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
  57 #include <rdma/ib_cm.h>
  58
  59 #include "xprt_rdma.h"
  60
  61 /*
  62  * Globals/Macros
  63  */
  64
  65 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  66 # define RPCDBG_FACILITY        RPCDBG_TRANS
  67 #endif
  68
  69 /*
  70  * internal functions
  71  */
  72 static void rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt);
  73 static void rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf);
  74 static void rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb);
  75
  76 static struct workqueue_struct *rpcrdma_receive_wq __read_mostly;
  77
  78 int
  79 rpcrdma_alloc_wq(void)
  80 {
  81         struct workqueue_struct *recv_wq;
  82
  83         recv_wq = alloc_workqueue("xprtrdma_receive",
  84                                   WQ_MEM_RECLAIM | WQ_UNBOUND | WQ_HIGHPRI,
  85                                   0);
  86         if (!recv_wq)
  87                 return -ENOMEM;
  88
  89         rpcrdma_receive_wq = recv_wq;
  90         return 0;
  91 }
  92
  93 void
  94 rpcrdma_destroy_wq(void)
  95 {
  96         struct workqueue_struct *wq;
  97
  98         if (rpcrdma_receive_wq) {
  99                 wq = rpcrdma_receive_wq;
 100                 rpcrdma_receive_wq = NULL;
 101                 destroy_workqueue(wq);
 102         }
 103 }
 104
 105 static void
 106 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
 107 {
 108         struct rpcrdma_ep *ep = context;
 109
 110         pr_err("rpcrdma: %s on device %s ep %p\n",
 111                ib_event_msg(event->event), event->device->name, context);
 112
 113         if (ep->rep_connected == 1) {
 114                 ep->rep_connected = -EIO;
 115                 rpcrdma_conn_func(ep);
 116                 wake_up_all(&ep->rep_connect_wait);
 117         }
 118 }
 119
 120 /**
 121  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
 122  * @cq: completion queue (ignored)
 123  * @wc: completed WR
 124  *
 125  */
 126 static void
 127 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
 128 {
 129         /* WARNING: Only wr_cqe and status are reliable at this point */
 130         if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
 131                 pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
 132                        ib_wc_status_msg(wc->status),
 133                        wc->status, wc->vendor_err);
 134 }
 135
 136 /* Perform basic sanity checking to avoid using garbage
 137  * to update the credit grant value.
 138  */
 139 static void
 140 rpcrdma_update_granted_credits(struct rpcrdma_rep *rep)
 141 {
 142         struct rpcrdma_buffer *buffer = &rep->rr_rxprt->rx_buf;
 143         __be32 *p = rep->rr_rdmabuf->rg_base;
 144         u32 credits;
 145
 146         credits = be32_to_cpup(p + 2);
 147         if (credits == 0)
 148                 credits = 1;    /* don't deadlock */
 149         else if (credits > buffer->rb_max_requests)
 150                 credits = buffer->rb_max_requests;
 151
 152         atomic_set(&buffer->rb_credits, credits);
 153 }
 154
 155 /**
 156  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
 157  * @cq: completion queue (ignored)
 158  * @wc: completed WR
 159  *
 160  */
 161 static void
 162 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
 163 {
 164         struct ib_cqe *cqe = wc->wr_cqe;
 165         struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
 166                                                rr_cqe);
 167
 168         /* WARNING: Only wr_id and status are reliable at this point */
 169         if (wc->status != IB_WC_SUCCESS)
 170                 goto out_fail;
 171
 172         /* status == SUCCESS means all fields in wc are trustworthy */
 173         dprintk("RPC:       %s: rep %p opcode 'recv', length %u: success\n",
 174                 __func__, rep, wc->byte_len);
 175
 176         rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
 177         rep->rr_wc_flags = wc->wc_flags;
 178         rep->rr_inv_rkey = wc->ex.invalidate_rkey;
 179
 180         ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
 181                                    rdmab_addr(rep->rr_rdmabuf),
 182                                    wc->byte_len, DMA_FROM_DEVICE);
 183
 184         if (wc->byte_len >= RPCRDMA_HDRLEN_ERR)
 185                 rpcrdma_update_granted_credits(rep);
 186
 187 out_schedule:
 188         queue_work(rpcrdma_receive_wq, &rep->rr_work);
 189         return;
 190
 191 out_fail:
 192         if (wc->status != IB_WC_WR_FLUSH_ERR)
 193                 pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
 194                        ib_wc_status_msg(wc->status),
 195                        wc->status, wc->vendor_err);
 196         rpcrdma_set_xdrlen(&rep->rr_hdrbuf, 0);
 197         goto out_schedule;
 198 }
 199
 200 static void
 201 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
 202                                struct rdma_conn_param *param)
 203 {
 204         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
 205         const struct rpcrdma_connect_private *pmsg = param->private_data;
 206         unsigned int rsize, wsize;
 207
 208         /* Default settings for RPC-over-RDMA Version One */
 209         r_xprt->rx_ia.ri_reminv_expected = false;
 210         r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
 211         rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
 212         wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
 213
 214         if (pmsg &&
 215             pmsg->cp_magic == rpcrdma_cmp_magic &&
 216             pmsg->cp_version == RPCRDMA_CMP_VERSION) {
 217                 r_xprt->rx_ia.ri_reminv_expected = true;
 218                 r_xprt->rx_ia.ri_implicit_roundup = true;
 219                 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
 220                 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
 221         }
 222
 223         if (rsize < cdata->inline_rsize)
 224                 cdata->inline_rsize = rsize;
 225         if (wsize < cdata->inline_wsize)
 226                 cdata->inline_wsize = wsize;
 227         dprintk("RPC:       %s: max send %u, max recv %u\n",
 228                 __func__, cdata->inline_wsize, cdata->inline_rsize);
 229         rpcrdma_set_max_header_sizes(r_xprt);
 230 }
 231
 232 static int
 233 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
 234 {
 235         struct rpcrdma_xprt *xprt = id->context;
 236         struct rpcrdma_ia *ia = &xprt->rx_ia;
 237         struct rpcrdma_ep *ep = &xprt->rx_ep;
 238 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 239         struct sockaddr *sap = (struct sockaddr *)&ep->rep_remote_addr;
 240 #endif
 241         int connstate = 0;
 242
 243         switch (event->event) {
 244         case RDMA_CM_EVENT_ADDR_RESOLVED:
 245         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 246                 ia->ri_async_rc = 0;
 247                 complete(&ia->ri_done);
 248                 break;
 249         case RDMA_CM_EVENT_ADDR_ERROR:
 250                 ia->ri_async_rc = -EHOSTUNREACH;
 251                 dprintk("RPC:       %s: CM address resolution error, ep 0x%p\n",
 252                         __func__, ep);
 253                 complete(&ia->ri_done);
 254                 break;
 255         case RDMA_CM_EVENT_ROUTE_ERROR:
 256                 ia->ri_async_rc = -ENETUNREACH;
 257                 dprintk("RPC:       %s: CM route resolution error, ep 0x%p\n",
 258                         __func__, ep);
 259                 complete(&ia->ri_done);
 260                 break;
 261         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 262 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 263                 pr_info("rpcrdma: removing device %s for %pIS:%u\n",
 264                         ia->ri_device->name,
 265                         sap, rpc_get_port(sap));
 266 #endif
 267                 init_completion(&ia->ri_remove_done);
 268                 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
 269                 ep->rep_connected = -ENODEV;
 270                 xprt_force_disconnect(&xprt->rx_xprt);
 271                 wait_for_completion(&ia->ri_remove_done);
 272
 273                 ia->ri_id = NULL;
 274                 ia->ri_device = NULL;
 275                 /* Return 1 to ensure the core destroys the id. */
 276                 return 1;
 277         case RDMA_CM_EVENT_ESTABLISHED:
 278                 connstate = 1;
 279                 rpcrdma_update_connect_private(xprt, &event->param.conn);
 280                 goto connected;
 281         case RDMA_CM_EVENT_CONNECT_ERROR:
 282                 connstate = -ENOTCONN;
 283                 goto connected;
 284         case RDMA_CM_EVENT_UNREACHABLE:
 285                 connstate = -ENETDOWN;
 286                 goto connected;
 287         case RDMA_CM_EVENT_REJECTED:
 288                 dprintk("rpcrdma: connection to %pIS:%u rejected: %s\n",
 289                         sap, rpc_get_port(sap),
 290                         rdma_reject_msg(id, event->status));
 291                 connstate = -ECONNREFUSED;
 292                 if (event->status == IB_CM_REJ_STALE_CONN)
 293                         connstate = -EAGAIN;
 294                 goto connected;
 295         case RDMA_CM_EVENT_DISCONNECTED:
 296                 connstate = -ECONNABORTED;
 297 connected:
 298                 atomic_set(&xprt->rx_buf.rb_credits, 1);
 299                 ep->rep_connected = connstate;
 300                 rpcrdma_conn_func(ep);
 301                 wake_up_all(&ep->rep_connect_wait);
 302                 /*FALLTHROUGH*/
 303         default:
 304                 dprintk("RPC:       %s: %pIS:%u on %s/%s (ep 0x%p): %s\n",
 305                         __func__, sap, rpc_get_port(sap),
 306                         ia->ri_device->name, ia->ri_ops->ro_displayname,
 307                         ep, rdma_event_msg(event->event));
 308                 break;
 309         }
 310
 311         return 0;
 312 }
 313
 314 static struct rdma_cm_id *
 315 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
 316                         struct rpcrdma_ia *ia, struct sockaddr *addr)
 317 {
 318         unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
 319         struct rdma_cm_id *id;
 320         int rc;
 321
 322         init_completion(&ia->ri_done);
 323
 324         id = rdma_create_id(&init_net, rpcrdma_conn_upcall, xprt, RDMA_PS_TCP,
 325                             IB_QPT_RC);
 326         if (IS_ERR(id)) {
 327                 rc = PTR_ERR(id);
 328                 dprintk("RPC:       %s: rdma_create_id() failed %i\n",
 329                         __func__, rc);
 330                 return id;
 331         }
 332
 333         ia->ri_async_rc = -ETIMEDOUT;
 334         rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
 335         if (rc) {
 336                 dprintk("RPC:       %s: rdma_resolve_addr() failed %i\n",
 337                         __func__, rc);
 338                 goto out;
 339         }
 340         rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
 341         if (rc < 0) {
 342                 dprintk("RPC:       %s: wait() exited: %i\n",
 343                         __func__, rc);
 344                 goto out;
 345         }
 346
 347         rc = ia->ri_async_rc;
 348         if (rc)
 349                 goto out;
 350
 351         ia->ri_async_rc = -ETIMEDOUT;
 352         rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
 353         if (rc) {
 354                 dprintk("RPC:       %s: rdma_resolve_route() failed %i\n",
 355                         __func__, rc);
 356                 goto out;
 357         }
 358         rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
 359         if (rc < 0) {
 360                 dprintk("RPC:       %s: wait() exited: %i\n",
 361                         __func__, rc);
 362                 goto out;
 363         }
 364         rc = ia->ri_async_rc;
 365         if (rc)
 366                 goto out;
 367
 368         return id;
 369
 370 out:
 371         rdma_destroy_id(id);
 372         return ERR_PTR(rc);
 373 }
 374
 375 /*
 376  * Exported functions.
 377  */
 378
 379 /**
 380  * rpcrdma_ia_open - Open and initialize an Interface Adapter.
 381  * @xprt: controlling transport
 382  * @addr: IP address of remote peer
 383  *
 384  * Returns 0 on success, negative errno if an appropriate
 385  * Interface Adapter could not be found and opened.
 386  */
 387 int
 388 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr)
 389 {
 390         struct rpcrdma_ia *ia = &xprt->rx_ia;
 391         int rc;
 392
 393         ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
 394         if (IS_ERR(ia->ri_id)) {
 395                 rc = PTR_ERR(ia->ri_id);
 396                 goto out_err;
 397         }
 398         ia->ri_device = ia->ri_id->device;
 399
 400         ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
 401         if (IS_ERR(ia->ri_pd)) {
 402                 rc = PTR_ERR(ia->ri_pd);
 403                 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
 404                 goto out_err;
 405         }
 406
 407         switch (xprt_rdma_memreg_strategy) {
 408         case RPCRDMA_FRMR:
 409                 if (frwr_is_supported(ia)) {
 410                         ia->ri_ops = &rpcrdma_frwr_memreg_ops;
 411                         break;
 412                 }
 413                 /*FALLTHROUGH*/
 414         case RPCRDMA_MTHCAFMR:
 415                 if (fmr_is_supported(ia)) {
 416                         ia->ri_ops = &rpcrdma_fmr_memreg_ops;
 417                         break;
 418                 }
 419                 /*FALLTHROUGH*/
 420         default:
 421                 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
 422                        ia->ri_device->name, xprt_rdma_memreg_strategy);
 423                 rc = -EINVAL;
 424                 goto out_err;
 425         }
 426
 427         return 0;
 428
 429 out_err:
 430         rpcrdma_ia_close(ia);
 431         return rc;
 432 }
 433
 434 /**
 435  * rpcrdma_ia_remove - Handle device driver unload
 436  * @ia: interface adapter being removed
 437  *
 438  * Divest transport H/W resources associated with this adapter,
 439  * but allow it to be restored later.
 440  */
 441 void
 442 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
 443 {
 444         struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
 445                                                    rx_ia);
 446         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 447         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 448         struct rpcrdma_req *req;
 449         struct rpcrdma_rep *rep;
 450
 451         cancel_delayed_work_sync(&buf->rb_refresh_worker);
 452
 453         /* This is similar to rpcrdma_ep_destroy, but:
 454          * - Don't cancel the connect worker.
 455          * - Don't call rpcrdma_ep_disconnect, which waits
 456          *   for another conn upcall, which will deadlock.
 457          * - rdma_disconnect is unneeded, the underlying
 458          *   connection is already gone.
 459          */
 460         if (ia->ri_id->qp) {
 461                 ib_drain_qp(ia->ri_id->qp);
 462                 rdma_destroy_qp(ia->ri_id);
 463                 ia->ri_id->qp = NULL;
 464         }
 465         ib_free_cq(ep->rep_attr.recv_cq);
 466         ep->rep_attr.recv_cq = NULL;
 467         ib_free_cq(ep->rep_attr.send_cq);
 468         ep->rep_attr.send_cq = NULL;
 469
 470         /* The ULP is responsible for ensuring all DMA
 471          * mappings and MRs are gone.
 472          */
 473         list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
 474                 rpcrdma_dma_unmap_regbuf(rep->rr_rdmabuf);
 475         list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
 476                 rpcrdma_dma_unmap_regbuf(req->rl_rdmabuf);
 477                 rpcrdma_dma_unmap_regbuf(req->rl_sendbuf);
 478                 rpcrdma_dma_unmap_regbuf(req->rl_recvbuf);
 479         }
 480         rpcrdma_destroy_mrs(buf);
 481         ib_dealloc_pd(ia->ri_pd);
 482         ia->ri_pd = NULL;
 483
 484         /* Allow waiters to continue */
 485         complete(&ia->ri_remove_done);
 486 }
 487
 488 /**
 489  * rpcrdma_ia_close - Clean up/close an IA.
 490  * @ia: interface adapter to close
 491  *
 492  */
 493 void
 494 rpcrdma_ia_close(struct rpcrdma_ia *ia)
 495 {
 496         dprintk("RPC:       %s: entering\n", __func__);
 497         if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
 498                 if (ia->ri_id->qp)
 499                         rdma_destroy_qp(ia->ri_id);
 500                 rdma_destroy_id(ia->ri_id);
 501         }
 502         ia->ri_id = NULL;
 503         ia->ri_device = NULL;
 504
 505         /* If the pd is still busy, xprtrdma missed freeing a resource */
 506         if (ia->ri_pd && !IS_ERR(ia->ri_pd))
 507                 ib_dealloc_pd(ia->ri_pd);
 508         ia->ri_pd = NULL;
 509 }
 510
 511 /*
 512  * Create unconnected endpoint.
 513  */
 514 int
 515 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
 516                   struct rpcrdma_create_data_internal *cdata)
 517 {
 518         struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
 519         unsigned int max_qp_wr, max_sge;
 520         struct ib_cq *sendcq, *recvcq;
 521         int rc;
 522
 523         max_sge = min_t(unsigned int, ia->ri_device->attrs.max_sge,
 524                         RPCRDMA_MAX_SEND_SGES);
 525         if (max_sge < RPCRDMA_MIN_SEND_SGES) {
 526                 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
 527                 return -ENOMEM;
 528         }
 529         ia->ri_max_send_sges = max_sge;
 530
 531         if (ia->ri_device->attrs.max_qp_wr <= RPCRDMA_BACKWARD_WRS) {
 532                 dprintk("RPC:       %s: insufficient wqe's available\n",
 533                         __func__);
 534                 return -ENOMEM;
 535         }
 536         max_qp_wr = ia->ri_device->attrs.max_qp_wr - RPCRDMA_BACKWARD_WRS - 1;
 537
 538         /* check provider's send/recv wr limits */
 539         if (cdata->max_requests > max_qp_wr)
 540                 cdata->max_requests = max_qp_wr;
 541
 542         ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
 543         ep->rep_attr.qp_context = ep;
 544         ep->rep_attr.srq = NULL;
 545         ep->rep_attr.cap.max_send_wr = cdata->max_requests;
 546         ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
 547         ep->rep_attr.cap.max_send_wr += 1;      /* drain cqe */
 548         rc = ia->ri_ops->ro_open(ia, ep, cdata);
 549         if (rc)
 550                 return rc;
 551         ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
 552         ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
 553         ep->rep_attr.cap.max_recv_wr += 1;      /* drain cqe */
 554         ep->rep_attr.cap.max_send_sge = max_sge;
 555         ep->rep_attr.cap.max_recv_sge = 1;
 556         ep->rep_attr.cap.max_inline_data = 0;
 557         ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 558         ep->rep_attr.qp_type = IB_QPT_RC;
 559         ep->rep_attr.port_num = ~0;
 560
 561         dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
 562                 "iovs: send %d recv %d\n",
 563                 __func__,
 564                 ep->rep_attr.cap.max_send_wr,
 565                 ep->rep_attr.cap.max_recv_wr,
 566                 ep->rep_attr.cap.max_send_sge,
 567                 ep->rep_attr.cap.max_recv_sge);
 568
 569         /* set trigger for requesting send completion */
 570         ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 - 1;
 571         if (ep->rep_cqinit <= 2)
 572                 ep->rep_cqinit = 0;     /* always signal? */
 573         rpcrdma_init_cqcount(ep, 0);
 574         init_waitqueue_head(&ep->rep_connect_wait);
 575         INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
 576
 577         sendcq = ib_alloc_cq(ia->ri_device, NULL,
 578                              ep->rep_attr.cap.max_send_wr + 1,
 579                              0, IB_POLL_SOFTIRQ);
 580         if (IS_ERR(sendcq)) {
 581                 rc = PTR_ERR(sendcq);
 582                 dprintk("RPC:       %s: failed to create send CQ: %i\n",
 583                         __func__, rc);
 584                 goto out1;
 585         }
 586
 587         recvcq = ib_alloc_cq(ia->ri_device, NULL,
 588                              ep->rep_attr.cap.max_recv_wr + 1,
 589                              0, IB_POLL_SOFTIRQ);
 590         if (IS_ERR(recvcq)) {
 591                 rc = PTR_ERR(recvcq);
 592                 dprintk("RPC:       %s: failed to create recv CQ: %i\n",
 593                         __func__, rc);
 594                 goto out2;
 595         }
 596
 597         ep->rep_attr.send_cq = sendcq;
 598         ep->rep_attr.recv_cq = recvcq;
 599
 600         /* Initialize cma parameters */
 601         memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
 602
 603         /* Prepare RDMA-CM private message */
 604         pmsg->cp_magic = rpcrdma_cmp_magic;
 605         pmsg->cp_version = RPCRDMA_CMP_VERSION;
 606         pmsg->cp_flags |= ia->ri_ops->ro_send_w_inv_ok;
 607         pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize);
 608         pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize);
 609         ep->rep_remote_cma.private_data = pmsg;
 610         ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
 611
 612         /* Client offers RDMA Read but does not initiate */
 613         ep->rep_remote_cma.initiator_depth = 0;
 614         if (ia->ri_device->attrs.max_qp_rd_atom > 32)   /* arbitrary but <= 255 */
 615                 ep->rep_remote_cma.responder_resources = 32;
 616         else
 617                 ep->rep_remote_cma.responder_resources =
 618                                                 ia->ri_device->attrs.max_qp_rd_atom;
 619
 620         /* Limit transport retries so client can detect server
 621          * GID changes quickly. RPC layer handles re-establishing
 622          * transport connection and retransmission.
 623          */
 624         ep->rep_remote_cma.retry_count = 6;
 625
 626         /* RPC-over-RDMA handles its own flow control. In addition,
 627          * make all RNR NAKs visible so we know that RPC-over-RDMA
 628          * flow control is working correctly (no NAKs should be seen).
 629          */
 630         ep->rep_remote_cma.flow_control = 0;
 631         ep->rep_remote_cma.rnr_retry_count = 0;
 632
 633         return 0;
 634
 635 out2:
 636         ib_free_cq(sendcq);
 637 out1:
 638         return rc;
 639 }
 640
 641 /*
 642  * rpcrdma_ep_destroy
 643  *
 644  * Disconnect and destroy endpoint. After this, the only
 645  * valid operations on the ep are to free it (if dynamically
 646  * allocated) or re-create it.
 647  */
 648 void
 649 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 650 {
 651         dprintk("RPC:       %s: entering, connected is %d\n",
 652                 __func__, ep->rep_connected);
 653
 654         cancel_delayed_work_sync(&ep->rep_connect_worker);
 655
 656         if (ia->ri_id && ia->ri_id->qp) {
 657                 rpcrdma_ep_disconnect(ep, ia);
 658                 rdma_destroy_qp(ia->ri_id);
 659                 ia->ri_id->qp = NULL;
 660         }
 661
 662         if (ep->rep_attr.recv_cq)
 663                 ib_free_cq(ep->rep_attr.recv_cq);
 664         if (ep->rep_attr.send_cq)
 665                 ib_free_cq(ep->rep_attr.send_cq);
 666 }
 667
 668 /* Re-establish a connection after a device removal event.
 669  * Unlike a normal reconnection, a fresh PD and a new set
 670  * of MRs and buffers is needed.
 671  */
 672 static int
 673 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
 674                          struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 675 {
 676         struct sockaddr *sap = (struct sockaddr *)&r_xprt->rx_data.addr;
 677         int rc, err;
 678
 679         pr_info("%s: r_xprt = %p\n", __func__, r_xprt);
 680
 681         rc = -EHOSTUNREACH;
 682         if (rpcrdma_ia_open(r_xprt, sap))
 683                 goto out1;
 684
 685         rc = -ENOMEM;
 686         err = rpcrdma_ep_create(ep, ia, &r_xprt->rx_data);
 687         if (err) {
 688                 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
 689                 goto out2;
 690         }
 691
 692         rc = -ENETUNREACH;
 693         err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
 694         if (err) {
 695                 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
 696                 goto out3;
 697         }
 698
 699         rpcrdma_create_mrs(r_xprt);
 700         return 0;
 701
 702 out3:
 703         rpcrdma_ep_destroy(ep, ia);
 704 out2:
 705         rpcrdma_ia_close(ia);
 706 out1:
 707         return rc;
 708 }
 709
 710 static int
 711 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
 712                      struct rpcrdma_ia *ia)
 713 {
 714         struct sockaddr *sap = (struct sockaddr *)&r_xprt->rx_data.addr;
 715         struct rdma_cm_id *id, *old;
 716         int err, rc;
 717
 718         dprintk("RPC:       %s: reconnecting...\n", __func__);
 719
 720         rpcrdma_ep_disconnect(ep, ia);
 721
 722         rc = -EHOSTUNREACH;
 723         id = rpcrdma_create_id(r_xprt, ia, sap);
 724         if (IS_ERR(id))
 725                 goto out;
 726
 727         /* As long as the new ID points to the same device as the
 728          * old ID, we can reuse the transport's existing PD and all
 729          * previously allocated MRs. Also, the same device means
 730          * the transport's previous DMA mappings are still valid.
 731          *
 732          * This is a sanity check only. There should be no way these
 733          * point to two different devices here.
 734          */
 735         old = id;
 736         rc = -ENETUNREACH;
 737         if (ia->ri_device != id->device) {
 738                 pr_err("rpcrdma: can't reconnect on different device!\n");
 739                 goto out_destroy;
 740         }
 741
 742         err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
 743         if (err) {
 744                 dprintk("RPC:       %s: rdma_create_qp returned %d\n",
 745                         __func__, err);
 746                 goto out_destroy;
 747         }
 748
 749         /* Atomically replace the transport's ID and QP. */
 750         rc = 0;
 751         old = ia->ri_id;
 752         ia->ri_id = id;
 753         rdma_destroy_qp(old);
 754
 755 out_destroy:
 756         rdma_destroy_id(old);
 757 out:
 758         return rc;
 759 }
 760
 761 /*
 762  * Connect unconnected endpoint.
 763  */
 764 int
 765 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 766 {
 767         struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
 768                                                    rx_ia);
 769         unsigned int extras;
 770         int rc;
 771
 772 retry:
 773         switch (ep->rep_connected) {
 774         case 0:
 775                 dprintk("RPC:       %s: connecting...\n", __func__);
 776                 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
 777                 if (rc) {
 778                         dprintk("RPC:       %s: rdma_create_qp failed %i\n",
 779                                 __func__, rc);
 780                         rc = -ENETUNREACH;
 781                         goto out_noupdate;
 782                 }
 783                 break;
 784         case -ENODEV:
 785                 rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
 786                 if (rc)
 787                         goto out_noupdate;
 788                 break;
 789         default:
 790                 rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
 791                 if (rc)
 792                         goto out;
 793         }
 794
 795         ep->rep_connected = 0;
 796
 797         rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
 798         if (rc) {
 799                 dprintk("RPC:       %s: rdma_connect() failed with %i\n",
 800                                 __func__, rc);
 801                 goto out;
 802         }
 803
 804         wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
 805         if (ep->rep_connected <= 0) {
 806                 if (ep->rep_connected == -EAGAIN)
 807                         goto retry;
 808                 rc = ep->rep_connected;
 809                 goto out;
 810         }
 811
 812         dprintk("RPC:       %s: connected\n", __func__);
 813         extras = r_xprt->rx_buf.rb_bc_srv_max_requests;
 814         if (extras)
 815                 rpcrdma_ep_post_extra_recv(r_xprt, extras);
 816
 817 out:
 818         if (rc)
 819                 ep->rep_connected = rc;
 820
 821 out_noupdate:
 822         return rc;
 823 }
 824
 825 /*
 826  * rpcrdma_ep_disconnect
 827  *
 828  * This is separate from destroy to facilitate the ability
 829  * to reconnect without recreating the endpoint.
 830  *
 831  * This call is not reentrant, and must not be made in parallel
 832  * on the same endpoint.
 833  */
 834 void
 835 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 836 {
 837         int rc;
 838
 839         rc = rdma_disconnect(ia->ri_id);
 840         if (!rc) {
 841                 /* returns without wait if not connected */
 842                 wait_event_interruptible(ep->rep_connect_wait,
 843                                                         ep->rep_connected != 1);
 844                 dprintk("RPC:       %s: after wait, %sconnected\n", __func__,
 845                         (ep->rep_connected == 1) ? "still " : "dis");
 846         } else {
 847                 dprintk("RPC:       %s: rdma_disconnect %i\n", __func__, rc);
 848                 ep->rep_connected = rc;
 849         }
 850
 851         ib_drain_qp(ia->ri_id->qp);
 852 }
 853
 854 static void
 855 rpcrdma_mr_recovery_worker(struct work_struct *work)
 856 {
 857         struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
 858                                                   rb_recovery_worker.work);
 859         struct rpcrdma_mw *mw;
 860
 861         spin_lock(&buf->rb_recovery_lock);
 862         while (!list_empty(&buf->rb_stale_mrs)) {
 863                 mw = rpcrdma_pop_mw(&buf->rb_stale_mrs);
 864                 spin_unlock(&buf->rb_recovery_lock);
 865
 866                 dprintk("RPC:       %s: recovering MR %p\n", __func__, mw);
 867                 mw->mw_xprt->rx_ia.ri_ops->ro_recover_mr(mw);
 868
 869                 spin_lock(&buf->rb_recovery_lock);
 870         }
 871         spin_unlock(&buf->rb_recovery_lock);
 872 }
 873
 874 void
 875 rpcrdma_defer_mr_recovery(struct rpcrdma_mw *mw)
 876 {
 877         struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
 878         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 879
 880         spin_lock(&buf->rb_recovery_lock);
 881         rpcrdma_push_mw(mw, &buf->rb_stale_mrs);
 882         spin_unlock(&buf->rb_recovery_lock);
 883
 884         schedule_delayed_work(&buf->rb_recovery_worker, 0);
 885 }
 886
 887 static void
 888 rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt)
 889 {
 890         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 891         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 892         unsigned int count;
 893         LIST_HEAD(free);
 894         LIST_HEAD(all);
 895
 896         for (count = 0; count < 32; count++) {
 897                 struct rpcrdma_mw *mw;
 898                 int rc;
 899
 900                 mw = kzalloc(sizeof(*mw), GFP_KERNEL);
 901                 if (!mw)
 902                         break;
 903
 904                 rc = ia->ri_ops->ro_init_mr(ia, mw);
 905                 if (rc) {
 906                         kfree(mw);
 907                         break;
 908                 }
 909
 910                 mw->mw_xprt = r_xprt;
 911
 912                 list_add(&mw->mw_list, &free);
 913                 list_add(&mw->mw_all, &all);
 914         }
 915
 916         spin_lock(&buf->rb_mwlock);
 917         list_splice(&free, &buf->rb_mws);
 918         list_splice(&all, &buf->rb_all);
 919         r_xprt->rx_stats.mrs_allocated += count;
 920         spin_unlock(&buf->rb_mwlock);
 921
 922         dprintk("RPC:       %s: created %u MRs\n", __func__, count);
 923 }
 924
 925 static void
 926 rpcrdma_mr_refresh_worker(struct work_struct *work)
 927 {
 928         struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
 929                                                   rb_refresh_worker.work);
 930         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
 931                                                    rx_buf);
 932
 933         rpcrdma_create_mrs(r_xprt);
 934 }
 935
 936 struct rpcrdma_req *
 937 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
 938 {
 939         struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
 940         struct rpcrdma_req *req;
 941
 942         req = kzalloc(sizeof(*req), GFP_KERNEL);
 943         if (req == NULL)
 944                 return ERR_PTR(-ENOMEM);
 945
 946         spin_lock(&buffer->rb_reqslock);
 947         list_add(&req->rl_all, &buffer->rb_allreqs);
 948         spin_unlock(&buffer->rb_reqslock);
 949         req->rl_cqe.done = rpcrdma_wc_send;
 950         req->rl_buffer = &r_xprt->rx_buf;
 951         INIT_LIST_HEAD(&req->rl_registered);
 952         req->rl_send_wr.next = NULL;
 953         req->rl_send_wr.wr_cqe = &req->rl_cqe;
 954         req->rl_send_wr.sg_list = req->rl_send_sge;
 955         req->rl_send_wr.opcode = IB_WR_SEND;
 956         return req;
 957 }
 958
 959 /**
 960  * rpcrdma_create_rep - Allocate an rpcrdma_rep object
 961  * @r_xprt: controlling transport
 962  *
 963  * Returns 0 on success or a negative errno on failure.
 964  */
 965 int
 966 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
 967 {
 968         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
 969         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 970         struct rpcrdma_rep *rep;
 971         int rc;
 972
 973         rc = -ENOMEM;
 974         rep = kzalloc(sizeof(*rep), GFP_KERNEL);
 975         if (rep == NULL)
 976                 goto out;
 977
 978         rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
 979                                                DMA_FROM_DEVICE, GFP_KERNEL);
 980         if (IS_ERR(rep->rr_rdmabuf)) {
 981                 rc = PTR_ERR(rep->rr_rdmabuf);
 982                 goto out_free;
 983         }
 984         xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base,
 985                      rdmab_length(rep->rr_rdmabuf));
 986
 987         rep->rr_cqe.done = rpcrdma_wc_receive;
 988         rep->rr_rxprt = r_xprt;
 989         INIT_WORK(&rep->rr_work, rpcrdma_reply_handler);
 990         rep->rr_recv_wr.next = NULL;
 991         rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
 992         rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
 993         rep->rr_recv_wr.num_sge = 1;
 994
 995         spin_lock(&buf->rb_lock);
 996         list_add(&rep->rr_list, &buf->rb_recv_bufs);
 997         spin_unlock(&buf->rb_lock);
 998         return 0;
 999
1000 out_free:
1001         kfree(rep);
1002 out:
1003         dprintk("RPC:       %s: reply buffer %d alloc failed\n",
1004                 __func__, rc);
1005         return rc;
1006 }
1007
1008 int
1009 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1010 {
1011         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1012         int i, rc;
1013
1014         buf->rb_max_requests = r_xprt->rx_data.max_requests;
1015         buf->rb_bc_srv_max_requests = 0;
1016         atomic_set(&buf->rb_credits, 1);
1017         spin_lock_init(&buf->rb_mwlock);
1018         spin_lock_init(&buf->rb_lock);
1019         spin_lock_init(&buf->rb_recovery_lock);
1020         INIT_LIST_HEAD(&buf->rb_mws);
1021         INIT_LIST_HEAD(&buf->rb_all);
1022         INIT_LIST_HEAD(&buf->rb_stale_mrs);
1023         INIT_DELAYED_WORK(&buf->rb_refresh_worker,
1024                           rpcrdma_mr_refresh_worker);
1025         INIT_DELAYED_WORK(&buf->rb_recovery_worker,
1026                           rpcrdma_mr_recovery_worker);
1027
1028         rpcrdma_create_mrs(r_xprt);
1029
1030         INIT_LIST_HEAD(&buf->rb_send_bufs);
1031         INIT_LIST_HEAD(&buf->rb_allreqs);
1032         spin_lock_init(&buf->rb_reqslock);
1033         for (i = 0; i < buf->rb_max_requests; i++) {
1034                 struct rpcrdma_req *req;
1035
1036                 req = rpcrdma_create_req(r_xprt);
1037                 if (IS_ERR(req)) {
1038                         dprintk("RPC:       %s: request buffer %d alloc"
1039                                 " failed\n", __func__, i);
1040                         rc = PTR_ERR(req);
1041                         goto out;
1042                 }
1043                 req->rl_backchannel = false;
1044                 list_add(&req->rl_list, &buf->rb_send_bufs);
1045         }
1046
1047         INIT_LIST_HEAD(&buf->rb_recv_bufs);
1048         for (i = 0; i <= buf->rb_max_requests; i++) {
1049                 rc = rpcrdma_create_rep(r_xprt);
1050                 if (rc)
1051                         goto out;
1052         }
1053
1054         return 0;
1055 out:
1056         rpcrdma_buffer_destroy(buf);
1057         return rc;
1058 }
1059
1060 static struct rpcrdma_req *
1061 rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer *buf)
1062 {
1063         struct rpcrdma_req *req;
1064
1065         req = list_first_entry(&buf->rb_send_bufs,
1066                                struct rpcrdma_req, rl_list);
1067         list_del_init(&req->rl_list);
1068         return req;
1069 }
1070
1071 static struct rpcrdma_rep *
1072 rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer *buf)
1073 {
1074         struct rpcrdma_rep *rep;
1075
1076         rep = list_first_entry(&buf->rb_recv_bufs,
1077                                struct rpcrdma_rep, rr_list);
1078         list_del(&rep->rr_list);
1079         return rep;
1080 }
1081
1082 static void
1083 rpcrdma_destroy_rep(struct rpcrdma_rep *rep)
1084 {
1085         rpcrdma_free_regbuf(rep->rr_rdmabuf);
1086         kfree(rep);
1087 }
1088
1089 void
1090 rpcrdma_destroy_req(struct rpcrdma_req *req)
1091 {
1092         rpcrdma_free_regbuf(req->rl_recvbuf);
1093         rpcrdma_free_regbuf(req->rl_sendbuf);
1094         rpcrdma_free_regbuf(req->rl_rdmabuf);
1095         kfree(req);
1096 }
1097
1098 static void
1099 rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf)
1100 {
1101         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1102                                                    rx_buf);
1103         struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1104         struct rpcrdma_mw *mw;
1105         unsigned int count;
1106
1107         count = 0;
1108         spin_lock(&buf->rb_mwlock);
1109         while (!list_empty(&buf->rb_all)) {
1110                 mw = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
1111                 list_del(&mw->mw_all);
1112
1113                 spin_unlock(&buf->rb_mwlock);
1114                 ia->ri_ops->ro_release_mr(mw);
1115                 count++;
1116                 spin_lock(&buf->rb_mwlock);
1117         }
1118         spin_unlock(&buf->rb_mwlock);
1119         r_xprt->rx_stats.mrs_allocated = 0;
1120
1121         dprintk("RPC:       %s: released %u MRs\n", __func__, count);
1122 }
1123
1124 void
1125 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1126 {
1127         cancel_delayed_work_sync(&buf->rb_recovery_worker);
1128         cancel_delayed_work_sync(&buf->rb_refresh_worker);
1129
1130         while (!list_empty(&buf->rb_recv_bufs)) {
1131                 struct rpcrdma_rep *rep;
1132
1133                 rep = rpcrdma_buffer_get_rep_locked(buf);
1134                 rpcrdma_destroy_rep(rep);
1135         }
1136         buf->rb_send_count = 0;
1137
1138         spin_lock(&buf->rb_reqslock);
1139         while (!list_empty(&buf->rb_allreqs)) {
1140                 struct rpcrdma_req *req;
1141
1142                 req = list_first_entry(&buf->rb_allreqs,
1143                                        struct rpcrdma_req, rl_all);
1144                 list_del(&req->rl_all);
1145
1146                 spin_unlock(&buf->rb_reqslock);
1147                 rpcrdma_destroy_req(req);
1148                 spin_lock(&buf->rb_reqslock);
1149         }
1150         spin_unlock(&buf->rb_reqslock);
1151         buf->rb_recv_count = 0;
1152
1153         rpcrdma_destroy_mrs(buf);
1154 }
1155
1156 struct rpcrdma_mw *
1157 rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt)
1158 {
1159         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1160         struct rpcrdma_mw *mw = NULL;
1161
1162         spin_lock(&buf->rb_mwlock);
1163         if (!list_empty(&buf->rb_mws))
1164                 mw = rpcrdma_pop_mw(&buf->rb_mws);
1165         spin_unlock(&buf->rb_mwlock);
1166
1167         if (!mw)
1168                 goto out_nomws;
1169         mw->mw_flags = 0;
1170         return mw;
1171
1172 out_nomws:
1173         dprintk("RPC:       %s: no MWs available\n", __func__);
1174         if (r_xprt->rx_ep.rep_connected != -ENODEV)
1175                 schedule_delayed_work(&buf->rb_refresh_worker, 0);
1176
1177         /* Allow the reply handler and refresh worker to run */
1178         cond_resched();
1179
1180         return NULL;
1181 }
1182
1183 void
1184 rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw)
1185 {
1186         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1187
1188         spin_lock(&buf->rb_mwlock);
1189         rpcrdma_push_mw(mw, &buf->rb_mws);
1190         spin_unlock(&buf->rb_mwlock);
1191 }
1192
1193 static struct rpcrdma_rep *
1194 rpcrdma_buffer_get_rep(struct rpcrdma_buffer *buffers)
1195 {
1196         /* If an RPC previously completed without a reply (say, a
1197          * credential problem or a soft timeout occurs) then hold off
1198          * on supplying more Receive buffers until the number of new
1199          * pending RPCs catches up to the number of posted Receives.
1200          */
1201         if (unlikely(buffers->rb_send_count < buffers->rb_recv_count))
1202                 return NULL;
1203
1204         if (unlikely(list_empty(&buffers->rb_recv_bufs)))
1205                 return NULL;
1206         buffers->rb_recv_count++;
1207         return rpcrdma_buffer_get_rep_locked(buffers);
1208 }
1209
1210 /*
1211  * Get a set of request/reply buffers.
1212  *
1213  * Reply buffer (if available) is attached to send buffer upon return.
1214  */
1215 struct rpcrdma_req *
1216 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1217 {
1218         struct rpcrdma_req *req;
1219
1220         spin_lock(&buffers->rb_lock);
1221         if (list_empty(&buffers->rb_send_bufs))
1222                 goto out_reqbuf;
1223         buffers->rb_send_count++;
1224         req = rpcrdma_buffer_get_req_locked(buffers);
1225         req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1226         spin_unlock(&buffers->rb_lock);
1227         return req;
1228
1229 out_reqbuf:
1230         spin_unlock(&buffers->rb_lock);
1231         pr_warn("RPC:       %s: out of request buffers\n", __func__);
1232         return NULL;
1233 }
1234
1235 /*
1236  * Put request/reply buffers back into pool.
1237  * Pre-decrement counter/array index.
1238  */
1239 void
1240 rpcrdma_buffer_put(struct rpcrdma_req *req)
1241 {
1242         struct rpcrdma_buffer *buffers = req->rl_buffer;
1243         struct rpcrdma_rep *rep = req->rl_reply;
1244
1245         req->rl_send_wr.num_sge = 0;
1246         req->rl_reply = NULL;
1247
1248         spin_lock(&buffers->rb_lock);
1249         buffers->rb_send_count--;
1250         list_add_tail(&req->rl_list, &buffers->rb_send_bufs);
1251         if (rep) {
1252                 buffers->rb_recv_count--;
1253                 list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1254         }
1255         spin_unlock(&buffers->rb_lock);
1256 }
1257
1258 /*
1259  * Recover reply buffers from pool.
1260  * This happens when recovering from disconnect.
1261  */
1262 void
1263 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1264 {
1265         struct rpcrdma_buffer *buffers = req->rl_buffer;
1266
1267         spin_lock(&buffers->rb_lock);
1268         req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1269         spin_unlock(&buffers->rb_lock);
1270 }
1271
1272 /*
1273  * Put reply buffers back into pool when not attached to
1274  * request. This happens in error conditions.
1275  */
1276 void
1277 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1278 {
1279         struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1280
1281         spin_lock(&buffers->rb_lock);
1282         buffers->rb_recv_count--;
1283         list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1284         spin_unlock(&buffers->rb_lock);
1285 }
1286
1287 /**
1288  * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1289  * @size: size of buffer to be allocated, in bytes
1290  * @direction: direction of data movement
1291  * @flags: GFP flags
1292  *
1293  * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
1294  * can be persistently DMA-mapped for I/O.
1295  *
1296  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1297  * receiving the payload of RDMA RECV operations. During Long Calls
1298  * or Replies they may be registered externally via ro_map.
1299  */
1300 struct rpcrdma_regbuf *
1301 rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
1302                      gfp_t flags)
1303 {
1304         struct rpcrdma_regbuf *rb;
1305
1306         rb = kmalloc(sizeof(*rb) + size, flags);
1307         if (rb == NULL)
1308                 return ERR_PTR(-ENOMEM);
1309
1310         rb->rg_device = NULL;
1311         rb->rg_direction = direction;
1312         rb->rg_iov.length = size;
1313
1314         return rb;
1315 }
1316
1317 /**
1318  * __rpcrdma_map_regbuf - DMA-map a regbuf
1319  * @ia: controlling rpcrdma_ia
1320  * @rb: regbuf to be mapped
1321  */
1322 bool
1323 __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1324 {
1325         struct ib_device *device = ia->ri_device;
1326
1327         if (rb->rg_direction == DMA_NONE)
1328                 return false;
1329
1330         rb->rg_iov.addr = ib_dma_map_single(device,
1331                                             (void *)rb->rg_base,
1332                                             rdmab_length(rb),
1333                                             rb->rg_direction);
1334         if (ib_dma_mapping_error(device, rdmab_addr(rb)))
1335                 return false;
1336
1337         rb->rg_device = device;
1338         rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey;
1339         return true;
1340 }
1341
1342 static void
1343 rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb)
1344 {
1345         if (!rb)
1346                 return;
1347
1348         if (!rpcrdma_regbuf_is_mapped(rb))
1349                 return;
1350
1351         ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb),
1352                             rdmab_length(rb), rb->rg_direction);
1353         rb->rg_device = NULL;
1354 }
1355
1356 /**
1357  * rpcrdma_free_regbuf - deregister and free registered buffer
1358  * @rb: regbuf to be deregistered and freed
1359  */
1360 void
1361 rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
1362 {
1363         rpcrdma_dma_unmap_regbuf(rb);
1364         kfree(rb);
1365 }
1366
1367 /*
1368  * Prepost any receive buffer, then post send.
1369  *
1370  * Receive buffer is donated to hardware, reclaimed upon recv completion.
1371  */
1372 int
1373 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1374                 struct rpcrdma_ep *ep,
1375                 struct rpcrdma_req *req)
1376 {
1377         struct ib_send_wr *send_wr = &req->rl_send_wr;
1378         struct ib_send_wr *send_wr_fail;
1379         int rc;
1380
1381         if (req->rl_reply) {
1382                 rc = rpcrdma_ep_post_recv(ia, req->rl_reply);
1383                 if (rc)
1384                         return rc;
1385                 req->rl_reply = NULL;
1386         }
1387
1388         dprintk("RPC:       %s: posting %d s/g entries\n",
1389                 __func__, send_wr->num_sge);
1390
1391         rpcrdma_set_signaled(ep, send_wr);
1392         rc = ib_post_send(ia->ri_id->qp, send_wr, &send_wr_fail);
1393         if (rc)
1394                 goto out_postsend_err;
1395         return 0;
1396
1397 out_postsend_err:
1398         pr_err("rpcrdma: RDMA Send ib_post_send returned %i\n", rc);
1399         return -ENOTCONN;
1400 }
1401
1402 int
1403 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1404                      struct rpcrdma_rep *rep)
1405 {
1406         struct ib_recv_wr *recv_wr_fail;
1407         int rc;
1408
1409         if (!rpcrdma_dma_map_regbuf(ia, rep->rr_rdmabuf))
1410                 goto out_map;
1411         rc = ib_post_recv(ia->ri_id->qp, &rep->rr_recv_wr, &recv_wr_fail);
1412         if (rc)
1413                 goto out_postrecv;
1414         return 0;
1415
1416 out_map:
1417         pr_err("rpcrdma: failed to DMA map the Receive buffer\n");
1418         return -EIO;
1419
1420 out_postrecv:
1421         pr_err("rpcrdma: ib_post_recv returned %i\n", rc);
1422         return -ENOTCONN;
1423 }
1424
1425 /**
1426  * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests
1427  * @r_xprt: transport associated with these backchannel resources
1428  * @min_reqs: minimum number of incoming requests expected
1429  *
1430  * Returns zero if all requested buffers were posted, or a negative errno.
1431  */
1432 int
1433 rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *r_xprt, unsigned int count)
1434 {
1435         struct rpcrdma_buffer *buffers = &r_xprt->rx_buf;
1436         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1437         struct rpcrdma_rep *rep;
1438         int rc;
1439
1440         while (count--) {
1441                 spin_lock(&buffers->rb_lock);
1442                 if (list_empty(&buffers->rb_recv_bufs))
1443                         goto out_reqbuf;
1444                 rep = rpcrdma_buffer_get_rep_locked(buffers);
1445                 spin_unlock(&buffers->rb_lock);
1446
1447                 rc = rpcrdma_ep_post_recv(ia, rep);
1448                 if (rc)
1449                         goto out_rc;
1450         }
1451
1452         return 0;
1453
1454 out_reqbuf:
1455         spin_unlock(&buffers->rb_lock);
1456         pr_warn("%s: no extra receive buffers\n", __func__);
1457         return -ENOMEM;
1458
1459 out_rc:
1460         rpcrdma_recv_buffer_put(rep);
1461         return rc;
1462 }