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