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                 atomic_dec(&xprt->sc_dma_used);
 159                 ib_dma_unmap_single(xprt->sc_cm_id->device,
 160                                     ctxt->sge[i].addr,
 161                                     ctxt->sge[i].length,
 162                                     ctxt->direction);
 163         }
 164 }
 165
 166 void svc_rdma_put_context(struct svc_rdma_op_ctxt *ctxt, int free_pages)
 167 {
 168         struct svcxprt_rdma *xprt;
 169         int i;
 170
 171         BUG_ON(!ctxt);
 172         xprt = ctxt->xprt;
 173         if (free_pages)
 174                 for (i = 0; i < ctxt->count; i++)
 175                         put_page(ctxt->pages[i]);
 176
 177         spin_lock_bh(&xprt->sc_ctxt_lock);
 178         list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
 179         spin_unlock_bh(&xprt->sc_ctxt_lock);
 180         atomic_dec(&xprt->sc_ctxt_used);
 181 }
 182
 183 /* Temporary NFS request map cache. Created in svc_rdma.c  */
 184 extern struct kmem_cache *svc_rdma_map_cachep;
 185
 186 /*
 187  * Temporary NFS req mappings are shared across all transport
 188  * instances. These are short lived and should be bounded by the number
 189  * of concurrent server threads * depth of the SQ.
 190  */
 191 struct svc_rdma_req_map *svc_rdma_get_req_map(void)
 192 {
 193         struct svc_rdma_req_map *map;
 194         while (1) {
 195                 map = kmem_cache_alloc(svc_rdma_map_cachep, GFP_KERNEL);
 196                 if (map)
 197                         break;
 198                 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
 199         }
 200         map->count = 0;
 201         return map;
 202 }
 203
 204 void svc_rdma_put_req_map(struct svc_rdma_req_map *map)
 205 {
 206         kmem_cache_free(svc_rdma_map_cachep, map);
 207 }
 208
 209 /* ib_cq event handler */
 210 static void cq_event_handler(struct ib_event *event, void *context)
 211 {
 212         struct svc_xprt *xprt = context;
 213         dprintk("svcrdma: received CQ event id=%d, context=%p\n",
 214                 event->event, context);
 215         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 216 }
 217
 218 /* QP event handler */
 219 static void qp_event_handler(struct ib_event *event, void *context)
 220 {
 221         struct svc_xprt *xprt = context;
 222
 223         switch (event->event) {
 224         /* These are considered benign events */
 225         case IB_EVENT_PATH_MIG:
 226         case IB_EVENT_COMM_EST:
 227         case IB_EVENT_SQ_DRAINED:
 228         case IB_EVENT_QP_LAST_WQE_REACHED:
 229                 dprintk("svcrdma: QP event %d received for QP=%p\n",
 230                         event->event, event->element.qp);
 231                 break;
 232         /* These are considered fatal events */
 233         case IB_EVENT_PATH_MIG_ERR:
 234         case IB_EVENT_QP_FATAL:
 235         case IB_EVENT_QP_REQ_ERR:
 236         case IB_EVENT_QP_ACCESS_ERR:
 237         case IB_EVENT_DEVICE_FATAL:
 238         default:
 239                 dprintk("svcrdma: QP ERROR event %d received for QP=%p, "
 240                         "closing transport\n",
 241                         event->event, event->element.qp);
 242                 set_bit(XPT_CLOSE, &xprt->xpt_flags);
 243                 break;
 244         }
 245 }
 246
 247 /*
 248  * Data Transfer Operation Tasklet
 249  *
 250  * Walks a list of transports with I/O pending, removing entries as
 251  * they are added to the server's I/O pending list. Two bits indicate
 252  * if SQ, RQ, or both have I/O pending. The dto_lock is an irqsave
 253  * spinlock that serializes access to the transport list with the RQ
 254  * and SQ interrupt handlers.
 255  */
 256 static void dto_tasklet_func(unsigned long data)
 257 {
 258         struct svcxprt_rdma *xprt;
 259         unsigned long flags;
 260
 261         spin_lock_irqsave(&dto_lock, flags);
 262         while (!list_empty(&dto_xprt_q)) {
 263                 xprt = list_entry(dto_xprt_q.next,
 264                                   struct svcxprt_rdma, sc_dto_q);
 265                 list_del_init(&xprt->sc_dto_q);
 266                 spin_unlock_irqrestore(&dto_lock, flags);
 267
 268                 rq_cq_reap(xprt);
 269                 sq_cq_reap(xprt);
 270
 271                 svc_xprt_put(&xprt->sc_xprt);
 272                 spin_lock_irqsave(&dto_lock, flags);
 273         }
 274         spin_unlock_irqrestore(&dto_lock, flags);
 275 }
 276
 277 /*
 278  * Receive Queue Completion Handler
 279  *
 280  * Since an RQ completion handler is called on interrupt context, we
 281  * need to defer the handling of the I/O to a tasklet
 282  */
 283 static void rq_comp_handler(struct ib_cq *cq, void *cq_context)
 284 {
 285         struct svcxprt_rdma *xprt = cq_context;
 286         unsigned long flags;
 287
 288         /* Guard against unconditional flush call for destroyed QP */
 289         if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
 290                 return;
 291
 292         /*
 293          * Set the bit regardless of whether or not it's on the list
 294          * because it may be on the list already due to an SQ
 295          * completion.
 296          */
 297         set_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags);
 298
 299         /*
 300          * If this transport is not already on the DTO transport queue,
 301          * add it
 302          */
 303         spin_lock_irqsave(&dto_lock, flags);
 304         if (list_empty(&xprt->sc_dto_q)) {
 305                 svc_xprt_get(&xprt->sc_xprt);
 306                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
 307         }
 308         spin_unlock_irqrestore(&dto_lock, flags);
 309
 310         /* Tasklet does all the work to avoid irqsave locks. */
 311         tasklet_schedule(&dto_tasklet);
 312 }
 313
 314 /*
 315  * rq_cq_reap - Process the RQ CQ.
 316  *
 317  * Take all completing WC off the CQE and enqueue the associated DTO
 318  * context on the dto_q for the transport.
 319  *
 320  * Note that caller must hold a transport reference.
 321  */
 322 static void rq_cq_reap(struct svcxprt_rdma *xprt)
 323 {
 324         int ret;
 325         struct ib_wc wc;
 326         struct svc_rdma_op_ctxt *ctxt = NULL;
 327
 328         if (!test_and_clear_bit(RDMAXPRT_RQ_PENDING, &xprt->sc_flags))
 329                 return;
 330
 331         ib_req_notify_cq(xprt->sc_rq_cq, IB_CQ_NEXT_COMP);
 332         atomic_inc(&rdma_stat_rq_poll);
 333
 334         while ((ret = ib_poll_cq(xprt->sc_rq_cq, 1, &wc)) > 0) {
 335                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
 336                 ctxt->wc_status = wc.status;
 337                 ctxt->byte_len = wc.byte_len;
 338                 svc_rdma_unmap_dma(ctxt);
 339                 if (wc.status != IB_WC_SUCCESS) {
 340                         /* Close the transport */
 341                         dprintk("svcrdma: transport closing putting ctxt %p\n", ctxt);
 342                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 343                         svc_rdma_put_context(ctxt, 1);
 344                         svc_xprt_put(&xprt->sc_xprt);
 345                         continue;
 346                 }
 347                 spin_lock_bh(&xprt->sc_rq_dto_lock);
 348                 list_add_tail(&ctxt->dto_q, &xprt->sc_rq_dto_q);
 349                 spin_unlock_bh(&xprt->sc_rq_dto_lock);
 350                 svc_xprt_put(&xprt->sc_xprt);
 351         }
 352
 353         if (ctxt)
 354                 atomic_inc(&rdma_stat_rq_prod);
 355
 356         set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
 357         /*
 358          * If data arrived before established event,
 359          * don't enqueue. This defers RPC I/O until the
 360          * RDMA connection is complete.
 361          */
 362         if (!test_bit(RDMAXPRT_CONN_PENDING, &xprt->sc_flags))
 363                 svc_xprt_enqueue(&xprt->sc_xprt);
 364 }
 365
 366 /*
 367  * Send Queue Completion Handler - potentially called on interrupt context.
 368  *
 369  * Note that caller must hold a transport reference.
 370  */
 371 static void sq_cq_reap(struct svcxprt_rdma *xprt)
 372 {
 373         struct svc_rdma_op_ctxt *ctxt = NULL;
 374         struct ib_wc wc;
 375         struct ib_cq *cq = xprt->sc_sq_cq;
 376         int ret;
 377
 378
 379         if (!test_and_clear_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags))
 380                 return;
 381
 382         ib_req_notify_cq(xprt->sc_sq_cq, IB_CQ_NEXT_COMP);
 383         atomic_inc(&rdma_stat_sq_poll);
 384         while ((ret = ib_poll_cq(cq, 1, &wc)) > 0) {
 385                 ctxt = (struct svc_rdma_op_ctxt *)(unsigned long)wc.wr_id;
 386                 xprt = ctxt->xprt;
 387
 388                 svc_rdma_unmap_dma(ctxt);
 389                 if (wc.status != IB_WC_SUCCESS)
 390                         /* Close the transport */
 391                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 392
 393                 /* Decrement used SQ WR count */
 394                 atomic_dec(&xprt->sc_sq_count);
 395                 wake_up(&xprt->sc_send_wait);
 396
 397                 switch (ctxt->wr_op) {
 398                 case IB_WR_SEND:
 399                         svc_rdma_put_context(ctxt, 1);
 400                         break;
 401
 402                 case IB_WR_RDMA_WRITE:
 403                         svc_rdma_put_context(ctxt, 0);
 404                         break;
 405
 406                 case IB_WR_RDMA_READ:
 407                         if (test_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags)) {
 408                                 struct svc_rdma_op_ctxt *read_hdr = ctxt->read_hdr;
 409                                 BUG_ON(!read_hdr);
 410                                 set_bit(XPT_DATA, &xprt->sc_xprt.xpt_flags);
 411                                 spin_lock_bh(&xprt->sc_read_complete_lock);
 412                                 list_add_tail(&read_hdr->dto_q,
 413                                               &xprt->sc_read_complete_q);
 414                                 spin_unlock_bh(&xprt->sc_read_complete_lock);
 415                                 svc_xprt_enqueue(&xprt->sc_xprt);
 416                         }
 417                         svc_rdma_put_context(ctxt, 0);
 418                         break;
 419
 420                 default:
 421                         printk(KERN_ERR "svcrdma: unexpected completion type, "
 422                                "opcode=%d, status=%d\n",
 423                                wc.opcode, wc.status);
 424                         break;
 425                 }
 426                 svc_xprt_put(&xprt->sc_xprt);
 427         }
 428
 429         if (ctxt)
 430                 atomic_inc(&rdma_stat_sq_prod);
 431 }
 432
 433 static void sq_comp_handler(struct ib_cq *cq, void *cq_context)
 434 {
 435         struct svcxprt_rdma *xprt = cq_context;
 436         unsigned long flags;
 437
 438         /* Guard against unconditional flush call for destroyed QP */
 439         if (atomic_read(&xprt->sc_xprt.xpt_ref.refcount)==0)
 440                 return;
 441
 442         /*
 443          * Set the bit regardless of whether or not it's on the list
 444          * because it may be on the list already due to an RQ
 445          * completion.
 446          */
 447         set_bit(RDMAXPRT_SQ_PENDING, &xprt->sc_flags);
 448
 449         /*
 450          * If this transport is not already on the DTO transport queue,
 451          * add it
 452          */
 453         spin_lock_irqsave(&dto_lock, flags);
 454         if (list_empty(&xprt->sc_dto_q)) {
 455                 svc_xprt_get(&xprt->sc_xprt);
 456                 list_add_tail(&xprt->sc_dto_q, &dto_xprt_q);
 457         }
 458         spin_unlock_irqrestore(&dto_lock, flags);
 459
 460         /* Tasklet does all the work to avoid irqsave locks. */
 461         tasklet_schedule(&dto_tasklet);
 462 }
 463
 464 static void create_context_cache(struct svcxprt_rdma *xprt,
 465                                  int ctxt_count, int ctxt_bump, int ctxt_max)
 466 {
 467         struct svc_rdma_op_ctxt *ctxt;
 468         int i;
 469
 470         xprt->sc_ctxt_max = ctxt_max;
 471         xprt->sc_ctxt_bump = ctxt_bump;
 472         xprt->sc_ctxt_cnt = 0;
 473         atomic_set(&xprt->sc_ctxt_used, 0);
 474
 475         INIT_LIST_HEAD(&xprt->sc_ctxt_free);
 476         for (i = 0; i < ctxt_count; i++) {
 477                 ctxt = kmalloc(sizeof(*ctxt), GFP_KERNEL);
 478                 if (ctxt) {
 479                         INIT_LIST_HEAD(&ctxt->free_list);
 480                         list_add(&ctxt->free_list, &xprt->sc_ctxt_free);
 481                         xprt->sc_ctxt_cnt++;
 482                 }
 483         }
 484 }
 485
 486 static void destroy_context_cache(struct svcxprt_rdma *xprt)
 487 {
 488         while (!list_empty(&xprt->sc_ctxt_free)) {
 489                 struct svc_rdma_op_ctxt *ctxt;
 490                 ctxt = list_entry(xprt->sc_ctxt_free.next,
 491                                   struct svc_rdma_op_ctxt,
 492                                   free_list);
 493                 list_del_init(&ctxt->free_list);
 494                 kfree(ctxt);
 495         }
 496 }
 497
 498 static struct svcxprt_rdma *rdma_create_xprt(struct svc_serv *serv,
 499                                              int listener)
 500 {
 501         struct svcxprt_rdma *cma_xprt = kzalloc(sizeof *cma_xprt, GFP_KERNEL);
 502
 503         if (!cma_xprt)
 504                 return NULL;
 505         svc_xprt_init(&svc_rdma_class, &cma_xprt->sc_xprt, serv);
 506         INIT_LIST_HEAD(&cma_xprt->sc_accept_q);
 507         INIT_LIST_HEAD(&cma_xprt->sc_dto_q);
 508         INIT_LIST_HEAD(&cma_xprt->sc_rq_dto_q);
 509         INIT_LIST_HEAD(&cma_xprt->sc_read_complete_q);
 510         init_waitqueue_head(&cma_xprt->sc_send_wait);
 511
 512         spin_lock_init(&cma_xprt->sc_lock);
 513         spin_lock_init(&cma_xprt->sc_read_complete_lock);
 514         spin_lock_init(&cma_xprt->sc_ctxt_lock);
 515         spin_lock_init(&cma_xprt->sc_rq_dto_lock);
 516
 517         cma_xprt->sc_ord = svcrdma_ord;
 518
 519         cma_xprt->sc_max_req_size = svcrdma_max_req_size;
 520         cma_xprt->sc_max_requests = svcrdma_max_requests;
 521         cma_xprt->sc_sq_depth = svcrdma_max_requests * RPCRDMA_SQ_DEPTH_MULT;
 522         atomic_set(&cma_xprt->sc_sq_count, 0);
 523         atomic_set(&cma_xprt->sc_ctxt_used, 0);
 524
 525         if (!listener) {
 526                 int reqs = cma_xprt->sc_max_requests;
 527                 create_context_cache(cma_xprt,
 528                                      reqs << 1, /* starting size */
 529                                      reqs,      /* bump amount */
 530                                      reqs +
 531                                      cma_xprt->sc_sq_depth +
 532                                      RPCRDMA_MAX_THREADS + 1); /* max */
 533                 if (list_empty(&cma_xprt->sc_ctxt_free)) {
 534                         kfree(cma_xprt);
 535                         return NULL;
 536                 }
 537                 clear_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
 538         } else
 539                 set_bit(XPT_LISTENER, &cma_xprt->sc_xprt.xpt_flags);
 540
 541         return cma_xprt;
 542 }
 543
 544 struct page *svc_rdma_get_page(void)
 545 {
 546         struct page *page;
 547
 548         while ((page = alloc_page(GFP_KERNEL)) == NULL) {
 549                 /* If we can't get memory, wait a bit and try again */
 550                 printk(KERN_INFO "svcrdma: out of memory...retrying in 1000 "
 551                        "jiffies.\n");
 552                 schedule_timeout_uninterruptible(msecs_to_jiffies(1000));
 553         }
 554         return page;
 555 }
 556
 557 int svc_rdma_post_recv(struct svcxprt_rdma *xprt)
 558 {
 559         struct ib_recv_wr recv_wr, *bad_recv_wr;
 560         struct svc_rdma_op_ctxt *ctxt;
 561         struct page *page;
 562         unsigned long pa;
 563         int sge_no;
 564         int buflen;
 565         int ret;
 566
 567         ctxt = svc_rdma_get_context(xprt);
 568         buflen = 0;
 569         ctxt->direction = DMA_FROM_DEVICE;
 570         for (sge_no = 0; buflen < xprt->sc_max_req_size; sge_no++) {
 571                 BUG_ON(sge_no >= xprt->sc_max_sge);
 572                 page = svc_rdma_get_page();
 573                 ctxt->pages[sge_no] = page;
 574                 atomic_inc(&xprt->sc_dma_used);
 575                 pa = ib_dma_map_page(xprt->sc_cm_id->device,
 576                                      page, 0, PAGE_SIZE,
 577                                      DMA_FROM_DEVICE);
 578                 ctxt->sge[sge_no].addr = pa;
 579                 ctxt->sge[sge_no].length = PAGE_SIZE;
 580                 ctxt->sge[sge_no].lkey = xprt->sc_phys_mr->lkey;
 581                 buflen += PAGE_SIZE;
 582         }
 583         ctxt->count = sge_no;
 584         recv_wr.next = NULL;
 585         recv_wr.sg_list = &ctxt->sge[0];
 586         recv_wr.num_sge = ctxt->count;
 587         recv_wr.wr_id = (u64)(unsigned long)ctxt;
 588
 589         svc_xprt_get(&xprt->sc_xprt);
 590         ret = ib_post_recv(xprt->sc_qp, &recv_wr, &bad_recv_wr);
 591         if (ret) {
 592                 svc_xprt_put(&xprt->sc_xprt);
 593                 svc_rdma_put_context(ctxt, 1);
 594         }
 595         return ret;
 596 }
 597
 598 /*
 599  * This function handles the CONNECT_REQUEST event on a listening
 600  * endpoint. It is passed the cma_id for the _new_ connection. The context in
 601  * this cma_id is inherited from the listening cma_id and is the svc_xprt
 602  * structure for the listening endpoint.
 603  *
 604  * This function creates a new xprt for the new connection and enqueues it on
 605  * the accept queue for the listent xprt. When the listen thread is kicked, it
 606  * will call the recvfrom method on the listen xprt which will accept the new
 607  * connection.
 608  */
 609 static void handle_connect_req(struct rdma_cm_id *new_cma_id)
 610 {
 611         struct svcxprt_rdma *listen_xprt = new_cma_id->context;
 612         struct svcxprt_rdma *newxprt;
 613         struct sockaddr *sa;
 614
 615         /* Create a new transport */
 616         newxprt = rdma_create_xprt(listen_xprt->sc_xprt.xpt_server, 0);
 617         if (!newxprt) {
 618                 dprintk("svcrdma: failed to create new transport\n");
 619                 return;
 620         }
 621         newxprt->sc_cm_id = new_cma_id;
 622         new_cma_id->context = newxprt;
 623         dprintk("svcrdma: Creating newxprt=%p, cm_id=%p, listenxprt=%p\n",
 624                 newxprt, newxprt->sc_cm_id, listen_xprt);
 625
 626         /* Set the local and remote addresses in the transport */
 627         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.dst_addr;
 628         svc_xprt_set_remote(&newxprt->sc_xprt, sa, svc_addr_len(sa));
 629         sa = (struct sockaddr *)&newxprt->sc_cm_id->route.addr.src_addr;
 630         svc_xprt_set_local(&newxprt->sc_xprt, sa, svc_addr_len(sa));
 631
 632         /*
 633          * Enqueue the new transport on the accept queue of the listening
 634          * transport
 635          */
 636         spin_lock_bh(&listen_xprt->sc_lock);
 637         list_add_tail(&newxprt->sc_accept_q, &listen_xprt->sc_accept_q);
 638         spin_unlock_bh(&listen_xprt->sc_lock);
 639
 640         /*
 641          * Can't use svc_xprt_received here because we are not on a
 642          * rqstp thread
 643         */
 644         set_bit(XPT_CONN, &listen_xprt->sc_xprt.xpt_flags);
 645         svc_xprt_enqueue(&listen_xprt->sc_xprt);
 646 }
 647
 648 /*
 649  * Handles events generated on the listening endpoint. These events will be
 650  * either be incoming connect requests or adapter removal  events.
 651  */
 652 static int rdma_listen_handler(struct rdma_cm_id *cma_id,
 653                                struct rdma_cm_event *event)
 654 {
 655         struct svcxprt_rdma *xprt = cma_id->context;
 656         int ret = 0;
 657
 658         switch (event->event) {
 659         case RDMA_CM_EVENT_CONNECT_REQUEST:
 660                 dprintk("svcrdma: Connect request on cma_id=%p, xprt = %p, "
 661                         "event=%d\n", cma_id, cma_id->context, event->event);
 662                 handle_connect_req(cma_id);
 663                 break;
 664
 665         case RDMA_CM_EVENT_ESTABLISHED:
 666                 /* Accept complete */
 667                 dprintk("svcrdma: Connection completed on LISTEN xprt=%p, "
 668                         "cm_id=%p\n", xprt, cma_id);
 669                 break;
 670
 671         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 672                 dprintk("svcrdma: Device removal xprt=%p, cm_id=%p\n",
 673                         xprt, cma_id);
 674                 if (xprt)
 675                         set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
 676                 break;
 677
 678         default:
 679                 dprintk("svcrdma: Unexpected event on listening endpoint %p, "
 680                         "event=%d\n", cma_id, event->event);
 681                 break;
 682         }
 683
 684         return ret;
 685 }
 686
 687 static int rdma_cma_handler(struct rdma_cm_id *cma_id,
 688                             struct rdma_cm_event *event)
 689 {
 690         struct svc_xprt *xprt = cma_id->context;
 691         struct svcxprt_rdma *rdma =
 692                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
 693         switch (event->event) {
 694         case RDMA_CM_EVENT_ESTABLISHED:
 695                 /* Accept complete */
 696                 svc_xprt_get(xprt);
 697                 dprintk("svcrdma: Connection completed on DTO xprt=%p, "
 698                         "cm_id=%p\n", xprt, cma_id);
 699                 clear_bit(RDMAXPRT_CONN_PENDING, &rdma->sc_flags);
 700                 svc_xprt_enqueue(xprt);
 701                 break;
 702         case RDMA_CM_EVENT_DISCONNECTED:
 703                 dprintk("svcrdma: Disconnect on DTO xprt=%p, cm_id=%p\n",
 704                         xprt, cma_id);
 705                 if (xprt) {
 706                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 707                         svc_xprt_enqueue(xprt);
 708                         svc_xprt_put(xprt);
 709                 }
 710                 break;
 711         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 712                 dprintk("svcrdma: Device removal cma_id=%p, xprt = %p, "
 713                         "event=%d\n", cma_id, xprt, event->event);
 714                 if (xprt) {
 715                         set_bit(XPT_CLOSE, &xprt->xpt_flags);
 716                         svc_xprt_enqueue(xprt);
 717                 }
 718                 break;
 719         default:
 720                 dprintk("svcrdma: Unexpected event on DTO endpoint %p, "
 721                         "event=%d\n", cma_id, event->event);
 722                 break;
 723         }
 724         return 0;
 725 }
 726
 727 /*
 728  * Create a listening RDMA service endpoint.
 729  */
 730 static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
 731                                         struct sockaddr *sa, int salen,
 732                                         int flags)
 733 {
 734         struct rdma_cm_id *listen_id;
 735         struct svcxprt_rdma *cma_xprt;
 736         struct svc_xprt *xprt;
 737         int ret;
 738
 739         dprintk("svcrdma: Creating RDMA socket\n");
 740
 741         cma_xprt = rdma_create_xprt(serv, 1);
 742         if (!cma_xprt)
 743                 return ERR_PTR(-ENOMEM);
 744         xprt = &cma_xprt->sc_xprt;
 745
 746         listen_id = rdma_create_id(rdma_listen_handler, cma_xprt, RDMA_PS_TCP);
 747         if (IS_ERR(listen_id)) {
 748                 ret = PTR_ERR(listen_id);
 749                 dprintk("svcrdma: rdma_create_id failed = %d\n", ret);
 750                 goto err0;
 751         }
 752
 753         ret = rdma_bind_addr(listen_id, sa);
 754         if (ret) {
 755                 dprintk("svcrdma: rdma_bind_addr failed = %d\n", ret);
 756                 goto err1;
 757         }
 758         cma_xprt->sc_cm_id = listen_id;
 759
 760         ret = rdma_listen(listen_id, RPCRDMA_LISTEN_BACKLOG);
 761         if (ret) {
 762                 dprintk("svcrdma: rdma_listen failed = %d\n", ret);
 763                 goto err1;
 764         }
 765
 766         /*
 767          * We need to use the address from the cm_id in case the
 768          * caller specified 0 for the port number.
 769          */
 770         sa = (struct sockaddr *)&cma_xprt->sc_cm_id->route.addr.src_addr;
 771         svc_xprt_set_local(&cma_xprt->sc_xprt, sa, salen);
 772
 773         return &cma_xprt->sc_xprt;
 774
 775  err1:
 776         rdma_destroy_id(listen_id);
 777  err0:
 778         kfree(cma_xprt);
 779         return ERR_PTR(ret);
 780 }
 781
 782 /*
 783  * This is the xpo_recvfrom function for listening endpoints. Its
 784  * purpose is to accept incoming connections. The CMA callback handler
 785  * has already created a new transport and attached it to the new CMA
 786  * ID.
 787  *
 788  * There is a queue of pending connections hung on the listening
 789  * transport. This queue contains the new svc_xprt structure. This
 790  * function takes svc_xprt structures off the accept_q and completes
 791  * the connection.
 792  */
 793 static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt)
 794 {
 795         struct svcxprt_rdma *listen_rdma;
 796         struct svcxprt_rdma *newxprt = NULL;
 797         struct rdma_conn_param conn_param;
 798         struct ib_qp_init_attr qp_attr;
 799         struct ib_device_attr devattr;
 800         int ret;
 801         int i;
 802
 803         listen_rdma = container_of(xprt, struct svcxprt_rdma, sc_xprt);
 804         clear_bit(XPT_CONN, &xprt->xpt_flags);
 805         /* Get the next entry off the accept list */
 806         spin_lock_bh(&listen_rdma->sc_lock);
 807         if (!list_empty(&listen_rdma->sc_accept_q)) {
 808                 newxprt = list_entry(listen_rdma->sc_accept_q.next,
 809                                      struct svcxprt_rdma, sc_accept_q);
 810                 list_del_init(&newxprt->sc_accept_q);
 811         }
 812         if (!list_empty(&listen_rdma->sc_accept_q))
 813                 set_bit(XPT_CONN, &listen_rdma->sc_xprt.xpt_flags);
 814         spin_unlock_bh(&listen_rdma->sc_lock);
 815         if (!newxprt)
 816                 return NULL;
 817
 818         dprintk("svcrdma: newxprt from accept queue = %p, cm_id=%p\n",
 819                 newxprt, newxprt->sc_cm_id);
 820
 821         ret = ib_query_device(newxprt->sc_cm_id->device, &devattr);
 822         if (ret) {
 823                 dprintk("svcrdma: could not query device attributes on "
 824                         "device %p, rc=%d\n", newxprt->sc_cm_id->device, ret);
 825                 goto errout;
 826         }
 827
 828         /* Qualify the transport resource defaults with the
 829          * capabilities of this particular device */
 830         newxprt->sc_max_sge = min((size_t)devattr.max_sge,
 831                                   (size_t)RPCSVC_MAXPAGES);
 832         newxprt->sc_max_requests = min((size_t)devattr.max_qp_wr,
 833                                    (size_t)svcrdma_max_requests);
 834         newxprt->sc_sq_depth = RPCRDMA_SQ_DEPTH_MULT * newxprt->sc_max_requests;
 835
 836         newxprt->sc_ord =  min((size_t)devattr.max_qp_rd_atom,
 837                                (size_t)svcrdma_ord);
 838
 839         newxprt->sc_pd = ib_alloc_pd(newxprt->sc_cm_id->device);
 840         if (IS_ERR(newxprt->sc_pd)) {
 841                 dprintk("svcrdma: error creating PD for connect request\n");
 842                 goto errout;
 843         }
 844         newxprt->sc_sq_cq = ib_create_cq(newxprt->sc_cm_id->device,
 845                                          sq_comp_handler,
 846                                          cq_event_handler,
 847                                          newxprt,
 848                                          newxprt->sc_sq_depth,
 849                                          0);
 850         if (IS_ERR(newxprt->sc_sq_cq)) {
 851                 dprintk("svcrdma: error creating SQ CQ for connect request\n");
 852                 goto errout;
 853         }
 854         newxprt->sc_rq_cq = ib_create_cq(newxprt->sc_cm_id->device,
 855                                          rq_comp_handler,
 856                                          cq_event_handler,
 857                                          newxprt,
 858                                          newxprt->sc_max_requests,
 859                                          0);
 860         if (IS_ERR(newxprt->sc_rq_cq)) {
 861                 dprintk("svcrdma: error creating RQ CQ for connect request\n");
 862                 goto errout;
 863         }
 864
 865         memset(&qp_attr, 0, sizeof qp_attr);
 866         qp_attr.event_handler = qp_event_handler;
 867         qp_attr.qp_context = &newxprt->sc_xprt;
 868         qp_attr.cap.max_send_wr = newxprt->sc_sq_depth;
 869         qp_attr.cap.max_recv_wr = newxprt->sc_max_requests;
 870         qp_attr.cap.max_send_sge = newxprt->sc_max_sge;
 871         qp_attr.cap.max_recv_sge = newxprt->sc_max_sge;
 872         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 873         qp_attr.qp_type = IB_QPT_RC;
 874         qp_attr.send_cq = newxprt->sc_sq_cq;
 875         qp_attr.recv_cq = newxprt->sc_rq_cq;
 876         dprintk("svcrdma: newxprt->sc_cm_id=%p, newxprt->sc_pd=%p\n"
 877                 "    cm_id->device=%p, sc_pd->device=%p\n"
 878                 "    cap.max_send_wr = %d\n"
 879                 "    cap.max_recv_wr = %d\n"
 880                 "    cap.max_send_sge = %d\n"
 881                 "    cap.max_recv_sge = %d\n",
 882                 newxprt->sc_cm_id, newxprt->sc_pd,
 883                 newxprt->sc_cm_id->device, newxprt->sc_pd->device,
 884                 qp_attr.cap.max_send_wr,
 885                 qp_attr.cap.max_recv_wr,
 886                 qp_attr.cap.max_send_sge,
 887                 qp_attr.cap.max_recv_sge);
 888
 889         ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd, &qp_attr);
 890         if (ret) {
 891                 /*
 892                  * XXX: This is a hack. We need a xx_request_qp interface
 893                  * that will adjust the qp_attr's with a best-effort
 894                  * number
 895                  */
 896                 qp_attr.cap.max_send_sge -= 2;
 897                 qp_attr.cap.max_recv_sge -= 2;
 898                 ret = rdma_create_qp(newxprt->sc_cm_id, newxprt->sc_pd,
 899                                      &qp_attr);
 900                 if (ret) {
 901                         dprintk("svcrdma: failed to create QP, ret=%d\n", ret);
 902                         goto errout;
 903                 }
 904                 newxprt->sc_max_sge = qp_attr.cap.max_send_sge;
 905                 newxprt->sc_max_sge = qp_attr.cap.max_recv_sge;
 906                 newxprt->sc_sq_depth = qp_attr.cap.max_send_wr;
 907                 newxprt->sc_max_requests = qp_attr.cap.max_recv_wr;
 908         }
 909         newxprt->sc_qp = newxprt->sc_cm_id->qp;
 910
 911         /* Register all of physical memory */
 912         newxprt->sc_phys_mr = ib_get_dma_mr(newxprt->sc_pd,
 913                                             IB_ACCESS_LOCAL_WRITE |
 914                                             IB_ACCESS_REMOTE_WRITE);
 915         if (IS_ERR(newxprt->sc_phys_mr)) {
 916                 dprintk("svcrdma: Failed to create DMA MR ret=%d\n", ret);
 917                 goto errout;
 918         }
 919
 920         /* Post receive buffers */
 921         for (i = 0; i < newxprt->sc_max_requests; i++) {
 922                 ret = svc_rdma_post_recv(newxprt);
 923                 if (ret) {
 924                         dprintk("svcrdma: failure posting receive buffers\n");
 925                         goto errout;
 926                 }
 927         }
 928
 929         /* Swap out the handler */
 930         newxprt->sc_cm_id->event_handler = rdma_cma_handler;
 931
 932         /*
 933          * Arm the CQs for the SQ and RQ before accepting so we can't
 934          * miss the first message
 935          */
 936         ib_req_notify_cq(newxprt->sc_sq_cq, IB_CQ_NEXT_COMP);
 937         ib_req_notify_cq(newxprt->sc_rq_cq, IB_CQ_NEXT_COMP);
 938
 939         /* Accept Connection */
 940         set_bit(RDMAXPRT_CONN_PENDING, &newxprt->sc_flags);
 941         memset(&conn_param, 0, sizeof conn_param);
 942         conn_param.responder_resources = 0;
 943         conn_param.initiator_depth = newxprt->sc_ord;
 944         ret = rdma_accept(newxprt->sc_cm_id, &conn_param);
 945         if (ret) {
 946                 dprintk("svcrdma: failed to accept new connection, ret=%d\n",
 947                        ret);
 948                 goto errout;
 949         }
 950
 951         dprintk("svcrdma: new connection %p accepted with the following "
 952                 "attributes:\n"
 953                 "    local_ip        : %d.%d.%d.%d\n"
 954                 "    local_port      : %d\n"
 955                 "    remote_ip       : %d.%d.%d.%d\n"
 956                 "    remote_port     : %d\n"
 957                 "    max_sge         : %d\n"
 958                 "    sq_depth        : %d\n"
 959                 "    max_requests    : %d\n"
 960                 "    ord             : %d\n",
 961                 newxprt,
 962                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
 963                          route.addr.src_addr)->sin_addr.s_addr),
 964                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
 965                        route.addr.src_addr)->sin_port),
 966                 NIPQUAD(((struct sockaddr_in *)&newxprt->sc_cm_id->
 967                          route.addr.dst_addr)->sin_addr.s_addr),
 968                 ntohs(((struct sockaddr_in *)&newxprt->sc_cm_id->
 969                        route.addr.dst_addr)->sin_port),
 970                 newxprt->sc_max_sge,
 971                 newxprt->sc_sq_depth,
 972                 newxprt->sc_max_requests,
 973                 newxprt->sc_ord);
 974
 975         return &newxprt->sc_xprt;
 976
 977  errout:
 978         dprintk("svcrdma: failure accepting new connection rc=%d.\n", ret);
 979         /* Take a reference in case the DTO handler runs */
 980         svc_xprt_get(&newxprt->sc_xprt);
 981         if (newxprt->sc_qp && !IS_ERR(newxprt->sc_qp))
 982                 ib_destroy_qp(newxprt->sc_qp);
 983         rdma_destroy_id(newxprt->sc_cm_id);
 984         /* This call to put will destroy the transport */
 985         svc_xprt_put(&newxprt->sc_xprt);
 986         return NULL;
 987 }
 988
 989 static void svc_rdma_release_rqst(struct svc_rqst *rqstp)
 990 {
 991 }
 992
 993 /*
 994  * When connected, an svc_xprt has at least two references:
 995  *
 996  * - A reference held by the cm_id between the ESTABLISHED and
 997  *   DISCONNECTED events. If the remote peer disconnected first, this
 998  *   reference could be gone.
 999  *
1000  * - A reference held by the svc_recv code that called this function
1001  *   as part of close processing.
1002  *
1003  * At a minimum one references should still be held.
1004  */
1005 static void svc_rdma_detach(struct svc_xprt *xprt)
1006 {
1007         struct svcxprt_rdma *rdma =
1008                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1009         dprintk("svc: svc_rdma_detach(%p)\n", xprt);
1010
1011         /* Disconnect and flush posted WQE */
1012         rdma_disconnect(rdma->sc_cm_id);
1013 }
1014
1015 static void __svc_rdma_free(struct work_struct *work)
1016 {
1017         struct svcxprt_rdma *rdma =
1018                 container_of(work, struct svcxprt_rdma, sc_work);
1019         dprintk("svcrdma: svc_rdma_free(%p)\n", rdma);
1020
1021         /* We should only be called from kref_put */
1022         BUG_ON(atomic_read(&rdma->sc_xprt.xpt_ref.refcount) != 0);
1023
1024         /*
1025          * Destroy queued, but not processed read completions. Note
1026          * that this cleanup has to be done before destroying the
1027          * cm_id because the device ptr is needed to unmap the dma in
1028          * svc_rdma_put_context.
1029          */
1030         spin_lock_bh(&rdma->sc_read_complete_lock);
1031         while (!list_empty(&rdma->sc_read_complete_q)) {
1032                 struct svc_rdma_op_ctxt *ctxt;
1033                 ctxt = list_entry(rdma->sc_read_complete_q.next,
1034                                   struct svc_rdma_op_ctxt,
1035                                   dto_q);
1036                 list_del_init(&ctxt->dto_q);
1037                 svc_rdma_put_context(ctxt, 1);
1038         }
1039         spin_unlock_bh(&rdma->sc_read_complete_lock);
1040
1041         /* Destroy queued, but not processed recv completions */
1042         spin_lock_bh(&rdma->sc_rq_dto_lock);
1043         while (!list_empty(&rdma->sc_rq_dto_q)) {
1044                 struct svc_rdma_op_ctxt *ctxt;
1045                 ctxt = list_entry(rdma->sc_rq_dto_q.next,
1046                                   struct svc_rdma_op_ctxt,
1047                                   dto_q);
1048                 list_del_init(&ctxt->dto_q);
1049                 svc_rdma_put_context(ctxt, 1);
1050         }
1051         spin_unlock_bh(&rdma->sc_rq_dto_lock);
1052
1053         /* Warn if we leaked a resource or under-referenced */
1054         WARN_ON(atomic_read(&rdma->sc_ctxt_used) != 0);
1055         WARN_ON(atomic_read(&rdma->sc_dma_used) != 0);
1056
1057         /* Destroy the QP if present (not a listener) */
1058         if (rdma->sc_qp && !IS_ERR(rdma->sc_qp))
1059                 ib_destroy_qp(rdma->sc_qp);
1060
1061         if (rdma->sc_sq_cq && !IS_ERR(rdma->sc_sq_cq))
1062                 ib_destroy_cq(rdma->sc_sq_cq);
1063
1064         if (rdma->sc_rq_cq && !IS_ERR(rdma->sc_rq_cq))
1065                 ib_destroy_cq(rdma->sc_rq_cq);
1066
1067         if (rdma->sc_phys_mr && !IS_ERR(rdma->sc_phys_mr))
1068                 ib_dereg_mr(rdma->sc_phys_mr);
1069
1070         if (rdma->sc_pd && !IS_ERR(rdma->sc_pd))
1071                 ib_dealloc_pd(rdma->sc_pd);
1072
1073         /* Destroy the CM ID */
1074         rdma_destroy_id(rdma->sc_cm_id);
1075
1076         destroy_context_cache(rdma);
1077         kfree(rdma);
1078 }
1079
1080 static void svc_rdma_free(struct svc_xprt *xprt)
1081 {
1082         struct svcxprt_rdma *rdma =
1083                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1084         INIT_WORK(&rdma->sc_work, __svc_rdma_free);
1085         schedule_work(&rdma->sc_work);
1086 }
1087
1088 static int svc_rdma_has_wspace(struct svc_xprt *xprt)
1089 {
1090         struct svcxprt_rdma *rdma =
1091                 container_of(xprt, struct svcxprt_rdma, sc_xprt);
1092
1093         /*
1094          * If there are fewer SQ WR available than required to send a
1095          * simple response, return false.
1096          */
1097         if ((rdma->sc_sq_depth - atomic_read(&rdma->sc_sq_count) < 3))
1098                 return 0;
1099
1100         /*
1101          * ...or there are already waiters on the SQ,
1102          * return false.
1103          */
1104         if (waitqueue_active(&rdma->sc_send_wait))
1105                 return 0;
1106
1107         /* Otherwise return true. */
1108         return 1;
1109 }
1110
1111 int svc_rdma_send(struct svcxprt_rdma *xprt, struct ib_send_wr *wr)
1112 {
1113         struct ib_send_wr *bad_wr;
1114         int ret;
1115
1116         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1117                 return -ENOTCONN;
1118
1119         BUG_ON(wr->send_flags != IB_SEND_SIGNALED);
1120         BUG_ON(((struct svc_rdma_op_ctxt *)(unsigned long)wr->wr_id)->wr_op !=
1121                 wr->opcode);
1122         /* If the SQ is full, wait until an SQ entry is available */
1123         while (1) {
1124                 spin_lock_bh(&xprt->sc_lock);
1125                 if (xprt->sc_sq_depth == atomic_read(&xprt->sc_sq_count)) {
1126                         spin_unlock_bh(&xprt->sc_lock);
1127                         atomic_inc(&rdma_stat_sq_starve);
1128
1129                         /* See if we can opportunistically reap SQ WR to make room */
1130                         sq_cq_reap(xprt);
1131
1132                         /* Wait until SQ WR available if SQ still full */
1133                         wait_event(xprt->sc_send_wait,
1134                                    atomic_read(&xprt->sc_sq_count) <
1135                                    xprt->sc_sq_depth);
1136                         if (test_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags))
1137                                 return 0;
1138                         continue;
1139                 }
1140                 /* Bumped used SQ WR count and post */
1141                 svc_xprt_get(&xprt->sc_xprt);
1142                 ret = ib_post_send(xprt->sc_qp, wr, &bad_wr);
1143                 if (!ret)
1144                         atomic_inc(&xprt->sc_sq_count);
1145                 else {
1146                         svc_xprt_put(&xprt->sc_xprt);
1147                         dprintk("svcrdma: failed to post SQ WR rc=%d, "
1148                                "sc_sq_count=%d, sc_sq_depth=%d\n",
1149                                ret, atomic_read(&xprt->sc_sq_count),
1150                                xprt->sc_sq_depth);
1151                 }
1152                 spin_unlock_bh(&xprt->sc_lock);
1153                 break;
1154         }
1155         return ret;
1156 }
1157
1158 void svc_rdma_send_error(struct svcxprt_rdma *xprt, struct rpcrdma_msg *rmsgp,
1159                          enum rpcrdma_errcode err)
1160 {
1161         struct ib_send_wr err_wr;
1162         struct ib_sge sge;
1163         struct page *p;
1164         struct svc_rdma_op_ctxt *ctxt;
1165         u32 *va;
1166         int length;
1167         int ret;
1168
1169         p = svc_rdma_get_page();
1170         va = page_address(p);
1171
1172         /* XDR encode error */
1173         length = svc_rdma_xdr_encode_error(xprt, rmsgp, err, va);
1174
1175         /* Prepare SGE for local address */
1176         atomic_inc(&xprt->sc_dma_used);
1177         sge.addr = ib_dma_map_page(xprt->sc_cm_id->device,
1178                                    p, 0, PAGE_SIZE, DMA_FROM_DEVICE);
1179         sge.lkey = xprt->sc_phys_mr->lkey;
1180         sge.length = length;
1181
1182         ctxt = svc_rdma_get_context(xprt);
1183         ctxt->count = 1;
1184         ctxt->pages[0] = p;
1185
1186         /* Prepare SEND WR */
1187         memset(&err_wr, 0, sizeof err_wr);
1188         ctxt->wr_op = IB_WR_SEND;
1189         err_wr.wr_id = (unsigned long)ctxt;
1190         err_wr.sg_list = &sge;
1191         err_wr.num_sge = 1;
1192         err_wr.opcode = IB_WR_SEND;
1193         err_wr.send_flags = IB_SEND_SIGNALED;
1194
1195         /* Post It */
1196         ret = svc_rdma_send(xprt, &err_wr);
1197         if (ret) {
1198                 dprintk("svcrdma: Error %d posting send for protocol error\n",
1199                         ret);
1200                 svc_rdma_put_context(ctxt, 1);
1201         }
1202 }