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