xprt->sc_ctxt_cnt);
 117         return at_least_one;
 118 }
 119
 120 struct svc_rdma_op_ctxt *svc_rdma_get_context(struct svcxprt_rdma *xprt)
 121 {
 122         struct svc_rdma_op_ctxt *ctxt;
 123
 124         while (1) {
 125                 spin_lock_bh(&xprt->sc_ctxt_lock);
 126                 if (unlikely(xprt->sc_ctxt_head == NULL)) {
 127                         /* Try to bump my cache. */
 128                         spin_unlock_bh(&xprt->sc_ctxt_lock);
 129
 130                         if (rdma_bump_context_cache(xprt))
 131                                 continue;
 132
 133                         printk(KERN_INFO "svcrdma: sleeping waiting for "
 134                                "context memory on xprt=%p\n",
 135                                xprt);
 136                         schedule_timeout_uninterruptible(msecs_to_jiffies(500));
 137                         continue;
 138                 }
 139                 ctxt = xprt->sc_ctxt_head;
 140                 xprt->sc_ctxt_head = ctxt->next;
 141                 spin_unlock_bh(&xprt->sc_ctxt_lock);
 142                 ctxt->xprt = xprt;
 143                 INIT_LIST_HEAD(&ctxt->dto_q);
 144                 ctxt->count = 0;
 145                 break;
 146         }
 147         return ctxt;
 148 }
 149
 150 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
 151 {
 152         struct svcxprt_rdma *xprt;
 153         int i;
 154
 155         BUG_ON(!ctxt);
 156         xprt = ctxt->xprt;
 157         if (free_pages)
 158                 for (i = 0; i < ctxt->count; i++)
 159                         put_page(ctxt->pages[i]);
 160
 161         for (i = 0; i < ctxt->count; i++)
 162                 dma_unmap_single(xprt->sc_cm_id->device->dma_device,
 163                                  ctxt->sge[i].addr,
 164                                  ctxt->sge[i].length,
 165                                  ctxt->direction);
 166         spin_lock_bh(&xprt->sc_ctxt_lock);
 167         ctxt->next = xprt->sc_ctxt_head;
 168         xprt->sc_ctxt_head = ctxt;
 169         spin_unlock_bh(&xprt->sc_ctxt_lock);
 170 }
 171
 172 /* ib_cq event handler */
 173 static void cq_event_handler(struct ib_event *event, void *context)
 174 {
 175         struct svc_xprt *xprt = context;
 176         dprintk("svcrdma: received CQ event id=%d, context=%p\n",
 177                 event->event, context);
 178         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 179 }
 180
 181 /* QP event handler */
 182 static void qp_event_handler(struct ib_event *event, void *context)
 183 {
 184         struct svc_xprt *xprt = context;
 185
 186         switch (event->event) {
 187         /* These are considered benign events */
 188         case IB_EVENT_PATH_MIG:
 189         case IB_EVENT_COMM_EST:
 190         case IB_EVENT_SQ_DRAINED:
 191         case IB_EVENT_QP_LAST_WQE_REACHED:
 192                 dprintk("svcrdma: QP event %d received for QP=%p\n",
 193                         event->event, event->element.qp);
 194                 break;
 195         /* These are considered fatal events */
 196         case IB_EVENT_PATH_MIG_ERR:
 197         case IB_EVENT_QP_FATAL:
 198         case IB_EVENT_QP_REQ_ERR:
 199         case IB_EVENT_QP_ACCESS_ERR:
 200         case IB_EVENT_DEVICE_FATAL:
 201         default:
 202                 dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
 203                         "closing transport\n",
 204                         event->event, event->element.qp);
 205                 set_bit(XPT_CLOSE, &xprt->xpt_flags);
 206                 break;
 207         }
 208 }
 209
 210 /*
 211  * Data Transfer Operation Tasklet
 212  *
 213  * Walks a list of transports with I/O pending, removing entries as
 214  * they are added to the server's I/O pending list. Two bits indicate
 215  * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
 216  * spinlock that serializes access to the transport list with the RQ
 217  * and SQ interrupt handlers.
 218  */
 219 static void dto_tasklet_func(unsigned long data)
 220 {
 221         struct svcxprt_rdma *xprt;
 222         unsigned long flags;
 223
 224         spin_lock_irqsave(&dto_lock, flags);
 225         while (!list_empty(&dto_xprt_q)) {
 226                 xprt = list_entry(dto_xprt_q.next,
 227                                   struct svcxprt_rdma, sc_dto_q);
 228                 list_del_init(&xprt->sc_dto_q);
 229                 spin_unlock_irqrestore(&dto_lock, flags);
 230
 231                 if (test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags)) {
 232                         ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
 233                         rq_cq_reap(xprt);
 234                         set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
 235                         /*
 236                          * If data arrived before established event,
 237                          * don't enqueue. This defers RPC I/O until the
 238                          * RDMA connection is complete.
 239                          */
 240                         if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
 241                                 svc_xprt_enqueue(&xprt->sc_xprt);
 242                 }
 243
 244                 if (test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags)) {
 245                         ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
 246                         sq_cq_reap(xprt);
 247                 }
 248
 249                 svc_xprt_put(&xprt->sc_xprt);
 250                 spin_lock_irqsave(&dto_lock, flags);
 251         }
 252         spin_unlock_irqrestore(&dto_lock, flags);
 253 }
 254
 255 /*
 256  * Receive Queue Completion Handler
 257  *
 258  * Since an RQ completion handler is called on interrupt context, we
 259  * need to defer the handling of the I/O to a tasklet
 260  */
 261 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
 262 {
 263         struct svcxprt_rdma *xprt = cq_context;
 264         unsigned long flags;
 265
 266         /*
 267          * Set the bit regardless of whether or not it's on the list
 268          * because it may be on the list already due to an SQ
 269          * completion.
 270         */
 271         set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
 272
 273         /*
 274          * If this transport is not already on the DTO transport queue,
 275          * add it
 276          */
 277         spin_lock_irqsave(&dto_lock, flags);
 278         if (list_empty(&xprt->sc_dto_q)) {
 279                 svc_xprt_get(&xprt->sc_xprt);
 280                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
 281         }
 282         spin_unlock_irqrestore(&dto_lock, flags);
 283
 284         /* Tasklet does all the work to avoid irqsave locks. */
 285         tasklet_schedule(&dto_tasklet);
 286 }
 287
 288 /*
 289  * rq_cq_reap - Process the RQ CQ.
 290  *
 291  * Take all completing WC off the CQE and enqueue the associated DTO
 292  * context on the dto_q for the transport.
 293  */
 294 static void rq_cq_reap(struct svcxprt_rdma *xprt)
 295 {
 296         int ret;
 297         struct ib_wc wc;
 298         struct svc_rdma_op_ctxt *ctxt = NULL;
 299
 300         atomic_inc(&rdma_stat_rq_poll);
 301
 302         spin_lock_bh(&xprt->sc_rq_dto_lock);
 303         while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
 304                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
 305                 ctxt->wc_status = wc.status;
 306                 ctxt->byte_len = wc.byte_len;
 307                 if (wc.status != IB_WC_SUCCESS) {
 308                         /* Close the transport */
 309                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 310                         svc_rdma_put_context(ctxt, 1);
 311                         continue;
 312                 }
 313                 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
 314         }
 315         spin_unlock_bh(&xprt->sc_rq_dto_lock);
 316
 317         if (ctxt)
 318                 atomic_inc(&rdma_stat_rq_prod);
 319 }
 320
 321 /*
 322  * Send Queue Completion Handler - potentially called on interrupt context.
 323  */
 324 static void sq_cq_reap(struct svcxprt_rdma *xprt)
 325 {
 326         struct svc_rdma_op_ctxt *ctxt = NULL;
 327         struct ib_wc wc;
 328         struct ib_cq *cq = xprt->sc_sq_cq;
 329         int ret;
 330
 331         atomic_inc(&rdma_stat_sq_poll);
 332         while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
 333                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
 334                 xprt = ctxt->xprt;
 335
 336                 if (wc.status != IB_WC_SUCCESS)
 337                         /* Close the transport */
 338                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 339
 340                 /* Decrement used SQ WR count */
 341                 atomic_dec(&xprt->sc_sq_count);
 342                 wake_up(&xprt->sc_send_wait);
 343
 344                 switch (ctxt->wr_op) {
 345                 case IB_WR_SEND:
 346                 case IB_WR_RDMA_WRITE:
 347                         svc_rdma_put_context(ctxt, 1);
 348                         break;
 349
 350                 case IB_WR_RDMA_READ:
 351                         if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
 352                                 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
 353                                 set_bit(RDMACTXT_F_READ_DONE, &ctxt->flags);
 354                                 spin_lock_bh(&xprt->sc_read_complete_lock);
 355                                 list_add_tail(&ctxt->dto_q,
 356                                               &xprt->sc_read_complete_q);
 357                                 spin_unlock_bh(&xprt->sc_read_complete_lock);
 358                                 svc_xprt_enqueue(&xprt->sc_xprt);
 359                         }
 360                         break;
 361
 362                 default:
 363                         printk(KERN_ERR "svcrdma: unexpected completion type, "
 364                                "opcode=%d, status=%d\n",
 365                                wc.opcode, wc.status);
 366                         break;
 367                 }
 368         }
 369
 370         if (ctxt)
 371                 atomic_inc(&rdma_stat_sq_prod);
 372 }
 373
 374 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
 375 {
 376         struct svcxprt_rdma *xprt = cq_context;
 377         unsigned long flags;
 378
 379         /*
 380          * Set the bit regardless of whether or not it's on the list
 381          * because it may be on the list already due to an RQ
 382          * completion.
 383         */
 384         set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
 385
 386         /*
 387          * If this transport is not already on the DTO transport queue,
 388          * add it
 389          */
 390         spin_lock_irqsave(&dto_lock, flags);
 391         if (list_empty(&xprt->sc_dto_q)) {
 392                 svc_xprt_get(&xprt->sc_xprt);
 393                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
 394         }
 395         spin_unlock_irqrestore(&dto_lock, flags);
 396
 397         /* Tasklet does all the work to avoid irqsave locks. */
 398         tasklet_schedule(&dto_tasklet);
 399 }
 400
 401 static void create_context_cache(struct svcxprt_rdma *xprt,
 402                                  int ctxt_count, int ctxt_bump, int ctxt_max)
 403 {
 404         struct svc_rdma_op_ctxt *ctxt;
 405         int i;
 406
 407         xprt->sc_ctxt_max = ctxt_max;
 408         xprt->sc_ctxt_bump = ctxt_bump;
 409         xprt->sc_ctxt_cnt = 0;
 410         xprt->sc_ctxt_head = NULL;
 411         for (i = 0; i < ctxt_count; i++) {
 412                 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
 413                 if (ctxt) {
 414                         ctxt->next = xprt->sc_ctxt_head;
 415                         xprt->sc_ctxt_head = ctxt;
 416                         xprt->sc_ctxt_cnt++;
 417                 }
 418         }
 419 }
 420
 421 static void destroy_context_cache(struct svc_rdma_op_ctxt *ctxt)
 422 {
 423         struct svc_rdma_op_ctxt *next;
 424         if (!ctxt)
 425                 return;
 426
 427         do {
 428                 next = ctxt->next;
 429                 kfree(ctxt);
 430                 ctxt = next;
 431         } while (next);
 432 }
 433
 434 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
 435                                              int listener)
 436 {
 437         struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
 438
 439         if (!cma_xprt)
 440                 return NULL;
 441         svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv);
 442         INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
 443         INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
 444         INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
 445         INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
 446         init_waitqueue_head(&cma_xprt->sc_send_wait);
 447
 448         spin_lock_init(&cma_xprt->sc_lock);
 449         spin_lock_init(&cma_xprt->sc_read_complete_lock);
 450         spin_lock_init(&cma_xprt->sc_ctxt_lock);
 451         spin_lock_init(&cma_xprt->sc_rq_dto_lock);
 452
 453         cma_xprt->sc_ord = svcrdma_ord;
 454
 455         cma_xprt->sc_max_req_size = svcrdma_max_req_size;
 456         cma_xprt->sc_max_requests = svcrdma_max_requests;
 457         cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
 458         atomic_set(&cma_xprt->sc_sq_count, 0);
 459
 460         if (!listener) {
 461                 int reqs = cma_xprt->sc_max_requests;
 462                 create_context_cache(cma_xprt,
 463                                      reqs << 1, /* starting size */
 464                                      reqs,      /* bump amount */
 465                                      reqs +
 466                                      cma_xprt->sc_sq_depth +
 467                                      RPCRDMA_MAX_THREADS + 1); /* max */
 468                 if (!cma_xprt->sc_ctxt_head) {
 469                         kfree(cma_xprt);
 470                         return NULL;
 471                 }
 472                 clear_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
 473         } else
 474                 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
 475
 476         return cma_xprt;
 477 }
 478
 479 struct page *svc_rdma_get_page(void)
 480 {
 481         struct page *page;
 482
 483         while ((page = alloc_page(GFP_KERNEL)) == NULL) {
 484                 /* If we can't get memory, wait a bit and try again */
 485                 printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 "
 486                        "jiffies.\n");
 487                 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
 488         }
 489         return page;
 490 }
 491
 492 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
 493 {
 494         struct ib_recv_wr recv_wr, *bad_recv_wr;
 495         struct svc_rdma_op_ctxt *ctxt;
 496         struct page *page;
 497         unsigned long pa;
 498         int sge_no;
 499         int buflen;
 500         int ret;
 501
 502         ctxt = svc_rdma_get_context(xprt);
 503         buflen = 0;
 504         ctxt->direction = DMA_FROM_DEVICE;
 505         for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
 506                 BUG_ON(sge_no >= xprt->sc_max_sge);
 507                 page = svc_rdma_get_page();
 508                 ctxt->pages[sge_no] = page;
 509                 pa = ib_dma_map_page(xprt->sc_cm_id->device,
 510                                      page, 0, PAGE_SIZE,
 511                                      DMA_FROM_DEVICE);
 512                 ctxt->sge[sge_no].addr = pa;
 513                 ctxt->sge[sge_no].length = PAGE_SIZE;
 514                 ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
 515                 buflen += PAGE_SIZE;
 516         }
 517         ctxt->count = sge_no;
 518         recv_wr.next = NULL;
 519         recv_wr.sg_list = &ctxt->sge[0];
 520         recv_wr.num_sge = ctxt->count;
 521         recv_wr.wr_id = (u64)(unsigned long)ctxt;
 522
 523         ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
 524         return ret;
 525 }
 526
 527 /*
 528  * This function handles the CONNECT_REQUEST event on a listening
 529  * endpoint. It is passed the cma_id for the _new_ connection. The context in
 530  * this cma_id is inherited from the listening cma_id and is the svc_xprt
 531  * structure for the listening endpoint.
 532  *
 533  * This function creates a new xprt for the new connection and enqueues it on
 534  * the accept queue for the listent xprt. When the listen thread is kicked, it
 535  * will call the recvfrom method on the listen xprt which will accept the new
 536  * connection.
 537  */
 538 static void handle_connect_req(struct rdma_cm_id *new_cma_id)
 539 {
 540         struct svcxprt_rdma *listen_xprt = new_cma_id->context;
 541         struct svcxprt_rdma *newxprt;
 542
 543         /* Create a new transport */
 544         newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
 545         if (!newxprt) {
 546                 dprintk("svcrdma: failed to create new transport\n");
 547                 return;
 548         }
 549         newxprt->sc_cm_id = new_cma_id;
 550         new_cma_id->context = newxprt;
 551         dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
 552                 newxprt, newxprt->sc_cm_id, listen_xprt);
 553
 554         /*
 555          * Enqueue the new transport on the accept queue of the listening
 556          * transport
 557          */
 558         spin_lock_bh(&listen_xprt->sc_lock);
 559         list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
 560         spin_unlock_bh(&listen_xprt->sc_lock);
 561
 562         /*
 563          * Can't use svc_xprt_received here because we are not on a
 564          * rqstp thread
 565         */
 566         set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
 567         svc_xprt_enqueue(&listen_xprt->sc_xprt);
 568 }
 569
 570 /*
 571  * Handles events generated on the listening endpoint. These events will be
 572  * either be incoming connect requests or adapter removal  events.
 573  */
 574 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
 575                                struct rdma_cm_event *event)
 576 {
 577         struct svcxprt_rdma *xprt = cma_id->context;
 578         int ret = 0;
 579
 580         switch (event->event) {
 581         case RDMA_CM_EVENT_CONNECT_REQUEST:
 582                 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
 583                         "event=%d\n", cma_id, cma_id->context, event->event);
 584                 handle_connect_req(cma_id);
 585                 break;
 586
 587         case RDMA_CM_EVENT_ESTABLISHED:
 588                 /* Accept complete */
 589                 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
 590                         "cm_id=%p\n", xprt, cma_id);
 591                 break;
 592
 593         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 594                 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
 595                         xprt, cma_id);
 596                 if (xprt)
 597                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 598                 break;
 599
 600         default:
 601                 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
 602                         "event=%d\n", cma_id, event->event);
 603                 break;
 604         }
 605
 606         return ret;
 607 }
 608
 609 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
 610                             struct rdma_cm_event *event)
 611 {
 612         struct svc_xprt *xprt = cma_id->context;
 613         struct svcxprt_rdma *rdma =
 614                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 615         switch (event->event) {
 616         case RDMA_CM_EVENT_ESTABLISHED:
 617                 /* Accept complete */
 618                 svc_xprt_get(xprt);
 619                 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
 620                         "cm_id=%p\n", xprt, cma_id);
 621                 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
 622                 svc_xprt_enqueue(xprt);
 623                 break;
 624         case RDMA_CM_EVENT_DISCONNECTED:
 625                 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
 626                         xprt, cma_id);
 627                 if (xprt) {
 628                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 629                         svc_xprt_enqueue(xprt);
 630                 }
 631                 break;
 632         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 633                 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
 634                         "event=%d\n", cma_id, xprt, event->event);
 635                 if (xprt) {
 636                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 637                         svc_xprt_enqueue(xprt);
 638                 }
 639                 break;
 640         default:
 641                 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
 642                         "event=%d\n", cma_id, event->event);
 643                 break;
 644         }
 645         return 0;
 646 }
 647
 648 /*
 649  * Create a listening RDMA service endpoint.
 650  */
 651 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
 652                                         struct sockaddr *sa, int salen,
 653                                         int flags)
 654 {
 655         struct rdma_cm_id *listen_id;
 656         struct svcxprt_rdma *cma_xprt;
 657         struct svc_xprt *xprt;
 658         int ret;
 659
 660         dprintk("svcrdma: Creating RDMA socket\n");
 661
 662         cma_xprt = rdma_create_xprt(serv, 1);
 663         if (!cma_xprt)
 664                 return ERR_PTR(ENOMEM);
 665         xprt = &cma_xprt->sc_xprt;
 666
 667         listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
 668         if (IS_ERR(listen_id)) {
 669                 svc_xprt_put(&cma_xprt->sc_xprt);
 670                 dprintk("svcrdma: rdma_create_id failed = %ld\n",
 671                         PTR_ERR(listen_id));
 672                 return (void *)listen_id;
 673         }
 674         ret = rdma_bind_addr(listen_id, sa);
 675         if (ret) {
 676                 rdma_destroy_id(listen_id);
 677                 svc_xprt_put(&cma_xprt->sc_xprt);
 678                 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
 679                 return ERR_PTR(ret);
 680         }
 681         cma_xprt->sc_cm_id = listen_id;
 682
 683         ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
 684         if (ret) {
 685                 rdma_destroy_id(listen_id);
 686                 svc_xprt_put(&cma_xprt->sc_xprt);
 687                 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
 688                 return ERR_PTR(ret);
 689         }
 690
 691         /*
 692          * We need to use the address from the cm_id in case the
 693          * caller specified 0 for the port number.
 694          */
 695         sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
 696         svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
 697
 698         return &cma_xprt->sc_xprt;
 699 }
 700
 701 /*
 702  * This is the xpo_recvfrom function for listening endpoints. Its
 703  * purpose is to accept incoming connections. The CMA callback handler
 704  * has already created a new transport and attached it to the new CMA
 705  * ID.
 706  *
 707  * There is a queue of pending connections hung on the listening
 708  * transport. This queue contains the new svc_xprt structure. This
 709  * function takes svc_xprt structures off the accept_q and completes
 710  * the connection.
 711  */
 712 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
 713 {
 714         struct svcxprt_rdma *listen_rdma;
 715         struct svcxprt_rdma *newxprt = NULL;
 716         struct rdma_conn_param conn_param;
 717         struct ib_qp_init_attr qp_attr;
 718         struct ib_device_attr devattr;
 719         struct sockaddr *sa;
 720         int ret;
 721         int i;
 722
 723         listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
 724         clear_bit(XPT_CONN, &xprt->xpt_flags);
 725         /* Get the next entry off the accept list */
 726         spin_lock_bh(&listen_rdma->sc_lock);
 727         if (!list_empty(&listen_rdma->sc_accept_q)) {
 728                 newxprt = list_entry(listen_rdma->sc_accept_q.next,
 729                                      struct svcxprt_rdma, sc_accept_q);
 730                 list_del_init(&newxprt->sc_accept_q);
 731         }
 732         if (!list_empty(&listen_rdma->sc_accept_q))
 733                 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
 734         spin_unlock_bh(&listen_rdma->sc_lock);
 735         if (!newxprt)
 736                 return NULL;
 737
 738         dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
 739                 newxprt, newxprt->sc_cm_id);
 740
 741         ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
 742         if (ret) {
 743                 dprintk("svcrdma: could not query device attributes on "
 744                         "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
 745                 goto errout;
 746         }
 747
 748         /* Qualify the transport resource defaults with the
 749          * capabilities of this particular device */
 750         newxprt->sc_max_sge = min((size_t)devattr.max_sge,
 751                                   (size_t)RPCSVC_MAXPAGES);
 752         newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
 753                                    (size_t)svcrdma_max_requests);
 754         newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
 755
 756         newxprt->sc_ord =  min((size_t)devattr.max_qp_rd_atom,
 757                                (size_t)svcrdma_ord);
 758
 759         newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
 760         if (IS_ERR(newxprt->sc_pd)) {
 761                 dprintk("svcrdma: error creating PD for connect request\n");
 762                 goto errout;
 763         }
 764         newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
 765                                          sq_comp_handler,
 766                                          cq_event_handler,
 767                                          newxprt,
 768                                          newxprt->sc_sq_depth,
 769                                          0);
 770         if (IS_ERR(newxprt->sc_sq_cq)) {
 771                 dprintk("svcrdma: error creating SQ CQ for connect request\n");
 772                 goto errout;
 773         }
 774         newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
 775                                          rq_comp_handler,
 776                                          cq_event_handler,
 777                                          newxprt,
 778                                          newxprt->sc_max_requests,
 779                                          0);
 780         if (IS_ERR(newxprt->sc_rq_cq)) {
 781                 dprintk("svcrdma: error creating RQ CQ for connect request\n");
 782                 goto errout;
 783         }
 784
 785         memset(&qp_attr, 0, sizeof qp_attr);
 786         qp_attr.event_handler = qp_event_handler;
 787         qp_attr.qp_context = &newxprt->sc_xprt;
 788         qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
 789         qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
 790         qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
 791         qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
 792         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 793         qp_attr.qp_type = IB_QPT_RC;
 794         qp_attr.send_cq = newxprt->sc_sq_cq;
 795         qp_attr.recv_cq = newxprt->sc_rq_cq;
 796         dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
 797                 "    cm_id->device=%p, sc_pd->device=%p\n"
 798                 "    cap.max_send_wr = %d\n"
 799                 "    cap.max_recv_wr = %d\n"
 800                 "    cap.max_send_sge = %d\n"
 801                 "    cap.max_recv_sge = %d\n",
 802                 newxprt->sc_cm_id, newxprt->sc_pd,
 803                 newxprt->sc_cm_id->device, newxprt->sc_pd->device,
 804                 qp_attr.cap.max_send_wr,
 805                 qp_attr.cap.max_recv_wr,
 806                 qp_attr.cap.max_send_sge,
 807                 qp_attr.cap.max_recv_sge);
 808
 809         ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
 810         if (ret) {
 811                 /*
 812                  * XXX: This is a hack. We need a xx_request_qp interface
 813                  * that will adjust the qp_attr's with a best-effort
 814                  * number
 815                  */
 816                 qp_attr.cap.max_send_sge -= 2;
 817                 qp_attr.cap.max_recv_sge -= 2;
 818                 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd,
 819                                      &qp_attr);
 820                 if (ret) {
 821                         dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
 822                         goto errout;
 823                 }
 824                 newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
 825                 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
 826                 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
 827                 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
 828         }
 829         svc_xprt_get(&newxprt->sc_xprt);
 830         newxprt->sc_qp = newxprt->sc_cm_id->qp;
 831
 832         /* Register all of physical memory */
 833         newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
 834                                             IB_ACCESS_LOCAL_WRITE |
 835                                             IB_ACCESS_REMOTE_WRITE);
 836         if (IS_ERR(newxprt->sc_phys_mr)) {
 837                 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
 838                 goto errout;
 839         }
 840
 841         /* Post receive buffers */
 842         for (i = 0; i < newxprt->sc_max_requests; i++) {
 843                 ret = svc_rdma_post_recv(newxprt);
 844                 if (ret) {
 845                         dprintk("svcrdma: failure posting receive buffers\n");
 846                         goto errout;
 847                 }
 848         }
 849
 850         /* Swap out the handler */
 851         newxprt->sc_cm_id->event_handler = rdma_cma_handler;
 852
 853         /* Accept Connection */
 854         set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
 855         memset(&conn_param, 0, sizeof conn_param);
 856         conn_param.responder_resources = 0;
 857         conn_param.initiator_depth = newxprt->sc_ord;
 858         ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
 859         if (ret) {
 860                 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
 861                        ret);
 862                 goto errout;
 863         }
 864
 865         dprintk("svcrdma: new connection %p accepted with the following "
 866                 "attributes:\n"
 867                 "    local_ip        : %d.%d.%d.%d\n"
 868                 "    local_port      : %d\n"
 869                 "    remote_ip       : %d.%d.%d.%d\n"
 870                 "    remote_port     : %d\n"
 871                 "    max_sge         : %d\n"
 872                 "    sq_depth        : %d\n"
 873                 "    max_requests    : %d\n"
 874                 "    ord             : %d\n",
 875                 newxprt,
 876                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
 877                          route.addr.src_addr)->sin_addr.s_addr),
 878                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
 879                        route.addr.src_addr)->sin_port),
 880                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
 881                          route.addr.dst_addr)->sin_addr.s_addr),
 882                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
 883                        route.addr.dst_addr)->sin_port),
 884                 newxprt->sc_max_sge,
 885                 newxprt->sc_sq_depth,
 886                 newxprt->sc_max_requests,
 887                 newxprt->sc_ord);
 888
 889         /* Set the local and remote addresses in the transport */
 890         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
 891         svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
 892         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
 893         svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
 894
 895         ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
 896         ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
 897         return &newxprt->sc_xprt;
 898
 899  errout:
 900         dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
 901         /* Take a reference in case the DTO handler runs */
 902         svc_xprt_get(&newxprt->sc_xprt);
 903         if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) {
 904                 ib_destroy_qp(newxprt->sc_qp);
 905                 svc_xprt_put(&newxprt->sc_xprt);
 906         }
 907         rdma_destroy_id(newxprt->sc_cm_id);
 908         /* This call to put will destroy the transport */
 909         svc_xprt_put(&newxprt->sc_xprt);
 910         return NULL;
 911 }
 912
 913 /*
 914  * Post an RQ WQE to the RQ when the rqst is being released. This
 915  * effectively returns an RQ credit to the client. The rq_xprt_ctxt
 916  * will be null if the request is deferred due to an RDMA_READ or the
 917  * transport had no data ready (EAGAIN). Note that an RPC deferred in
 918  * svc_process will still return the credit, this is because the data
 919  * is copied and no longer consume a WQE/WC.
 920  */
 921 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
 922 {
 923         int err;
 924         struct svcxprt_rdma *rdma =
 925                 container_of(rqstp->rq_xprt, struct svcxprt_rdma, sc_xprt);
 926         if (rqstp->rq_xprt_ctxt) {
 927                 BUG_ON(rqstp->rq_xprt_ctxt != rdma);
 928                 err = svc_rdma_post_recv(rdma);
 929                 if (err)
 930                         dprintk("svcrdma: failed to post an RQ WQE error=%d\n",
 931                                 err);
 932         }
 933         rqstp->rq_xprt_ctxt = NULL;
 934 }
 935
 936 /*
 937  * When connected, an svc_xprt has at least three references:
 938  *
 939  * - A reference held by the QP. We still hold that here because this
 940  *   code deletes the QP and puts the reference.
 941  *
 942  * - A reference held by the cm_id between the ESTABLISHED and
 943  *   DISCONNECTED events. If the remote peer disconnected first, this
 944  *   reference could be gone.
 945  *
 946  * - A reference held by the svc_recv code that called this function
 947  *   as part of close processing.
 948  *
 949  * At a minimum two references should still be held.
 950  */
 951 static void svc_rdma_detach(struct svc_xprt *xprt)
 952 {
 953         struct svcxprt_rdma *rdma =
 954                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 955         dprintk("svc: svc_rdma_detach(%p)\n", xprt);
 956
 957         /* Disconnect and flush posted WQE */
 958         rdma_disconnect(rdma->sc_cm_id);
 959
 960         /* Destroy the QP if present (not a listener) */
 961         if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) {
 962                 ib_destroy_qp(rdma->sc_qp);
 963                 svc_xprt_put(xprt);
 964         }
 965
 966         /* Destroy the CM ID */
 967         rdma_destroy_id(rdma->sc_cm_id);
 968 }
 969
 970 static void svc_rdma_free(struct svc_xprt *xprt)
 971 {
 972         struct svcxprt_rdma *rdma = (struct svcxprt_rdma *)xprt;
 973         dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
 974         /* We should only be called from kref_put */
 975         BUG_ON(atomic_read(&xprt->xpt_ref.refcount) != 0);
 976         if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
 977                 ib_destroy_cq(rdma->sc_sq_cq);
 978
 979         if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
 980                 ib_destroy_cq(rdma->sc_rq_cq);
 981
 982         if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
 983                 ib_dereg_mr(rdma->sc_phys_mr);
 984
 985         if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
 986                 ib_dealloc_pd(rdma->sc_pd);
 987
 988         destroy_context_cache(rdma->sc_ctxt_head);
 989         kfree(rdma);
 990 }
 991
 992 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
 993 {
 994         struct svcxprt_rdma *rdma =
 995                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 996
 997         /*
 998          * If there are fewer SQ WR available than required to send a
 999          * simple response, return false.
1000          */
1001         if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
1002                 return 0;
1003
1004         /*
1005          * ...or there are already waiters on the SQ,
1006          * return false.
1007          */
1008         if (waitqueue_active(&rdma->sc_send_wait))
1009                 return 0;
1010
1011         /* Otherwise return true. */
1012         return 1;
1013 }
1014
1015 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1016 {
1017         struct ib_send_wr *bad_wr;
1018         int ret;
1019
1020         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1021                 return 0;
1022
1023         BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
1024         BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
1025                 wr->opcode);
1026         /* If the SQ is full, wait until an SQ entry is available */
1027         while (1) {
1028                 spin_lock_bh(&xprt->sc_lock);
1029                 if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
1030                         spin_unlock_bh(&xprt->sc_lock);
1031                         atomic_inc(&rdma_stat_sq_starve);
1032                         /* See if we can reap some SQ WR */
1033                         sq_cq_reap(xprt);
1034
1035                         /* Wait until SQ WR available if SQ still full */
1036                         wait_event(xprt->sc_send_wait,
1037                                    atomic_read(&xprt->sc_sq_count) <
1038                                    xprt->sc_sq_depth);
1039                         continue;
1040                 }
1041                 /* Bumped used SQ WR count and post */
1042                 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1043                 if (!ret)
1044                         atomic_inc(&xprt->sc_sq_count);
1045                 else
1046                         dprintk("svcrdma: failed to post SQ WR rc=%d, "
1047                                "sc_sq_count=%d, sc_sq_depth=%d\n",
1048                                ret, atomic_read(&xprt->sc_sq_count),
1049                                xprt->sc_sq_depth);
1050                 spin_unlock_bh(&xprt->sc_lock);
1051                 break;
1052         }
1053         return ret;
1054 }
1055
1056 int svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1057                         enum rpcrdma_errcode err)
1058 {
1059         struct ib_send_wr err_wr;
1060         struct ib_sge sge;
1061         struct page *p;
1062         struct svc_rdma_op_ctxt *ctxt;
1063         u32 *va;
1064         int length;
1065         int ret;
1066
1067         p = svc_rdma_get_page();
1068         va = page_address(p);
1069
1070         /* XDR encode error */
1071         length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1072
1073         /* Prepare SGE for local address */
1074         sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
1075                                    p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
1076         sge.lkey = xprt->sc_phys_mr->lkey;
1077         sge.length = length;
1078
1079         ctxt = svc_rdma_get_context(xprt);
1080         ctxt->count = 1;
1081         ctxt->pages[0] = p;
1082
1083         /* Prepare SEND WR */
1084         memset(&err_wr, 0, sizeof err_wr);
1085         ctxt->wr_op = IB_WR_SEND;
1086         err_wr.wr_id = (unsigned long)ctxt;
1087         err_wr.sg_list = &sge;
1088         err_wr.num_sge = 1;
1089         err_wr.opcode = IB_WR_SEND;
1090         err_wr.send_flags = IB_SEND_SIGNALED;
1091
1092         /* Post It */
1093         ret = svc_rdma_send(xprt, &err_wr);
1094         if (ret) {
1095                 dprintk("svcrdma: Error posting send = %d\n", ret);
1096                 svc_rdma_put_context(ctxt, 1);
1097         }
1098
1099         return ret;
1100 }