net/sunrpc/xprtrdma/verbs.c

   1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
   2 /*
   3  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
   4  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the BSD-type
  10  * license below:
  11  *
  12  * Redistribution and use in source and binary forms, with or without
  13  * modification, are permitted provided that the following conditions
  14  * are met:
  15  *
  16  *      Redistributions of source code must retain the above copyright
  17  *      notice, this list of conditions and the following disclaimer.
  18  *
  19  *      Redistributions in binary form must reproduce the above
  20  *      copyright notice, this list of conditions and the following
  21  *      disclaimer in the documentation and/or other materials provided
  22  *      with the distribution.
  23  *
  24  *      Neither the name of the Network Appliance, Inc. nor the names of
  25  *      its contributors may be used to endorse or promote products
  26  *      derived from this software without specific prior written
  27  *      permission.
  28  *
  29  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  30  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  31  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  32  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  33  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  34  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  35  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  36  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  37  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  38  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  39  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  40  */
  41
  42 /*
  43  * verbs.c
  44  *
  45  * Encapsulates the major functions managing:
  46  *  o adapters
  47  *  o endpoints
  48  *  o connections
  49  *  o buffer memory
  50  */
  51
  52 #include <linux/interrupt.h>
  53 #include <linux/slab.h>
  54 #include <linux/sunrpc/addr.h>
  55 #include <linux/sunrpc/svc_rdma.h>
  56
  57 #include <asm-generic/barrier.h>
  58 #include <asm/bitops.h>
  59
  60 #include <rdma/ib_cm.h>
  61
  62 #include "xprt_rdma.h"
  63 #include <trace/events/rpcrdma.h>
  64
  65 /*
  66  * Globals/Macros
  67  */
  68
  69 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  70 # define RPCDBG_FACILITY        RPCDBG_TRANS
  71 #endif
  72
  73 /*
  74  * internal functions
  75  */
  76 static void rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc);
  77 static void rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt);
  78 static void rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf);
  79 static int rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt, bool temp);
  80 static void rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb);
  81
  82 struct workqueue_struct *rpcrdma_receive_wq __read_mostly;
  83
  84 int
  85 rpcrdma_alloc_wq(void)
  86 {
  87         struct workqueue_struct *recv_wq;
  88
  89         recv_wq = alloc_workqueue("xprtrdma_receive",
  90                                   WQ_MEM_RECLAIM | WQ_HIGHPRI,
  91                                   0);
  92         if (!recv_wq)
  93                 return -ENOMEM;
  94
  95         rpcrdma_receive_wq = recv_wq;
  96         return 0;
  97 }
  98
  99 void
 100 rpcrdma_destroy_wq(void)
 101 {
 102         struct workqueue_struct *wq;
 103
 104         if (rpcrdma_receive_wq) {
 105                 wq = rpcrdma_receive_wq;
 106                 rpcrdma_receive_wq = NULL;
 107                 destroy_workqueue(wq);
 108         }
 109 }
 110
 111 static void
 112 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
 113 {
 114         struct rpcrdma_ep *ep = context;
 115         struct rpcrdma_xprt *r_xprt = container_of(ep, struct rpcrdma_xprt,
 116                                                    rx_ep);
 117
 118         trace_xprtrdma_qp_error(r_xprt, event);
 119         pr_err("rpcrdma: %s on device %s ep %p\n",
 120                ib_event_msg(event->event), event->device->name, context);
 121
 122         if (ep->rep_connected == 1) {
 123                 ep->rep_connected = -EIO;
 124                 rpcrdma_conn_func(ep);
 125                 wake_up_all(&ep->rep_connect_wait);
 126         }
 127 }
 128
 129 /**
 130  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
 131  * @cq: completion queue (ignored)
 132  * @wc: completed WR
 133  *
 134  */
 135 static void
 136 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
 137 {
 138         struct ib_cqe *cqe = wc->wr_cqe;
 139         struct rpcrdma_sendctx *sc =
 140                 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
 141
 142         /* WARNING: Only wr_cqe and status are reliable at this point */
 143         trace_xprtrdma_wc_send(sc, wc);
 144         if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
 145                 pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
 146                        ib_wc_status_msg(wc->status),
 147                        wc->status, wc->vendor_err);
 148
 149         rpcrdma_sendctx_put_locked(sc);
 150 }
 151
 152 /**
 153  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
 154  * @cq: completion queue (ignored)
 155  * @wc: completed WR
 156  *
 157  */
 158 static void
 159 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
 160 {
 161         struct ib_cqe *cqe = wc->wr_cqe;
 162         struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
 163                                                rr_cqe);
 164
 165         /* WARNING: Only wr_id and status are reliable at this point */
 166         trace_xprtrdma_wc_receive(wc);
 167         if (wc->status != IB_WC_SUCCESS)
 168                 goto out_fail;
 169
 170         /* status == SUCCESS means all fields in wc are trustworthy */
 171         rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
 172         rep->rr_wc_flags = wc->wc_flags;
 173         rep->rr_inv_rkey = wc->ex.invalidate_rkey;
 174
 175         ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
 176                                    rdmab_addr(rep->rr_rdmabuf),
 177                                    wc->byte_len, DMA_FROM_DEVICE);
 178
 179 out_schedule:
 180         rpcrdma_reply_handler(rep);
 181         return;
 182
 183 out_fail:
 184         if (wc->status != IB_WC_WR_FLUSH_ERR)
 185                 pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
 186                        ib_wc_status_msg(wc->status),
 187                        wc->status, wc->vendor_err);
 188         rpcrdma_set_xdrlen(&rep->rr_hdrbuf, 0);
 189         goto out_schedule;
 190 }
 191
 192 static void
 193 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
 194                                struct rdma_conn_param *param)
 195 {
 196         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
 197         const struct rpcrdma_connect_private *pmsg = param->private_data;
 198         unsigned int rsize, wsize;
 199
 200         /* Default settings for RPC-over-RDMA Version One */
 201         r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
 202         rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
 203         wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
 204
 205         if (pmsg &&
 206             pmsg->cp_magic == rpcrdma_cmp_magic &&
 207             pmsg->cp_version == RPCRDMA_CMP_VERSION) {
 208                 r_xprt->rx_ia.ri_implicit_roundup = true;
 209                 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
 210                 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
 211         }
 212
 213         if (rsize < cdata->inline_rsize)
 214                 cdata->inline_rsize = rsize;
 215         if (wsize < cdata->inline_wsize)
 216                 cdata->inline_wsize = wsize;
 217         dprintk("RPC:       %s: max send %u, max recv %u\n",
 218                 __func__, cdata->inline_wsize, cdata->inline_rsize);
 219         rpcrdma_set_max_header_sizes(r_xprt);
 220 }
 221
 222 static int
 223 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
 224 {
 225         struct rpcrdma_xprt *xprt = id->context;
 226         struct rpcrdma_ia *ia = &xprt->rx_ia;
 227         struct rpcrdma_ep *ep = &xprt->rx_ep;
 228         int connstate = 0;
 229
 230         trace_xprtrdma_conn_upcall(xprt, event);
 231         switch (event->event) {
 232         case RDMA_CM_EVENT_ADDR_RESOLVED:
 233         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 234                 ia->ri_async_rc = 0;
 235                 complete(&ia->ri_done);
 236                 break;
 237         case RDMA_CM_EVENT_ADDR_ERROR:
 238                 ia->ri_async_rc = -EPROTO;
 239                 complete(&ia->ri_done);
 240                 break;
 241         case RDMA_CM_EVENT_ROUTE_ERROR:
 242                 ia->ri_async_rc = -ENETUNREACH;
 243                 complete(&ia->ri_done);
 244                 break;
 245         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 246 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 247                 pr_info("rpcrdma: removing device %s for %s:%s\n",
 248                         ia->ri_device->name,
 249                         rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt));
 250 #endif
 251                 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
 252                 ep->rep_connected = -ENODEV;
 253                 xprt_force_disconnect(&xprt->rx_xprt);
 254                 wait_for_completion(&ia->ri_remove_done);
 255
 256                 ia->ri_id = NULL;
 257                 ia->ri_device = NULL;
 258                 /* Return 1 to ensure the core destroys the id. */
 259                 return 1;
 260         case RDMA_CM_EVENT_ESTABLISHED:
 261                 ++xprt->rx_xprt.connect_cookie;
 262                 connstate = 1;
 263                 rpcrdma_update_connect_private(xprt, &event->param.conn);
 264                 goto connected;
 265         case RDMA_CM_EVENT_CONNECT_ERROR:
 266                 connstate = -ENOTCONN;
 267                 goto connected;
 268         case RDMA_CM_EVENT_UNREACHABLE:
 269                 connstate = -ENETUNREACH;
 270                 goto connected;
 271         case RDMA_CM_EVENT_REJECTED:
 272                 dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
 273                         rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
 274                         rdma_reject_msg(id, event->status));
 275                 connstate = -ECONNREFUSED;
 276                 if (event->status == IB_CM_REJ_STALE_CONN)
 277                         connstate = -EAGAIN;
 278                 goto connected;
 279         case RDMA_CM_EVENT_DISCONNECTED:
 280                 ++xprt->rx_xprt.connect_cookie;
 281                 connstate = -ECONNABORTED;
 282 connected:
 283                 xprt->rx_buf.rb_credits = 1;
 284                 ep->rep_connected = connstate;
 285                 rpcrdma_conn_func(ep);
 286                 wake_up_all(&ep->rep_connect_wait);
 287                 /*FALLTHROUGH*/
 288         default:
 289                 dprintk("RPC:       %s: %s:%s on %s/%s (ep 0x%p): %s\n",
 290                         __func__,
 291                         rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
 292                         ia->ri_device->name, ia->ri_ops->ro_displayname,
 293                         ep, rdma_event_msg(event->event));
 294                 break;
 295         }
 296
 297         return 0;
 298 }
 299
 300 static struct rdma_cm_id *
 301 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
 302 {
 303         unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
 304         struct rdma_cm_id *id;
 305         int rc;
 306
 307         trace_xprtrdma_conn_start(xprt);
 308
 309         init_completion(&ia->ri_done);
 310         init_completion(&ia->ri_remove_done);
 311
 312         id = rdma_create_id(xprt->rx_xprt.xprt_net, rpcrdma_conn_upcall,
 313                             xprt, RDMA_PS_TCP, IB_QPT_RC);
 314         if (IS_ERR(id)) {
 315                 rc = PTR_ERR(id);
 316                 dprintk("RPC:       %s: rdma_create_id() failed %i\n",
 317                         __func__, rc);
 318                 return id;
 319         }
 320
 321         ia->ri_async_rc = -ETIMEDOUT;
 322         rc = rdma_resolve_addr(id, NULL,
 323                                (struct sockaddr *)&xprt->rx_xprt.addr,
 324                                RDMA_RESOLVE_TIMEOUT);
 325         if (rc) {
 326                 dprintk("RPC:       %s: rdma_resolve_addr() failed %i\n",
 327                         __func__, rc);
 328                 goto out;
 329         }
 330         rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
 331         if (rc < 0) {
 332                 trace_xprtrdma_conn_tout(xprt);
 333                 goto out;
 334         }
 335
 336         rc = ia->ri_async_rc;
 337         if (rc)
 338                 goto out;
 339
 340         ia->ri_async_rc = -ETIMEDOUT;
 341         rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
 342         if (rc) {
 343                 dprintk("RPC:       %s: rdma_resolve_route() failed %i\n",
 344                         __func__, rc);
 345                 goto out;
 346         }
 347         rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
 348         if (rc < 0) {
 349                 trace_xprtrdma_conn_tout(xprt);
 350                 goto out;
 351         }
 352         rc = ia->ri_async_rc;
 353         if (rc)
 354                 goto out;
 355
 356         return id;
 357
 358 out:
 359         rdma_destroy_id(id);
 360         return ERR_PTR(rc);
 361 }
 362
 363 /*
 364  * Exported functions.
 365  */
 366
 367 /**
 368  * rpcrdma_ia_open - Open and initialize an Interface Adapter.
 369  * @xprt: transport with IA to (re)initialize
 370  *
 371  * Returns 0 on success, negative errno if an appropriate
 372  * Interface Adapter could not be found and opened.
 373  */
 374 int
 375 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
 376 {
 377         struct rpcrdma_ia *ia = &xprt->rx_ia;
 378         int rc;
 379
 380         ia->ri_id = rpcrdma_create_id(xprt, ia);
 381         if (IS_ERR(ia->ri_id)) {
 382                 rc = PTR_ERR(ia->ri_id);
 383                 goto out_err;
 384         }
 385         ia->ri_device = ia->ri_id->device;
 386
 387         ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
 388         if (IS_ERR(ia->ri_pd)) {
 389                 rc = PTR_ERR(ia->ri_pd);
 390                 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
 391                 goto out_err;
 392         }
 393
 394         switch (xprt_rdma_memreg_strategy) {
 395         case RPCRDMA_FRWR:
 396                 if (frwr_is_supported(ia)) {
 397                         ia->ri_ops = &rpcrdma_frwr_memreg_ops;
 398                         break;
 399                 }
 400                 /*FALLTHROUGH*/
 401         case RPCRDMA_MTHCAFMR:
 402                 if (fmr_is_supported(ia)) {
 403                         ia->ri_ops = &rpcrdma_fmr_memreg_ops;
 404                         break;
 405                 }
 406                 /*FALLTHROUGH*/
 407         default:
 408                 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
 409                        ia->ri_device->name, xprt_rdma_memreg_strategy);
 410                 rc = -EINVAL;
 411                 goto out_err;
 412         }
 413
 414         return 0;
 415
 416 out_err:
 417         rpcrdma_ia_close(ia);
 418         return rc;
 419 }
 420
 421 /**
 422  * rpcrdma_ia_remove - Handle device driver unload
 423  * @ia: interface adapter being removed
 424  *
 425  * Divest transport H/W resources associated with this adapter,
 426  * but allow it to be restored later.
 427  */
 428 void
 429 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
 430 {
 431         struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
 432                                                    rx_ia);
 433         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 434         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 435         struct rpcrdma_req *req;
 436         struct rpcrdma_rep *rep;
 437
 438         cancel_delayed_work_sync(&buf->rb_refresh_worker);
 439
 440         /* This is similar to rpcrdma_ep_destroy, but:
 441          * - Don't cancel the connect worker.
 442          * - Don't call rpcrdma_ep_disconnect, which waits
 443          *   for another conn upcall, which will deadlock.
 444          * - rdma_disconnect is unneeded, the underlying
 445          *   connection is already gone.
 446          */
 447         if (ia->ri_id->qp) {
 448                 ib_drain_qp(ia->ri_id->qp);
 449                 rdma_destroy_qp(ia->ri_id);
 450                 ia->ri_id->qp = NULL;
 451         }
 452         ib_free_cq(ep->rep_attr.recv_cq);
 453         ep->rep_attr.recv_cq = NULL;
 454         ib_free_cq(ep->rep_attr.send_cq);
 455         ep->rep_attr.send_cq = NULL;
 456
 457         /* The ULP is responsible for ensuring all DMA
 458          * mappings and MRs are gone.
 459          */
 460         list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
 461                 rpcrdma_dma_unmap_regbuf(rep->rr_rdmabuf);
 462         list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
 463                 rpcrdma_dma_unmap_regbuf(req->rl_rdmabuf);
 464                 rpcrdma_dma_unmap_regbuf(req->rl_sendbuf);
 465                 rpcrdma_dma_unmap_regbuf(req->rl_recvbuf);
 466         }
 467         rpcrdma_mrs_destroy(buf);
 468         ib_dealloc_pd(ia->ri_pd);
 469         ia->ri_pd = NULL;
 470
 471         /* Allow waiters to continue */
 472         complete(&ia->ri_remove_done);
 473
 474         trace_xprtrdma_remove(r_xprt);
 475 }
 476
 477 /**
 478  * rpcrdma_ia_close - Clean up/close an IA.
 479  * @ia: interface adapter to close
 480  *
 481  */
 482 void
 483 rpcrdma_ia_close(struct rpcrdma_ia *ia)
 484 {
 485         if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
 486                 if (ia->ri_id->qp)
 487                         rdma_destroy_qp(ia->ri_id);
 488                 rdma_destroy_id(ia->ri_id);
 489         }
 490         ia->ri_id = NULL;
 491         ia->ri_device = NULL;
 492
 493         /* If the pd is still busy, xprtrdma missed freeing a resource */
 494         if (ia->ri_pd && !IS_ERR(ia->ri_pd))
 495                 ib_dealloc_pd(ia->ri_pd);
 496         ia->ri_pd = NULL;
 497 }
 498
 499 /*
 500  * Create unconnected endpoint.
 501  */
 502 int
 503 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
 504                   struct rpcrdma_create_data_internal *cdata)
 505 {
 506         struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
 507         struct ib_cq *sendcq, *recvcq;
 508         unsigned int max_sge;
 509         int rc;
 510
 511         max_sge = min_t(unsigned int, ia->ri_device->attrs.max_send_sge,
 512                         RPCRDMA_MAX_SEND_SGES);
 513         if (max_sge < RPCRDMA_MIN_SEND_SGES) {
 514                 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
 515                 return -ENOMEM;
 516         }
 517         ia->ri_max_send_sges = max_sge;
 518
 519         rc = ia->ri_ops->ro_open(ia, ep, cdata);
 520         if (rc)
 521                 return rc;
 522
 523         ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
 524         ep->rep_attr.qp_context = ep;
 525         ep->rep_attr.srq = NULL;
 526         ep->rep_attr.cap.max_send_sge = max_sge;
 527         ep->rep_attr.cap.max_recv_sge = 1;
 528         ep->rep_attr.cap.max_inline_data = 0;
 529         ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 530         ep->rep_attr.qp_type = IB_QPT_RC;
 531         ep->rep_attr.port_num = ~0;
 532
 533         dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
 534                 "iovs: send %d recv %d\n",
 535                 __func__,
 536                 ep->rep_attr.cap.max_send_wr,
 537                 ep->rep_attr.cap.max_recv_wr,
 538                 ep->rep_attr.cap.max_send_sge,
 539                 ep->rep_attr.cap.max_recv_sge);
 540
 541         /* set trigger for requesting send completion */
 542         ep->rep_send_batch = min_t(unsigned int, RPCRDMA_MAX_SEND_BATCH,
 543                                    cdata->max_requests >> 2);
 544         ep->rep_send_count = ep->rep_send_batch;
 545         init_waitqueue_head(&ep->rep_connect_wait);
 546         INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
 547
 548         sendcq = ib_alloc_cq(ia->ri_device, NULL,
 549                              ep->rep_attr.cap.max_send_wr + 1,
 550                              1, IB_POLL_WORKQUEUE);
 551         if (IS_ERR(sendcq)) {
 552                 rc = PTR_ERR(sendcq);
 553                 dprintk("RPC:       %s: failed to create send CQ: %i\n",
 554                         __func__, rc);
 555                 goto out1;
 556         }
 557
 558         recvcq = ib_alloc_cq(ia->ri_device, NULL,
 559                              ep->rep_attr.cap.max_recv_wr + 1,
 560                              0, IB_POLL_WORKQUEUE);
 561         if (IS_ERR(recvcq)) {
 562                 rc = PTR_ERR(recvcq);
 563                 dprintk("RPC:       %s: failed to create recv CQ: %i\n",
 564                         __func__, rc);
 565                 goto out2;
 566         }
 567
 568         ep->rep_attr.send_cq = sendcq;
 569         ep->rep_attr.recv_cq = recvcq;
 570
 571         /* Initialize cma parameters */
 572         memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
 573
 574         /* Prepare RDMA-CM private message */
 575         pmsg->cp_magic = rpcrdma_cmp_magic;
 576         pmsg->cp_version = RPCRDMA_CMP_VERSION;
 577         pmsg->cp_flags |= ia->ri_ops->ro_send_w_inv_ok;
 578         pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize);
 579         pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize);
 580         ep->rep_remote_cma.private_data = pmsg;
 581         ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
 582
 583         /* Client offers RDMA Read but does not initiate */
 584         ep->rep_remote_cma.initiator_depth = 0;
 585         ep->rep_remote_cma.responder_resources =
 586                 min_t(int, U8_MAX, ia->ri_device->attrs.max_qp_rd_atom);
 587
 588         /* Limit transport retries so client can detect server
 589          * GID changes quickly. RPC layer handles re-establishing
 590          * transport connection and retransmission.
 591          */
 592         ep->rep_remote_cma.retry_count = 6;
 593
 594         /* RPC-over-RDMA handles its own flow control. In addition,
 595          * make all RNR NAKs visible so we know that RPC-over-RDMA
 596          * flow control is working correctly (no NAKs should be seen).
 597          */
 598         ep->rep_remote_cma.flow_control = 0;
 599         ep->rep_remote_cma.rnr_retry_count = 0;
 600
 601         return 0;
 602
 603 out2:
 604         ib_free_cq(sendcq);
 605 out1:
 606         return rc;
 607 }
 608
 609 /*
 610  * rpcrdma_ep_destroy
 611  *
 612  * Disconnect and destroy endpoint. After this, the only
 613  * valid operations on the ep are to free it (if dynamically
 614  * allocated) or re-create it.
 615  */
 616 void
 617 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 618 {
 619         cancel_delayed_work_sync(&ep->rep_connect_worker);
 620
 621         if (ia->ri_id && ia->ri_id->qp) {
 622                 rpcrdma_ep_disconnect(ep, ia);
 623                 rdma_destroy_qp(ia->ri_id);
 624                 ia->ri_id->qp = NULL;
 625         }
 626
 627         if (ep->rep_attr.recv_cq)
 628                 ib_free_cq(ep->rep_attr.recv_cq);
 629         if (ep->rep_attr.send_cq)
 630                 ib_free_cq(ep->rep_attr.send_cq);
 631 }
 632
 633 /* Re-establish a connection after a device removal event.
 634  * Unlike a normal reconnection, a fresh PD and a new set
 635  * of MRs and buffers is needed.
 636  */
 637 static int
 638 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
 639                          struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 640 {
 641         int rc, err;
 642
 643         trace_xprtrdma_reinsert(r_xprt);
 644
 645         rc = -EHOSTUNREACH;
 646         if (rpcrdma_ia_open(r_xprt))
 647                 goto out1;
 648
 649         rc = -ENOMEM;
 650         err = rpcrdma_ep_create(ep, ia, &r_xprt->rx_data);
 651         if (err) {
 652                 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
 653                 goto out2;
 654         }
 655
 656         rc = -ENETUNREACH;
 657         err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
 658         if (err) {
 659                 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
 660                 goto out3;
 661         }
 662
 663         rpcrdma_mrs_create(r_xprt);
 664         return 0;
 665
 666 out3:
 667         rpcrdma_ep_destroy(ep, ia);
 668 out2:
 669         rpcrdma_ia_close(ia);
 670 out1:
 671         return rc;
 672 }
 673
 674 static int
 675 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
 676                      struct rpcrdma_ia *ia)
 677 {
 678         struct rdma_cm_id *id, *old;
 679         int err, rc;
 680
 681         trace_xprtrdma_reconnect(r_xprt);
 682
 683         rpcrdma_ep_disconnect(ep, ia);
 684
 685         rc = -EHOSTUNREACH;
 686         id = rpcrdma_create_id(r_xprt, ia);
 687         if (IS_ERR(id))
 688                 goto out;
 689
 690         /* As long as the new ID points to the same device as the
 691          * old ID, we can reuse the transport's existing PD and all
 692          * previously allocated MRs. Also, the same device means
 693          * the transport's previous DMA mappings are still valid.
 694          *
 695          * This is a sanity check only. There should be no way these
 696          * point to two different devices here.
 697          */
 698         old = id;
 699         rc = -ENETUNREACH;
 700         if (ia->ri_device != id->device) {
 701                 pr_err("rpcrdma: can't reconnect on different device!\n");
 702                 goto out_destroy;
 703         }
 704
 705         err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
 706         if (err) {
 707                 dprintk("RPC:       %s: rdma_create_qp returned %d\n",
 708                         __func__, err);
 709                 goto out_destroy;
 710         }
 711
 712         /* Atomically replace the transport's ID and QP. */
 713         rc = 0;
 714         old = ia->ri_id;
 715         ia->ri_id = id;
 716         rdma_destroy_qp(old);
 717
 718 out_destroy:
 719         rdma_destroy_id(old);
 720 out:
 721         return rc;
 722 }
 723
 724 /*
 725  * Connect unconnected endpoint.
 726  */
 727 int
 728 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 729 {
 730         struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
 731                                                    rx_ia);
 732         int rc;
 733
 734 retry:
 735         switch (ep->rep_connected) {
 736         case 0:
 737                 dprintk("RPC:       %s: connecting...\n", __func__);
 738                 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
 739                 if (rc) {
 740                         dprintk("RPC:       %s: rdma_create_qp failed %i\n",
 741                                 __func__, rc);
 742                         rc = -ENETUNREACH;
 743                         goto out_noupdate;
 744                 }
 745                 break;
 746         case -ENODEV:
 747                 rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
 748                 if (rc)
 749                         goto out_noupdate;
 750                 break;
 751         default:
 752                 rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
 753                 if (rc)
 754                         goto out;
 755         }
 756
 757         ep->rep_connected = 0;
 758
 759         rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
 760         if (rc) {
 761                 dprintk("RPC:       %s: rdma_connect() failed with %i\n",
 762                                 __func__, rc);
 763                 goto out;
 764         }
 765
 766         wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
 767         if (ep->rep_connected <= 0) {
 768                 if (ep->rep_connected == -EAGAIN)
 769                         goto retry;
 770                 rc = ep->rep_connected;
 771                 goto out;
 772         }
 773
 774         dprintk("RPC:       %s: connected\n", __func__);
 775
 776         rpcrdma_post_recvs(r_xprt, true);
 777
 778 out:
 779         if (rc)
 780                 ep->rep_connected = rc;
 781
 782 out_noupdate:
 783         return rc;
 784 }
 785
 786 /*
 787  * rpcrdma_ep_disconnect
 788  *
 789  * This is separate from destroy to facilitate the ability
 790  * to reconnect without recreating the endpoint.
 791  *
 792  * This call is not reentrant, and must not be made in parallel
 793  * on the same endpoint.
 794  */
 795 void
 796 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 797 {
 798         int rc;
 799
 800         rc = rdma_disconnect(ia->ri_id);
 801         if (!rc)
 802                 /* returns without wait if not connected */
 803                 wait_event_interruptible(ep->rep_connect_wait,
 804                                                         ep->rep_connected != 1);
 805         else
 806                 ep->rep_connected = rc;
 807         trace_xprtrdma_disconnect(container_of(ep, struct rpcrdma_xprt,
 808                                                rx_ep), rc);
 809
 810         ib_drain_qp(ia->ri_id->qp);
 811 }
 812
 813 /* Fixed-size circular FIFO queue. This implementation is wait-free and
 814  * lock-free.
 815  *
 816  * Consumer is the code path that posts Sends. This path dequeues a
 817  * sendctx for use by a Send operation. Multiple consumer threads
 818  * are serialized by the RPC transport lock, which allows only one
 819  * ->send_request call at a time.
 820  *
 821  * Producer is the code path that handles Send completions. This path
 822  * enqueues a sendctx that has been completed. Multiple producer
 823  * threads are serialized by the ib_poll_cq() function.
 824  */
 825
 826 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
 827  * queue activity, and ib_drain_qp has flushed all remaining Send
 828  * requests.
 829  */
 830 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
 831 {
 832         unsigned long i;
 833
 834         for (i = 0; i <= buf->rb_sc_last; i++)
 835                 kfree(buf->rb_sc_ctxs[i]);
 836         kfree(buf->rb_sc_ctxs);
 837 }
 838
 839 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
 840 {
 841         struct rpcrdma_sendctx *sc;
 842
 843         sc = kzalloc(sizeof(*sc) +
 844                      ia->ri_max_send_sges * sizeof(struct ib_sge),
 845                      GFP_KERNEL);
 846         if (!sc)
 847                 return NULL;
 848
 849         sc->sc_wr.wr_cqe = &sc->sc_cqe;
 850         sc->sc_wr.sg_list = sc->sc_sges;
 851         sc->sc_wr.opcode = IB_WR_SEND;
 852         sc->sc_cqe.done = rpcrdma_wc_send;
 853         return sc;
 854 }
 855
 856 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
 857 {
 858         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 859         struct rpcrdma_sendctx *sc;
 860         unsigned long i;
 861
 862         /* Maximum number of concurrent outstanding Send WRs. Capping
 863          * the circular queue size stops Send Queue overflow by causing
 864          * the ->send_request call to fail temporarily before too many
 865          * Sends are posted.
 866          */
 867         i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
 868         dprintk("RPC:       %s: allocating %lu send_ctxs\n", __func__, i);
 869         buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
 870         if (!buf->rb_sc_ctxs)
 871                 return -ENOMEM;
 872
 873         buf->rb_sc_last = i - 1;
 874         for (i = 0; i <= buf->rb_sc_last; i++) {
 875                 sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
 876                 if (!sc)
 877                         goto out_destroy;
 878
 879                 sc->sc_xprt = r_xprt;
 880                 buf->rb_sc_ctxs[i] = sc;
 881         }
 882         buf->rb_flags = 0;
 883
 884         return 0;
 885
 886 out_destroy:
 887         rpcrdma_sendctxs_destroy(buf);
 888         return -ENOMEM;
 889 }
 890
 891 /* The sendctx queue is not guaranteed to have a size that is a
 892  * power of two, thus the helpers in circ_buf.h cannot be used.
 893  * The other option is to use modulus (%), which can be expensive.
 894  */
 895 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
 896                                           unsigned long item)
 897 {
 898         return likely(item < buf->rb_sc_last) ? item + 1 : 0;
 899 }
 900
 901 /**
 902  * rpcrdma_sendctx_get_locked - Acquire a send context
 903  * @buf: transport buffers from which to acquire an unused context
 904  *
 905  * Returns pointer to a free send completion context; or NULL if
 906  * the queue is empty.
 907  *
 908  * Usage: Called to acquire an SGE array before preparing a Send WR.
 909  *
 910  * The caller serializes calls to this function (per rpcrdma_buffer),
 911  * and provides an effective memory barrier that flushes the new value
 912  * of rb_sc_head.
 913  */
 914 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf)
 915 {
 916         struct rpcrdma_xprt *r_xprt;
 917         struct rpcrdma_sendctx *sc;
 918         unsigned long next_head;
 919
 920         next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
 921
 922         if (next_head == READ_ONCE(buf->rb_sc_tail))
 923                 goto out_emptyq;
 924
 925         /* ORDER: item must be accessed _before_ head is updated */
 926         sc = buf->rb_sc_ctxs[next_head];
 927
 928         /* Releasing the lock in the caller acts as a memory
 929          * barrier that flushes rb_sc_head.
 930          */
 931         buf->rb_sc_head = next_head;
 932
 933         return sc;
 934
 935 out_emptyq:
 936         /* The queue is "empty" if there have not been enough Send
 937          * completions recently. This is a sign the Send Queue is
 938          * backing up. Cause the caller to pause and try again.
 939          */
 940         set_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags);
 941         r_xprt = container_of(buf, struct rpcrdma_xprt, rx_buf);
 942         r_xprt->rx_stats.empty_sendctx_q++;
 943         return NULL;
 944 }
 945
 946 /**
 947  * rpcrdma_sendctx_put_locked - Release a send context
 948  * @sc: send context to release
 949  *
 950  * Usage: Called from Send completion to return a sendctxt
 951  * to the queue.
 952  *
 953  * The caller serializes calls to this function (per rpcrdma_buffer).
 954  */
 955 static void
 956 rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc)
 957 {
 958         struct rpcrdma_buffer *buf = &sc->sc_xprt->rx_buf;
 959         unsigned long next_tail;
 960
 961         /* Unmap SGEs of previously completed by unsignaled
 962          * Sends by walking up the queue until @sc is found.
 963          */
 964         next_tail = buf->rb_sc_tail;
 965         do {
 966                 next_tail = rpcrdma_sendctx_next(buf, next_tail);
 967
 968                 /* ORDER: item must be accessed _before_ tail is updated */
 969                 rpcrdma_unmap_sendctx(buf->rb_sc_ctxs[next_tail]);
 970
 971         } while (buf->rb_sc_ctxs[next_tail] != sc);
 972
 973         /* Paired with READ_ONCE */
 974         smp_store_release(&buf->rb_sc_tail, next_tail);
 975
 976         if (test_and_clear_bit(RPCRDMA_BUF_F_EMPTY_SCQ, &buf->rb_flags)) {
 977                 smp_mb__after_atomic();
 978                 xprt_write_space(&sc->sc_xprt->rx_xprt);
 979         }
 980 }
 981
 982 static void
 983 rpcrdma_mr_recovery_worker(struct work_struct *work)
 984 {
 985         struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
 986                                                   rb_recovery_worker.work);
 987         struct rpcrdma_mr *mr;
 988
 989         spin_lock(&buf->rb_recovery_lock);
 990         while (!list_empty(&buf->rb_stale_mrs)) {
 991                 mr = rpcrdma_mr_pop(&buf->rb_stale_mrs);
 992                 spin_unlock(&buf->rb_recovery_lock);
 993
 994                 trace_xprtrdma_recover_mr(mr);
 995                 mr->mr_xprt->rx_ia.ri_ops->ro_recover_mr(mr);
 996
 997                 spin_lock(&buf->rb_recovery_lock);
 998         }
 999         spin_unlock(&buf->rb_recovery_lock);
1000 }
1001
1002 void
1003 rpcrdma_mr_defer_recovery(struct rpcrdma_mr *mr)
1004 {
1005         struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1006         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1007
1008         spin_lock(&buf->rb_recovery_lock);
1009         rpcrdma_mr_push(mr, &buf->rb_stale_mrs);
1010         spin_unlock(&buf->rb_recovery_lock);
1011
1012         schedule_delayed_work(&buf->rb_recovery_worker, 0);
1013 }
1014
1015 static void
1016 rpcrdma_mrs_create(struct rpcrdma_xprt *r_xprt)
1017 {
1018         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1019         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1020         unsigned int count;
1021         LIST_HEAD(free);
1022         LIST_HEAD(all);
1023
1024         for (count = 0; count < 3; count++) {
1025                 struct rpcrdma_mr *mr;
1026                 int rc;
1027
1028                 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
1029                 if (!mr)
1030                         break;
1031
1032                 rc = ia->ri_ops->ro_init_mr(ia, mr);
1033                 if (rc) {
1034                         kfree(mr);
1035                         break;
1036                 }
1037
1038                 mr->mr_xprt = r_xprt;
1039
1040                 list_add(&mr->mr_list, &free);
1041                 list_add(&mr->mr_all, &all);
1042         }
1043
1044         spin_lock(&buf->rb_mrlock);
1045         list_splice(&free, &buf->rb_mrs);
1046         list_splice(&all, &buf->rb_all);
1047         r_xprt->rx_stats.mrs_allocated += count;
1048         spin_unlock(&buf->rb_mrlock);
1049         trace_xprtrdma_createmrs(r_xprt, count);
1050
1051         xprt_write_space(&r_xprt->rx_xprt);
1052 }
1053
1054 static void
1055 rpcrdma_mr_refresh_worker(struct work_struct *work)
1056 {
1057         struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
1058                                                   rb_refresh_worker.work);
1059         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1060                                                    rx_buf);
1061
1062         rpcrdma_mrs_create(r_xprt);
1063 }
1064
1065 struct rpcrdma_req *
1066 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
1067 {
1068         struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1069         struct rpcrdma_regbuf *rb;
1070         struct rpcrdma_req *req;
1071
1072         req = kzalloc(sizeof(*req), GFP_KERNEL);
1073         if (req == NULL)
1074                 return ERR_PTR(-ENOMEM);
1075
1076         rb = rpcrdma_alloc_regbuf(RPCRDMA_HDRBUF_SIZE,
1077                                   DMA_TO_DEVICE, GFP_KERNEL);
1078         if (IS_ERR(rb)) {
1079                 kfree(req);
1080                 return ERR_PTR(-ENOMEM);
1081         }
1082         req->rl_rdmabuf = rb;
1083         xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
1084         req->rl_buffer = buffer;
1085         INIT_LIST_HEAD(&req->rl_registered);
1086
1087         spin_lock(&buffer->rb_reqslock);
1088         list_add(&req->rl_all, &buffer->rb_allreqs);
1089         spin_unlock(&buffer->rb_reqslock);
1090         return req;
1091 }
1092
1093 static int
1094 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt, bool temp)
1095 {
1096         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1097         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1098         struct rpcrdma_rep *rep;
1099         int rc;
1100
1101         rc = -ENOMEM;
1102         rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1103         if (rep == NULL)
1104                 goto out;
1105
1106         rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
1107                                                DMA_FROM_DEVICE, GFP_KERNEL);
1108         if (IS_ERR(rep->rr_rdmabuf)) {
1109                 rc = PTR_ERR(rep->rr_rdmabuf);
1110                 goto out_free;
1111         }
1112         xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base,
1113                      rdmab_length(rep->rr_rdmabuf));
1114
1115         rep->rr_cqe.done = rpcrdma_wc_receive;
1116         rep->rr_rxprt = r_xprt;
1117         INIT_WORK(&rep->rr_work, rpcrdma_deferred_completion);
1118         rep->rr_recv_wr.next = NULL;
1119         rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
1120         rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1121         rep->rr_recv_wr.num_sge = 1;
1122         rep->rr_temp = temp;
1123
1124         spin_lock(&buf->rb_lock);
1125         list_add(&rep->rr_list, &buf->rb_recv_bufs);
1126         spin_unlock(&buf->rb_lock);
1127         return 0;
1128
1129 out_free:
1130         kfree(rep);
1131 out:
1132         dprintk("RPC:       %s: reply buffer %d alloc failed\n",
1133                 __func__, rc);
1134         return rc;
1135 }
1136
1137 int
1138 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1139 {
1140         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1141         int i, rc;
1142
1143         buf->rb_max_requests = r_xprt->rx_data.max_requests;
1144         buf->rb_bc_srv_max_requests = 0;
1145         spin_lock_init(&buf->rb_mrlock);
1146         spin_lock_init(&buf->rb_lock);
1147         spin_lock_init(&buf->rb_recovery_lock);
1148         INIT_LIST_HEAD(&buf->rb_mrs);
1149         INIT_LIST_HEAD(&buf->rb_all);
1150         INIT_LIST_HEAD(&buf->rb_stale_mrs);
1151         INIT_DELAYED_WORK(&buf->rb_refresh_worker,
1152                           rpcrdma_mr_refresh_worker);
1153         INIT_DELAYED_WORK(&buf->rb_recovery_worker,
1154                           rpcrdma_mr_recovery_worker);
1155
1156         rpcrdma_mrs_create(r_xprt);
1157
1158         INIT_LIST_HEAD(&buf->rb_send_bufs);
1159         INIT_LIST_HEAD(&buf->rb_allreqs);
1160         spin_lock_init(&buf->rb_reqslock);
1161         for (i = 0; i < buf->rb_max_requests; i++) {
1162                 struct rpcrdma_req *req;
1163
1164                 req = rpcrdma_create_req(r_xprt);
1165                 if (IS_ERR(req)) {
1166                         dprintk("RPC:       %s: request buffer %d alloc"
1167                                 " failed\n", __func__, i);
1168                         rc = PTR_ERR(req);
1169                         goto out;
1170                 }
1171                 list_add(&req->rl_list, &buf->rb_send_bufs);
1172         }
1173
1174         buf->rb_posted_receives = 0;
1175         INIT_LIST_HEAD(&buf->rb_recv_bufs);
1176
1177         rc = rpcrdma_sendctxs_create(r_xprt);
1178         if (rc)
1179                 goto out;
1180
1181         return 0;
1182 out:
1183         rpcrdma_buffer_destroy(buf);
1184         return rc;
1185 }
1186
1187 static void
1188 rpcrdma_destroy_rep(struct rpcrdma_rep *rep)
1189 {
1190         rpcrdma_free_regbuf(rep->rr_rdmabuf);
1191         kfree(rep);
1192 }
1193
1194 void
1195 rpcrdma_destroy_req(struct rpcrdma_req *req)
1196 {
1197         rpcrdma_free_regbuf(req->rl_recvbuf);
1198         rpcrdma_free_regbuf(req->rl_sendbuf);
1199         rpcrdma_free_regbuf(req->rl_rdmabuf);
1200         kfree(req);
1201 }
1202
1203 static void
1204 rpcrdma_mrs_destroy(struct rpcrdma_buffer *buf)
1205 {
1206         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1207                                                    rx_buf);
1208         struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1209         struct rpcrdma_mr *mr;
1210         unsigned int count;
1211
1212         count = 0;
1213         spin_lock(&buf->rb_mrlock);
1214         while (!list_empty(&buf->rb_all)) {
1215                 mr = list_entry(buf->rb_all.next, struct rpcrdma_mr, mr_all);
1216                 list_del(&mr->mr_all);
1217
1218                 spin_unlock(&buf->rb_mrlock);
1219
1220                 /* Ensure MW is not on any rl_registered list */
1221                 if (!list_empty(&mr->mr_list))
1222                         list_del(&mr->mr_list);
1223
1224                 ia->ri_ops->ro_release_mr(mr);
1225                 count++;
1226                 spin_lock(&buf->rb_mrlock);
1227         }
1228         spin_unlock(&buf->rb_mrlock);
1229         r_xprt->rx_stats.mrs_allocated = 0;
1230
1231         dprintk("RPC:       %s: released %u MRs\n", __func__, count);
1232 }
1233
1234 void
1235 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1236 {
1237         cancel_delayed_work_sync(&buf->rb_recovery_worker);
1238         cancel_delayed_work_sync(&buf->rb_refresh_worker);
1239
1240         rpcrdma_sendctxs_destroy(buf);
1241
1242         while (!list_empty(&buf->rb_recv_bufs)) {
1243                 struct rpcrdma_rep *rep;
1244
1245                 rep = list_first_entry(&buf->rb_recv_bufs,
1246                                        struct rpcrdma_rep, rr_list);
1247                 list_del(&rep->rr_list);
1248                 rpcrdma_destroy_rep(rep);
1249         }
1250
1251         spin_lock(&buf->rb_reqslock);
1252         while (!list_empty(&buf->rb_allreqs)) {
1253                 struct rpcrdma_req *req;
1254
1255                 req = list_first_entry(&buf->rb_allreqs,
1256                                        struct rpcrdma_req, rl_all);
1257                 list_del(&req->rl_all);
1258
1259                 spin_unlock(&buf->rb_reqslock);
1260                 rpcrdma_destroy_req(req);
1261                 spin_lock(&buf->rb_reqslock);
1262         }
1263         spin_unlock(&buf->rb_reqslock);
1264
1265         rpcrdma_mrs_destroy(buf);
1266 }
1267
1268 /**
1269  * rpcrdma_mr_get - Allocate an rpcrdma_mr object
1270  * @r_xprt: controlling transport
1271  *
1272  * Returns an initialized rpcrdma_mr or NULL if no free
1273  * rpcrdma_mr objects are available.
1274  */
1275 struct rpcrdma_mr *
1276 rpcrdma_mr_get(struct rpcrdma_xprt *r_xprt)
1277 {
1278         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1279         struct rpcrdma_mr *mr = NULL;
1280
1281         spin_lock(&buf->rb_mrlock);
1282         if (!list_empty(&buf->rb_mrs))
1283                 mr = rpcrdma_mr_pop(&buf->rb_mrs);
1284         spin_unlock(&buf->rb_mrlock);
1285
1286         if (!mr)
1287                 goto out_nomrs;
1288         return mr;
1289
1290 out_nomrs:
1291         trace_xprtrdma_nomrs(r_xprt);
1292         if (r_xprt->rx_ep.rep_connected != -ENODEV)
1293                 schedule_delayed_work(&buf->rb_refresh_worker, 0);
1294
1295         /* Allow the reply handler and refresh worker to run */
1296         cond_resched();
1297
1298         return NULL;
1299 }
1300
1301 static void
1302 __rpcrdma_mr_put(struct rpcrdma_buffer *buf, struct rpcrdma_mr *mr)
1303 {
1304         spin_lock(&buf->rb_mrlock);
1305         rpcrdma_mr_push(mr, &buf->rb_mrs);
1306         spin_unlock(&buf->rb_mrlock);
1307 }
1308
1309 /**
1310  * rpcrdma_mr_put - Release an rpcrdma_mr object
1311  * @mr: object to release
1312  *
1313  */
1314 void
1315 rpcrdma_mr_put(struct rpcrdma_mr *mr)
1316 {
1317         __rpcrdma_mr_put(&mr->mr_xprt->rx_buf, mr);
1318 }
1319
1320 /**
1321  * rpcrdma_mr_unmap_and_put - DMA unmap an MR and release it
1322  * @mr: object to release
1323  *
1324  */
1325 void
1326 rpcrdma_mr_unmap_and_put(struct rpcrdma_mr *mr)
1327 {
1328         struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
1329
1330         trace_xprtrdma_dma_unmap(mr);
1331         ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
1332                         mr->mr_sg, mr->mr_nents, mr->mr_dir);
1333         __rpcrdma_mr_put(&r_xprt->rx_buf, mr);
1334 }
1335
1336 /**
1337  * rpcrdma_buffer_get - Get a request buffer
1338  * @buffers: Buffer pool from which to obtain a buffer
1339  *
1340  * Returns a fresh rpcrdma_req, or NULL if none are available.
1341  */
1342 struct rpcrdma_req *
1343 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1344 {
1345         struct rpcrdma_req *req;
1346
1347         spin_lock(&buffers->rb_lock);
1348         req = list_first_entry_or_null(&buffers->rb_send_bufs,
1349                                        struct rpcrdma_req, rl_list);
1350         if (req)
1351                 list_del_init(&req->rl_list);
1352         spin_unlock(&buffers->rb_lock);
1353         return req;
1354 }
1355
1356 /**
1357  * rpcrdma_buffer_put - Put request/reply buffers back into pool
1358  * @req: object to return
1359  *
1360  */
1361 void
1362 rpcrdma_buffer_put(struct rpcrdma_req *req)
1363 {
1364         struct rpcrdma_buffer *buffers = req->rl_buffer;
1365         struct rpcrdma_rep *rep = req->rl_reply;
1366
1367         req->rl_reply = NULL;
1368
1369         spin_lock(&buffers->rb_lock);
1370         list_add(&req->rl_list, &buffers->rb_send_bufs);
1371         if (rep) {
1372                 if (!rep->rr_temp) {
1373                         list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1374                         rep = NULL;
1375                 }
1376         }
1377         spin_unlock(&buffers->rb_lock);
1378         if (rep)
1379                 rpcrdma_destroy_rep(rep);
1380 }
1381
1382 /*
1383  * Put reply buffers back into pool when not attached to
1384  * request. This happens in error conditions.
1385  */
1386 void
1387 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1388 {
1389         struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1390
1391         if (!rep->rr_temp) {
1392                 spin_lock(&buffers->rb_lock);
1393                 list_add(&rep->rr_list, &buffers->rb_recv_bufs);
1394                 spin_unlock(&buffers->rb_lock);
1395         } else {
1396                 rpcrdma_destroy_rep(rep);
1397         }
1398 }
1399
1400 /**
1401  * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1402  * @size: size of buffer to be allocated, in bytes
1403  * @direction: direction of data movement
1404  * @flags: GFP flags
1405  *
1406  * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
1407  * can be persistently DMA-mapped for I/O.
1408  *
1409  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1410  * receiving the payload of RDMA RECV operations. During Long Calls
1411  * or Replies they may be registered externally via ro_map.
1412  */
1413 struct rpcrdma_regbuf *
1414 rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
1415                      gfp_t flags)
1416 {
1417         struct rpcrdma_regbuf *rb;
1418
1419         rb = kmalloc(sizeof(*rb) + size, flags);
1420         if (rb == NULL)
1421                 return ERR_PTR(-ENOMEM);
1422
1423         rb->rg_device = NULL;
1424         rb->rg_direction = direction;
1425         rb->rg_iov.length = size;
1426
1427         return rb;
1428 }
1429
1430 /**
1431  * __rpcrdma_map_regbuf - DMA-map a regbuf
1432  * @ia: controlling rpcrdma_ia
1433  * @rb: regbuf to be mapped
1434  */
1435 bool
1436 __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1437 {
1438         struct ib_device *device = ia->ri_device;
1439
1440         if (rb->rg_direction == DMA_NONE)
1441                 return false;
1442
1443         rb->rg_iov.addr = ib_dma_map_single(device,
1444                                             (void *)rb->rg_base,
1445                                             rdmab_length(rb),
1446                                             rb->rg_direction);
1447         if (ib_dma_mapping_error(device, rdmab_addr(rb)))
1448                 return false;
1449
1450         rb->rg_device = device;
1451         rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey;
1452         return true;
1453 }
1454
1455 static void
1456 rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb)
1457 {
1458         if (!rb)
1459                 return;
1460
1461         if (!rpcrdma_regbuf_is_mapped(rb))
1462                 return;
1463
1464         ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb),
1465                             rdmab_length(rb), rb->rg_direction);
1466         rb->rg_device = NULL;
1467 }
1468
1469 /**
1470  * rpcrdma_free_regbuf - deregister and free registered buffer
1471  * @rb: regbuf to be deregistered and freed
1472  */
1473 void
1474 rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
1475 {
1476         rpcrdma_dma_unmap_regbuf(rb);
1477         kfree(rb);
1478 }
1479
1480 /*
1481  * Prepost any receive buffer, then post send.
1482  *
1483  * Receive buffer is donated to hardware, reclaimed upon recv completion.
1484  */
1485 int
1486 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1487                 struct rpcrdma_ep *ep,
1488                 struct rpcrdma_req *req)
1489 {
1490         struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
1491         int rc;
1492
1493         if (!ep->rep_send_count ||
1494             test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags)) {
1495                 send_wr->send_flags |= IB_SEND_SIGNALED;
1496                 ep->rep_send_count = ep->rep_send_batch;
1497         } else {
1498                 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1499                 --ep->rep_send_count;
1500         }
1501
1502         rc = ia->ri_ops->ro_send(ia, req);
1503         trace_xprtrdma_post_send(req, rc);
1504         if (rc)
1505                 return -ENOTCONN;
1506         return 0;
1507 }
1508
1509 /**
1510  * rpcrdma_post_recvs - Maybe post some Receive buffers
1511  * @r_xprt: controlling transport
1512  * @temp: when true, allocate temp rpcrdma_rep objects
1513  *
1514  */
1515 void
1516 rpcrdma_post_recvs(struct rpcrdma_xprt *r_xprt, bool temp)
1517 {
1518         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1519         struct ib_recv_wr *wr, *bad_wr;
1520         int needed, count, rc;
1521
1522         needed = buf->rb_credits + (buf->rb_bc_srv_max_requests << 1);
1523         if (buf->rb_posted_receives > needed)
1524                 return;
1525         needed -= buf->rb_posted_receives;
1526
1527         count = 0;
1528         wr = NULL;
1529         while (needed) {
1530                 struct rpcrdma_regbuf *rb;
1531                 struct rpcrdma_rep *rep;
1532
1533                 spin_lock(&buf->rb_lock);
1534                 rep = list_first_entry_or_null(&buf->rb_recv_bufs,
1535                                                struct rpcrdma_rep, rr_list);
1536                 if (likely(rep))
1537                         list_del(&rep->rr_list);
1538                 spin_unlock(&buf->rb_lock);
1539                 if (!rep) {
1540                         if (rpcrdma_create_rep(r_xprt, temp))
1541                                 break;
1542                         continue;
1543                 }
1544
1545                 rb = rep->rr_rdmabuf;
1546                 if (!rpcrdma_regbuf_is_mapped(rb)) {
1547                         if (!__rpcrdma_dma_map_regbuf(&r_xprt->rx_ia, rb)) {
1548                                 rpcrdma_recv_buffer_put(rep);
1549                                 break;
1550                         }
1551                 }
1552
1553                 trace_xprtrdma_post_recv(rep->rr_recv_wr.wr_cqe);
1554                 rep->rr_recv_wr.next = wr;
1555                 wr = &rep->rr_recv_wr;
1556                 ++count;
1557                 --needed;
1558         }
1559         if (!count)
1560                 return;
1561
1562         rc = ib_post_recv(r_xprt->rx_ia.ri_id->qp, wr,
1563                           (const struct ib_recv_wr **)&bad_wr);
1564         if (rc) {
1565                 for (wr = bad_wr; wr; wr = wr->next) {
1566                         struct rpcrdma_rep *rep;
1567
1568                         rep = container_of(wr, struct rpcrdma_rep, rr_recv_wr);
1569                         rpcrdma_recv_buffer_put(rep);
1570                         --count;
1571                 }
1572         }
1573         buf->rb_posted_receives += count;
1574         trace_xprtrdma_post_recvs(r_xprt, count, rc);
1575 }