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                         INIT_LIST_HEAD(&ctxt->free_list);
 107                         list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
 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(list_empty(&xprt->sc_ctxt_free))) {
 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 = list_entry(xprt->sc_ctxt_free.next,
 140                                   struct svc_rdma_op_ctxt,
 141                                   free_list);
 142                 list_del_init(&ctxt->free_list);
 143                 spin_unlock_bh(&xprt->sc_ctxt_lock);
 144                 ctxt->xprt = xprt;
 145                 INIT_LIST_HEAD(&ctxt->dto_q);
 146                 ctxt->count = 0;
 147                 atomic_inc(&xprt->sc_ctxt_used);
 148                 break;
 149         }
 150         return ctxt;
 151 }
 152
 153 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
 154 {
 155         struct svcxprt_rdma *xprt;
 156         int i;
 157
 158         BUG_ON(!ctxt);
 159         xprt = ctxt->xprt;
 160         if (free_pages)
 161                 for (i = 0; i < ctxt->count; i++)
 162                         put_page(ctxt->pages[i]);
 163
 164         for (i = 0; i < ctxt->count; i++)
 165                 ib_dma_unmap_single(xprt->sc_cm_id->device,
 166                                     ctxt->sge[i].addr,
 167                                     ctxt->sge[i].length,
 168                                     ctxt->direction);
 169
 170         spin_lock_bh(&xprt->sc_ctxt_lock);
 171         list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
 172         spin_unlock_bh(&xprt->sc_ctxt_lock);
 173         atomic_dec(&xprt->sc_ctxt_used);
 174 }
 175
 176 /* Temporary NFS request map cache. Created in svc_rdma.c  */
 177 extern struct kmem_cache *svc_rdma_map_cachep;
 178
 179 /*
 180  * Temporary NFS req mappings are shared across all transport
 181  * instances. These are short lived and should be bounded by the number
 182  * of concurrent server threads * depth of the SQ.
 183  */
 184 struct svc_rdma_req_map *svc_rdma_get_req_map(void)
 185 {
 186         struct svc_rdma_req_map *map;
 187         while (1) {
 188                 map = kmem_cache_alloc(svc_rdma_map_cachep, GFP_KERNEL);
 189                 if (map)
 190                         break;
 191                 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
 192         }
 193         map->count = 0;
 194         return map;
 195 }
 196
 197 void svc_rdma_put_req_map(struct svc_rdma_req_map *map)
 198 {
 199         kmem_cache_free(svc_rdma_map_cachep, map);
 200 }
 201
 202 /* ib_cq event handler */
 203 static void cq_event_handler(struct ib_event *event, void *context)
 204 {
 205         struct svc_xprt *xprt = context;
 206         dprintk("svcrdma: received CQ event id=%d, context=%p\n",
 207                 event->event, context);
 208         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 209 }
 210
 211 /* QP event handler */
 212 static void qp_event_handler(struct ib_event *event, void *context)
 213 {
 214         struct svc_xprt *xprt = context;
 215
 216         switch (event->event) {
 217         /* These are considered benign events */
 218         case IB_EVENT_PATH_MIG:
 219         case IB_EVENT_COMM_EST:
 220         case IB_EVENT_SQ_DRAINED:
 221         case IB_EVENT_QP_LAST_WQE_REACHED:
 222                 dprintk("svcrdma: QP event %d received for QP=%p\n",
 223                         event->event, event->element.qp);
 224                 break;
 225         /* These are considered fatal events */
 226         case IB_EVENT_PATH_MIG_ERR:
 227         case IB_EVENT_QP_FATAL:
 228         case IB_EVENT_QP_REQ_ERR:
 229         case IB_EVENT_QP_ACCESS_ERR:
 230         case IB_EVENT_DEVICE_FATAL:
 231         default:
 232                 dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
 233                         "closing transport\n",
 234                         event->event, event->element.qp);
 235                 set_bit(XPT_CLOSE, &xprt->xpt_flags);
 236                 break;
 237         }
 238 }
 239
 240 /*
 241  * Data Transfer Operation Tasklet
 242  *
 243  * Walks a list of transports with I/O pending, removing entries as
 244  * they are added to the server's I/O pending list. Two bits indicate
 245  * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
 246  * spinlock that serializes access to the transport list with the RQ
 247  * and SQ interrupt handlers.
 248  */
 249 static void dto_tasklet_func(unsigned long data)
 250 {
 251         struct svcxprt_rdma *xprt;
 252         unsigned long flags;
 253
 254         spin_lock_irqsave(&dto_lock, flags);
 255         while (!list_empty(&dto_xprt_q)) {
 256                 xprt = list_entry(dto_xprt_q.next,
 257                                   struct svcxprt_rdma, sc_dto_q);
 258                 list_del_init(&xprt->sc_dto_q);
 259                 spin_unlock_irqrestore(&dto_lock, flags);
 260
 261                 rq_cq_reap(xprt);
 262                 sq_cq_reap(xprt);
 263
 264                 svc_xprt_put(&xprt->sc_xprt);
 265                 spin_lock_irqsave(&dto_lock, flags);
 266         }
 267         spin_unlock_irqrestore(&dto_lock, flags);
 268 }
 269
 270 /*
 271  * Receive Queue Completion Handler
 272  *
 273  * Since an RQ completion handler is called on interrupt context, we
 274  * need to defer the handling of the I/O to a tasklet
 275  */
 276 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
 277 {
 278         struct svcxprt_rdma *xprt = cq_context;
 279         unsigned long flags;
 280
 281         /* Guard against unconditional flush call for destroyed QP */
 282         if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
 283                 return;
 284
 285         /*
 286          * Set the bit regardless of whether or not it's on the list
 287          * because it may be on the list already due to an SQ
 288          * completion.
 289          */
 290         set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
 291
 292         /*
 293          * If this transport is not already on the DTO transport queue,
 294          * add it
 295          */
 296         spin_lock_irqsave(&dto_lock, flags);
 297         if (list_empty(&xprt->sc_dto_q)) {
 298                 svc_xprt_get(&xprt->sc_xprt);
 299                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
 300         }
 301         spin_unlock_irqrestore(&dto_lock, flags);
 302
 303         /* Tasklet does all the work to avoid irqsave locks. */
 304         tasklet_schedule(&dto_tasklet);
 305 }
 306
 307 /*
 308  * rq_cq_reap - Process the RQ CQ.
 309  *
 310  * Take all completing WC off the CQE and enqueue the associated DTO
 311  * context on the dto_q for the transport.
 312  *
 313  * Note that caller must hold a transport reference.
 314  */
 315 static void rq_cq_reap(struct svcxprt_rdma *xprt)
 316 {
 317         int ret;
 318         struct ib_wc wc;
 319         struct svc_rdma_op_ctxt *ctxt = NULL;
 320
 321         if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
 322                 return;
 323
 324         ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
 325         atomic_inc(&rdma_stat_rq_poll);
 326
 327         while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
 328                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
 329                 ctxt->wc_status = wc.status;
 330                 ctxt->byte_len = wc.byte_len;
 331                 if (wc.status != IB_WC_SUCCESS) {
 332                         /* Close the transport */
 333                         dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt);
 334                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 335                         svc_rdma_put_context(ctxt, 1);
 336                         svc_xprt_put(&xprt->sc_xprt);
 337                         continue;
 338                 }
 339                 spin_lock_bh(&xprt->sc_rq_dto_lock);
 340                 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
 341                 spin_unlock_bh(&xprt->sc_rq_dto_lock);
 342                 svc_xprt_put(&xprt->sc_xprt);
 343         }
 344
 345         if (ctxt)
 346                 atomic_inc(&rdma_stat_rq_prod);
 347
 348         set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
 349         /*
 350          * If data arrived before established event,
 351          * don't enqueue. This defers RPC I/O until the
 352          * RDMA connection is complete.
 353          */
 354         if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
 355                 svc_xprt_enqueue(&xprt->sc_xprt);
 356 }
 357
 358 /*
 359  * Send Queue Completion Handler - potentially called on interrupt context.
 360  *
 361  * Note that caller must hold a transport reference.
 362  */
 363 static void sq_cq_reap(struct svcxprt_rdma *xprt)
 364 {
 365         struct svc_rdma_op_ctxt *ctxt = NULL;
 366         struct ib_wc wc;
 367         struct ib_cq *cq = xprt->sc_sq_cq;
 368         int ret;
 369
 370
 371         if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
 372                 return;
 373
 374         ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
 375         atomic_inc(&rdma_stat_sq_poll);
 376         while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
 377                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
 378                 xprt = ctxt->xprt;
 379
 380                 if (wc.status != IB_WC_SUCCESS)
 381                         /* Close the transport */
 382                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 383
 384                 /* Decrement used SQ WR count */
 385                 atomic_dec(&xprt->sc_sq_count);
 386                 wake_up(&xprt->sc_send_wait);
 387
 388                 switch (ctxt->wr_op) {
 389                 case IB_WR_SEND:
 390                         svc_rdma_put_context(ctxt, 1);
 391                         break;
 392
 393                 case IB_WR_RDMA_WRITE:
 394                         svc_rdma_put_context(ctxt, 0);
 395                         break;
 396
 397                 case IB_WR_RDMA_READ:
 398                         if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
 399                                 struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr;
 400                                 BUG_ON(!read_hdr);
 401                                 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
 402                                 spin_lock_bh(&xprt->sc_read_complete_lock);
 403                                 list_add_tail(&read_hdr->dto_q,
 404                                               &xprt->sc_read_complete_q);
 405                                 spin_unlock_bh(&xprt->sc_read_complete_lock);
 406                                 svc_xprt_enqueue(&xprt->sc_xprt);
 407                         }
 408                         svc_rdma_put_context(ctxt, 0);
 409                         break;
 410
 411                 default:
 412                         printk(KERN_ERR "svcrdma: unexpected completion type, "
 413                                "opcode=%d, status=%d\n",
 414                                wc.opcode, wc.status);
 415                         break;
 416                 }
 417                 svc_xprt_put(&xprt->sc_xprt);
 418         }
 419
 420         if (ctxt)
 421                 atomic_inc(&rdma_stat_sq_prod);
 422 }
 423
 424 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
 425 {
 426         struct svcxprt_rdma *xprt = cq_context;
 427         unsigned long flags;
 428
 429         /* Guard against unconditional flush call for destroyed QP */
 430         if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
 431                 return;
 432
 433         /*
 434          * Set the bit regardless of whether or not it's on the list
 435          * because it may be on the list already due to an RQ
 436          * completion.
 437          */
 438         set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
 439
 440         /*
 441          * If this transport is not already on the DTO transport queue,
 442          * add it
 443          */
 444         spin_lock_irqsave(&dto_lock, flags);
 445         if (list_empty(&xprt->sc_dto_q)) {
 446                 svc_xprt_get(&xprt->sc_xprt);
 447                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
 448         }
 449         spin_unlock_irqrestore(&dto_lock, flags);
 450
 451         /* Tasklet does all the work to avoid irqsave locks. */
 452         tasklet_schedule(&dto_tasklet);
 453 }
 454
 455 static void create_context_cache(struct svcxprt_rdma *xprt,
 456                                  int ctxt_count, int ctxt_bump, int ctxt_max)
 457 {
 458         struct svc_rdma_op_ctxt *ctxt;
 459         int i;
 460
 461         xprt->sc_ctxt_max = ctxt_max;
 462         xprt->sc_ctxt_bump = ctxt_bump;
 463         xprt->sc_ctxt_cnt = 0;
 464         atomic_set(&xprt->sc_ctxt_used, 0);
 465
 466         INIT_LIST_HEAD(&xprt->sc_ctxt_free);
 467         for (i = 0; i < ctxt_count; i++) {
 468                 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
 469                 if (ctxt) {
 470                         INIT_LIST_HEAD(&ctxt->free_list);
 471                         list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
 472                         xprt->sc_ctxt_cnt++;
 473                 }
 474         }
 475 }
 476
 477 static void destroy_context_cache(struct svcxprt_rdma *xprt)
 478 {
 479         while (!list_empty(&xprt->sc_ctxt_free)) {
 480                 struct svc_rdma_op_ctxt *ctxt;
 481                 ctxt = list_entry(xprt->sc_ctxt_free.next,
 482                                   struct svc_rdma_op_ctxt,
 483                                   free_list);
 484                 list_del_init(&ctxt->free_list);
 485                 kfree(ctxt);
 486         }
 487 }
 488
 489 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
 490                                              int listener)
 491 {
 492         struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
 493
 494         if (!cma_xprt)
 495                 return NULL;
 496         svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv);
 497         INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
 498         INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
 499         INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
 500         INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
 501         init_waitqueue_head(&cma_xprt->sc_send_wait);
 502
 503         spin_lock_init(&cma_xprt->sc_lock);
 504         spin_lock_init(&cma_xprt->sc_read_complete_lock);
 505         spin_lock_init(&cma_xprt->sc_ctxt_lock);
 506         spin_lock_init(&cma_xprt->sc_rq_dto_lock);
 507
 508         cma_xprt->sc_ord = svcrdma_ord;
 509
 510         cma_xprt->sc_max_req_size = svcrdma_max_req_size;
 511         cma_xprt->sc_max_requests = svcrdma_max_requests;
 512         cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
 513         atomic_set(&cma_xprt->sc_sq_count, 0);
 514
 515         if (!listener) {
 516                 int reqs = cma_xprt->sc_max_requests;
 517                 create_context_cache(cma_xprt,
 518                                      reqs << 1, /* starting size */
 519                                      reqs,      /* bump amount */
 520                                      reqs +
 521                                      cma_xprt->sc_sq_depth +
 522                                      RPCRDMA_MAX_THREADS + 1); /* max */
 523                 if (list_empty(&cma_xprt->sc_ctxt_free)) {
 524                         kfree(cma_xprt);
 525                         return NULL;
 526                 }
 527                 clear_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
 528         } else
 529                 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
 530
 531         return cma_xprt;
 532 }
 533
 534 struct page *svc_rdma_get_page(void)
 535 {
 536         struct page *page;
 537
 538         while ((page = alloc_page(GFP_KERNEL)) == NULL) {
 539                 /* If we can't get memory, wait a bit and try again */
 540                 printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 "
 541                        "jiffies.\n");
 542                 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
 543         }
 544         return page;
 545 }
 546
 547 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
 548 {
 549         struct ib_recv_wr recv_wr, *bad_recv_wr;
 550         struct svc_rdma_op_ctxt *ctxt;
 551         struct page *page;
 552         unsigned long pa;
 553         int sge_no;
 554         int buflen;
 555         int ret;
 556
 557         ctxt = svc_rdma_get_context(xprt);
 558         buflen = 0;
 559         ctxt->direction = DMA_FROM_DEVICE;
 560         for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
 561                 BUG_ON(sge_no >= xprt->sc_max_sge);
 562                 page = svc_rdma_get_page();
 563                 ctxt->pages[sge_no] = page;
 564                 pa = ib_dma_map_page(xprt->sc_cm_id->device,
 565                                      page, 0, PAGE_SIZE,
 566                                      DMA_FROM_DEVICE);
 567                 ctxt->sge[sge_no].addr = pa;
 568                 ctxt->sge[sge_no].length = PAGE_SIZE;
 569                 ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
 570                 buflen += PAGE_SIZE;
 571         }
 572         ctxt->count = sge_no;
 573         recv_wr.next = NULL;
 574         recv_wr.sg_list = &ctxt->sge[0];
 575         recv_wr.num_sge = ctxt->count;
 576         recv_wr.wr_id = (u64)(unsigned long)ctxt;
 577
 578         svc_xprt_get(&xprt->sc_xprt);
 579         ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
 580         if (ret) {
 581                 svc_xprt_put(&xprt->sc_xprt);
 582                 svc_rdma_put_context(ctxt, 1);
 583         }
 584         return ret;
 585 }
 586
 587 /*
 588  * This function handles the CONNECT_REQUEST event on a listening
 589  * endpoint. It is passed the cma_id for the _new_ connection. The context in
 590  * this cma_id is inherited from the listening cma_id and is the svc_xprt
 591  * structure for the listening endpoint.
 592  *
 593  * This function creates a new xprt for the new connection and enqueues it on
 594  * the accept queue for the listent xprt. When the listen thread is kicked, it
 595  * will call the recvfrom method on the listen xprt which will accept the new
 596  * connection.
 597  */
 598 static void handle_connect_req(struct rdma_cm_id *new_cma_id)
 599 {
 600         struct svcxprt_rdma *listen_xprt = new_cma_id->context;
 601         struct svcxprt_rdma *newxprt;
 602         struct sockaddr *sa;
 603
 604         /* Create a new transport */
 605         newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
 606         if (!newxprt) {
 607                 dprintk("svcrdma: failed to create new transport\n");
 608                 return;
 609         }
 610         newxprt->sc_cm_id = new_cma_id;
 611         new_cma_id->context = newxprt;
 612         dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
 613                 newxprt, newxprt->sc_cm_id, listen_xprt);
 614
 615         /* Set the local and remote addresses in the transport */
 616         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
 617         svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
 618         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
 619         svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
 620
 621         /*
 622          * Enqueue the new transport on the accept queue of the listening
 623          * transport
 624          */
 625         spin_lock_bh(&listen_xprt->sc_lock);
 626         list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
 627         spin_unlock_bh(&listen_xprt->sc_lock);
 628
 629         /*
 630          * Can't use svc_xprt_received here because we are not on a
 631          * rqstp thread
 632         */
 633         set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
 634         svc_xprt_enqueue(&listen_xprt->sc_xprt);
 635 }
 636
 637 /*
 638  * Handles events generated on the listening endpoint. These events will be
 639  * either be incoming connect requests or adapter removal  events.
 640  */
 641 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
 642                                struct rdma_cm_event *event)
 643 {
 644         struct svcxprt_rdma *xprt = cma_id->context;
 645         int ret = 0;
 646
 647         switch (event->event) {
 648         case RDMA_CM_EVENT_CONNECT_REQUEST:
 649                 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
 650                         "event=%d\n", cma_id, cma_id->context, event->event);
 651                 handle_connect_req(cma_id);
 652                 break;
 653
 654         case RDMA_CM_EVENT_ESTABLISHED:
 655                 /* Accept complete */
 656                 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
 657                         "cm_id=%p\n", xprt, cma_id);
 658                 break;
 659
 660         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 661                 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
 662                         xprt, cma_id);
 663                 if (xprt)
 664                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 665                 break;
 666
 667         default:
 668                 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
 669                         "event=%d\n", cma_id, event->event);
 670                 break;
 671         }
 672
 673         return ret;
 674 }
 675
 676 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
 677                             struct rdma_cm_event *event)
 678 {
 679         struct svc_xprt *xprt = cma_id->context;
 680         struct svcxprt_rdma *rdma =
 681                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 682         switch (event->event) {
 683         case RDMA_CM_EVENT_ESTABLISHED:
 684                 /* Accept complete */
 685                 svc_xprt_get(xprt);
 686                 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
 687                         "cm_id=%p\n", xprt, cma_id);
 688                 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
 689                 svc_xprt_enqueue(xprt);
 690                 break;
 691         case RDMA_CM_EVENT_DISCONNECTED:
 692                 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
 693                         xprt, cma_id);
 694                 if (xprt) {
 695                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 696                         svc_xprt_enqueue(xprt);
 697                         svc_xprt_put(xprt);
 698                 }
 699                 break;
 700         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 701                 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
 702                         "event=%d\n", cma_id, xprt, event->event);
 703                 if (xprt) {
 704                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 705                         svc_xprt_enqueue(xprt);
 706                 }
 707                 break;
 708         default:
 709                 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
 710                         "event=%d\n", cma_id, event->event);
 711                 break;
 712         }
 713         return 0;
 714 }
 715
 716 /*
 717  * Create a listening RDMA service endpoint.
 718  */
 719 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
 720                                         struct sockaddr *sa, int salen,
 721                                         int flags)
 722 {
 723         struct rdma_cm_id *listen_id;
 724         struct svcxprt_rdma *cma_xprt;
 725         struct svc_xprt *xprt;
 726         int ret;
 727
 728         dprintk("svcrdma: Creating RDMA socket\n");
 729
 730         cma_xprt = rdma_create_xprt(serv, 1);
 731         if (!cma_xprt)
 732                 return ERR_PTR(-ENOMEM);
 733         xprt = &cma_xprt->sc_xprt;
 734
 735         listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
 736         if (IS_ERR(listen_id)) {
 737                 ret = PTR_ERR(listen_id);
 738                 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
 739                 goto err0;
 740         }
 741
 742         ret = rdma_bind_addr(listen_id, sa);
 743         if (ret) {
 744                 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
 745                 goto err1;
 746         }
 747         cma_xprt->sc_cm_id = listen_id;
 748
 749         ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
 750         if (ret) {
 751                 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
 752                 goto err1;
 753         }
 754
 755         /*
 756          * We need to use the address from the cm_id in case the
 757          * caller specified 0 for the port number.
 758          */
 759         sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
 760         svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
 761
 762         return &cma_xprt->sc_xprt;
 763
 764  err1:
 765         rdma_destroy_id(listen_id);
 766  err0:
 767         kfree(cma_xprt);
 768         return ERR_PTR(ret);
 769 }
 770
 771 /*
 772  * This is the xpo_recvfrom function for listening endpoints. Its
 773  * purpose is to accept incoming connections. The CMA callback handler
 774  * has already created a new transport and attached it to the new CMA
 775  * ID.
 776  *
 777  * There is a queue of pending connections hung on the listening
 778  * transport. This queue contains the new svc_xprt structure. This
 779  * function takes svc_xprt structures off the accept_q and completes
 780  * the connection.
 781  */
 782 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
 783 {
 784         struct svcxprt_rdma *listen_rdma;
 785         struct svcxprt_rdma *newxprt = NULL;
 786         struct rdma_conn_param conn_param;
 787         struct ib_qp_init_attr qp_attr;
 788         struct ib_device_attr devattr;
 789         int ret;
 790         int i;
 791
 792         listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
 793         clear_bit(XPT_CONN, &xprt->xpt_flags);
 794         /* Get the next entry off the accept list */
 795         spin_lock_bh(&listen_rdma->sc_lock);
 796         if (!list_empty(&listen_rdma->sc_accept_q)) {
 797                 newxprt = list_entry(listen_rdma->sc_accept_q.next,
 798                                      struct svcxprt_rdma, sc_accept_q);
 799                 list_del_init(&newxprt->sc_accept_q);
 800         }
 801         if (!list_empty(&listen_rdma->sc_accept_q))
 802                 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
 803         spin_unlock_bh(&listen_rdma->sc_lock);
 804         if (!newxprt)
 805                 return NULL;
 806
 807         dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
 808                 newxprt, newxprt->sc_cm_id);
 809
 810         ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
 811         if (ret) {
 812                 dprintk("svcrdma: could not query device attributes on "
 813                         "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
 814                 goto errout;
 815         }
 816
 817         /* Qualify the transport resource defaults with the
 818          * capabilities of this particular device */
 819         newxprt->sc_max_sge = min((size_t)devattr.max_sge,
 820                                   (size_t)RPCSVC_MAXPAGES);
 821         newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
 822                                    (size_t)svcrdma_max_requests);
 823         newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
 824
 825         newxprt->sc_ord =  min((size_t)devattr.max_qp_rd_atom,
 826                                (size_t)svcrdma_ord);
 827
 828         newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
 829         if (IS_ERR(newxprt->sc_pd)) {
 830                 dprintk("svcrdma: error creating PD for connect request\n");
 831                 goto errout;
 832         }
 833         newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
 834                                          sq_comp_handler,
 835                                          cq_event_handler,
 836                                          newxprt,
 837                                          newxprt->sc_sq_depth,
 838                                          0);
 839         if (IS_ERR(newxprt->sc_sq_cq)) {
 840                 dprintk("svcrdma: error creating SQ CQ for connect request\n");
 841                 goto errout;
 842         }
 843         newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
 844                                          rq_comp_handler,
 845                                          cq_event_handler,
 846                                          newxprt,
 847                                          newxprt->sc_max_requests,
 848                                          0);
 849         if (IS_ERR(newxprt->sc_rq_cq)) {
 850                 dprintk("svcrdma: error creating RQ CQ for connect request\n");
 851                 goto errout;
 852         }
 853
 854         memset(&qp_attr, 0, sizeof qp_attr);
 855         qp_attr.event_handler = qp_event_handler;
 856         qp_attr.qp_context = &newxprt->sc_xprt;
 857         qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
 858         qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
 859         qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
 860         qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
 861         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 862         qp_attr.qp_type = IB_QPT_RC;
 863         qp_attr.send_cq = newxprt->sc_sq_cq;
 864         qp_attr.recv_cq = newxprt->sc_rq_cq;
 865         dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
 866                 "    cm_id->device=%p, sc_pd->device=%p\n"
 867                 "    cap.max_send_wr = %d\n"
 868                 "    cap.max_recv_wr = %d\n"
 869                 "    cap.max_send_sge = %d\n"
 870                 "    cap.max_recv_sge = %d\n",
 871                 newxprt->sc_cm_id, newxprt->sc_pd,
 872                 newxprt->sc_cm_id->device, newxprt->sc_pd->device,
 873                 qp_attr.cap.max_send_wr,
 874                 qp_attr.cap.max_recv_wr,
 875                 qp_attr.cap.max_send_sge,
 876                 qp_attr.cap.max_recv_sge);
 877
 878         ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
 879         if (ret) {
 880                 /*
 881                  * XXX: This is a hack. We need a xx_request_qp interface
 882                  * that will adjust the qp_attr's with a best-effort
 883                  * number
 884                  */
 885                 qp_attr.cap.max_send_sge -= 2;
 886                 qp_attr.cap.max_recv_sge -= 2;
 887                 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd,
 888                                      &qp_attr);
 889                 if (ret) {
 890                         dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
 891                         goto errout;
 892                 }
 893                 newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
 894                 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
 895                 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
 896                 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
 897         }
 898         newxprt->sc_qp = newxprt->sc_cm_id->qp;
 899
 900         /* Register all of physical memory */
 901         newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
 902                                             IB_ACCESS_LOCAL_WRITE |
 903                                             IB_ACCESS_REMOTE_WRITE);
 904         if (IS_ERR(newxprt->sc_phys_mr)) {
 905                 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
 906                 goto errout;
 907         }
 908
 909         /* Post receive buffers */
 910         for (i = 0; i < newxprt->sc_max_requests; i++) {
 911                 ret = svc_rdma_post_recv(newxprt);
 912                 if (ret) {
 913                         dprintk("svcrdma: failure posting receive buffers\n");
 914                         goto errout;
 915                 }
 916         }
 917
 918         /* Swap out the handler */
 919         newxprt->sc_cm_id->event_handler = rdma_cma_handler;
 920
 921         /*
 922          * Arm the CQs for the SQ and RQ before accepting so we can't
 923          * miss the first message
 924          */
 925         ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
 926         ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
 927
 928         /* Accept Connection */
 929         set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
 930         memset(&conn_param, 0, sizeof conn_param);
 931         conn_param.responder_resources = 0;
 932         conn_param.initiator_depth = newxprt->sc_ord;
 933         ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
 934         if (ret) {
 935                 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
 936                        ret);
 937                 goto errout;
 938         }
 939
 940         dprintk("svcrdma: new connection %p accepted with the following "
 941                 "attributes:\n"
 942                 "    local_ip        : %d.%d.%d.%d\n"
 943                 "    local_port      : %d\n"
 944                 "    remote_ip       : %d.%d.%d.%d\n"
 945                 "    remote_port     : %d\n"
 946                 "    max_sge         : %d\n"
 947                 "    sq_depth        : %d\n"
 948                 "    max_requests    : %d\n"
 949                 "    ord             : %d\n",
 950                 newxprt,
 951                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
 952                          route.addr.src_addr)->sin_addr.s_addr),
 953                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
 954                        route.addr.src_addr)->sin_port),
 955                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
 956                          route.addr.dst_addr)->sin_addr.s_addr),
 957                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
 958                        route.addr.dst_addr)->sin_port),
 959                 newxprt->sc_max_sge,
 960                 newxprt->sc_sq_depth,
 961                 newxprt->sc_max_requests,
 962                 newxprt->sc_ord);
 963
 964         return &newxprt->sc_xprt;
 965
 966  errout:
 967         dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
 968         /* Take a reference in case the DTO handler runs */
 969         svc_xprt_get(&newxprt->sc_xprt);
 970         if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
 971                 ib_destroy_qp(newxprt->sc_qp);
 972         rdma_destroy_id(newxprt->sc_cm_id);
 973         /* This call to put will destroy the transport */
 974         svc_xprt_put(&newxprt->sc_xprt);
 975         return NULL;
 976 }
 977
 978 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
 979 {
 980 }
 981
 982 /*
 983  * When connected, an svc_xprt has at least two references:
 984  *
 985  * - A reference held by the cm_id between the ESTABLISHED and
 986  *   DISCONNECTED events. If the remote peer disconnected first, this
 987  *   reference could be gone.
 988  *
 989  * - A reference held by the svc_recv code that called this function
 990  *   as part of close processing.
 991  *
 992  * At a minimum one references should still be held.
 993  */
 994 static void svc_rdma_detach(struct svc_xprt *xprt)
 995 {
 996         struct svcxprt_rdma *rdma =
 997                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 998         dprintk("svc: svc_rdma_detach(%p)\n", xprt);
 999
1000         /* Disconnect and flush posted WQE */
1001         rdma_disconnect(rdma->sc_cm_id);
1002 }
1003
1004 static void __svc_rdma_free(struct work_struct *work)
1005 {
1006         struct svcxprt_rdma *rdma =
1007                 container_of(work, struct svcxprt_rdma, sc_work);
1008         dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
1009
1010         /* We should only be called from kref_put */
1011         BUG_ON(atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0);
1012
1013         /*
1014          * Destroy queued, but not processed read completions. Note
1015          * that this cleanup has to be done before destroying the
1016          * cm_id because the device ptr is needed to unmap the dma in
1017          * svc_rdma_put_context.
1018          */
1019         spin_lock_bh(&rdma->sc_read_complete_lock);
1020         while (!list_empty(&rdma->sc_read_complete_q)) {
1021                 struct svc_rdma_op_ctxt *ctxt;
1022                 ctxt = list_entry(rdma->sc_read_complete_q.next,
1023                                   struct svc_rdma_op_ctxt,
1024                                   dto_q);
1025                 list_del_init(&ctxt->dto_q);
1026                 svc_rdma_put_context(ctxt, 1);
1027         }
1028         spin_unlock_bh(&rdma->sc_read_complete_lock);
1029
1030         /* Destroy queued, but not processed recv completions */
1031         spin_lock_bh(&rdma->sc_rq_dto_lock);
1032         while (!list_empty(&rdma->sc_rq_dto_q)) {
1033                 struct svc_rdma_op_ctxt *ctxt;
1034                 ctxt = list_entry(rdma->sc_rq_dto_q.next,
1035                                   struct svc_rdma_op_ctxt,
1036                                   dto_q);
1037                 list_del_init(&ctxt->dto_q);
1038                 svc_rdma_put_context(ctxt, 1);
1039         }
1040         spin_unlock_bh(&rdma->sc_rq_dto_lock);
1041
1042         /* Warn if we leaked a resource or under-referenced */
1043         WARN_ON(atomic_read(&rdma->sc_ctxt_used) != 0);
1044
1045         /* Destroy the QP if present (not a listener) */
1046         if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
1047                 ib_destroy_qp(rdma->sc_qp);
1048
1049         if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
1050                 ib_destroy_cq(rdma->sc_sq_cq);
1051
1052         if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
1053                 ib_destroy_cq(rdma->sc_rq_cq);
1054
1055         if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
1056                 ib_dereg_mr(rdma->sc_phys_mr);
1057
1058         if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
1059                 ib_dealloc_pd(rdma->sc_pd);
1060
1061         /* Destroy the CM ID */
1062         rdma_destroy_id(rdma->sc_cm_id);
1063
1064         destroy_context_cache(rdma);
1065         kfree(rdma);
1066 }
1067
1068 static void svc_rdma_free(struct svc_xprt *xprt)
1069 {
1070         struct svcxprt_rdma *rdma =
1071                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1072         INIT_WORK(&rdma->sc_work, __svc_rdma_free);
1073         schedule_work(&rdma->sc_work);
1074 }
1075
1076 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
1077 {
1078         struct svcxprt_rdma *rdma =
1079                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1080
1081         /*
1082          * If there are fewer SQ WR available than required to send a
1083          * simple response, return false.
1084          */
1085         if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
1086                 return 0;
1087
1088         /*
1089          * ...or there are already waiters on the SQ,
1090          * return false.
1091          */
1092         if (waitqueue_active(&rdma->sc_send_wait))
1093                 return 0;
1094
1095         /* Otherwise return true. */
1096         return 1;
1097 }
1098
1099 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1100 {
1101         struct ib_send_wr *bad_wr;
1102         int ret;
1103
1104         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1105                 return -ENOTCONN;
1106
1107         BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
1108         BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
1109                 wr->opcode);
1110         /* If the SQ is full, wait until an SQ entry is available */
1111         while (1) {
1112                 spin_lock_bh(&xprt->sc_lock);
1113                 if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
1114                         spin_unlock_bh(&xprt->sc_lock);
1115                         atomic_inc(&rdma_stat_sq_starve);
1116
1117                         /* See if we can opportunistically reap SQ WR to make room */
1118                         sq_cq_reap(xprt);
1119
1120                         /* Wait until SQ WR available if SQ still full */
1121                         wait_event(xprt->sc_send_wait,
1122                                    atomic_read(&xprt->sc_sq_count) <
1123                                    xprt->sc_sq_depth);
1124                         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1125                                 return 0;
1126                         continue;
1127                 }
1128                 /* Bumped used SQ WR count and post */
1129                 svc_xprt_get(&xprt->sc_xprt);
1130                 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1131                 if (!ret)
1132                         atomic_inc(&xprt->sc_sq_count);
1133                 else {
1134                         svc_xprt_put(&xprt->sc_xprt);
1135                         dprintk("svcrdma: failed to post SQ WR rc=%d, "
1136                                "sc_sq_count=%d, sc_sq_depth=%d\n",
1137                                ret, atomic_read(&xprt->sc_sq_count),
1138                                xprt->sc_sq_depth);
1139                 }
1140                 spin_unlock_bh(&xprt->sc_lock);
1141                 break;
1142         }
1143         return ret;
1144 }
1145
1146 void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1147                          enum rpcrdma_errcode err)
1148 {
1149         struct ib_send_wr err_wr;
1150         struct ib_sge sge;
1151         struct page *p;
1152         struct svc_rdma_op_ctxt *ctxt;
1153         u32 *va;
1154         int length;
1155         int ret;
1156
1157         p = svc_rdma_get_page();
1158         va = page_address(p);
1159
1160         /* XDR encode error */
1161         length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1162
1163         /* Prepare SGE for local address */
1164         sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
1165                                    p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
1166         sge.lkey = xprt->sc_phys_mr->lkey;
1167         sge.length = length;
1168
1169         ctxt = svc_rdma_get_context(xprt);
1170         ctxt->count = 1;
1171         ctxt->pages[0] = p;
1172
1173         /* Prepare SEND WR */
1174         memset(&err_wr, 0, sizeof err_wr);
1175         ctxt->wr_op = IB_WR_SEND;
1176         err_wr.wr_id = (unsigned long)ctxt;
1177         err_wr.sg_list = &sge;
1178         err_wr.num_sge = 1;
1179         err_wr.opcode = IB_WR_SEND;
1180         err_wr.send_flags = IB_SEND_SIGNALED;
1181
1182         /* Post It */
1183         ret = svc_rdma_send(xprt, &err_wr);
1184         if (ret) {
1185                 dprintk("svcrdma: Error %d posting send for protocol error\n",
1186                         ret);
1187                 svc_rdma_put_context(ctxt, 1);
1188         }
1189 }