net/sunrpc/xprtrdma/verbs.c

   1 /*
   2  * Copyright (c) 2003-2007 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
  40 /*
  41  * verbs.c
  42  *
  43  * Encapsulates the major functions managing:
  44  *  o adapters
  45  *  o endpoints
  46  *  o connections
  47  *  o buffer memory
  48  */
  49
  50 #include <linux/interrupt.h>
  51 #include <linux/slab.h>
  52 #include <linux/prefetch.h>
  53 #include <linux/sunrpc/addr.h>
  54 #include <linux/sunrpc/svc_rdma.h>
  55 #include <asm/bitops.h>
  56 #include <linux/module.h> /* try_module_get()/module_put() */
  57
  58 #include "xprt_rdma.h"
  59
  60 /*
  61  * Globals/Macros
  62  */
  63
  64 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  65 # define RPCDBG_FACILITY        RPCDBG_TRANS
  66 #endif
  67
  68 /*
  69  * internal functions
  70  */
  71
  72 static struct workqueue_struct *rpcrdma_receive_wq;
  73
  74 int
  75 rpcrdma_alloc_wq(void)
  76 {
  77         struct workqueue_struct *recv_wq;
  78
  79         recv_wq = alloc_workqueue("xprtrdma_receive",
  80                                   WQ_MEM_RECLAIM | WQ_UNBOUND | WQ_HIGHPRI,
  81                                   0);
  82         if (!recv_wq)
  83                 return -ENOMEM;
  84
  85         rpcrdma_receive_wq = recv_wq;
  86         return 0;
  87 }
  88
  89 void
  90 rpcrdma_destroy_wq(void)
  91 {
  92         struct workqueue_struct *wq;
  93
  94         if (rpcrdma_receive_wq) {
  95                 wq = rpcrdma_receive_wq;
  96                 rpcrdma_receive_wq = NULL;
  97                 destroy_workqueue(wq);
  98         }
  99 }
 100
 101 static void
 102 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
 103 {
 104         struct rpcrdma_ep *ep = context;
 105
 106         pr_err("RPC:       %s: %s on device %s ep %p\n",
 107                __func__, ib_event_msg(event->event),
 108                 event->device->name, context);
 109         if (ep->rep_connected == 1) {
 110                 ep->rep_connected = -EIO;
 111                 rpcrdma_conn_func(ep);
 112                 wake_up_all(&ep->rep_connect_wait);
 113         }
 114 }
 115
 116 /**
 117  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
 118  * @cq: completion queue (ignored)
 119  * @wc: completed WR
 120  *
 121  */
 122 static void
 123 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
 124 {
 125         /* WARNING: Only wr_cqe and status are reliable at this point */
 126         if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
 127                 pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
 128                        ib_wc_status_msg(wc->status),
 129                        wc->status, wc->vendor_err);
 130 }
 131
 132 static void
 133 rpcrdma_receive_worker(struct work_struct *work)
 134 {
 135         struct rpcrdma_rep *rep =
 136                         container_of(work, struct rpcrdma_rep, rr_work);
 137
 138         rpcrdma_reply_handler(rep);
 139 }
 140
 141 /* Perform basic sanity checking to avoid using garbage
 142  * to update the credit grant value.
 143  */
 144 static void
 145 rpcrdma_update_granted_credits(struct rpcrdma_rep *rep)
 146 {
 147         struct rpcrdma_msg *rmsgp = rdmab_to_msg(rep->rr_rdmabuf);
 148         struct rpcrdma_buffer *buffer = &rep->rr_rxprt->rx_buf;
 149         u32 credits;
 150
 151         if (rep->rr_len < RPCRDMA_HDRLEN_ERR)
 152                 return;
 153
 154         credits = be32_to_cpu(rmsgp->rm_credit);
 155         if (credits == 0)
 156                 credits = 1;    /* don't deadlock */
 157         else if (credits > buffer->rb_max_requests)
 158                 credits = buffer->rb_max_requests;
 159
 160         atomic_set(&buffer->rb_credits, credits);
 161 }
 162
 163 /**
 164  * rpcrdma_receive_wc - Invoked by RDMA provider for each polled Receive WC
 165  * @cq: completion queue (ignored)
 166  * @wc: completed WR
 167  *
 168  */
 169 static void
 170 rpcrdma_receive_wc(struct ib_cq *cq, struct ib_wc *wc)
 171 {
 172         struct ib_cqe *cqe = wc->wr_cqe;
 173         struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
 174                                                rr_cqe);
 175
 176         /* WARNING: Only wr_id and status are reliable at this point */
 177         if (wc->status != IB_WC_SUCCESS)
 178                 goto out_fail;
 179
 180         /* status == SUCCESS means all fields in wc are trustworthy */
 181         if (wc->opcode != IB_WC_RECV)
 182                 return;
 183
 184         dprintk("RPC:       %s: rep %p opcode 'recv', length %u: success\n",
 185                 __func__, rep, wc->byte_len);
 186
 187         rep->rr_len = wc->byte_len;
 188         ib_dma_sync_single_for_cpu(rep->rr_device,
 189                                    rdmab_addr(rep->rr_rdmabuf),
 190                                    rep->rr_len, DMA_FROM_DEVICE);
 191
 192         rpcrdma_update_granted_credits(rep);
 193
 194 out_schedule:
 195         queue_work(rpcrdma_receive_wq, &rep->rr_work);
 196         return;
 197
 198 out_fail:
 199         if (wc->status != IB_WC_WR_FLUSH_ERR)
 200                 pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
 201                        ib_wc_status_msg(wc->status),
 202                        wc->status, wc->vendor_err);
 203         rep->rr_len = RPCRDMA_BAD_LEN;
 204         goto out_schedule;
 205 }
 206
 207 static int
 208 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
 209 {
 210         struct rpcrdma_xprt *xprt = id->context;
 211         struct rpcrdma_ia *ia = &xprt->rx_ia;
 212         struct rpcrdma_ep *ep = &xprt->rx_ep;
 213 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 214         struct sockaddr *sap = (struct sockaddr *)&ep->rep_remote_addr;
 215 #endif
 216         struct ib_qp_attr *attr = &ia->ri_qp_attr;
 217         struct ib_qp_init_attr *iattr = &ia->ri_qp_init_attr;
 218         int connstate = 0;
 219
 220         switch (event->event) {
 221         case RDMA_CM_EVENT_ADDR_RESOLVED:
 222         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 223                 ia->ri_async_rc = 0;
 224                 complete(&ia->ri_done);
 225                 break;
 226         case RDMA_CM_EVENT_ADDR_ERROR:
 227                 ia->ri_async_rc = -EHOSTUNREACH;
 228                 dprintk("RPC:       %s: CM address resolution error, ep 0x%p\n",
 229                         __func__, ep);
 230                 complete(&ia->ri_done);
 231                 break;
 232         case RDMA_CM_EVENT_ROUTE_ERROR:
 233                 ia->ri_async_rc = -ENETUNREACH;
 234                 dprintk("RPC:       %s: CM route resolution error, ep 0x%p\n",
 235                         __func__, ep);
 236                 complete(&ia->ri_done);
 237                 break;
 238         case RDMA_CM_EVENT_ESTABLISHED:
 239                 connstate = 1;
 240                 ib_query_qp(ia->ri_id->qp, attr,
 241                             IB_QP_MAX_QP_RD_ATOMIC | IB_QP_MAX_DEST_RD_ATOMIC,
 242                             iattr);
 243                 dprintk("RPC:       %s: %d responder resources"
 244                         " (%d initiator)\n",
 245                         __func__, attr->max_dest_rd_atomic,
 246                         attr->max_rd_atomic);
 247                 goto connected;
 248         case RDMA_CM_EVENT_CONNECT_ERROR:
 249                 connstate = -ENOTCONN;
 250                 goto connected;
 251         case RDMA_CM_EVENT_UNREACHABLE:
 252                 connstate = -ENETDOWN;
 253                 goto connected;
 254         case RDMA_CM_EVENT_REJECTED:
 255                 connstate = -ECONNREFUSED;
 256                 goto connected;
 257         case RDMA_CM_EVENT_DISCONNECTED:
 258                 connstate = -ECONNABORTED;
 259                 goto connected;
 260         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 261                 connstate = -ENODEV;
 262 connected:
 263                 dprintk("RPC:       %s: %sconnected\n",
 264                                         __func__, connstate > 0 ? "" : "dis");
 265                 atomic_set(&xprt->rx_buf.rb_credits, 1);
 266                 ep->rep_connected = connstate;
 267                 rpcrdma_conn_func(ep);
 268                 wake_up_all(&ep->rep_connect_wait);
 269                 /*FALLTHROUGH*/
 270         default:
 271                 dprintk("RPC:       %s: %pIS:%u (ep 0x%p): %s\n",
 272                         __func__, sap, rpc_get_port(sap), ep,
 273                         rdma_event_msg(event->event));
 274                 break;
 275         }
 276
 277 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 278         if (connstate == 1) {
 279                 int ird = attr->max_dest_rd_atomic;
 280                 int tird = ep->rep_remote_cma.responder_resources;
 281
 282                 pr_info("rpcrdma: connection to %pIS:%u on %s, memreg '%s', %d credits, %d responders%s\n",
 283                         sap, rpc_get_port(sap),
 284                         ia->ri_device->name,
 285                         ia->ri_ops->ro_displayname,
 286                         xprt->rx_buf.rb_max_requests,
 287                         ird, ird < 4 && ird < tird / 2 ? " (low!)" : "");
 288         } else if (connstate < 0) {
 289                 pr_info("rpcrdma: connection to %pIS:%u closed (%d)\n",
 290                         sap, rpc_get_port(sap), connstate);
 291         }
 292 #endif
 293
 294         return 0;
 295 }
 296
 297 static void rpcrdma_destroy_id(struct rdma_cm_id *id)
 298 {
 299         if (id) {
 300                 module_put(id->device->owner);
 301                 rdma_destroy_id(id);
 302         }
 303 }
 304
 305 static struct rdma_cm_id *
 306 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
 307                         struct rpcrdma_ia *ia, struct sockaddr *addr)
 308 {
 309         struct rdma_cm_id *id;
 310         int rc;
 311
 312         init_completion(&ia->ri_done);
 313
 314         id = rdma_create_id(&init_net, rpcrdma_conn_upcall, xprt, RDMA_PS_TCP,
 315                             IB_QPT_RC);
 316         if (IS_ERR(id)) {
 317                 rc = PTR_ERR(id);
 318                 dprintk("RPC:       %s: rdma_create_id() failed %i\n",
 319                         __func__, rc);
 320                 return id;
 321         }
 322
 323         ia->ri_async_rc = -ETIMEDOUT;
 324         rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
 325         if (rc) {
 326                 dprintk("RPC:       %s: rdma_resolve_addr() failed %i\n",
 327                         __func__, rc);
 328                 goto out;
 329         }
 330         wait_for_completion_interruptible_timeout(&ia->ri_done,
 331                                 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
 332
 333         /* FIXME:
 334          * Until xprtrdma supports DEVICE_REMOVAL, the provider must
 335          * be pinned while there are active NFS/RDMA mounts to prevent
 336          * hangs and crashes at umount time.
 337          */
 338         if (!ia->ri_async_rc && !try_module_get(id->device->owner)) {
 339                 dprintk("RPC:       %s: Failed to get device module\n",
 340                         __func__);
 341                 ia->ri_async_rc = -ENODEV;
 342         }
 343         rc = ia->ri_async_rc;
 344         if (rc)
 345                 goto out;
 346
 347         ia->ri_async_rc = -ETIMEDOUT;
 348         rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
 349         if (rc) {
 350                 dprintk("RPC:       %s: rdma_resolve_route() failed %i\n",
 351                         __func__, rc);
 352                 goto put;
 353         }
 354         wait_for_completion_interruptible_timeout(&ia->ri_done,
 355                                 msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1);
 356         rc = ia->ri_async_rc;
 357         if (rc)
 358                 goto put;
 359
 360         return id;
 361 put:
 362         module_put(id->device->owner);
 363 out:
 364         rdma_destroy_id(id);
 365         return ERR_PTR(rc);
 366 }
 367
 368 /*
 369  * Exported functions.
 370  */
 371
 372 /*
 373  * Open and initialize an Interface Adapter.
 374  *  o initializes fields of struct rpcrdma_ia, including
 375  *    interface and provider attributes and protection zone.
 376  */
 377 int
 378 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr, int memreg)
 379 {
 380         struct rpcrdma_ia *ia = &xprt->rx_ia;
 381         int rc;
 382
 383         ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
 384         if (IS_ERR(ia->ri_id)) {
 385                 rc = PTR_ERR(ia->ri_id);
 386                 goto out1;
 387         }
 388         ia->ri_device = ia->ri_id->device;
 389
 390         ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
 391         if (IS_ERR(ia->ri_pd)) {
 392                 rc = PTR_ERR(ia->ri_pd);
 393                 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
 394                 goto out2;
 395         }
 396
 397         switch (memreg) {
 398         case RPCRDMA_FRMR:
 399                 if (frwr_is_supported(ia)) {
 400                         ia->ri_ops = &rpcrdma_frwr_memreg_ops;
 401                         break;
 402                 }
 403                 /*FALLTHROUGH*/
 404         case RPCRDMA_MTHCAFMR:
 405                 if (fmr_is_supported(ia)) {
 406                         ia->ri_ops = &rpcrdma_fmr_memreg_ops;
 407                         break;
 408                 }
 409                 /*FALLTHROUGH*/
 410         default:
 411                 pr_err("rpcrdma: Unsupported memory registration mode: %d\n",
 412                        memreg);
 413                 rc = -EINVAL;
 414                 goto out3;
 415         }
 416
 417         return 0;
 418
 419 out3:
 420         ib_dealloc_pd(ia->ri_pd);
 421         ia->ri_pd = NULL;
 422 out2:
 423         rpcrdma_destroy_id(ia->ri_id);
 424         ia->ri_id = NULL;
 425 out1:
 426         return rc;
 427 }
 428
 429 /*
 430  * Clean up/close an IA.
 431  *   o if event handles and PD have been initialized, free them.
 432  *   o close the IA
 433  */
 434 void
 435 rpcrdma_ia_close(struct rpcrdma_ia *ia)
 436 {
 437         dprintk("RPC:       %s: entering\n", __func__);
 438         if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
 439                 if (ia->ri_id->qp)
 440                         rdma_destroy_qp(ia->ri_id);
 441                 rpcrdma_destroy_id(ia->ri_id);
 442                 ia->ri_id = NULL;
 443         }
 444
 445         /* If the pd is still busy, xprtrdma missed freeing a resource */
 446         if (ia->ri_pd && !IS_ERR(ia->ri_pd))
 447                 ib_dealloc_pd(ia->ri_pd);
 448 }
 449
 450 /*
 451  * Create unconnected endpoint.
 452  */
 453 int
 454 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
 455                                 struct rpcrdma_create_data_internal *cdata)
 456 {
 457         struct ib_cq *sendcq, *recvcq;
 458         unsigned int max_qp_wr;
 459         int rc;
 460
 461         if (ia->ri_device->attrs.max_sge < RPCRDMA_MAX_IOVS) {
 462                 dprintk("RPC:       %s: insufficient sge's available\n",
 463                         __func__);
 464                 return -ENOMEM;
 465         }
 466
 467         if (ia->ri_device->attrs.max_qp_wr <= RPCRDMA_BACKWARD_WRS) {
 468                 dprintk("RPC:       %s: insufficient wqe's available\n",
 469                         __func__);
 470                 return -ENOMEM;
 471         }
 472         max_qp_wr = ia->ri_device->attrs.max_qp_wr - RPCRDMA_BACKWARD_WRS - 1;
 473
 474         /* check provider's send/recv wr limits */
 475         if (cdata->max_requests > max_qp_wr)
 476                 cdata->max_requests = max_qp_wr;
 477
 478         ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
 479         ep->rep_attr.qp_context = ep;
 480         ep->rep_attr.srq = NULL;
 481         ep->rep_attr.cap.max_send_wr = cdata->max_requests;
 482         ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
 483         ep->rep_attr.cap.max_send_wr += 1;      /* drain cqe */
 484         rc = ia->ri_ops->ro_open(ia, ep, cdata);
 485         if (rc)
 486                 return rc;
 487         ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
 488         ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
 489         ep->rep_attr.cap.max_recv_wr += 1;      /* drain cqe */
 490         ep->rep_attr.cap.max_send_sge = RPCRDMA_MAX_IOVS;
 491         ep->rep_attr.cap.max_recv_sge = 1;
 492         ep->rep_attr.cap.max_inline_data = 0;
 493         ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 494         ep->rep_attr.qp_type = IB_QPT_RC;
 495         ep->rep_attr.port_num = ~0;
 496
 497         dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
 498                 "iovs: send %d recv %d\n",
 499                 __func__,
 500                 ep->rep_attr.cap.max_send_wr,
 501                 ep->rep_attr.cap.max_recv_wr,
 502                 ep->rep_attr.cap.max_send_sge,
 503                 ep->rep_attr.cap.max_recv_sge);
 504
 505         /* set trigger for requesting send completion */
 506         ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 - 1;
 507         if (ep->rep_cqinit <= 2)
 508                 ep->rep_cqinit = 0;     /* always signal? */
 509         INIT_CQCOUNT(ep);
 510         init_waitqueue_head(&ep->rep_connect_wait);
 511         INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
 512
 513         sendcq = ib_alloc_cq(ia->ri_device, NULL,
 514                              ep->rep_attr.cap.max_send_wr + 1,
 515                              0, IB_POLL_SOFTIRQ);
 516         if (IS_ERR(sendcq)) {
 517                 rc = PTR_ERR(sendcq);
 518                 dprintk("RPC:       %s: failed to create send CQ: %i\n",
 519                         __func__, rc);
 520                 goto out1;
 521         }
 522
 523         recvcq = ib_alloc_cq(ia->ri_device, NULL,
 524                              ep->rep_attr.cap.max_recv_wr + 1,
 525                              0, IB_POLL_SOFTIRQ);
 526         if (IS_ERR(recvcq)) {
 527                 rc = PTR_ERR(recvcq);
 528                 dprintk("RPC:       %s: failed to create recv CQ: %i\n",
 529                         __func__, rc);
 530                 goto out2;
 531         }
 532
 533         ep->rep_attr.send_cq = sendcq;
 534         ep->rep_attr.recv_cq = recvcq;
 535
 536         /* Initialize cma parameters */
 537         memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
 538
 539         /* RPC/RDMA does not use private data */
 540         ep->rep_remote_cma.private_data = NULL;
 541         ep->rep_remote_cma.private_data_len = 0;
 542
 543         /* Client offers RDMA Read but does not initiate */
 544         ep->rep_remote_cma.initiator_depth = 0;
 545         if (ia->ri_device->attrs.max_qp_rd_atom > 32)   /* arbitrary but <= 255 */
 546                 ep->rep_remote_cma.responder_resources = 32;
 547         else
 548                 ep->rep_remote_cma.responder_resources =
 549                                                 ia->ri_device->attrs.max_qp_rd_atom;
 550
 551         /* Limit transport retries so client can detect server
 552          * GID changes quickly. RPC layer handles re-establishing
 553          * transport connection and retransmission.
 554          */
 555         ep->rep_remote_cma.retry_count = 6;
 556
 557         /* RPC-over-RDMA handles its own flow control. In addition,
 558          * make all RNR NAKs visible so we know that RPC-over-RDMA
 559          * flow control is working correctly (no NAKs should be seen).
 560          */
 561         ep->rep_remote_cma.flow_control = 0;
 562         ep->rep_remote_cma.rnr_retry_count = 0;
 563
 564         return 0;
 565
 566 out2:
 567         ib_free_cq(sendcq);
 568 out1:
 569         return rc;
 570 }
 571
 572 /*
 573  * rpcrdma_ep_destroy
 574  *
 575  * Disconnect and destroy endpoint. After this, the only
 576  * valid operations on the ep are to free it (if dynamically
 577  * allocated) or re-create it.
 578  */
 579 void
 580 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 581 {
 582         dprintk("RPC:       %s: entering, connected is %d\n",
 583                 __func__, ep->rep_connected);
 584
 585         cancel_delayed_work_sync(&ep->rep_connect_worker);
 586
 587         if (ia->ri_id->qp) {
 588                 rpcrdma_ep_disconnect(ep, ia);
 589                 rdma_destroy_qp(ia->ri_id);
 590                 ia->ri_id->qp = NULL;
 591         }
 592
 593         ib_free_cq(ep->rep_attr.recv_cq);
 594         ib_free_cq(ep->rep_attr.send_cq);
 595 }
 596
 597 /*
 598  * Connect unconnected endpoint.
 599  */
 600 int
 601 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 602 {
 603         struct rdma_cm_id *id, *old;
 604         int rc = 0;
 605         int retry_count = 0;
 606
 607         if (ep->rep_connected != 0) {
 608                 struct rpcrdma_xprt *xprt;
 609 retry:
 610                 dprintk("RPC:       %s: reconnecting...\n", __func__);
 611
 612                 rpcrdma_ep_disconnect(ep, ia);
 613
 614                 xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
 615                 id = rpcrdma_create_id(xprt, ia,
 616                                 (struct sockaddr *)&xprt->rx_data.addr);
 617                 if (IS_ERR(id)) {
 618                         rc = -EHOSTUNREACH;
 619                         goto out;
 620                 }
 621                 /* TEMP TEMP TEMP - fail if new device:
 622                  * Deregister/remarshal *all* requests!
 623                  * Close and recreate adapter, pd, etc!
 624                  * Re-determine all attributes still sane!
 625                  * More stuff I haven't thought of!
 626                  * Rrrgh!
 627                  */
 628                 if (ia->ri_device != id->device) {
 629                         printk("RPC:       %s: can't reconnect on "
 630                                 "different device!\n", __func__);
 631                         rpcrdma_destroy_id(id);
 632                         rc = -ENETUNREACH;
 633                         goto out;
 634                 }
 635                 /* END TEMP */
 636                 rc = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
 637                 if (rc) {
 638                         dprintk("RPC:       %s: rdma_create_qp failed %i\n",
 639                                 __func__, rc);
 640                         rpcrdma_destroy_id(id);
 641                         rc = -ENETUNREACH;
 642                         goto out;
 643                 }
 644
 645                 old = ia->ri_id;
 646                 ia->ri_id = id;
 647
 648                 rdma_destroy_qp(old);
 649                 rpcrdma_destroy_id(old);
 650         } else {
 651                 dprintk("RPC:       %s: connecting...\n", __func__);
 652                 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
 653                 if (rc) {
 654                         dprintk("RPC:       %s: rdma_create_qp failed %i\n",
 655                                 __func__, rc);
 656                         /* do not update ep->rep_connected */
 657                         return -ENETUNREACH;
 658                 }
 659         }
 660
 661         ep->rep_connected = 0;
 662
 663         rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
 664         if (rc) {
 665                 dprintk("RPC:       %s: rdma_connect() failed with %i\n",
 666                                 __func__, rc);
 667                 goto out;
 668         }
 669
 670         wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
 671
 672         /*
 673          * Check state. A non-peer reject indicates no listener
 674          * (ECONNREFUSED), which may be a transient state. All
 675          * others indicate a transport condition which has already
 676          * undergone a best-effort.
 677          */
 678         if (ep->rep_connected == -ECONNREFUSED &&
 679             ++retry_count <= RDMA_CONNECT_RETRY_MAX) {
 680                 dprintk("RPC:       %s: non-peer_reject, retry\n", __func__);
 681                 goto retry;
 682         }
 683         if (ep->rep_connected <= 0) {
 684                 /* Sometimes, the only way to reliably connect to remote
 685                  * CMs is to use same nonzero values for ORD and IRD. */
 686                 if (retry_count++ <= RDMA_CONNECT_RETRY_MAX + 1 &&
 687                     (ep->rep_remote_cma.responder_resources == 0 ||
 688                      ep->rep_remote_cma.initiator_depth !=
 689                                 ep->rep_remote_cma.responder_resources)) {
 690                         if (ep->rep_remote_cma.responder_resources == 0)
 691                                 ep->rep_remote_cma.responder_resources = 1;
 692                         ep->rep_remote_cma.initiator_depth =
 693                                 ep->rep_remote_cma.responder_resources;
 694                         goto retry;
 695                 }
 696                 rc = ep->rep_connected;
 697         } else {
 698                 struct rpcrdma_xprt *r_xprt;
 699                 unsigned int extras;
 700
 701                 dprintk("RPC:       %s: connected\n", __func__);
 702
 703                 r_xprt = container_of(ia, struct rpcrdma_xprt, rx_ia);
 704                 extras = r_xprt->rx_buf.rb_bc_srv_max_requests;
 705
 706                 if (extras) {
 707                         rc = rpcrdma_ep_post_extra_recv(r_xprt, extras);
 708                         if (rc) {
 709                                 pr_warn("%s: rpcrdma_ep_post_extra_recv: %i\n",
 710                                         __func__, rc);
 711                                 rc = 0;
 712                         }
 713                 }
 714         }
 715
 716 out:
 717         if (rc)
 718                 ep->rep_connected = rc;
 719         return rc;
 720 }
 721
 722 /*
 723  * rpcrdma_ep_disconnect
 724  *
 725  * This is separate from destroy to facilitate the ability
 726  * to reconnect without recreating the endpoint.
 727  *
 728  * This call is not reentrant, and must not be made in parallel
 729  * on the same endpoint.
 730  */
 731 void
 732 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 733 {
 734         int rc;
 735
 736         rc = rdma_disconnect(ia->ri_id);
 737         if (!rc) {
 738                 /* returns without wait if not connected */
 739                 wait_event_interruptible(ep->rep_connect_wait,
 740                                                         ep->rep_connected != 1);
 741                 dprintk("RPC:       %s: after wait, %sconnected\n", __func__,
 742                         (ep->rep_connected == 1) ? "still " : "dis");
 743         } else {
 744                 dprintk("RPC:       %s: rdma_disconnect %i\n", __func__, rc);
 745                 ep->rep_connected = rc;
 746         }
 747
 748         ib_drain_qp(ia->ri_id->qp);
 749 }
 750
 751 static void
 752 rpcrdma_mr_recovery_worker(struct work_struct *work)
 753 {
 754         struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
 755                                                   rb_recovery_worker.work);
 756         struct rpcrdma_mw *mw;
 757
 758         spin_lock(&buf->rb_recovery_lock);
 759         while (!list_empty(&buf->rb_stale_mrs)) {
 760                 mw = list_first_entry(&buf->rb_stale_mrs,
 761                                       struct rpcrdma_mw, mw_list);
 762                 list_del_init(&mw->mw_list);
 763                 spin_unlock(&buf->rb_recovery_lock);
 764
 765                 dprintk("RPC:       %s: recovering MR %p\n", __func__, mw);
 766                 mw->mw_xprt->rx_ia.ri_ops->ro_recover_mr(mw);
 767
 768                 spin_lock(&buf->rb_recovery_lock);
 769         }
 770         spin_unlock(&buf->rb_recovery_lock);
 771 }
 772
 773 void
 774 rpcrdma_defer_mr_recovery(struct rpcrdma_mw *mw)
 775 {
 776         struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
 777         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 778
 779         spin_lock(&buf->rb_recovery_lock);
 780         list_add(&mw->mw_list, &buf->rb_stale_mrs);
 781         spin_unlock(&buf->rb_recovery_lock);
 782
 783         schedule_delayed_work(&buf->rb_recovery_worker, 0);
 784 }
 785
 786 static void
 787 rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt)
 788 {
 789         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 790         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 791         unsigned int count;
 792         LIST_HEAD(free);
 793         LIST_HEAD(all);
 794
 795         for (count = 0; count < 32; count++) {
 796                 struct rpcrdma_mw *mw;
 797                 int rc;
 798
 799                 mw = kzalloc(sizeof(*mw), GFP_KERNEL);
 800                 if (!mw)
 801                         break;
 802
 803                 rc = ia->ri_ops->ro_init_mr(ia, mw);
 804                 if (rc) {
 805                         kfree(mw);
 806                         break;
 807                 }
 808
 809                 mw->mw_xprt = r_xprt;
 810
 811                 list_add(&mw->mw_list, &free);
 812                 list_add(&mw->mw_all, &all);
 813         }
 814
 815         spin_lock(&buf->rb_mwlock);
 816         list_splice(&free, &buf->rb_mws);
 817         list_splice(&all, &buf->rb_all);
 818         r_xprt->rx_stats.mrs_allocated += count;
 819         spin_unlock(&buf->rb_mwlock);
 820
 821         dprintk("RPC:       %s: created %u MRs\n", __func__, count);
 822 }
 823
 824 static void
 825 rpcrdma_mr_refresh_worker(struct work_struct *work)
 826 {
 827         struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
 828                                                   rb_refresh_worker.work);
 829         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
 830                                                    rx_buf);
 831
 832         rpcrdma_create_mrs(r_xprt);
 833 }
 834
 835 struct rpcrdma_req *
 836 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
 837 {
 838         struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
 839         struct rpcrdma_req *req;
 840
 841         req = kzalloc(sizeof(*req), GFP_KERNEL);
 842         if (req == NULL)
 843                 return ERR_PTR(-ENOMEM);
 844
 845         INIT_LIST_HEAD(&req->rl_free);
 846         spin_lock(&buffer->rb_reqslock);
 847         list_add(&req->rl_all, &buffer->rb_allreqs);
 848         spin_unlock(&buffer->rb_reqslock);
 849         req->rl_cqe.done = rpcrdma_wc_send;
 850         req->rl_buffer = &r_xprt->rx_buf;
 851         INIT_LIST_HEAD(&req->rl_registered);
 852         return req;
 853 }
 854
 855 struct rpcrdma_rep *
 856 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
 857 {
 858         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
 859         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 860         struct rpcrdma_rep *rep;
 861         int rc;
 862
 863         rc = -ENOMEM;
 864         rep = kzalloc(sizeof(*rep), GFP_KERNEL);
 865         if (rep == NULL)
 866                 goto out;
 867
 868         rep->rr_rdmabuf = rpcrdma_alloc_regbuf(ia, cdata->inline_rsize,
 869                                                GFP_KERNEL);
 870         if (IS_ERR(rep->rr_rdmabuf)) {
 871                 rc = PTR_ERR(rep->rr_rdmabuf);
 872                 goto out_free;
 873         }
 874
 875         rep->rr_device = ia->ri_device;
 876         rep->rr_cqe.done = rpcrdma_receive_wc;
 877         rep->rr_rxprt = r_xprt;
 878         INIT_WORK(&rep->rr_work, rpcrdma_receive_worker);
 879         return rep;
 880
 881 out_free:
 882         kfree(rep);
 883 out:
 884         return ERR_PTR(rc);
 885 }
 886
 887 int
 888 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
 889 {
 890         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 891         int i, rc;
 892
 893         buf->rb_max_requests = r_xprt->rx_data.max_requests;
 894         buf->rb_bc_srv_max_requests = 0;
 895         atomic_set(&buf->rb_credits, 1);
 896         spin_lock_init(&buf->rb_mwlock);
 897         spin_lock_init(&buf->rb_lock);
 898         spin_lock_init(&buf->rb_recovery_lock);
 899         INIT_LIST_HEAD(&buf->rb_mws);
 900         INIT_LIST_HEAD(&buf->rb_all);
 901         INIT_LIST_HEAD(&buf->rb_stale_mrs);
 902         INIT_DELAYED_WORK(&buf->rb_refresh_worker,
 903                           rpcrdma_mr_refresh_worker);
 904         INIT_DELAYED_WORK(&buf->rb_recovery_worker,
 905                           rpcrdma_mr_recovery_worker);
 906
 907         rpcrdma_create_mrs(r_xprt);
 908
 909         INIT_LIST_HEAD(&buf->rb_send_bufs);
 910         INIT_LIST_HEAD(&buf->rb_allreqs);
 911         spin_lock_init(&buf->rb_reqslock);
 912         for (i = 0; i < buf->rb_max_requests; i++) {
 913                 struct rpcrdma_req *req;
 914
 915                 req = rpcrdma_create_req(r_xprt);
 916                 if (IS_ERR(req)) {
 917                         dprintk("RPC:       %s: request buffer %d alloc"
 918                                 " failed\n", __func__, i);
 919                         rc = PTR_ERR(req);
 920                         goto out;
 921                 }
 922                 req->rl_backchannel = false;
 923                 list_add(&req->rl_free, &buf->rb_send_bufs);
 924         }
 925
 926         INIT_LIST_HEAD(&buf->rb_recv_bufs);
 927         for (i = 0; i < buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS; i++) {
 928                 struct rpcrdma_rep *rep;
 929
 930                 rep = rpcrdma_create_rep(r_xprt);
 931                 if (IS_ERR(rep)) {
 932                         dprintk("RPC:       %s: reply buffer %d alloc failed\n",
 933                                 __func__, i);
 934                         rc = PTR_ERR(rep);
 935                         goto out;
 936                 }
 937                 list_add(&rep->rr_list, &buf->rb_recv_bufs);
 938         }
 939
 940         return 0;
 941 out:
 942         rpcrdma_buffer_destroy(buf);
 943         return rc;
 944 }
 945
 946 static struct rpcrdma_req *
 947 rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer *buf)
 948 {
 949         struct rpcrdma_req *req;
 950
 951         req = list_first_entry(&buf->rb_send_bufs,
 952                                struct rpcrdma_req, rl_free);
 953         list_del(&req->rl_free);
 954         return req;
 955 }
 956
 957 static struct rpcrdma_rep *
 958 rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer *buf)
 959 {
 960         struct rpcrdma_rep *rep;
 961
 962         rep = list_first_entry(&buf->rb_recv_bufs,
 963                                struct rpcrdma_rep, rr_list);
 964         list_del(&rep->rr_list);
 965         return rep;
 966 }
 967
 968 static void
 969 rpcrdma_destroy_rep(struct rpcrdma_ia *ia, struct rpcrdma_rep *rep)
 970 {
 971         rpcrdma_free_regbuf(ia, rep->rr_rdmabuf);
 972         kfree(rep);
 973 }
 974
 975 void
 976 rpcrdma_destroy_req(struct rpcrdma_ia *ia, struct rpcrdma_req *req)
 977 {
 978         rpcrdma_free_regbuf(ia, req->rl_sendbuf);
 979         rpcrdma_free_regbuf(ia, req->rl_rdmabuf);
 980         kfree(req);
 981 }
 982
 983 static void
 984 rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf)
 985 {
 986         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
 987                                                    rx_buf);
 988         struct rpcrdma_ia *ia = rdmab_to_ia(buf);
 989         struct rpcrdma_mw *mw;
 990         unsigned int count;
 991
 992         count = 0;
 993         spin_lock(&buf->rb_mwlock);
 994         while (!list_empty(&buf->rb_all)) {
 995                 mw = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
 996                 list_del(&mw->mw_all);
 997
 998                 spin_unlock(&buf->rb_mwlock);
 999                 ia->ri_ops->ro_release_mr(mw);
1000                 count++;
1001                 spin_lock(&buf->rb_mwlock);
1002         }
1003         spin_unlock(&buf->rb_mwlock);
1004         r_xprt->rx_stats.mrs_allocated = 0;
1005
1006         dprintk("RPC:       %s: released %u MRs\n", __func__, count);
1007 }
1008
1009 void
1010 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1011 {
1012         struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1013
1014         cancel_delayed_work_sync(&buf->rb_recovery_worker);
1015
1016         while (!list_empty(&buf->rb_recv_bufs)) {
1017                 struct rpcrdma_rep *rep;
1018
1019                 rep = rpcrdma_buffer_get_rep_locked(buf);
1020                 rpcrdma_destroy_rep(ia, rep);
1021         }
1022         buf->rb_send_count = 0;
1023
1024         spin_lock(&buf->rb_reqslock);
1025         while (!list_empty(&buf->rb_allreqs)) {
1026                 struct rpcrdma_req *req;
1027
1028                 req = list_first_entry(&buf->rb_allreqs,
1029                                        struct rpcrdma_req, rl_all);
1030                 list_del(&req->rl_all);
1031
1032                 spin_unlock(&buf->rb_reqslock);
1033                 rpcrdma_destroy_req(ia, req);
1034                 spin_lock(&buf->rb_reqslock);
1035         }
1036         spin_unlock(&buf->rb_reqslock);
1037         buf->rb_recv_count = 0;
1038
1039         rpcrdma_destroy_mrs(buf);
1040 }
1041
1042 struct rpcrdma_mw *
1043 rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt)
1044 {
1045         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1046         struct rpcrdma_mw *mw = NULL;
1047
1048         spin_lock(&buf->rb_mwlock);
1049         if (!list_empty(&buf->rb_mws)) {
1050                 mw = list_first_entry(&buf->rb_mws,
1051                                       struct rpcrdma_mw, mw_list);
1052                 list_del_init(&mw->mw_list);
1053         }
1054         spin_unlock(&buf->rb_mwlock);
1055
1056         if (!mw)
1057                 goto out_nomws;
1058         return mw;
1059
1060 out_nomws:
1061         dprintk("RPC:       %s: no MWs available\n", __func__);
1062         schedule_delayed_work(&buf->rb_refresh_worker, 0);
1063
1064         /* Allow the reply handler and refresh worker to run */
1065         cond_resched();
1066
1067         return NULL;
1068 }
1069
1070 void
1071 rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw)
1072 {
1073         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1074
1075         spin_lock(&buf->rb_mwlock);
1076         list_add_tail(&mw->mw_list, &buf->rb_mws);
1077         spin_unlock(&buf->rb_mwlock);
1078 }
1079
1080 static struct rpcrdma_rep *
1081 rpcrdma_buffer_get_rep(struct rpcrdma_buffer *buffers)
1082 {
1083         /* If an RPC previously completed without a reply (say, a
1084          * credential problem or a soft timeout occurs) then hold off
1085          * on supplying more Receive buffers until the number of new
1086          * pending RPCs catches up to the number of posted Receives.
1087          */
1088         if (unlikely(buffers->rb_send_count < buffers->rb_recv_count))
1089                 return NULL;
1090
1091         if (unlikely(list_empty(&buffers->rb_recv_bufs)))
1092                 return NULL;
1093         buffers->rb_recv_count++;
1094         return rpcrdma_buffer_get_rep_locked(buffers);
1095 }
1096
1097 /*
1098  * Get a set of request/reply buffers.
1099  *
1100  * Reply buffer (if available) is attached to send buffer upon return.
1101  */
1102 struct rpcrdma_req *
1103 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1104 {
1105         struct rpcrdma_req *req;
1106
1107         spin_lock(&buffers->rb_lock);
1108         if (list_empty(&buffers->rb_send_bufs))
1109                 goto out_reqbuf;
1110         buffers->rb_send_count++;
1111         req = rpcrdma_buffer_get_req_locked(buffers);
1112         req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1113         spin_unlock(&buffers->rb_lock);
1114         return req;
1115
1116 out_reqbuf:
1117         spin_unlock(&buffers->rb_lock);
1118         pr_warn("RPC:       %s: out of request buffers\n", __func__);
1119         return NULL;
1120 }
1121
1122 /*
1123  * Put request/reply buffers back into pool.
1124  * Pre-decrement counter/array index.
1125  */
1126 void
1127 rpcrdma_buffer_put(struct rpcrdma_req *req)
1128 {
1129         struct rpcrdma_buffer *buffers = req->rl_buffer;
1130         struct rpcrdma_rep *rep = req->rl_reply;
1131
1132         req->rl_niovs = 0;
1133         req->rl_reply = NULL;
1134
1135         spin_lock(&buffers->rb_lock);
1136         buffers->rb_send_count--;
1137         list_add_tail(&req->rl_free, &buffers->rb_send_bufs);
1138         if (rep) {
1139                 buffers->rb_recv_count--;
1140                 list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1141         }
1142         spin_unlock(&buffers->rb_lock);
1143 }
1144
1145 /*
1146  * Recover reply buffers from pool.
1147  * This happens when recovering from disconnect.
1148  */
1149 void
1150 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1151 {
1152         struct rpcrdma_buffer *buffers = req->rl_buffer;
1153
1154         spin_lock(&buffers->rb_lock);
1155         req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1156         spin_unlock(&buffers->rb_lock);
1157 }
1158
1159 /*
1160  * Put reply buffers back into pool when not attached to
1161  * request. This happens in error conditions.
1162  */
1163 void
1164 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1165 {
1166         struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1167
1168         spin_lock(&buffers->rb_lock);
1169         buffers->rb_recv_count--;
1170         list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1171         spin_unlock(&buffers->rb_lock);
1172 }
1173
1174 /*
1175  * Wrappers for internal-use kmalloc memory registration, used by buffer code.
1176  */
1177
1178 /**
1179  * rpcrdma_alloc_regbuf - kmalloc and register memory for SEND/RECV buffers
1180  * @ia: controlling rpcrdma_ia
1181  * @size: size of buffer to be allocated, in bytes
1182  * @flags: GFP flags
1183  *
1184  * Returns pointer to private header of an area of internally
1185  * registered memory, or an ERR_PTR. The registered buffer follows
1186  * the end of the private header.
1187  *
1188  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1189  * receiving the payload of RDMA RECV operations. regbufs are not
1190  * used for RDMA READ/WRITE operations, thus are registered only for
1191  * LOCAL access.
1192  */
1193 struct rpcrdma_regbuf *
1194 rpcrdma_alloc_regbuf(struct rpcrdma_ia *ia, size_t size, gfp_t flags)
1195 {
1196         struct rpcrdma_regbuf *rb;
1197         struct ib_sge *iov;
1198
1199         rb = kmalloc(sizeof(*rb) + size, flags);
1200         if (rb == NULL)
1201                 goto out;
1202
1203         iov = &rb->rg_iov;
1204         iov->addr = ib_dma_map_single(ia->ri_device,
1205                                       (void *)rb->rg_base, size,
1206                                       DMA_BIDIRECTIONAL);
1207         if (ib_dma_mapping_error(ia->ri_device, iov->addr))
1208                 goto out_free;
1209
1210         iov->length = size;
1211         iov->lkey = ia->ri_pd->local_dma_lkey;
1212         rb->rg_size = size;
1213         rb->rg_owner = NULL;
1214         return rb;
1215
1216 out_free:
1217         kfree(rb);
1218 out:
1219         return ERR_PTR(-ENOMEM);
1220 }
1221
1222 /**
1223  * rpcrdma_free_regbuf - deregister and free registered buffer
1224  * @ia: controlling rpcrdma_ia
1225  * @rb: regbuf to be deregistered and freed
1226  */
1227 void
1228 rpcrdma_free_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1229 {
1230         struct ib_sge *iov;
1231
1232         if (!rb)
1233                 return;
1234
1235         iov = &rb->rg_iov;
1236         ib_dma_unmap_single(ia->ri_device,
1237                             iov->addr, iov->length, DMA_BIDIRECTIONAL);
1238         kfree(rb);
1239 }
1240
1241 /*
1242  * Prepost any receive buffer, then post send.
1243  *
1244  * Receive buffer is donated to hardware, reclaimed upon recv completion.
1245  */
1246 int
1247 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1248                 struct rpcrdma_ep *ep,
1249                 struct rpcrdma_req *req)
1250 {
1251         struct ib_device *device = ia->ri_device;
1252         struct ib_send_wr send_wr, *send_wr_fail;
1253         struct rpcrdma_rep *rep = req->rl_reply;
1254         struct ib_sge *iov = req->rl_send_iov;
1255         int i, rc;
1256
1257         if (rep) {
1258                 rc = rpcrdma_ep_post_recv(ia, ep, rep);
1259                 if (rc)
1260                         return rc;
1261                 req->rl_reply = NULL;
1262         }
1263
1264         send_wr.next = NULL;
1265         send_wr.wr_cqe = &req->rl_cqe;
1266         send_wr.sg_list = iov;
1267         send_wr.num_sge = req->rl_niovs;
1268         send_wr.opcode = IB_WR_SEND;
1269
1270         for (i = 0; i < send_wr.num_sge; i++)
1271                 ib_dma_sync_single_for_device(device, iov[i].addr,
1272                                               iov[i].length, DMA_TO_DEVICE);
1273         dprintk("RPC:       %s: posting %d s/g entries\n",
1274                 __func__, send_wr.num_sge);
1275
1276         if (DECR_CQCOUNT(ep) > 0)
1277                 send_wr.send_flags = 0;
1278         else { /* Provider must take a send completion every now and then */
1279                 INIT_CQCOUNT(ep);
1280                 send_wr.send_flags = IB_SEND_SIGNALED;
1281         }
1282
1283         rc = ib_post_send(ia->ri_id->qp, &send_wr, &send_wr_fail);
1284         if (rc)
1285                 goto out_postsend_err;
1286         return 0;
1287
1288 out_postsend_err:
1289         pr_err("rpcrdma: RDMA Send ib_post_send returned %i\n", rc);
1290         return -ENOTCONN;
1291 }
1292
1293 /*
1294  * (Re)post a receive buffer.
1295  */
1296 int
1297 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1298                      struct rpcrdma_ep *ep,
1299                      struct rpcrdma_rep *rep)
1300 {
1301         struct ib_recv_wr recv_wr, *recv_wr_fail;
1302         int rc;
1303
1304         recv_wr.next = NULL;
1305         recv_wr.wr_cqe = &rep->rr_cqe;
1306         recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1307         recv_wr.num_sge = 1;
1308
1309         ib_dma_sync_single_for_cpu(ia->ri_device,
1310                                    rdmab_addr(rep->rr_rdmabuf),
1311                                    rdmab_length(rep->rr_rdmabuf),
1312                                    DMA_BIDIRECTIONAL);
1313
1314         rc = ib_post_recv(ia->ri_id->qp, &recv_wr, &recv_wr_fail);
1315         if (rc)
1316                 goto out_postrecv;
1317         return 0;
1318
1319 out_postrecv:
1320         pr_err("rpcrdma: ib_post_recv returned %i\n", rc);
1321         return -ENOTCONN;
1322 }
1323
1324 /**
1325  * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests
1326  * @r_xprt: transport associated with these backchannel resources
1327  * @min_reqs: minimum number of incoming requests expected
1328  *
1329  * Returns zero if all requested buffers were posted, or a negative errno.
1330  */
1331 int
1332 rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *r_xprt, unsigned int count)
1333 {
1334         struct rpcrdma_buffer *buffers = &r_xprt->rx_buf;
1335         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1336         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
1337         struct rpcrdma_rep *rep;
1338         int rc;
1339
1340         while (count--) {
1341                 spin_lock(&buffers->rb_lock);
1342                 if (list_empty(&buffers->rb_recv_bufs))
1343                         goto out_reqbuf;
1344                 rep = rpcrdma_buffer_get_rep_locked(buffers);
1345                 spin_unlock(&buffers->rb_lock);
1346
1347                 rc = rpcrdma_ep_post_recv(ia, ep, rep);
1348                 if (rc)
1349                         goto out_rc;
1350         }
1351
1352         return 0;
1353
1354 out_reqbuf:
1355         spin_unlock(&buffers->rb_lock);
1356         pr_warn("%s: no extra receive buffers\n", __func__);
1357         return -ENOMEM;
1358
1359 out_rc:
1360         rpcrdma_recv_buffer_put(rep);
1361         return rc;
1362 }