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                 rq_cq_reap(xprt);
 232                 sq_cq_reap(xprt);
 233
 234                 svc_xprt_put(&xprt->sc_xprt);
 235                 spin_lock_irqsave(&dto_lock, flags);
 236         }
 237         spin_unlock_irqrestore(&dto_lock, flags);
 238 }
 239
 240 /*
 241  * Receive Queue Completion Handler
 242  *
 243  * Since an RQ completion handler is called on interrupt context, we
 244  * need to defer the handling of the I/O to a tasklet
 245  */
 246 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
 247 {
 248         struct svcxprt_rdma *xprt = cq_context;
 249         unsigned long flags;
 250
 251         /*
 252          * Set the bit regardless of whether or not it's on the list
 253          * because it may be on the list already due to an SQ
 254          * completion.
 255         */
 256         set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
 257
 258         /*
 259          * If this transport is not already on the DTO transport queue,
 260          * add it
 261          */
 262         spin_lock_irqsave(&dto_lock, flags);
 263         if (list_empty(&xprt->sc_dto_q)) {
 264                 svc_xprt_get(&xprt->sc_xprt);
 265                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
 266         }
 267         spin_unlock_irqrestore(&dto_lock, flags);
 268
 269         /* Tasklet does all the work to avoid irqsave locks. */
 270         tasklet_schedule(&dto_tasklet);
 271 }
 272
 273 /*
 274  * rq_cq_reap - Process the RQ CQ.
 275  *
 276  * Take all completing WC off the CQE and enqueue the associated DTO
 277  * context on the dto_q for the transport.
 278  */
 279 static void rq_cq_reap(struct svcxprt_rdma *xprt)
 280 {
 281         int ret;
 282         struct ib_wc wc;
 283         struct svc_rdma_op_ctxt *ctxt = NULL;
 284
 285         if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
 286                 return;
 287
 288         ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
 289         atomic_inc(&rdma_stat_rq_poll);
 290
 291         spin_lock_bh(&xprt->sc_rq_dto_lock);
 292         while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
 293                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
 294                 ctxt->wc_status = wc.status;
 295                 ctxt->byte_len = wc.byte_len;
 296                 if (wc.status != IB_WC_SUCCESS) {
 297                         /* Close the transport */
 298                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 299                         svc_rdma_put_context(ctxt, 1);
 300                         continue;
 301                 }
 302                 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
 303         }
 304         spin_unlock_bh(&xprt->sc_rq_dto_lock);
 305
 306         if (ctxt)
 307                 atomic_inc(&rdma_stat_rq_prod);
 308
 309         set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
 310         /*
 311          * If data arrived before established event,
 312          * don't enqueue. This defers RPC I/O until the
 313          * RDMA connection is complete.
 314          */
 315         if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
 316                 svc_xprt_enqueue(&xprt->sc_xprt);
 317 }
 318
 319 /*
 320  * Send Queue Completion Handler - potentially called on interrupt context.
 321  */
 322 static void sq_cq_reap(struct svcxprt_rdma *xprt)
 323 {
 324         struct svc_rdma_op_ctxt *ctxt = NULL;
 325         struct ib_wc wc;
 326         struct ib_cq *cq = xprt->sc_sq_cq;
 327         int ret;
 328
 329
 330         if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
 331                 return;
 332
 333         ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
 334         atomic_inc(&rdma_stat_sq_poll);
 335         while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
 336                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
 337                 xprt = ctxt->xprt;
 338
 339                 if (wc.status != IB_WC_SUCCESS)
 340                         /* Close the transport */
 341                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 342
 343                 /* Decrement used SQ WR count */
 344                 atomic_dec(&xprt->sc_sq_count);
 345                 wake_up(&xprt->sc_send_wait);
 346
 347                 switch (ctxt->wr_op) {
 348                 case IB_WR_SEND:
 349                 case IB_WR_RDMA_WRITE:
 350                         svc_rdma_put_context(ctxt, 1);
 351                         break;
 352
 353                 case IB_WR_RDMA_READ:
 354                         if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
 355                                 struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr;
 356                                 BUG_ON(!read_hdr);
 357                                 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
 358                                 spin_lock_bh(&xprt->sc_read_complete_lock);
 359                                 list_add_tail(&read_hdr->dto_q,
 360                                               &xprt->sc_read_complete_q);
 361                                 spin_unlock_bh(&xprt->sc_read_complete_lock);
 362                                 svc_xprt_enqueue(&xprt->sc_xprt);
 363                         }
 364                         svc_rdma_put_context(ctxt, 0);
 365                         break;
 366
 367                 default:
 368                         printk(KERN_ERR "svcrdma: unexpected completion type, "
 369                                "opcode=%d, status=%d\n",
 370                                wc.opcode, wc.status);
 371                         break;
 372                 }
 373         }
 374
 375         if (ctxt)
 376                 atomic_inc(&rdma_stat_sq_prod);
 377 }
 378
 379 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
 380 {
 381         struct svcxprt_rdma *xprt = cq_context;
 382         unsigned long flags;
 383
 384         /*
 385          * Set the bit regardless of whether or not it's on the list
 386          * because it may be on the list already due to an RQ
 387          * completion.
 388         */
 389         set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
 390
 391         /*
 392          * If this transport is not already on the DTO transport queue,
 393          * add it
 394          */
 395         spin_lock_irqsave(&dto_lock, flags);
 396         if (list_empty(&xprt->sc_dto_q)) {
 397                 svc_xprt_get(&xprt->sc_xprt);
 398                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
 399         }
 400         spin_unlock_irqrestore(&dto_lock, flags);
 401
 402         /* Tasklet does all the work to avoid irqsave locks. */
 403         tasklet_schedule(&dto_tasklet);
 404 }
 405
 406 static void create_context_cache(struct svcxprt_rdma *xprt,
 407                                  int ctxt_count, int ctxt_bump, int ctxt_max)
 408 {
 409         struct svc_rdma_op_ctxt *ctxt;
 410         int i;
 411
 412         xprt->sc_ctxt_max = ctxt_max;
 413         xprt->sc_ctxt_bump = ctxt_bump;
 414         xprt->sc_ctxt_cnt = 0;
 415         xprt->sc_ctxt_head = NULL;
 416         for (i = 0; i < ctxt_count; i++) {
 417                 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
 418                 if (ctxt) {
 419                         ctxt->next = xprt->sc_ctxt_head;
 420                         xprt->sc_ctxt_head = ctxt;
 421                         xprt->sc_ctxt_cnt++;
 422                 }
 423         }
 424 }
 425
 426 static void destroy_context_cache(struct svc_rdma_op_ctxt *ctxt)
 427 {
 428         struct svc_rdma_op_ctxt *next;
 429         if (!ctxt)
 430                 return;
 431
 432         do {
 433                 next = ctxt->next;
 434                 kfree(ctxt);
 435                 ctxt = next;
 436         } while (next);
 437 }
 438
 439 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
 440                                              int listener)
 441 {
 442         struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
 443
 444         if (!cma_xprt)
 445                 return NULL;
 446         svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv);
 447         INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
 448         INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
 449         INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
 450         INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
 451         init_waitqueue_head(&cma_xprt->sc_send_wait);
 452
 453         spin_lock_init(&cma_xprt->sc_lock);
 454         spin_lock_init(&cma_xprt->sc_read_complete_lock);
 455         spin_lock_init(&cma_xprt->sc_ctxt_lock);
 456         spin_lock_init(&cma_xprt->sc_rq_dto_lock);
 457
 458         cma_xprt->sc_ord = svcrdma_ord;
 459
 460         cma_xprt->sc_max_req_size = svcrdma_max_req_size;
 461         cma_xprt->sc_max_requests = svcrdma_max_requests;
 462         cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
 463         atomic_set(&cma_xprt->sc_sq_count, 0);
 464
 465         if (!listener) {
 466                 int reqs = cma_xprt->sc_max_requests;
 467                 create_context_cache(cma_xprt,
 468                                      reqs << 1, /* starting size */
 469                                      reqs,      /* bump amount */
 470                                      reqs +
 471                                      cma_xprt->sc_sq_depth +
 472                                      RPCRDMA_MAX_THREADS + 1); /* max */
 473                 if (!cma_xprt->sc_ctxt_head) {
 474                         kfree(cma_xprt);
 475                         return NULL;
 476                 }
 477                 clear_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
 478         } else
 479                 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
 480
 481         return cma_xprt;
 482 }
 483
 484 struct page *svc_rdma_get_page(void)
 485 {
 486         struct page *page;
 487
 488         while ((page = alloc_page(GFP_KERNEL)) == NULL) {
 489                 /* If we can't get memory, wait a bit and try again */
 490                 printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 "
 491                        "jiffies.\n");
 492                 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
 493         }
 494         return page;
 495 }
 496
 497 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
 498 {
 499         struct ib_recv_wr recv_wr, *bad_recv_wr;
 500         struct svc_rdma_op_ctxt *ctxt;
 501         struct page *page;
 502         unsigned long pa;
 503         int sge_no;
 504         int buflen;
 505         int ret;
 506
 507         ctxt = svc_rdma_get_context(xprt);
 508         buflen = 0;
 509         ctxt->direction = DMA_FROM_DEVICE;
 510         for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
 511                 BUG_ON(sge_no >= xprt->sc_max_sge);
 512                 page = svc_rdma_get_page();
 513                 ctxt->pages[sge_no] = page;
 514                 pa = ib_dma_map_page(xprt->sc_cm_id->device,
 515                                      page, 0, PAGE_SIZE,
 516                                      DMA_FROM_DEVICE);
 517                 ctxt->sge[sge_no].addr = pa;
 518                 ctxt->sge[sge_no].length = PAGE_SIZE;
 519                 ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
 520                 buflen += PAGE_SIZE;
 521         }
 522         ctxt->count = sge_no;
 523         recv_wr.next = NULL;
 524         recv_wr.sg_list = &ctxt->sge[0];
 525         recv_wr.num_sge = ctxt->count;
 526         recv_wr.wr_id = (u64)(unsigned long)ctxt;
 527
 528         ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
 529         if (ret)
 530                 svc_rdma_put_context(ctxt, 1);
 531         return ret;
 532 }
 533
 534 /*
 535  * This function handles the CONNECT_REQUEST event on a listening
 536  * endpoint. It is passed the cma_id for the _new_ connection. The context in
 537  * this cma_id is inherited from the listening cma_id and is the svc_xprt
 538  * structure for the listening endpoint.
 539  *
 540  * This function creates a new xprt for the new connection and enqueues it on
 541  * the accept queue for the listent xprt. When the listen thread is kicked, it
 542  * will call the recvfrom method on the listen xprt which will accept the new
 543  * connection.
 544  */
 545 static void handle_connect_req(struct rdma_cm_id *new_cma_id)
 546 {
 547         struct svcxprt_rdma *listen_xprt = new_cma_id->context;
 548         struct svcxprt_rdma *newxprt;
 549
 550         /* Create a new transport */
 551         newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
 552         if (!newxprt) {
 553                 dprintk("svcrdma: failed to create new transport\n");
 554                 return;
 555         }
 556         newxprt->sc_cm_id = new_cma_id;
 557         new_cma_id->context = newxprt;
 558         dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
 559                 newxprt, newxprt->sc_cm_id, listen_xprt);
 560
 561         /*
 562          * Enqueue the new transport on the accept queue of the listening
 563          * transport
 564          */
 565         spin_lock_bh(&listen_xprt->sc_lock);
 566         list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
 567         spin_unlock_bh(&listen_xprt->sc_lock);
 568
 569         /*
 570          * Can't use svc_xprt_received here because we are not on a
 571          * rqstp thread
 572         */
 573         set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
 574         svc_xprt_enqueue(&listen_xprt->sc_xprt);
 575 }
 576
 577 /*
 578  * Handles events generated on the listening endpoint. These events will be
 579  * either be incoming connect requests or adapter removal  events.
 580  */
 581 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
 582                                struct rdma_cm_event *event)
 583 {
 584         struct svcxprt_rdma *xprt = cma_id->context;
 585         int ret = 0;
 586
 587         switch (event->event) {
 588         case RDMA_CM_EVENT_CONNECT_REQUEST:
 589                 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
 590                         "event=%d\n", cma_id, cma_id->context, event->event);
 591                 handle_connect_req(cma_id);
 592                 break;
 593
 594         case RDMA_CM_EVENT_ESTABLISHED:
 595                 /* Accept complete */
 596                 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
 597                         "cm_id=%p\n", xprt, cma_id);
 598                 break;
 599
 600         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 601                 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
 602                         xprt, cma_id);
 603                 if (xprt)
 604                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 605                 break;
 606
 607         default:
 608                 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
 609                         "event=%d\n", cma_id, event->event);
 610                 break;
 611         }
 612
 613         return ret;
 614 }
 615
 616 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
 617                             struct rdma_cm_event *event)
 618 {
 619         struct svc_xprt *xprt = cma_id->context;
 620         struct svcxprt_rdma *rdma =
 621                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 622         switch (event->event) {
 623         case RDMA_CM_EVENT_ESTABLISHED:
 624                 /* Accept complete */
 625                 svc_xprt_get(xprt);
 626                 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
 627                         "cm_id=%p\n", xprt, cma_id);
 628                 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
 629                 svc_xprt_enqueue(xprt);
 630                 break;
 631         case RDMA_CM_EVENT_DISCONNECTED:
 632                 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
 633                         xprt, cma_id);
 634                 if (xprt) {
 635                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 636                         svc_xprt_enqueue(xprt);
 637                         svc_xprt_put(xprt);
 638                 }
 639                 break;
 640         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 641                 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
 642                         "event=%d\n", cma_id, xprt, event->event);
 643                 if (xprt) {
 644                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 645                         svc_xprt_enqueue(xprt);
 646                 }
 647                 break;
 648         default:
 649                 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
 650                         "event=%d\n", cma_id, event->event);
 651                 break;
 652         }
 653         return 0;
 654 }
 655
 656 /*
 657  * Create a listening RDMA service endpoint.
 658  */
 659 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
 660                                         struct sockaddr *sa, int salen,
 661                                         int flags)
 662 {
 663         struct rdma_cm_id *listen_id;
 664         struct svcxprt_rdma *cma_xprt;
 665         struct svc_xprt *xprt;
 666         int ret;
 667
 668         dprintk("svcrdma: Creating RDMA socket\n");
 669
 670         cma_xprt = rdma_create_xprt(serv, 1);
 671         if (!cma_xprt)
 672                 return ERR_PTR(-ENOMEM);
 673         xprt = &cma_xprt->sc_xprt;
 674
 675         listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
 676         if (IS_ERR(listen_id)) {
 677                 ret = PTR_ERR(listen_id);
 678                 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
 679                 goto err0;
 680         }
 681
 682         ret = rdma_bind_addr(listen_id, sa);
 683         if (ret) {
 684                 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
 685                 goto err1;
 686         }
 687         cma_xprt->sc_cm_id = listen_id;
 688
 689         ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
 690         if (ret) {
 691                 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
 692                 goto err1;
 693         }
 694
 695         /*
 696          * We need to use the address from the cm_id in case the
 697          * caller specified 0 for the port number.
 698          */
 699         sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
 700         svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
 701
 702         return &cma_xprt->sc_xprt;
 703
 704  err1:
 705         rdma_destroy_id(listen_id);
 706  err0:
 707         kfree(cma_xprt);
 708         return ERR_PTR(ret);
 709 }
 710
 711 /*
 712  * This is the xpo_recvfrom function for listening endpoints. Its
 713  * purpose is to accept incoming connections. The CMA callback handler
 714  * has already created a new transport and attached it to the new CMA
 715  * ID.
 716  *
 717  * There is a queue of pending connections hung on the listening
 718  * transport. This queue contains the new svc_xprt structure. This
 719  * function takes svc_xprt structures off the accept_q and completes
 720  * the connection.
 721  */
 722 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
 723 {
 724         struct svcxprt_rdma *listen_rdma;
 725         struct svcxprt_rdma *newxprt = NULL;
 726         struct rdma_conn_param conn_param;
 727         struct ib_qp_init_attr qp_attr;
 728         struct ib_device_attr devattr;
 729         struct sockaddr *sa;
 730         int ret;
 731         int i;
 732
 733         listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
 734         clear_bit(XPT_CONN, &xprt->xpt_flags);
 735         /* Get the next entry off the accept list */
 736         spin_lock_bh(&listen_rdma->sc_lock);
 737         if (!list_empty(&listen_rdma->sc_accept_q)) {
 738                 newxprt = list_entry(listen_rdma->sc_accept_q.next,
 739                                      struct svcxprt_rdma, sc_accept_q);
 740                 list_del_init(&newxprt->sc_accept_q);
 741         }
 742         if (!list_empty(&listen_rdma->sc_accept_q))
 743                 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
 744         spin_unlock_bh(&listen_rdma->sc_lock);
 745         if (!newxprt)
 746                 return NULL;
 747
 748         dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
 749                 newxprt, newxprt->sc_cm_id);
 750
 751         ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
 752         if (ret) {
 753                 dprintk("svcrdma: could not query device attributes on "
 754                         "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
 755                 goto errout;
 756         }
 757
 758         /* Qualify the transport resource defaults with the
 759          * capabilities of this particular device */
 760         newxprt->sc_max_sge = min((size_t)devattr.max_sge,
 761                                   (size_t)RPCSVC_MAXPAGES);
 762         newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
 763                                    (size_t)svcrdma_max_requests);
 764         newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
 765
 766         newxprt->sc_ord =  min((size_t)devattr.max_qp_rd_atom,
 767                                (size_t)svcrdma_ord);
 768
 769         newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
 770         if (IS_ERR(newxprt->sc_pd)) {
 771                 dprintk("svcrdma: error creating PD for connect request\n");
 772                 goto errout;
 773         }
 774         newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
 775                                          sq_comp_handler,
 776                                          cq_event_handler,
 777                                          newxprt,
 778                                          newxprt->sc_sq_depth,
 779                                          0);
 780         if (IS_ERR(newxprt->sc_sq_cq)) {
 781                 dprintk("svcrdma: error creating SQ CQ for connect request\n");
 782                 goto errout;
 783         }
 784         newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
 785                                          rq_comp_handler,
 786                                          cq_event_handler,
 787                                          newxprt,
 788                                          newxprt->sc_max_requests,
 789                                          0);
 790         if (IS_ERR(newxprt->sc_rq_cq)) {
 791                 dprintk("svcrdma: error creating RQ CQ for connect request\n");
 792                 goto errout;
 793         }
 794
 795         memset(&qp_attr, 0, sizeof qp_attr);
 796         qp_attr.event_handler = qp_event_handler;
 797         qp_attr.qp_context = &newxprt->sc_xprt;
 798         qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
 799         qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
 800         qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
 801         qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
 802         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 803         qp_attr.qp_type = IB_QPT_RC;
 804         qp_attr.send_cq = newxprt->sc_sq_cq;
 805         qp_attr.recv_cq = newxprt->sc_rq_cq;
 806         dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
 807                 "    cm_id->device=%p, sc_pd->device=%p\n"
 808                 "    cap.max_send_wr = %d\n"
 809                 "    cap.max_recv_wr = %d\n"
 810                 "    cap.max_send_sge = %d\n"
 811                 "    cap.max_recv_sge = %d\n",
 812                 newxprt->sc_cm_id, newxprt->sc_pd,
 813                 newxprt->sc_cm_id->device, newxprt->sc_pd->device,
 814                 qp_attr.cap.max_send_wr,
 815                 qp_attr.cap.max_recv_wr,
 816                 qp_attr.cap.max_send_sge,
 817                 qp_attr.cap.max_recv_sge);
 818
 819         ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
 820         if (ret) {
 821                 /*
 822                  * XXX: This is a hack. We need a xx_request_qp interface
 823                  * that will adjust the qp_attr's with a best-effort
 824                  * number
 825                  */
 826                 qp_attr.cap.max_send_sge -= 2;
 827                 qp_attr.cap.max_recv_sge -= 2;
 828                 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd,
 829                                      &qp_attr);
 830                 if (ret) {
 831                         dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
 832                         goto errout;
 833                 }
 834                 newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
 835                 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
 836                 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
 837                 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
 838         }
 839         svc_xprt_get(&newxprt->sc_xprt);
 840         newxprt->sc_qp = newxprt->sc_cm_id->qp;
 841
 842         /* Register all of physical memory */
 843         newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
 844                                             IB_ACCESS_LOCAL_WRITE |
 845                                             IB_ACCESS_REMOTE_WRITE);
 846         if (IS_ERR(newxprt->sc_phys_mr)) {
 847                 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
 848                 goto errout;
 849         }
 850
 851         /* Post receive buffers */
 852         for (i = 0; i < newxprt->sc_max_requests; i++) {
 853                 ret = svc_rdma_post_recv(newxprt);
 854                 if (ret) {
 855                         dprintk("svcrdma: failure posting receive buffers\n");
 856                         goto errout;
 857                 }
 858         }
 859
 860         /* Swap out the handler */
 861         newxprt->sc_cm_id->event_handler = rdma_cma_handler;
 862
 863         /* Accept Connection */
 864         set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
 865         memset(&conn_param, 0, sizeof conn_param);
 866         conn_param.responder_resources = 0;
 867         conn_param.initiator_depth = newxprt->sc_ord;
 868         ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
 869         if (ret) {
 870                 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
 871                        ret);
 872                 goto errout;
 873         }
 874
 875         dprintk("svcrdma: new connection %p accepted with the following "
 876                 "attributes:\n"
 877                 "    local_ip        : %d.%d.%d.%d\n"
 878                 "    local_port      : %d\n"
 879                 "    remote_ip       : %d.%d.%d.%d\n"
 880                 "    remote_port     : %d\n"
 881                 "    max_sge         : %d\n"
 882                 "    sq_depth        : %d\n"
 883                 "    max_requests    : %d\n"
 884                 "    ord             : %d\n",
 885                 newxprt,
 886                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
 887                          route.addr.src_addr)->sin_addr.s_addr),
 888                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
 889                        route.addr.src_addr)->sin_port),
 890                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
 891                          route.addr.dst_addr)->sin_addr.s_addr),
 892                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
 893                        route.addr.dst_addr)->sin_port),
 894                 newxprt->sc_max_sge,
 895                 newxprt->sc_sq_depth,
 896                 newxprt->sc_max_requests,
 897                 newxprt->sc_ord);
 898
 899         /* Set the local and remote addresses in the transport */
 900         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
 901         svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
 902         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
 903         svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
 904
 905         ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
 906         ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
 907         return &newxprt->sc_xprt;
 908
 909  errout:
 910         dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
 911         /* Take a reference in case the DTO handler runs */
 912         svc_xprt_get(&newxprt->sc_xprt);
 913         if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp)) {
 914                 ib_destroy_qp(newxprt->sc_qp);
 915                 svc_xprt_put(&newxprt->sc_xprt);
 916         }
 917         rdma_destroy_id(newxprt->sc_cm_id);
 918         /* This call to put will destroy the transport */
 919         svc_xprt_put(&newxprt->sc_xprt);
 920         return NULL;
 921 }
 922
 923 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
 924 {
 925 }
 926
 927 /*
 928  * When connected, an svc_xprt has at least three references:
 929  *
 930  * - A reference held by the QP. We still hold that here because this
 931  *   code deletes the QP and puts the reference.
 932  *
 933  * - A reference held by the cm_id between the ESTABLISHED and
 934  *   DISCONNECTED events. If the remote peer disconnected first, this
 935  *   reference could be gone.
 936  *
 937  * - A reference held by the svc_recv code that called this function
 938  *   as part of close processing.
 939  *
 940  * At a minimum two references should still be held.
 941  */
 942 static void svc_rdma_detach(struct svc_xprt *xprt)
 943 {
 944         struct svcxprt_rdma *rdma =
 945                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 946         dprintk("svc: svc_rdma_detach(%p)\n", xprt);
 947
 948         /* Disconnect and flush posted WQE */
 949         rdma_disconnect(rdma->sc_cm_id);
 950
 951         /* Destroy the QP if present (not a listener) */
 952         if (rdma->sc_qp && !IS_ERR(rdma->sc_qp)) {
 953                 ib_destroy_qp(rdma->sc_qp);
 954                 svc_xprt_put(xprt);
 955         }
 956
 957         /* Destroy the CM ID */
 958         rdma_destroy_id(rdma->sc_cm_id);
 959 }
 960
 961 static void svc_rdma_free(struct svc_xprt *xprt)
 962 {
 963         struct svcxprt_rdma *rdma = (struct svcxprt_rdma *)xprt;
 964         dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
 965         /* We should only be called from kref_put */
 966         BUG_ON(atomic_read(&xprt->xpt_ref.refcount) != 0);
 967         if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
 968                 ib_destroy_cq(rdma->sc_sq_cq);
 969
 970         if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
 971                 ib_destroy_cq(rdma->sc_rq_cq);
 972
 973         if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
 974                 ib_dereg_mr(rdma->sc_phys_mr);
 975
 976         if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
 977                 ib_dealloc_pd(rdma->sc_pd);
 978
 979         destroy_context_cache(rdma->sc_ctxt_head);
 980         kfree(rdma);
 981 }
 982
 983 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
 984 {
 985         struct svcxprt_rdma *rdma =
 986                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 987
 988         /*
 989          * If there are fewer SQ WR available than required to send a
 990          * simple response, return false.
 991          */
 992         if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
 993                 return 0;
 994
 995         /*
 996          * ...or there are already waiters on the SQ,
 997          * return false.
 998          */
 999         if (waitqueue_active(&rdma->sc_send_wait))
1000                 return 0;
1001
1002         /* Otherwise return true. */
1003         return 1;
1004 }
1005
1006 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1007 {
1008         struct ib_send_wr *bad_wr;
1009         int ret;
1010
1011         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1012                 return -ENOTCONN;
1013
1014         BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
1015         BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
1016                 wr->opcode);
1017         /* If the SQ is full, wait until an SQ entry is available */
1018         while (1) {
1019                 spin_lock_bh(&xprt->sc_lock);
1020                 if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
1021                         spin_unlock_bh(&xprt->sc_lock);
1022                         atomic_inc(&rdma_stat_sq_starve);
1023
1024                         /* See if we can opportunistically reap SQ WR to make room */
1025                         sq_cq_reap(xprt);
1026
1027                         /* Wait until SQ WR available if SQ still full */
1028                         wait_event(xprt->sc_send_wait,
1029                                    atomic_read(&xprt->sc_sq_count) <
1030                                    xprt->sc_sq_depth);
1031                         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1032                                 return 0;
1033                         continue;
1034                 }
1035                 /* Bumped used SQ WR count and post */
1036                 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1037                 if (!ret)
1038                         atomic_inc(&xprt->sc_sq_count);
1039                 else
1040                         dprintk("svcrdma: failed to post SQ WR rc=%d, "
1041                                "sc_sq_count=%d, sc_sq_depth=%d\n",
1042                                ret, atomic_read(&xprt->sc_sq_count),
1043                                xprt->sc_sq_depth);
1044                 spin_unlock_bh(&xprt->sc_lock);
1045                 break;
1046         }
1047         return ret;
1048 }
1049
1050 int svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1051                         enum rpcrdma_errcode err)
1052 {
1053         struct ib_send_wr err_wr;
1054         struct ib_sge sge;
1055         struct page *p;
1056         struct svc_rdma_op_ctxt *ctxt;
1057         u32 *va;
1058         int length;
1059         int ret;
1060
1061         p = svc_rdma_get_page();
1062         va = page_address(p);
1063
1064         /* XDR encode error */
1065         length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1066
1067         /* Prepare SGE for local address */
1068         sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
1069                                    p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
1070         sge.lkey = xprt->sc_phys_mr->lkey;
1071         sge.length = length;
1072
1073         ctxt = svc_rdma_get_context(xprt);
1074         ctxt->count = 1;
1075         ctxt->pages[0] = p;
1076
1077         /* Prepare SEND WR */
1078         memset(&err_wr, 0, sizeof err_wr);
1079         ctxt->wr_op = IB_WR_SEND;
1080         err_wr.wr_id = (unsigned long)ctxt;
1081         err_wr.sg_list = &sge;
1082         err_wr.num_sge = 1;
1083         err_wr.opcode = IB_WR_SEND;
1084         err_wr.send_flags = IB_SEND_SIGNALED;
1085
1086         /* Post It */
1087         ret = svc_rdma_send(xprt, &err_wr);
1088         if (ret) {
1089                 dprintk("svcrdma: Error posting send = %d\n", ret);
1090                 svc_rdma_put_context(ctxt, 1);
1091         }
1092
1093         return ret;
1094 }