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 static void svc_rdma_unmap_dma(struct svc_rdma_op_ctxt *ctxt)
 154 {
 155         struct svcxprt_rdma *xprt = ctxt->xprt;
 156         int i;
 157         for (i = 0; i < ctxt->count && ctxt->sge[i].length; i++) {
 158                 ib_dma_unmap_single(xprt->sc_cm_id->device,
 159                                     ctxt->sge[i].addr,
 160                                     ctxt->sge[i].length,
 161                                     ctxt->direction);
 162         }
 163 }
 164
 165 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
 166 {
 167         struct svcxprt_rdma *xprt;
 168         int i;
 169
 170         BUG_ON(!ctxt);
 171         xprt = ctxt->xprt;
 172         if (free_pages)
 173                 for (i = 0; i < ctxt->count; i++)
 174                         put_page(ctxt->pages[i]);
 175
 176         spin_lock_bh(&xprt->sc_ctxt_lock);
 177         list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
 178         spin_unlock_bh(&xprt->sc_ctxt_lock);
 179         atomic_dec(&xprt->sc_ctxt_used);
 180 }
 181
 182 /* Temporary NFS request map cache. Created in svc_rdma.c  */
 183 extern struct kmem_cache *svc_rdma_map_cachep;
 184
 185 /*
 186  * Temporary NFS req mappings are shared across all transport
 187  * instances. These are short lived and should be bounded by the number
 188  * of concurrent server threads * depth of the SQ.
 189  */
 190 struct svc_rdma_req_map *svc_rdma_get_req_map(void)
 191 {
 192         struct svc_rdma_req_map *map;
 193         while (1) {
 194                 map = kmem_cache_alloc(svc_rdma_map_cachep, GFP_KERNEL);
 195                 if (map)
 196                         break;
 197                 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
 198         }
 199         map->count = 0;
 200         return map;
 201 }
 202
 203 void svc_rdma_put_req_map(struct svc_rdma_req_map *map)
 204 {
 205         kmem_cache_free(svc_rdma_map_cachep, map);
 206 }
 207
 208 /* ib_cq event handler */
 209 static void cq_event_handler(struct ib_event *event, void *context)
 210 {
 211         struct svc_xprt *xprt = context;
 212         dprintk("svcrdma: received CQ event id=%d, context=%p\n",
 213                 event->event, context);
 214         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 215 }
 216
 217 /* QP event handler */
 218 static void qp_event_handler(struct ib_event *event, void *context)
 219 {
 220         struct svc_xprt *xprt = context;
 221
 222         switch (event->event) {
 223         /* These are considered benign events */
 224         case IB_EVENT_PATH_MIG:
 225         case IB_EVENT_COMM_EST:
 226         case IB_EVENT_SQ_DRAINED:
 227         case IB_EVENT_QP_LAST_WQE_REACHED:
 228                 dprintk("svcrdma: QP event %d received for QP=%p\n",
 229                         event->event, event->element.qp);
 230                 break;
 231         /* These are considered fatal events */
 232         case IB_EVENT_PATH_MIG_ERR:
 233         case IB_EVENT_QP_FATAL:
 234         case IB_EVENT_QP_REQ_ERR:
 235         case IB_EVENT_QP_ACCESS_ERR:
 236         case IB_EVENT_DEVICE_FATAL:
 237         default:
 238                 dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
 239                         "closing transport\n",
 240                         event->event, event->element.qp);
 241                 set_bit(XPT_CLOSE, &xprt->xpt_flags);
 242                 break;
 243         }
 244 }
 245
 246 /*
 247  * Data Transfer Operation Tasklet
 248  *
 249  * Walks a list of transports with I/O pending, removing entries as
 250  * they are added to the server's I/O pending list. Two bits indicate
 251  * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
 252  * spinlock that serializes access to the transport list with the RQ
 253  * and SQ interrupt handlers.
 254  */
 255 static void dto_tasklet_func(unsigned long data)
 256 {
 257         struct svcxprt_rdma *xprt;
 258         unsigned long flags;
 259
 260         spin_lock_irqsave(&dto_lock, flags);
 261         while (!list_empty(&dto_xprt_q)) {
 262                 xprt = list_entry(dto_xprt_q.next,
 263                                   struct svcxprt_rdma, sc_dto_q);
 264                 list_del_init(&xprt->sc_dto_q);
 265                 spin_unlock_irqrestore(&dto_lock, flags);
 266
 267                 rq_cq_reap(xprt);
 268                 sq_cq_reap(xprt);
 269
 270                 svc_xprt_put(&xprt->sc_xprt);
 271                 spin_lock_irqsave(&dto_lock, flags);
 272         }
 273         spin_unlock_irqrestore(&dto_lock, flags);
 274 }
 275
 276 /*
 277  * Receive Queue Completion Handler
 278  *
 279  * Since an RQ completion handler is called on interrupt context, we
 280  * need to defer the handling of the I/O to a tasklet
 281  */
 282 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
 283 {
 284         struct svcxprt_rdma *xprt = cq_context;
 285         unsigned long flags;
 286
 287         /* Guard against unconditional flush call for destroyed QP */
 288         if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
 289                 return;
 290
 291         /*
 292          * Set the bit regardless of whether or not it's on the list
 293          * because it may be on the list already due to an SQ
 294          * completion.
 295          */
 296         set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
 297
 298         /*
 299          * If this transport is not already on the DTO transport queue,
 300          * add it
 301          */
 302         spin_lock_irqsave(&dto_lock, flags);
 303         if (list_empty(&xprt->sc_dto_q)) {
 304                 svc_xprt_get(&xprt->sc_xprt);
 305                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
 306         }
 307         spin_unlock_irqrestore(&dto_lock, flags);
 308
 309         /* Tasklet does all the work to avoid irqsave locks. */
 310         tasklet_schedule(&dto_tasklet);
 311 }
 312
 313 /*
 314  * rq_cq_reap - Process the RQ CQ.
 315  *
 316  * Take all completing WC off the CQE and enqueue the associated DTO
 317  * context on the dto_q for the transport.
 318  *
 319  * Note that caller must hold a transport reference.
 320  */
 321 static void rq_cq_reap(struct svcxprt_rdma *xprt)
 322 {
 323         int ret;
 324         struct ib_wc wc;
 325         struct svc_rdma_op_ctxt *ctxt = NULL;
 326
 327         if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
 328                 return;
 329
 330         ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
 331         atomic_inc(&rdma_stat_rq_poll);
 332
 333         while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
 334                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
 335                 ctxt->wc_status = wc.status;
 336                 ctxt->byte_len = wc.byte_len;
 337                 svc_rdma_unmap_dma(ctxt);
 338                 if (wc.status != IB_WC_SUCCESS) {
 339                         /* Close the transport */
 340                         dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt);
 341                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 342                         svc_rdma_put_context(ctxt, 1);
 343                         svc_xprt_put(&xprt->sc_xprt);
 344                         continue;
 345                 }
 346                 spin_lock_bh(&xprt->sc_rq_dto_lock);
 347                 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
 348                 spin_unlock_bh(&xprt->sc_rq_dto_lock);
 349                 svc_xprt_put(&xprt->sc_xprt);
 350         }
 351
 352         if (ctxt)
 353                 atomic_inc(&rdma_stat_rq_prod);
 354
 355         set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
 356         /*
 357          * If data arrived before established event,
 358          * don't enqueue. This defers RPC I/O until the
 359          * RDMA connection is complete.
 360          */
 361         if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
 362                 svc_xprt_enqueue(&xprt->sc_xprt);
 363 }
 364
 365 /*
 366  * Send Queue Completion Handler - potentially called on interrupt context.
 367  *
 368  * Note that caller must hold a transport reference.
 369  */
 370 static void sq_cq_reap(struct svcxprt_rdma *xprt)
 371 {
 372         struct svc_rdma_op_ctxt *ctxt = NULL;
 373         struct ib_wc wc;
 374         struct ib_cq *cq = xprt->sc_sq_cq;
 375         int ret;
 376
 377
 378         if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
 379                 return;
 380
 381         ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
 382         atomic_inc(&rdma_stat_sq_poll);
 383         while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
 384                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
 385                 xprt = ctxt->xprt;
 386
 387                 svc_rdma_unmap_dma(ctxt);
 388                 if (wc.status != IB_WC_SUCCESS)
 389                         /* Close the transport */
 390                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 391
 392                 /* Decrement used SQ WR count */
 393                 atomic_dec(&xprt->sc_sq_count);
 394                 wake_up(&xprt->sc_send_wait);
 395
 396                 switch (ctxt->wr_op) {
 397                 case IB_WR_SEND:
 398                         svc_rdma_put_context(ctxt, 1);
 399                         break;
 400
 401                 case IB_WR_RDMA_WRITE:
 402                         svc_rdma_put_context(ctxt, 0);
 403                         break;
 404
 405                 case IB_WR_RDMA_READ:
 406                         if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
 407                                 struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr;
 408                                 BUG_ON(!read_hdr);
 409                                 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
 410                                 spin_lock_bh(&xprt->sc_read_complete_lock);
 411                                 list_add_tail(&read_hdr->dto_q,
 412                                               &xprt->sc_read_complete_q);
 413                                 spin_unlock_bh(&xprt->sc_read_complete_lock);
 414                                 svc_xprt_enqueue(&xprt->sc_xprt);
 415                         }
 416                         svc_rdma_put_context(ctxt, 0);
 417                         break;
 418
 419                 default:
 420                         printk(KERN_ERR "svcrdma: unexpected completion type, "
 421                                "opcode=%d, status=%d\n",
 422                                wc.opcode, wc.status);
 423                         break;
 424                 }
 425                 svc_xprt_put(&xprt->sc_xprt);
 426         }
 427
 428         if (ctxt)
 429                 atomic_inc(&rdma_stat_sq_prod);
 430 }
 431
 432 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
 433 {
 434         struct svcxprt_rdma *xprt = cq_context;
 435         unsigned long flags;
 436
 437         /* Guard against unconditional flush call for destroyed QP */
 438         if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
 439                 return;
 440
 441         /*
 442          * Set the bit regardless of whether or not it's on the list
 443          * because it may be on the list already due to an RQ
 444          * completion.
 445          */
 446         set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
 447
 448         /*
 449          * If this transport is not already on the DTO transport queue,
 450          * add it
 451          */
 452         spin_lock_irqsave(&dto_lock, flags);
 453         if (list_empty(&xprt->sc_dto_q)) {
 454                 svc_xprt_get(&xprt->sc_xprt);
 455                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
 456         }
 457         spin_unlock_irqrestore(&dto_lock, flags);
 458
 459         /* Tasklet does all the work to avoid irqsave locks. */
 460         tasklet_schedule(&dto_tasklet);
 461 }
 462
 463 static void create_context_cache(struct svcxprt_rdma *xprt,
 464                                  int ctxt_count, int ctxt_bump, int ctxt_max)
 465 {
 466         struct svc_rdma_op_ctxt *ctxt;
 467         int i;
 468
 469         xprt->sc_ctxt_max = ctxt_max;
 470         xprt->sc_ctxt_bump = ctxt_bump;
 471         xprt->sc_ctxt_cnt = 0;
 472         atomic_set(&xprt->sc_ctxt_used, 0);
 473
 474         INIT_LIST_HEAD(&xprt->sc_ctxt_free);
 475         for (i = 0; i < ctxt_count; i++) {
 476                 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
 477                 if (ctxt) {
 478                         INIT_LIST_HEAD(&ctxt->free_list);
 479                         list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
 480                         xprt->sc_ctxt_cnt++;
 481                 }
 482         }
 483 }
 484
 485 static void destroy_context_cache(struct svcxprt_rdma *xprt)
 486 {
 487         while (!list_empty(&xprt->sc_ctxt_free)) {
 488                 struct svc_rdma_op_ctxt *ctxt;
 489                 ctxt = list_entry(xprt->sc_ctxt_free.next,
 490                                   struct svc_rdma_op_ctxt,
 491                                   free_list);
 492                 list_del_init(&ctxt->free_list);
 493                 kfree(ctxt);
 494         }
 495 }
 496
 497 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
 498                                              int listener)
 499 {
 500         struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
 501
 502         if (!cma_xprt)
 503                 return NULL;
 504         svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv);
 505         INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
 506         INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
 507         INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
 508         INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
 509         init_waitqueue_head(&cma_xprt->sc_send_wait);
 510
 511         spin_lock_init(&cma_xprt->sc_lock);
 512         spin_lock_init(&cma_xprt->sc_read_complete_lock);
 513         spin_lock_init(&cma_xprt->sc_ctxt_lock);
 514         spin_lock_init(&cma_xprt->sc_rq_dto_lock);
 515
 516         cma_xprt->sc_ord = svcrdma_ord;
 517
 518         cma_xprt->sc_max_req_size = svcrdma_max_req_size;
 519         cma_xprt->sc_max_requests = svcrdma_max_requests;
 520         cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
 521         atomic_set(&cma_xprt->sc_sq_count, 0);
 522
 523         if (!listener) {
 524                 int reqs = cma_xprt->sc_max_requests;
 525                 create_context_cache(cma_xprt,
 526                                      reqs << 1, /* starting size */
 527                                      reqs,      /* bump amount */
 528                                      reqs +
 529                                      cma_xprt->sc_sq_depth +
 530                                      RPCRDMA_MAX_THREADS + 1); /* max */
 531                 if (list_empty(&cma_xprt->sc_ctxt_free)) {
 532                         kfree(cma_xprt);
 533                         return NULL;
 534                 }
 535                 clear_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
 536         } else
 537                 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
 538
 539         return cma_xprt;
 540 }
 541
 542 struct page *svc_rdma_get_page(void)
 543 {
 544         struct page *page;
 545
 546         while ((page = alloc_page(GFP_KERNEL)) == NULL) {
 547                 /* If we can't get memory, wait a bit and try again */
 548                 printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 "
 549                        "jiffies.\n");
 550                 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
 551         }
 552         return page;
 553 }
 554
 555 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
 556 {
 557         struct ib_recv_wr recv_wr, *bad_recv_wr;
 558         struct svc_rdma_op_ctxt *ctxt;
 559         struct page *page;
 560         unsigned long pa;
 561         int sge_no;
 562         int buflen;
 563         int ret;
 564
 565         ctxt = svc_rdma_get_context(xprt);
 566         buflen = 0;
 567         ctxt->direction = DMA_FROM_DEVICE;
 568         for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
 569                 BUG_ON(sge_no >= xprt->sc_max_sge);
 570                 page = svc_rdma_get_page();
 571                 ctxt->pages[sge_no] = page;
 572                 pa = ib_dma_map_page(xprt->sc_cm_id->device,
 573                                      page, 0, PAGE_SIZE,
 574                                      DMA_FROM_DEVICE);
 575                 ctxt->sge[sge_no].addr = pa;
 576                 ctxt->sge[sge_no].length = PAGE_SIZE;
 577                 ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
 578                 buflen += PAGE_SIZE;
 579         }
 580         ctxt->count = sge_no;
 581         recv_wr.next = NULL;
 582         recv_wr.sg_list = &ctxt->sge[0];
 583         recv_wr.num_sge = ctxt->count;
 584         recv_wr.wr_id = (u64)(unsigned long)ctxt;
 585
 586         svc_xprt_get(&xprt->sc_xprt);
 587         ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
 588         if (ret) {
 589                 svc_xprt_put(&xprt->sc_xprt);
 590                 svc_rdma_put_context(ctxt, 1);
 591         }
 592         return ret;
 593 }
 594
 595 /*
 596  * This function handles the CONNECT_REQUEST event on a listening
 597  * endpoint. It is passed the cma_id for the _new_ connection. The context in
 598  * this cma_id is inherited from the listening cma_id and is the svc_xprt
 599  * structure for the listening endpoint.
 600  *
 601  * This function creates a new xprt for the new connection and enqueues it on
 602  * the accept queue for the listent xprt. When the listen thread is kicked, it
 603  * will call the recvfrom method on the listen xprt which will accept the new
 604  * connection.
 605  */
 606 static void handle_connect_req(struct rdma_cm_id *new_cma_id)
 607 {
 608         struct svcxprt_rdma *listen_xprt = new_cma_id->context;
 609         struct svcxprt_rdma *newxprt;
 610         struct sockaddr *sa;
 611
 612         /* Create a new transport */
 613         newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
 614         if (!newxprt) {
 615                 dprintk("svcrdma: failed to create new transport\n");
 616                 return;
 617         }
 618         newxprt->sc_cm_id = new_cma_id;
 619         new_cma_id->context = newxprt;
 620         dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
 621                 newxprt, newxprt->sc_cm_id, listen_xprt);
 622
 623         /* Set the local and remote addresses in the transport */
 624         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
 625         svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
 626         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
 627         svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
 628
 629         /*
 630          * Enqueue the new transport on the accept queue of the listening
 631          * transport
 632          */
 633         spin_lock_bh(&listen_xprt->sc_lock);
 634         list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
 635         spin_unlock_bh(&listen_xprt->sc_lock);
 636
 637         /*
 638          * Can't use svc_xprt_received here because we are not on a
 639          * rqstp thread
 640         */
 641         set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
 642         svc_xprt_enqueue(&listen_xprt->sc_xprt);
 643 }
 644
 645 /*
 646  * Handles events generated on the listening endpoint. These events will be
 647  * either be incoming connect requests or adapter removal  events.
 648  */
 649 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
 650                                struct rdma_cm_event *event)
 651 {
 652         struct svcxprt_rdma *xprt = cma_id->context;
 653         int ret = 0;
 654
 655         switch (event->event) {
 656         case RDMA_CM_EVENT_CONNECT_REQUEST:
 657                 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
 658                         "event=%d\n", cma_id, cma_id->context, event->event);
 659                 handle_connect_req(cma_id);
 660                 break;
 661
 662         case RDMA_CM_EVENT_ESTABLISHED:
 663                 /* Accept complete */
 664                 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
 665                         "cm_id=%p\n", xprt, cma_id);
 666                 break;
 667
 668         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 669                 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
 670                         xprt, cma_id);
 671                 if (xprt)
 672                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 673                 break;
 674
 675         default:
 676                 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
 677                         "event=%d\n", cma_id, event->event);
 678                 break;
 679         }
 680
 681         return ret;
 682 }
 683
 684 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
 685                             struct rdma_cm_event *event)
 686 {
 687         struct svc_xprt *xprt = cma_id->context;
 688         struct svcxprt_rdma *rdma =
 689                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 690         switch (event->event) {
 691         case RDMA_CM_EVENT_ESTABLISHED:
 692                 /* Accept complete */
 693                 svc_xprt_get(xprt);
 694                 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
 695                         "cm_id=%p\n", xprt, cma_id);
 696                 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
 697                 svc_xprt_enqueue(xprt);
 698                 break;
 699         case RDMA_CM_EVENT_DISCONNECTED:
 700                 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
 701                         xprt, cma_id);
 702                 if (xprt) {
 703                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 704                         svc_xprt_enqueue(xprt);
 705                         svc_xprt_put(xprt);
 706                 }
 707                 break;
 708         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 709                 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
 710                         "event=%d\n", cma_id, xprt, event->event);
 711                 if (xprt) {
 712                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 713                         svc_xprt_enqueue(xprt);
 714                 }
 715                 break;
 716         default:
 717                 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
 718                         "event=%d\n", cma_id, event->event);
 719                 break;
 720         }
 721         return 0;
 722 }
 723
 724 /*
 725  * Create a listening RDMA service endpoint.
 726  */
 727 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
 728                                         struct sockaddr *sa, int salen,
 729                                         int flags)
 730 {
 731         struct rdma_cm_id *listen_id;
 732         struct svcxprt_rdma *cma_xprt;
 733         struct svc_xprt *xprt;
 734         int ret;
 735
 736         dprintk("svcrdma: Creating RDMA socket\n");
 737
 738         cma_xprt = rdma_create_xprt(serv, 1);
 739         if (!cma_xprt)
 740                 return ERR_PTR(-ENOMEM);
 741         xprt = &cma_xprt->sc_xprt;
 742
 743         listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
 744         if (IS_ERR(listen_id)) {
 745                 ret = PTR_ERR(listen_id);
 746                 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
 747                 goto err0;
 748         }
 749
 750         ret = rdma_bind_addr(listen_id, sa);
 751         if (ret) {
 752                 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
 753                 goto err1;
 754         }
 755         cma_xprt->sc_cm_id = listen_id;
 756
 757         ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
 758         if (ret) {
 759                 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
 760                 goto err1;
 761         }
 762
 763         /*
 764          * We need to use the address from the cm_id in case the
 765          * caller specified 0 for the port number.
 766          */
 767         sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
 768         svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
 769
 770         return &cma_xprt->sc_xprt;
 771
 772  err1:
 773         rdma_destroy_id(listen_id);
 774  err0:
 775         kfree(cma_xprt);
 776         return ERR_PTR(ret);
 777 }
 778
 779 /*
 780  * This is the xpo_recvfrom function for listening endpoints. Its
 781  * purpose is to accept incoming connections. The CMA callback handler
 782  * has already created a new transport and attached it to the new CMA
 783  * ID.
 784  *
 785  * There is a queue of pending connections hung on the listening
 786  * transport. This queue contains the new svc_xprt structure. This
 787  * function takes svc_xprt structures off the accept_q and completes
 788  * the connection.
 789  */
 790 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
 791 {
 792         struct svcxprt_rdma *listen_rdma;
 793         struct svcxprt_rdma *newxprt = NULL;
 794         struct rdma_conn_param conn_param;
 795         struct ib_qp_init_attr qp_attr;
 796         struct ib_device_attr devattr;
 797         int ret;
 798         int i;
 799
 800         listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
 801         clear_bit(XPT_CONN, &xprt->xpt_flags);
 802         /* Get the next entry off the accept list */
 803         spin_lock_bh(&listen_rdma->sc_lock);
 804         if (!list_empty(&listen_rdma->sc_accept_q)) {
 805                 newxprt = list_entry(listen_rdma->sc_accept_q.next,
 806                                      struct svcxprt_rdma, sc_accept_q);
 807                 list_del_init(&newxprt->sc_accept_q);
 808         }
 809         if (!list_empty(&listen_rdma->sc_accept_q))
 810                 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
 811         spin_unlock_bh(&listen_rdma->sc_lock);
 812         if (!newxprt)
 813                 return NULL;
 814
 815         dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
 816                 newxprt, newxprt->sc_cm_id);
 817
 818         ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
 819         if (ret) {
 820                 dprintk("svcrdma: could not query device attributes on "
 821                         "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
 822                 goto errout;
 823         }
 824
 825         /* Qualify the transport resource defaults with the
 826          * capabilities of this particular device */
 827         newxprt->sc_max_sge = min((size_t)devattr.max_sge,
 828                                   (size_t)RPCSVC_MAXPAGES);
 829         newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
 830                                    (size_t)svcrdma_max_requests);
 831         newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
 832
 833         newxprt->sc_ord =  min((size_t)devattr.max_qp_rd_atom,
 834                                (size_t)svcrdma_ord);
 835
 836         newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
 837         if (IS_ERR(newxprt->sc_pd)) {
 838                 dprintk("svcrdma: error creating PD for connect request\n");
 839                 goto errout;
 840         }
 841         newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
 842                                          sq_comp_handler,
 843                                          cq_event_handler,
 844                                          newxprt,
 845                                          newxprt->sc_sq_depth,
 846                                          0);
 847         if (IS_ERR(newxprt->sc_sq_cq)) {
 848                 dprintk("svcrdma: error creating SQ CQ for connect request\n");
 849                 goto errout;
 850         }
 851         newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
 852                                          rq_comp_handler,
 853                                          cq_event_handler,
 854                                          newxprt,
 855                                          newxprt->sc_max_requests,
 856                                          0);
 857         if (IS_ERR(newxprt->sc_rq_cq)) {
 858                 dprintk("svcrdma: error creating RQ CQ for connect request\n");
 859                 goto errout;
 860         }
 861
 862         memset(&qp_attr, 0, sizeof qp_attr);
 863         qp_attr.event_handler = qp_event_handler;
 864         qp_attr.qp_context = &newxprt->sc_xprt;
 865         qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
 866         qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
 867         qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
 868         qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
 869         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 870         qp_attr.qp_type = IB_QPT_RC;
 871         qp_attr.send_cq = newxprt->sc_sq_cq;
 872         qp_attr.recv_cq = newxprt->sc_rq_cq;
 873         dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
 874                 "    cm_id->device=%p, sc_pd->device=%p\n"
 875                 "    cap.max_send_wr = %d\n"
 876                 "    cap.max_recv_wr = %d\n"
 877                 "    cap.max_send_sge = %d\n"
 878                 "    cap.max_recv_sge = %d\n",
 879                 newxprt->sc_cm_id, newxprt->sc_pd,
 880                 newxprt->sc_cm_id->device, newxprt->sc_pd->device,
 881                 qp_attr.cap.max_send_wr,
 882                 qp_attr.cap.max_recv_wr,
 883                 qp_attr.cap.max_send_sge,
 884                 qp_attr.cap.max_recv_sge);
 885
 886         ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
 887         if (ret) {
 888                 /*
 889                  * XXX: This is a hack. We need a xx_request_qp interface
 890                  * that will adjust the qp_attr's with a best-effort
 891                  * number
 892                  */
 893                 qp_attr.cap.max_send_sge -= 2;
 894                 qp_attr.cap.max_recv_sge -= 2;
 895                 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd,
 896                                      &qp_attr);
 897                 if (ret) {
 898                         dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
 899                         goto errout;
 900                 }
 901                 newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
 902                 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
 903                 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
 904                 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
 905         }
 906         newxprt->sc_qp = newxprt->sc_cm_id->qp;
 907
 908         /* Register all of physical memory */
 909         newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
 910                                             IB_ACCESS_LOCAL_WRITE |
 911                                             IB_ACCESS_REMOTE_WRITE);
 912         if (IS_ERR(newxprt->sc_phys_mr)) {
 913                 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
 914                 goto errout;
 915         }
 916
 917         /* Post receive buffers */
 918         for (i = 0; i < newxprt->sc_max_requests; i++) {
 919                 ret = svc_rdma_post_recv(newxprt);
 920                 if (ret) {
 921                         dprintk("svcrdma: failure posting receive buffers\n");
 922                         goto errout;
 923                 }
 924         }
 925
 926         /* Swap out the handler */
 927         newxprt->sc_cm_id->event_handler = rdma_cma_handler;
 928
 929         /*
 930          * Arm the CQs for the SQ and RQ before accepting so we can't
 931          * miss the first message
 932          */
 933         ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
 934         ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
 935
 936         /* Accept Connection */
 937         set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
 938         memset(&conn_param, 0, sizeof conn_param);
 939         conn_param.responder_resources = 0;
 940         conn_param.initiator_depth = newxprt->sc_ord;
 941         ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
 942         if (ret) {
 943                 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
 944                        ret);
 945                 goto errout;
 946         }
 947
 948         dprintk("svcrdma: new connection %p accepted with the following "
 949                 "attributes:\n"
 950                 "    local_ip        : %d.%d.%d.%d\n"
 951                 "    local_port      : %d\n"
 952                 "    remote_ip       : %d.%d.%d.%d\n"
 953                 "    remote_port     : %d\n"
 954                 "    max_sge         : %d\n"
 955                 "    sq_depth        : %d\n"
 956                 "    max_requests    : %d\n"
 957                 "    ord             : %d\n",
 958                 newxprt,
 959                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
 960                          route.addr.src_addr)->sin_addr.s_addr),
 961                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
 962                        route.addr.src_addr)->sin_port),
 963                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
 964                          route.addr.dst_addr)->sin_addr.s_addr),
 965                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
 966                        route.addr.dst_addr)->sin_port),
 967                 newxprt->sc_max_sge,
 968                 newxprt->sc_sq_depth,
 969                 newxprt->sc_max_requests,
 970                 newxprt->sc_ord);
 971
 972         return &newxprt->sc_xprt;
 973
 974  errout:
 975         dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
 976         /* Take a reference in case the DTO handler runs */
 977         svc_xprt_get(&newxprt->sc_xprt);
 978         if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
 979                 ib_destroy_qp(newxprt->sc_qp);
 980         rdma_destroy_id(newxprt->sc_cm_id);
 981         /* This call to put will destroy the transport */
 982         svc_xprt_put(&newxprt->sc_xprt);
 983         return NULL;
 984 }
 985
 986 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
 987 {
 988 }
 989
 990 /*
 991  * When connected, an svc_xprt has at least two references:
 992  *
 993  * - A reference held by the cm_id between the ESTABLISHED and
 994  *   DISCONNECTED events. If the remote peer disconnected first, this
 995  *   reference could be gone.
 996  *
 997  * - A reference held by the svc_recv code that called this function
 998  *   as part of close processing.
 999  *
1000  * At a minimum one references should still be held.
1001  */
1002 static void svc_rdma_detach(struct svc_xprt *xprt)
1003 {
1004         struct svcxprt_rdma *rdma =
1005                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1006         dprintk("svc: svc_rdma_detach(%p)\n", xprt);
1007
1008         /* Disconnect and flush posted WQE */
1009         rdma_disconnect(rdma->sc_cm_id);
1010 }
1011
1012 static void __svc_rdma_free(struct work_struct *work)
1013 {
1014         struct svcxprt_rdma *rdma =
1015                 container_of(work, struct svcxprt_rdma, sc_work);
1016         dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
1017
1018         /* We should only be called from kref_put */
1019         BUG_ON(atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0);
1020
1021         /*
1022          * Destroy queued, but not processed read completions. Note
1023          * that this cleanup has to be done before destroying the
1024          * cm_id because the device ptr is needed to unmap the dma in
1025          * svc_rdma_put_context.
1026          */
1027         spin_lock_bh(&rdma->sc_read_complete_lock);
1028         while (!list_empty(&rdma->sc_read_complete_q)) {
1029                 struct svc_rdma_op_ctxt *ctxt;
1030                 ctxt = list_entry(rdma->sc_read_complete_q.next,
1031                                   struct svc_rdma_op_ctxt,
1032                                   dto_q);
1033                 list_del_init(&ctxt->dto_q);
1034                 svc_rdma_put_context(ctxt, 1);
1035         }
1036         spin_unlock_bh(&rdma->sc_read_complete_lock);
1037
1038         /* Destroy queued, but not processed recv completions */
1039         spin_lock_bh(&rdma->sc_rq_dto_lock);
1040         while (!list_empty(&rdma->sc_rq_dto_q)) {
1041                 struct svc_rdma_op_ctxt *ctxt;
1042                 ctxt = list_entry(rdma->sc_rq_dto_q.next,
1043                                   struct svc_rdma_op_ctxt,
1044                                   dto_q);
1045                 list_del_init(&ctxt->dto_q);
1046                 svc_rdma_put_context(ctxt, 1);
1047         }
1048         spin_unlock_bh(&rdma->sc_rq_dto_lock);
1049
1050         /* Warn if we leaked a resource or under-referenced */
1051         WARN_ON(atomic_read(&rdma->sc_ctxt_used) != 0);
1052
1053         /* Destroy the QP if present (not a listener) */
1054         if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
1055                 ib_destroy_qp(rdma->sc_qp);
1056
1057         if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
1058                 ib_destroy_cq(rdma->sc_sq_cq);
1059
1060         if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
1061                 ib_destroy_cq(rdma->sc_rq_cq);
1062
1063         if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
1064                 ib_dereg_mr(rdma->sc_phys_mr);
1065
1066         if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
1067                 ib_dealloc_pd(rdma->sc_pd);
1068
1069         /* Destroy the CM ID */
1070         rdma_destroy_id(rdma->sc_cm_id);
1071
1072         destroy_context_cache(rdma);
1073         kfree(rdma);
1074 }
1075
1076 static void svc_rdma_free(struct svc_xprt *xprt)
1077 {
1078         struct svcxprt_rdma *rdma =
1079                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1080         INIT_WORK(&rdma->sc_work, __svc_rdma_free);
1081         schedule_work(&rdma->sc_work);
1082 }
1083
1084 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
1085 {
1086         struct svcxprt_rdma *rdma =
1087                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1088
1089         /*
1090          * If there are fewer SQ WR available than required to send a
1091          * simple response, return false.
1092          */
1093         if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
1094                 return 0;
1095
1096         /*
1097          * ...or there are already waiters on the SQ,
1098          * return false.
1099          */
1100         if (waitqueue_active(&rdma->sc_send_wait))
1101                 return 0;
1102
1103         /* Otherwise return true. */
1104         return 1;
1105 }
1106
1107 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1108 {
1109         struct ib_send_wr *bad_wr;
1110         int ret;
1111
1112         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1113                 return -ENOTCONN;
1114
1115         BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
1116         BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
1117                 wr->opcode);
1118         /* If the SQ is full, wait until an SQ entry is available */
1119         while (1) {
1120                 spin_lock_bh(&xprt->sc_lock);
1121                 if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
1122                         spin_unlock_bh(&xprt->sc_lock);
1123                         atomic_inc(&rdma_stat_sq_starve);
1124
1125                         /* See if we can opportunistically reap SQ WR to make room */
1126                         sq_cq_reap(xprt);
1127
1128                         /* Wait until SQ WR available if SQ still full */
1129                         wait_event(xprt->sc_send_wait,
1130                                    atomic_read(&xprt->sc_sq_count) <
1131                                    xprt->sc_sq_depth);
1132                         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1133                                 return 0;
1134                         continue;
1135                 }
1136                 /* Bumped used SQ WR count and post */
1137                 svc_xprt_get(&xprt->sc_xprt);
1138                 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1139                 if (!ret)
1140                         atomic_inc(&xprt->sc_sq_count);
1141                 else {
1142                         svc_xprt_put(&xprt->sc_xprt);
1143                         dprintk("svcrdma: failed to post SQ WR rc=%d, "
1144                                "sc_sq_count=%d, sc_sq_depth=%d\n",
1145                                ret, atomic_read(&xprt->sc_sq_count),
1146                                xprt->sc_sq_depth);
1147                 }
1148                 spin_unlock_bh(&xprt->sc_lock);
1149                 break;
1150         }
1151         return ret;
1152 }
1153
1154 void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1155                          enum rpcrdma_errcode err)
1156 {
1157         struct ib_send_wr err_wr;
1158         struct ib_sge sge;
1159         struct page *p;
1160         struct svc_rdma_op_ctxt *ctxt;
1161         u32 *va;
1162         int length;
1163         int ret;
1164
1165         p = svc_rdma_get_page();
1166         va = page_address(p);
1167
1168         /* XDR encode error */
1169         length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1170
1171         /* Prepare SGE for local address */
1172         sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
1173                                    p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
1174         sge.lkey = xprt->sc_phys_mr->lkey;
1175         sge.length = length;
1176
1177         ctxt = svc_rdma_get_context(xprt);
1178         ctxt->count = 1;
1179         ctxt->pages[0] = p;
1180
1181         /* Prepare SEND WR */
1182         memset(&err_wr, 0, sizeof err_wr);
1183         ctxt->wr_op = IB_WR_SEND;
1184         err_wr.wr_id = (unsigned long)ctxt;
1185         err_wr.sg_list = &sge;
1186         err_wr.num_sge = 1;
1187         err_wr.opcode = IB_WR_SEND;
1188         err_wr.send_flags = IB_SEND_SIGNALED;
1189
1190         /* Post It */
1191         ret = svc_rdma_send(xprt, &err_wr);
1192         if (ret) {
1193                 dprintk("svcrdma: Error %d posting send for protocol error\n",
1194                         ret);
1195                 svc_rdma_put_context(ctxt, 1);
1196         }
1197 }