net/sunrpc/xprtrdma/verbs.c

   1 /*
   2  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
   3  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the BSD-type
   9  * license below:
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  *
  15  *      Redistributions of source code must retain the above copyright
  16  *      notice, this list of conditions and the following disclaimer.
  17  *
  18  *      Redistributions in binary form must reproduce the above
  19  *      copyright notice, this list of conditions and the following
  20  *      disclaimer in the documentation and/or other materials provided
  21  *      with the distribution.
  22  *
  23  *      Neither the name of the Network Appliance, Inc. nor the names of
  24  *      its contributors may be used to endorse or promote products
  25  *      derived from this software without specific prior written
  26  *      permission.
  27  *
  28  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  29  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  30  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  31  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  32  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  33  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  34  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  35  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  36  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  37  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  38  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  39  */
  40
  41 /*
  42  * verbs.c
  43  *
  44  * Encapsulates the major functions managing:
  45  *  o adapters
  46  *  o endpoints
  47  *  o connections
  48  *  o buffer memory
  49  */
  50
  51 #include <linux/interrupt.h>
  52 #include <linux/slab.h>
  53 #include <linux/sunrpc/addr.h>
  54 #include <linux/sunrpc/svc_rdma.h>
  55
  56 #include <asm-generic/barrier.h>
  57 #include <asm/bitops.h>
  58
  59 #include <rdma/ib_cm.h>
  60
  61 #include "xprt_rdma.h"
  62
  63 /*
  64  * Globals/Macros
  65  */
  66
  67 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  68 # define RPCDBG_FACILITY        RPCDBG_TRANS
  69 #endif
  70
  71 /*
  72  * internal functions
  73  */
  74 static void rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt);
  75 static void rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf);
  76 static void rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb);
  77
  78 struct workqueue_struct *rpcrdma_receive_wq __read_mostly;
  79
  80 int
  81 rpcrdma_alloc_wq(void)
  82 {
  83         struct workqueue_struct *recv_wq;
  84
  85         recv_wq = alloc_workqueue("xprtrdma_receive",
  86                                   WQ_MEM_RECLAIM | WQ_HIGHPRI,
  87                                   0);
  88         if (!recv_wq)
  89                 return -ENOMEM;
  90
  91         rpcrdma_receive_wq = recv_wq;
  92         return 0;
  93 }
  94
  95 void
  96 rpcrdma_destroy_wq(void)
  97 {
  98         struct workqueue_struct *wq;
  99
 100         if (rpcrdma_receive_wq) {
 101                 wq = rpcrdma_receive_wq;
 102                 rpcrdma_receive_wq = NULL;
 103                 destroy_workqueue(wq);
 104         }
 105 }
 106
 107 static void
 108 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
 109 {
 110         struct rpcrdma_ep *ep = context;
 111
 112         pr_err("rpcrdma: %s on device %s ep %p\n",
 113                ib_event_msg(event->event), event->device->name, context);
 114
 115         if (ep->rep_connected == 1) {
 116                 ep->rep_connected = -EIO;
 117                 rpcrdma_conn_func(ep);
 118                 wake_up_all(&ep->rep_connect_wait);
 119         }
 120 }
 121
 122 /**
 123  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
 124  * @cq: completion queue (ignored)
 125  * @wc: completed WR
 126  *
 127  */
 128 static void
 129 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
 130 {
 131         struct ib_cqe *cqe = wc->wr_cqe;
 132         struct rpcrdma_sendctx *sc =
 133                 container_of(cqe, struct rpcrdma_sendctx, sc_cqe);
 134
 135         /* WARNING: Only wr_cqe and status are reliable at this point */
 136         if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
 137                 pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
 138                        ib_wc_status_msg(wc->status),
 139                        wc->status, wc->vendor_err);
 140
 141         rpcrdma_sendctx_put_locked(sc);
 142 }
 143
 144 /**
 145  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
 146  * @cq: completion queue (ignored)
 147  * @wc: completed WR
 148  *
 149  */
 150 static void
 151 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
 152 {
 153         struct ib_cqe *cqe = wc->wr_cqe;
 154         struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
 155                                                rr_cqe);
 156
 157         /* WARNING: Only wr_id and status are reliable at this point */
 158         if (wc->status != IB_WC_SUCCESS)
 159                 goto out_fail;
 160
 161         /* status == SUCCESS means all fields in wc are trustworthy */
 162         dprintk("RPC:       %s: rep %p opcode 'recv', length %u: success\n",
 163                 __func__, rep, wc->byte_len);
 164
 165         rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
 166         rep->rr_wc_flags = wc->wc_flags;
 167         rep->rr_inv_rkey = wc->ex.invalidate_rkey;
 168
 169         ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
 170                                    rdmab_addr(rep->rr_rdmabuf),
 171                                    wc->byte_len, DMA_FROM_DEVICE);
 172
 173 out_schedule:
 174         rpcrdma_reply_handler(rep);
 175         return;
 176
 177 out_fail:
 178         if (wc->status != IB_WC_WR_FLUSH_ERR)
 179                 pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
 180                        ib_wc_status_msg(wc->status),
 181                        wc->status, wc->vendor_err);
 182         rpcrdma_set_xdrlen(&rep->rr_hdrbuf, 0);
 183         goto out_schedule;
 184 }
 185
 186 static void
 187 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
 188                                struct rdma_conn_param *param)
 189 {
 190         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
 191         const struct rpcrdma_connect_private *pmsg = param->private_data;
 192         unsigned int rsize, wsize;
 193
 194         /* Default settings for RPC-over-RDMA Version One */
 195         r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
 196         rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
 197         wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
 198
 199         if (pmsg &&
 200             pmsg->cp_magic == rpcrdma_cmp_magic &&
 201             pmsg->cp_version == RPCRDMA_CMP_VERSION) {
 202                 r_xprt->rx_ia.ri_implicit_roundup = true;
 203                 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
 204                 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
 205         }
 206
 207         if (rsize < cdata->inline_rsize)
 208                 cdata->inline_rsize = rsize;
 209         if (wsize < cdata->inline_wsize)
 210                 cdata->inline_wsize = wsize;
 211         dprintk("RPC:       %s: max send %u, max recv %u\n",
 212                 __func__, cdata->inline_wsize, cdata->inline_rsize);
 213         rpcrdma_set_max_header_sizes(r_xprt);
 214 }
 215
 216 static int
 217 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
 218 {
 219         struct rpcrdma_xprt *xprt = id->context;
 220         struct rpcrdma_ia *ia = &xprt->rx_ia;
 221         struct rpcrdma_ep *ep = &xprt->rx_ep;
 222         int connstate = 0;
 223
 224         switch (event->event) {
 225         case RDMA_CM_EVENT_ADDR_RESOLVED:
 226         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 227                 ia->ri_async_rc = 0;
 228                 complete(&ia->ri_done);
 229                 break;
 230         case RDMA_CM_EVENT_ADDR_ERROR:
 231                 ia->ri_async_rc = -EHOSTUNREACH;
 232                 dprintk("RPC:       %s: CM address resolution error, ep 0x%p\n",
 233                         __func__, ep);
 234                 complete(&ia->ri_done);
 235                 break;
 236         case RDMA_CM_EVENT_ROUTE_ERROR:
 237                 ia->ri_async_rc = -ENETUNREACH;
 238                 dprintk("RPC:       %s: CM route resolution error, ep 0x%p\n",
 239                         __func__, ep);
 240                 complete(&ia->ri_done);
 241                 break;
 242         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 243 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 244                 pr_info("rpcrdma: removing device %s for %s:%s\n",
 245                         ia->ri_device->name,
 246                         rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt));
 247 #endif
 248                 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
 249                 ep->rep_connected = -ENODEV;
 250                 xprt_force_disconnect(&xprt->rx_xprt);
 251                 wait_for_completion(&ia->ri_remove_done);
 252
 253                 ia->ri_id = NULL;
 254                 ia->ri_pd = NULL;
 255                 ia->ri_device = NULL;
 256                 /* Return 1 to ensure the core destroys the id. */
 257                 return 1;
 258         case RDMA_CM_EVENT_ESTABLISHED:
 259                 connstate = 1;
 260                 rpcrdma_update_connect_private(xprt, &event->param.conn);
 261                 goto connected;
 262         case RDMA_CM_EVENT_CONNECT_ERROR:
 263                 connstate = -ENOTCONN;
 264                 goto connected;
 265         case RDMA_CM_EVENT_UNREACHABLE:
 266                 connstate = -ENETDOWN;
 267                 goto connected;
 268         case RDMA_CM_EVENT_REJECTED:
 269                 dprintk("rpcrdma: connection to %s:%s rejected: %s\n",
 270                         rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
 271                         rdma_reject_msg(id, event->status));
 272                 connstate = -ECONNREFUSED;
 273                 if (event->status == IB_CM_REJ_STALE_CONN)
 274                         connstate = -EAGAIN;
 275                 goto connected;
 276         case RDMA_CM_EVENT_DISCONNECTED:
 277                 connstate = -ECONNABORTED;
 278 connected:
 279                 xprt->rx_buf.rb_credits = 1;
 280                 ep->rep_connected = connstate;
 281                 rpcrdma_conn_func(ep);
 282                 wake_up_all(&ep->rep_connect_wait);
 283                 /*FALLTHROUGH*/
 284         default:
 285                 dprintk("RPC:       %s: %s:%s on %s/%s (ep 0x%p): %s\n",
 286                         __func__,
 287                         rpcrdma_addrstr(xprt), rpcrdma_portstr(xprt),
 288                         ia->ri_device->name, ia->ri_ops->ro_displayname,
 289                         ep, rdma_event_msg(event->event));
 290                 break;
 291         }
 292
 293         return 0;
 294 }
 295
 296 static struct rdma_cm_id *
 297 rpcrdma_create_id(struct rpcrdma_xprt *xprt, struct rpcrdma_ia *ia)
 298 {
 299         unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
 300         struct rdma_cm_id *id;
 301         int rc;
 302
 303         init_completion(&ia->ri_done);
 304         init_completion(&ia->ri_remove_done);
 305
 306         id = rdma_create_id(&init_net, rpcrdma_conn_upcall, xprt, RDMA_PS_TCP,
 307                             IB_QPT_RC);
 308         if (IS_ERR(id)) {
 309                 rc = PTR_ERR(id);
 310                 dprintk("RPC:       %s: rdma_create_id() failed %i\n",
 311                         __func__, rc);
 312                 return id;
 313         }
 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                 dprintk("RPC:       %s: rdma_resolve_addr() failed %i\n",
 321                         __func__, rc);
 322                 goto out;
 323         }
 324         rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
 325         if (rc < 0) {
 326                 dprintk("RPC:       %s: wait() exited: %i\n",
 327                         __func__, rc);
 328                 goto out;
 329         }
 330
 331         rc = ia->ri_async_rc;
 332         if (rc)
 333                 goto out;
 334
 335         ia->ri_async_rc = -ETIMEDOUT;
 336         rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
 337         if (rc) {
 338                 dprintk("RPC:       %s: rdma_resolve_route() failed %i\n",
 339                         __func__, rc);
 340                 goto out;
 341         }
 342         rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
 343         if (rc < 0) {
 344                 dprintk("RPC:       %s: wait() exited: %i\n",
 345                         __func__, rc);
 346                 goto out;
 347         }
 348         rc = ia->ri_async_rc;
 349         if (rc)
 350                 goto out;
 351
 352         return id;
 353
 354 out:
 355         rdma_destroy_id(id);
 356         return ERR_PTR(rc);
 357 }
 358
 359 /*
 360  * Exported functions.
 361  */
 362
 363 /**
 364  * rpcrdma_ia_open - Open and initialize an Interface Adapter.
 365  * @xprt: transport with IA to (re)initialize
 366  *
 367  * Returns 0 on success, negative errno if an appropriate
 368  * Interface Adapter could not be found and opened.
 369  */
 370 int
 371 rpcrdma_ia_open(struct rpcrdma_xprt *xprt)
 372 {
 373         struct rpcrdma_ia *ia = &xprt->rx_ia;
 374         int rc;
 375
 376         ia->ri_id = rpcrdma_create_id(xprt, ia);
 377         if (IS_ERR(ia->ri_id)) {
 378                 rc = PTR_ERR(ia->ri_id);
 379                 goto out_err;
 380         }
 381         ia->ri_device = ia->ri_id->device;
 382
 383         ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
 384         if (IS_ERR(ia->ri_pd)) {
 385                 rc = PTR_ERR(ia->ri_pd);
 386                 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
 387                 goto out_err;
 388         }
 389
 390         switch (xprt_rdma_memreg_strategy) {
 391         case RPCRDMA_FRWR:
 392                 if (frwr_is_supported(ia)) {
 393                         ia->ri_ops = &rpcrdma_frwr_memreg_ops;
 394                         break;
 395                 }
 396                 /*FALLTHROUGH*/
 397         case RPCRDMA_MTHCAFMR:
 398                 if (fmr_is_supported(ia)) {
 399                         ia->ri_ops = &rpcrdma_fmr_memreg_ops;
 400                         break;
 401                 }
 402                 /*FALLTHROUGH*/
 403         default:
 404                 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
 405                        ia->ri_device->name, xprt_rdma_memreg_strategy);
 406                 rc = -EINVAL;
 407                 goto out_err;
 408         }
 409
 410         return 0;
 411
 412 out_err:
 413         rpcrdma_ia_close(ia);
 414         return rc;
 415 }
 416
 417 /**
 418  * rpcrdma_ia_remove - Handle device driver unload
 419  * @ia: interface adapter being removed
 420  *
 421  * Divest transport H/W resources associated with this adapter,
 422  * but allow it to be restored later.
 423  */
 424 void
 425 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
 426 {
 427         struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
 428                                                    rx_ia);
 429         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 430         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 431         struct rpcrdma_req *req;
 432         struct rpcrdma_rep *rep;
 433
 434         cancel_delayed_work_sync(&buf->rb_refresh_worker);
 435
 436         /* This is similar to rpcrdma_ep_destroy, but:
 437          * - Don't cancel the connect worker.
 438          * - Don't call rpcrdma_ep_disconnect, which waits
 439          *   for another conn upcall, which will deadlock.
 440          * - rdma_disconnect is unneeded, the underlying
 441          *   connection is already gone.
 442          */
 443         if (ia->ri_id->qp) {
 444                 ib_drain_qp(ia->ri_id->qp);
 445                 rdma_destroy_qp(ia->ri_id);
 446                 ia->ri_id->qp = NULL;
 447         }
 448         ib_free_cq(ep->rep_attr.recv_cq);
 449         ib_free_cq(ep->rep_attr.send_cq);
 450
 451         /* The ULP is responsible for ensuring all DMA
 452          * mappings and MRs are gone.
 453          */
 454         list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
 455                 rpcrdma_dma_unmap_regbuf(rep->rr_rdmabuf);
 456         list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
 457                 rpcrdma_dma_unmap_regbuf(req->rl_rdmabuf);
 458                 rpcrdma_dma_unmap_regbuf(req->rl_sendbuf);
 459                 rpcrdma_dma_unmap_regbuf(req->rl_recvbuf);
 460         }
 461         rpcrdma_destroy_mrs(buf);
 462
 463         /* Allow waiters to continue */
 464         complete(&ia->ri_remove_done);
 465 }
 466
 467 /**
 468  * rpcrdma_ia_close - Clean up/close an IA.
 469  * @ia: interface adapter to close
 470  *
 471  */
 472 void
 473 rpcrdma_ia_close(struct rpcrdma_ia *ia)
 474 {
 475         dprintk("RPC:       %s: entering\n", __func__);
 476         if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
 477                 if (ia->ri_id->qp)
 478                         rdma_destroy_qp(ia->ri_id);
 479                 rdma_destroy_id(ia->ri_id);
 480         }
 481         ia->ri_id = NULL;
 482         ia->ri_device = NULL;
 483
 484         /* If the pd is still busy, xprtrdma missed freeing a resource */
 485         if (ia->ri_pd && !IS_ERR(ia->ri_pd))
 486                 ib_dealloc_pd(ia->ri_pd);
 487         ia->ri_pd = NULL;
 488 }
 489
 490 /*
 491  * Create unconnected endpoint.
 492  */
 493 int
 494 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
 495                   struct rpcrdma_create_data_internal *cdata)
 496 {
 497         struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
 498         unsigned int max_qp_wr, max_sge;
 499         struct ib_cq *sendcq, *recvcq;
 500         int rc;
 501
 502         max_sge = min_t(unsigned int, ia->ri_device->attrs.max_sge,
 503                         RPCRDMA_MAX_SEND_SGES);
 504         if (max_sge < RPCRDMA_MIN_SEND_SGES) {
 505                 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
 506                 return -ENOMEM;
 507         }
 508         ia->ri_max_send_sges = max_sge - RPCRDMA_MIN_SEND_SGES;
 509
 510         if (ia->ri_device->attrs.max_qp_wr <= RPCRDMA_BACKWARD_WRS) {
 511                 dprintk("RPC:       %s: insufficient wqe's available\n",
 512                         __func__);
 513                 return -ENOMEM;
 514         }
 515         max_qp_wr = ia->ri_device->attrs.max_qp_wr - RPCRDMA_BACKWARD_WRS - 1;
 516
 517         /* check provider's send/recv wr limits */
 518         if (cdata->max_requests > max_qp_wr)
 519                 cdata->max_requests = max_qp_wr;
 520
 521         ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
 522         ep->rep_attr.qp_context = ep;
 523         ep->rep_attr.srq = NULL;
 524         ep->rep_attr.cap.max_send_wr = cdata->max_requests;
 525         ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
 526         ep->rep_attr.cap.max_send_wr += 1;      /* drain cqe */
 527         rc = ia->ri_ops->ro_open(ia, ep, cdata);
 528         if (rc)
 529                 return rc;
 530         ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
 531         ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
 532         ep->rep_attr.cap.max_recv_wr += 1;      /* drain cqe */
 533         ep->rep_attr.cap.max_send_sge = max_sge;
 534         ep->rep_attr.cap.max_recv_sge = 1;
 535         ep->rep_attr.cap.max_inline_data = 0;
 536         ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 537         ep->rep_attr.qp_type = IB_QPT_RC;
 538         ep->rep_attr.port_num = ~0;
 539
 540         dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
 541                 "iovs: send %d recv %d\n",
 542                 __func__,
 543                 ep->rep_attr.cap.max_send_wr,
 544                 ep->rep_attr.cap.max_recv_wr,
 545                 ep->rep_attr.cap.max_send_sge,
 546                 ep->rep_attr.cap.max_recv_sge);
 547
 548         /* set trigger for requesting send completion */
 549         ep->rep_send_batch = min_t(unsigned int, RPCRDMA_MAX_SEND_BATCH,
 550                                    cdata->max_requests >> 2);
 551         ep->rep_send_count = ep->rep_send_batch;
 552         init_waitqueue_head(&ep->rep_connect_wait);
 553         INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
 554
 555         sendcq = ib_alloc_cq(ia->ri_device, NULL,
 556                              ep->rep_attr.cap.max_send_wr + 1,
 557                              1, IB_POLL_WORKQUEUE);
 558         if (IS_ERR(sendcq)) {
 559                 rc = PTR_ERR(sendcq);
 560                 dprintk("RPC:       %s: failed to create send CQ: %i\n",
 561                         __func__, rc);
 562                 goto out1;
 563         }
 564
 565         recvcq = ib_alloc_cq(ia->ri_device, NULL,
 566                              ep->rep_attr.cap.max_recv_wr + 1,
 567                              0, IB_POLL_WORKQUEUE);
 568         if (IS_ERR(recvcq)) {
 569                 rc = PTR_ERR(recvcq);
 570                 dprintk("RPC:       %s: failed to create recv CQ: %i\n",
 571                         __func__, rc);
 572                 goto out2;
 573         }
 574
 575         ep->rep_attr.send_cq = sendcq;
 576         ep->rep_attr.recv_cq = recvcq;
 577
 578         /* Initialize cma parameters */
 579         memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
 580
 581         /* Prepare RDMA-CM private message */
 582         pmsg->cp_magic = rpcrdma_cmp_magic;
 583         pmsg->cp_version = RPCRDMA_CMP_VERSION;
 584         pmsg->cp_flags |= ia->ri_ops->ro_send_w_inv_ok;
 585         pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize);
 586         pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize);
 587         ep->rep_remote_cma.private_data = pmsg;
 588         ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
 589
 590         /* Client offers RDMA Read but does not initiate */
 591         ep->rep_remote_cma.initiator_depth = 0;
 592         if (ia->ri_device->attrs.max_qp_rd_atom > 32)   /* arbitrary but <= 255 */
 593                 ep->rep_remote_cma.responder_resources = 32;
 594         else
 595                 ep->rep_remote_cma.responder_resources =
 596                                                 ia->ri_device->attrs.max_qp_rd_atom;
 597
 598         /* Limit transport retries so client can detect server
 599          * GID changes quickly. RPC layer handles re-establishing
 600          * transport connection and retransmission.
 601          */
 602         ep->rep_remote_cma.retry_count = 6;
 603
 604         /* RPC-over-RDMA handles its own flow control. In addition,
 605          * make all RNR NAKs visible so we know that RPC-over-RDMA
 606          * flow control is working correctly (no NAKs should be seen).
 607          */
 608         ep->rep_remote_cma.flow_control = 0;
 609         ep->rep_remote_cma.rnr_retry_count = 0;
 610
 611         return 0;
 612
 613 out2:
 614         ib_free_cq(sendcq);
 615 out1:
 616         return rc;
 617 }
 618
 619 /*
 620  * rpcrdma_ep_destroy
 621  *
 622  * Disconnect and destroy endpoint. After this, the only
 623  * valid operations on the ep are to free it (if dynamically
 624  * allocated) or re-create it.
 625  */
 626 void
 627 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 628 {
 629         dprintk("RPC:       %s: entering, connected is %d\n",
 630                 __func__, ep->rep_connected);
 631
 632         cancel_delayed_work_sync(&ep->rep_connect_worker);
 633
 634         if (ia->ri_id->qp) {
 635                 rpcrdma_ep_disconnect(ep, ia);
 636                 rdma_destroy_qp(ia->ri_id);
 637                 ia->ri_id->qp = NULL;
 638         }
 639
 640         ib_free_cq(ep->rep_attr.recv_cq);
 641         ib_free_cq(ep->rep_attr.send_cq);
 642 }
 643
 644 /* Re-establish a connection after a device removal event.
 645  * Unlike a normal reconnection, a fresh PD and a new set
 646  * of MRs and buffers is needed.
 647  */
 648 static int
 649 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
 650                          struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 651 {
 652         int rc, err;
 653
 654         pr_info("%s: r_xprt = %p\n", __func__, r_xprt);
 655
 656         rc = -EHOSTUNREACH;
 657         if (rpcrdma_ia_open(r_xprt))
 658                 goto out1;
 659
 660         rc = -ENOMEM;
 661         err = rpcrdma_ep_create(ep, ia, &r_xprt->rx_data);
 662         if (err) {
 663                 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
 664                 goto out2;
 665         }
 666
 667         rc = -ENETUNREACH;
 668         err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
 669         if (err) {
 670                 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
 671                 goto out3;
 672         }
 673
 674         rpcrdma_create_mrs(r_xprt);
 675         return 0;
 676
 677 out3:
 678         rpcrdma_ep_destroy(ep, ia);
 679 out2:
 680         rpcrdma_ia_close(ia);
 681 out1:
 682         return rc;
 683 }
 684
 685 static int
 686 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
 687                      struct rpcrdma_ia *ia)
 688 {
 689         struct rdma_cm_id *id, *old;
 690         int err, rc;
 691
 692         dprintk("RPC:       %s: reconnecting...\n", __func__);
 693
 694         rpcrdma_ep_disconnect(ep, ia);
 695
 696         rc = -EHOSTUNREACH;
 697         id = rpcrdma_create_id(r_xprt, ia);
 698         if (IS_ERR(id))
 699                 goto out;
 700
 701         /* As long as the new ID points to the same device as the
 702          * old ID, we can reuse the transport's existing PD and all
 703          * previously allocated MRs. Also, the same device means
 704          * the transport's previous DMA mappings are still valid.
 705          *
 706          * This is a sanity check only. There should be no way these
 707          * point to two different devices here.
 708          */
 709         old = id;
 710         rc = -ENETUNREACH;
 711         if (ia->ri_device != id->device) {
 712                 pr_err("rpcrdma: can't reconnect on different device!\n");
 713                 goto out_destroy;
 714         }
 715
 716         err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
 717         if (err) {
 718                 dprintk("RPC:       %s: rdma_create_qp returned %d\n",
 719                         __func__, err);
 720                 goto out_destroy;
 721         }
 722
 723         /* Atomically replace the transport's ID and QP. */
 724         rc = 0;
 725         old = ia->ri_id;
 726         ia->ri_id = id;
 727         rdma_destroy_qp(old);
 728
 729 out_destroy:
 730         rdma_destroy_id(old);
 731 out:
 732         return rc;
 733 }
 734
 735 /*
 736  * Connect unconnected endpoint.
 737  */
 738 int
 739 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 740 {
 741         struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
 742                                                    rx_ia);
 743         unsigned int extras;
 744         int rc;
 745
 746 retry:
 747         switch (ep->rep_connected) {
 748         case 0:
 749                 dprintk("RPC:       %s: connecting...\n", __func__);
 750                 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
 751                 if (rc) {
 752                         dprintk("RPC:       %s: rdma_create_qp failed %i\n",
 753                                 __func__, rc);
 754                         rc = -ENETUNREACH;
 755                         goto out_noupdate;
 756                 }
 757                 break;
 758         case -ENODEV:
 759                 rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
 760                 if (rc)
 761                         goto out_noupdate;
 762                 break;
 763         default:
 764                 rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
 765                 if (rc)
 766                         goto out;
 767         }
 768
 769         ep->rep_connected = 0;
 770
 771         rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
 772         if (rc) {
 773                 dprintk("RPC:       %s: rdma_connect() failed with %i\n",
 774                                 __func__, rc);
 775                 goto out;
 776         }
 777
 778         wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
 779         if (ep->rep_connected <= 0) {
 780                 if (ep->rep_connected == -EAGAIN)
 781                         goto retry;
 782                 rc = ep->rep_connected;
 783                 goto out;
 784         }
 785
 786         dprintk("RPC:       %s: connected\n", __func__);
 787         extras = r_xprt->rx_buf.rb_bc_srv_max_requests;
 788         if (extras)
 789                 rpcrdma_ep_post_extra_recv(r_xprt, extras);
 790
 791 out:
 792         if (rc)
 793                 ep->rep_connected = rc;
 794
 795 out_noupdate:
 796         return rc;
 797 }
 798
 799 /*
 800  * rpcrdma_ep_disconnect
 801  *
 802  * This is separate from destroy to facilitate the ability
 803  * to reconnect without recreating the endpoint.
 804  *
 805  * This call is not reentrant, and must not be made in parallel
 806  * on the same endpoint.
 807  */
 808 void
 809 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 810 {
 811         int rc;
 812
 813         rc = rdma_disconnect(ia->ri_id);
 814         if (!rc) {
 815                 /* returns without wait if not connected */
 816                 wait_event_interruptible(ep->rep_connect_wait,
 817                                                         ep->rep_connected != 1);
 818                 dprintk("RPC:       %s: after wait, %sconnected\n", __func__,
 819                         (ep->rep_connected == 1) ? "still " : "dis");
 820         } else {
 821                 dprintk("RPC:       %s: rdma_disconnect %i\n", __func__, rc);
 822                 ep->rep_connected = rc;
 823         }
 824
 825         ib_drain_qp(ia->ri_id->qp);
 826 }
 827
 828 /* Fixed-size circular FIFO queue. This implementation is wait-free and
 829  * lock-free.
 830  *
 831  * Consumer is the code path that posts Sends. This path dequeues a
 832  * sendctx for use by a Send operation. Multiple consumer threads
 833  * are serialized by the RPC transport lock, which allows only one
 834  * ->send_request call at a time.
 835  *
 836  * Producer is the code path that handles Send completions. This path
 837  * enqueues a sendctx that has been completed. Multiple producer
 838  * threads are serialized by the ib_poll_cq() function.
 839  */
 840
 841 /* rpcrdma_sendctxs_destroy() assumes caller has already quiesced
 842  * queue activity, and ib_drain_qp has flushed all remaining Send
 843  * requests.
 844  */
 845 static void rpcrdma_sendctxs_destroy(struct rpcrdma_buffer *buf)
 846 {
 847         unsigned long i;
 848
 849         for (i = 0; i <= buf->rb_sc_last; i++)
 850                 kfree(buf->rb_sc_ctxs[i]);
 851         kfree(buf->rb_sc_ctxs);
 852 }
 853
 854 static struct rpcrdma_sendctx *rpcrdma_sendctx_create(struct rpcrdma_ia *ia)
 855 {
 856         struct rpcrdma_sendctx *sc;
 857
 858         sc = kzalloc(sizeof(*sc) +
 859                      ia->ri_max_send_sges * sizeof(struct ib_sge),
 860                      GFP_KERNEL);
 861         if (!sc)
 862                 return NULL;
 863
 864         sc->sc_wr.wr_cqe = &sc->sc_cqe;
 865         sc->sc_wr.sg_list = sc->sc_sges;
 866         sc->sc_wr.opcode = IB_WR_SEND;
 867         sc->sc_cqe.done = rpcrdma_wc_send;
 868         return sc;
 869 }
 870
 871 static int rpcrdma_sendctxs_create(struct rpcrdma_xprt *r_xprt)
 872 {
 873         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 874         struct rpcrdma_sendctx *sc;
 875         unsigned long i;
 876
 877         /* Maximum number of concurrent outstanding Send WRs. Capping
 878          * the circular queue size stops Send Queue overflow by causing
 879          * the ->send_request call to fail temporarily before too many
 880          * Sends are posted.
 881          */
 882         i = buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS;
 883         dprintk("RPC:       %s: allocating %lu send_ctxs\n", __func__, i);
 884         buf->rb_sc_ctxs = kcalloc(i, sizeof(sc), GFP_KERNEL);
 885         if (!buf->rb_sc_ctxs)
 886                 return -ENOMEM;
 887
 888         buf->rb_sc_last = i - 1;
 889         for (i = 0; i <= buf->rb_sc_last; i++) {
 890                 sc = rpcrdma_sendctx_create(&r_xprt->rx_ia);
 891                 if (!sc)
 892                         goto out_destroy;
 893
 894                 sc->sc_xprt = r_xprt;
 895                 buf->rb_sc_ctxs[i] = sc;
 896         }
 897
 898         return 0;
 899
 900 out_destroy:
 901         rpcrdma_sendctxs_destroy(buf);
 902         return -ENOMEM;
 903 }
 904
 905 /* The sendctx queue is not guaranteed to have a size that is a
 906  * power of two, thus the helpers in circ_buf.h cannot be used.
 907  * The other option is to use modulus (%), which can be expensive.
 908  */
 909 static unsigned long rpcrdma_sendctx_next(struct rpcrdma_buffer *buf,
 910                                           unsigned long item)
 911 {
 912         return likely(item < buf->rb_sc_last) ? item + 1 : 0;
 913 }
 914
 915 /**
 916  * rpcrdma_sendctx_get_locked - Acquire a send context
 917  * @buf: transport buffers from which to acquire an unused context
 918  *
 919  * Returns pointer to a free send completion context; or NULL if
 920  * the queue is empty.
 921  *
 922  * Usage: Called to acquire an SGE array before preparing a Send WR.
 923  *
 924  * The caller serializes calls to this function (per rpcrdma_buffer),
 925  * and provides an effective memory barrier that flushes the new value
 926  * of rb_sc_head.
 927  */
 928 struct rpcrdma_sendctx *rpcrdma_sendctx_get_locked(struct rpcrdma_buffer *buf)
 929 {
 930         struct rpcrdma_xprt *r_xprt;
 931         struct rpcrdma_sendctx *sc;
 932         unsigned long next_head;
 933
 934         next_head = rpcrdma_sendctx_next(buf, buf->rb_sc_head);
 935
 936         if (next_head == READ_ONCE(buf->rb_sc_tail))
 937                 goto out_emptyq;
 938
 939         /* ORDER: item must be accessed _before_ head is updated */
 940         sc = buf->rb_sc_ctxs[next_head];
 941
 942         /* Releasing the lock in the caller acts as a memory
 943          * barrier that flushes rb_sc_head.
 944          */
 945         buf->rb_sc_head = next_head;
 946
 947         return sc;
 948
 949 out_emptyq:
 950         /* The queue is "empty" if there have not been enough Send
 951          * completions recently. This is a sign the Send Queue is
 952          * backing up. Cause the caller to pause and try again.
 953          */
 954         dprintk("RPC:       %s: empty sendctx queue\n", __func__);
 955         r_xprt = container_of(buf, struct rpcrdma_xprt, rx_buf);
 956         r_xprt->rx_stats.empty_sendctx_q++;
 957         return NULL;
 958 }
 959
 960 /**
 961  * rpcrdma_sendctx_put_locked - Release a send context
 962  * @sc: send context to release
 963  *
 964  * Usage: Called from Send completion to return a sendctxt
 965  * to the queue.
 966  *
 967  * The caller serializes calls to this function (per rpcrdma_buffer).
 968  */
 969 void rpcrdma_sendctx_put_locked(struct rpcrdma_sendctx *sc)
 970 {
 971         struct rpcrdma_buffer *buf = &sc->sc_xprt->rx_buf;
 972         unsigned long next_tail;
 973
 974         /* Unmap SGEs of previously completed by unsignaled
 975          * Sends by walking up the queue until @sc is found.
 976          */
 977         next_tail = buf->rb_sc_tail;
 978         do {
 979                 next_tail = rpcrdma_sendctx_next(buf, next_tail);
 980
 981                 /* ORDER: item must be accessed _before_ tail is updated */
 982                 rpcrdma_unmap_sendctx(buf->rb_sc_ctxs[next_tail]);
 983
 984         } while (buf->rb_sc_ctxs[next_tail] != sc);
 985
 986         /* Paired with READ_ONCE */
 987         smp_store_release(&buf->rb_sc_tail, next_tail);
 988 }
 989
 990 static void
 991 rpcrdma_mr_recovery_worker(struct work_struct *work)
 992 {
 993         struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
 994                                                   rb_recovery_worker.work);
 995         struct rpcrdma_mw *mw;
 996
 997         spin_lock(&buf->rb_recovery_lock);
 998         while (!list_empty(&buf->rb_stale_mrs)) {
 999                 mw = rpcrdma_pop_mw(&buf->rb_stale_mrs);
1000                 spin_unlock(&buf->rb_recovery_lock);
1001
1002                 dprintk("RPC:       %s: recovering MR %p\n", __func__, mw);
1003                 mw->mw_xprt->rx_ia.ri_ops->ro_recover_mr(mw);
1004
1005                 spin_lock(&buf->rb_recovery_lock);
1006         }
1007         spin_unlock(&buf->rb_recovery_lock);
1008 }
1009
1010 void
1011 rpcrdma_defer_mr_recovery(struct rpcrdma_mw *mw)
1012 {
1013         struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
1014         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1015
1016         spin_lock(&buf->rb_recovery_lock);
1017         rpcrdma_push_mw(mw, &buf->rb_stale_mrs);
1018         spin_unlock(&buf->rb_recovery_lock);
1019
1020         schedule_delayed_work(&buf->rb_recovery_worker, 0);
1021 }
1022
1023 static void
1024 rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt)
1025 {
1026         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1027         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1028         unsigned int count;
1029         LIST_HEAD(free);
1030         LIST_HEAD(all);
1031
1032         for (count = 0; count < 32; count++) {
1033                 struct rpcrdma_mw *mw;
1034                 int rc;
1035
1036                 mw = kzalloc(sizeof(*mw), GFP_KERNEL);
1037                 if (!mw)
1038                         break;
1039
1040                 rc = ia->ri_ops->ro_init_mr(ia, mw);
1041                 if (rc) {
1042                         kfree(mw);
1043                         break;
1044                 }
1045
1046                 mw->mw_xprt = r_xprt;
1047
1048                 list_add(&mw->mw_list, &free);
1049                 list_add(&mw->mw_all, &all);
1050         }
1051
1052         spin_lock(&buf->rb_mwlock);
1053         list_splice(&free, &buf->rb_mws);
1054         list_splice(&all, &buf->rb_all);
1055         r_xprt->rx_stats.mrs_allocated += count;
1056         spin_unlock(&buf->rb_mwlock);
1057
1058         dprintk("RPC:       %s: created %u MRs\n", __func__, count);
1059 }
1060
1061 static void
1062 rpcrdma_mr_refresh_worker(struct work_struct *work)
1063 {
1064         struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
1065                                                   rb_refresh_worker.work);
1066         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1067                                                    rx_buf);
1068
1069         rpcrdma_create_mrs(r_xprt);
1070 }
1071
1072 struct rpcrdma_req *
1073 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
1074 {
1075         struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
1076         struct rpcrdma_req *req;
1077
1078         req = kzalloc(sizeof(*req), GFP_KERNEL);
1079         if (req == NULL)
1080                 return ERR_PTR(-ENOMEM);
1081
1082         spin_lock(&buffer->rb_reqslock);
1083         list_add(&req->rl_all, &buffer->rb_allreqs);
1084         spin_unlock(&buffer->rb_reqslock);
1085         req->rl_buffer = &r_xprt->rx_buf;
1086         INIT_LIST_HEAD(&req->rl_registered);
1087         return req;
1088 }
1089
1090 /**
1091  * rpcrdma_create_rep - Allocate an rpcrdma_rep object
1092  * @r_xprt: controlling transport
1093  *
1094  * Returns 0 on success or a negative errno on failure.
1095  */
1096 int
1097 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
1098 {
1099         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
1100         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1101         struct rpcrdma_rep *rep;
1102         int rc;
1103
1104         rc = -ENOMEM;
1105         rep = kzalloc(sizeof(*rep), GFP_KERNEL);
1106         if (rep == NULL)
1107                 goto out;
1108
1109         rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
1110                                                DMA_FROM_DEVICE, GFP_KERNEL);
1111         if (IS_ERR(rep->rr_rdmabuf)) {
1112                 rc = PTR_ERR(rep->rr_rdmabuf);
1113                 goto out_free;
1114         }
1115         xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base,
1116                      rdmab_length(rep->rr_rdmabuf));
1117
1118         rep->rr_cqe.done = rpcrdma_wc_receive;
1119         rep->rr_rxprt = r_xprt;
1120         INIT_WORK(&rep->rr_work, rpcrdma_deferred_completion);
1121         rep->rr_recv_wr.next = NULL;
1122         rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
1123         rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
1124         rep->rr_recv_wr.num_sge = 1;
1125
1126         spin_lock(&buf->rb_lock);
1127         list_add(&rep->rr_list, &buf->rb_recv_bufs);
1128         spin_unlock(&buf->rb_lock);
1129         return 0;
1130
1131 out_free:
1132         kfree(rep);
1133 out:
1134         dprintk("RPC:       %s: reply buffer %d alloc failed\n",
1135                 __func__, rc);
1136         return rc;
1137 }
1138
1139 int
1140 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
1141 {
1142         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1143         int i, rc;
1144
1145         buf->rb_max_requests = r_xprt->rx_data.max_requests;
1146         buf->rb_bc_srv_max_requests = 0;
1147         spin_lock_init(&buf->rb_mwlock);
1148         spin_lock_init(&buf->rb_lock);
1149         spin_lock_init(&buf->rb_recovery_lock);
1150         INIT_LIST_HEAD(&buf->rb_mws);
1151         INIT_LIST_HEAD(&buf->rb_all);
1152         INIT_LIST_HEAD(&buf->rb_stale_mrs);
1153         INIT_DELAYED_WORK(&buf->rb_refresh_worker,
1154                           rpcrdma_mr_refresh_worker);
1155         INIT_DELAYED_WORK(&buf->rb_recovery_worker,
1156                           rpcrdma_mr_recovery_worker);
1157
1158         rpcrdma_create_mrs(r_xprt);
1159
1160         INIT_LIST_HEAD(&buf->rb_send_bufs);
1161         INIT_LIST_HEAD(&buf->rb_allreqs);
1162         spin_lock_init(&buf->rb_reqslock);
1163         for (i = 0; i < buf->rb_max_requests; i++) {
1164                 struct rpcrdma_req *req;
1165
1166                 req = rpcrdma_create_req(r_xprt);
1167                 if (IS_ERR(req)) {
1168                         dprintk("RPC:       %s: request buffer %d alloc"
1169                                 " failed\n", __func__, i);
1170                         rc = PTR_ERR(req);
1171                         goto out;
1172                 }
1173                 list_add(&req->rl_list, &buf->rb_send_bufs);
1174         }
1175
1176         INIT_LIST_HEAD(&buf->rb_recv_bufs);
1177         for (i = 0; i <= buf->rb_max_requests; i++) {
1178                 rc = rpcrdma_create_rep(r_xprt);
1179                 if (rc)
1180                         goto out;
1181         }
1182
1183         rc = rpcrdma_sendctxs_create(r_xprt);
1184         if (rc)
1185                 goto out;
1186
1187         return 0;
1188 out:
1189         rpcrdma_buffer_destroy(buf);
1190         return rc;
1191 }
1192
1193 static struct rpcrdma_req *
1194 rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer *buf)
1195 {
1196         struct rpcrdma_req *req;
1197
1198         req = list_first_entry(&buf->rb_send_bufs,
1199                                struct rpcrdma_req, rl_list);
1200         list_del_init(&req->rl_list);
1201         return req;
1202 }
1203
1204 static struct rpcrdma_rep *
1205 rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer *buf)
1206 {
1207         struct rpcrdma_rep *rep;
1208
1209         rep = list_first_entry(&buf->rb_recv_bufs,
1210                                struct rpcrdma_rep, rr_list);
1211         list_del(&rep->rr_list);
1212         return rep;
1213 }
1214
1215 static void
1216 rpcrdma_destroy_rep(struct rpcrdma_rep *rep)
1217 {
1218         rpcrdma_free_regbuf(rep->rr_rdmabuf);
1219         kfree(rep);
1220 }
1221
1222 void
1223 rpcrdma_destroy_req(struct rpcrdma_req *req)
1224 {
1225         rpcrdma_free_regbuf(req->rl_recvbuf);
1226         rpcrdma_free_regbuf(req->rl_sendbuf);
1227         rpcrdma_free_regbuf(req->rl_rdmabuf);
1228         kfree(req);
1229 }
1230
1231 static void
1232 rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf)
1233 {
1234         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1235                                                    rx_buf);
1236         struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1237         struct rpcrdma_mw *mw;
1238         unsigned int count;
1239
1240         count = 0;
1241         spin_lock(&buf->rb_mwlock);
1242         while (!list_empty(&buf->rb_all)) {
1243                 mw = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
1244                 list_del(&mw->mw_all);
1245
1246                 spin_unlock(&buf->rb_mwlock);
1247                 ia->ri_ops->ro_release_mr(mw);
1248                 count++;
1249                 spin_lock(&buf->rb_mwlock);
1250         }
1251         spin_unlock(&buf->rb_mwlock);
1252         r_xprt->rx_stats.mrs_allocated = 0;
1253
1254         dprintk("RPC:       %s: released %u MRs\n", __func__, count);
1255 }
1256
1257 void
1258 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1259 {
1260         cancel_delayed_work_sync(&buf->rb_recovery_worker);
1261         cancel_delayed_work_sync(&buf->rb_refresh_worker);
1262
1263         rpcrdma_sendctxs_destroy(buf);
1264
1265         while (!list_empty(&buf->rb_recv_bufs)) {
1266                 struct rpcrdma_rep *rep;
1267
1268                 rep = rpcrdma_buffer_get_rep_locked(buf);
1269                 rpcrdma_destroy_rep(rep);
1270         }
1271         buf->rb_send_count = 0;
1272
1273         spin_lock(&buf->rb_reqslock);
1274         while (!list_empty(&buf->rb_allreqs)) {
1275                 struct rpcrdma_req *req;
1276
1277                 req = list_first_entry(&buf->rb_allreqs,
1278                                        struct rpcrdma_req, rl_all);
1279                 list_del(&req->rl_all);
1280
1281                 spin_unlock(&buf->rb_reqslock);
1282                 rpcrdma_destroy_req(req);
1283                 spin_lock(&buf->rb_reqslock);
1284         }
1285         spin_unlock(&buf->rb_reqslock);
1286         buf->rb_recv_count = 0;
1287
1288         rpcrdma_destroy_mrs(buf);
1289 }
1290
1291 struct rpcrdma_mw *
1292 rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt)
1293 {
1294         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1295         struct rpcrdma_mw *mw = NULL;
1296
1297         spin_lock(&buf->rb_mwlock);
1298         if (!list_empty(&buf->rb_mws))
1299                 mw = rpcrdma_pop_mw(&buf->rb_mws);
1300         spin_unlock(&buf->rb_mwlock);
1301
1302         if (!mw)
1303                 goto out_nomws;
1304         return mw;
1305
1306 out_nomws:
1307         dprintk("RPC:       %s: no MWs available\n", __func__);
1308         if (r_xprt->rx_ep.rep_connected != -ENODEV)
1309                 schedule_delayed_work(&buf->rb_refresh_worker, 0);
1310
1311         /* Allow the reply handler and refresh worker to run */
1312         cond_resched();
1313
1314         return NULL;
1315 }
1316
1317 void
1318 rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw)
1319 {
1320         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1321
1322         spin_lock(&buf->rb_mwlock);
1323         rpcrdma_push_mw(mw, &buf->rb_mws);
1324         spin_unlock(&buf->rb_mwlock);
1325 }
1326
1327 static struct rpcrdma_rep *
1328 rpcrdma_buffer_get_rep(struct rpcrdma_buffer *buffers)
1329 {
1330         /* If an RPC previously completed without a reply (say, a
1331          * credential problem or a soft timeout occurs) then hold off
1332          * on supplying more Receive buffers until the number of new
1333          * pending RPCs catches up to the number of posted Receives.
1334          */
1335         if (unlikely(buffers->rb_send_count < buffers->rb_recv_count))
1336                 return NULL;
1337
1338         if (unlikely(list_empty(&buffers->rb_recv_bufs)))
1339                 return NULL;
1340         buffers->rb_recv_count++;
1341         return rpcrdma_buffer_get_rep_locked(buffers);
1342 }
1343
1344 /*
1345  * Get a set of request/reply buffers.
1346  *
1347  * Reply buffer (if available) is attached to send buffer upon return.
1348  */
1349 struct rpcrdma_req *
1350 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1351 {
1352         struct rpcrdma_req *req;
1353
1354         spin_lock(&buffers->rb_lock);
1355         if (list_empty(&buffers->rb_send_bufs))
1356                 goto out_reqbuf;
1357         buffers->rb_send_count++;
1358         req = rpcrdma_buffer_get_req_locked(buffers);
1359         req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1360         spin_unlock(&buffers->rb_lock);
1361         return req;
1362
1363 out_reqbuf:
1364         spin_unlock(&buffers->rb_lock);
1365         pr_warn("RPC:       %s: out of request buffers\n", __func__);
1366         return NULL;
1367 }
1368
1369 /*
1370  * Put request/reply buffers back into pool.
1371  * Pre-decrement counter/array index.
1372  */
1373 void
1374 rpcrdma_buffer_put(struct rpcrdma_req *req)
1375 {
1376         struct rpcrdma_buffer *buffers = req->rl_buffer;
1377         struct rpcrdma_rep *rep = req->rl_reply;
1378
1379         req->rl_reply = NULL;
1380
1381         spin_lock(&buffers->rb_lock);
1382         buffers->rb_send_count--;
1383         list_add_tail(&req->rl_list, &buffers->rb_send_bufs);
1384         if (rep) {
1385                 buffers->rb_recv_count--;
1386                 list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1387         }
1388         spin_unlock(&buffers->rb_lock);
1389 }
1390
1391 /*
1392  * Recover reply buffers from pool.
1393  * This happens when recovering from disconnect.
1394  */
1395 void
1396 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1397 {
1398         struct rpcrdma_buffer *buffers = req->rl_buffer;
1399
1400         spin_lock(&buffers->rb_lock);
1401         req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1402         spin_unlock(&buffers->rb_lock);
1403 }
1404
1405 /*
1406  * Put reply buffers back into pool when not attached to
1407  * request. This happens in error conditions.
1408  */
1409 void
1410 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1411 {
1412         struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1413
1414         spin_lock(&buffers->rb_lock);
1415         buffers->rb_recv_count--;
1416         list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1417         spin_unlock(&buffers->rb_lock);
1418 }
1419
1420 /**
1421  * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1422  * @size: size of buffer to be allocated, in bytes
1423  * @direction: direction of data movement
1424  * @flags: GFP flags
1425  *
1426  * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
1427  * can be persistently DMA-mapped for I/O.
1428  *
1429  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1430  * receiving the payload of RDMA RECV operations. During Long Calls
1431  * or Replies they may be registered externally via ro_map.
1432  */
1433 struct rpcrdma_regbuf *
1434 rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
1435                      gfp_t flags)
1436 {
1437         struct rpcrdma_regbuf *rb;
1438
1439         rb = kmalloc(sizeof(*rb) + size, flags);
1440         if (rb == NULL)
1441                 return ERR_PTR(-ENOMEM);
1442
1443         rb->rg_device = NULL;
1444         rb->rg_direction = direction;
1445         rb->rg_iov.length = size;
1446
1447         return rb;
1448 }
1449
1450 /**
1451  * __rpcrdma_map_regbuf - DMA-map a regbuf
1452  * @ia: controlling rpcrdma_ia
1453  * @rb: regbuf to be mapped
1454  */
1455 bool
1456 __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1457 {
1458         struct ib_device *device = ia->ri_device;
1459
1460         if (rb->rg_direction == DMA_NONE)
1461                 return false;
1462
1463         rb->rg_iov.addr = ib_dma_map_single(device,
1464                                             (void *)rb->rg_base,
1465                                             rdmab_length(rb),
1466                                             rb->rg_direction);
1467         if (ib_dma_mapping_error(device, rdmab_addr(rb)))
1468                 return false;
1469
1470         rb->rg_device = device;
1471         rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey;
1472         return true;
1473 }
1474
1475 static void
1476 rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb)
1477 {
1478         if (!rpcrdma_regbuf_is_mapped(rb))
1479                 return;
1480
1481         ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb),
1482                             rdmab_length(rb), rb->rg_direction);
1483         rb->rg_device = NULL;
1484 }
1485
1486 /**
1487  * rpcrdma_free_regbuf - deregister and free registered buffer
1488  * @rb: regbuf to be deregistered and freed
1489  */
1490 void
1491 rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
1492 {
1493         if (!rb)
1494                 return;
1495
1496         rpcrdma_dma_unmap_regbuf(rb);
1497         kfree(rb);
1498 }
1499
1500 /*
1501  * Prepost any receive buffer, then post send.
1502  *
1503  * Receive buffer is donated to hardware, reclaimed upon recv completion.
1504  */
1505 int
1506 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1507                 struct rpcrdma_ep *ep,
1508                 struct rpcrdma_req *req)
1509 {
1510         struct ib_send_wr *send_wr = &req->rl_sendctx->sc_wr;
1511         struct ib_send_wr *send_wr_fail;
1512         int rc;
1513
1514         if (req->rl_reply) {
1515                 rc = rpcrdma_ep_post_recv(ia, req->rl_reply);
1516                 if (rc)
1517                         return rc;
1518                 req->rl_reply = NULL;
1519         }
1520
1521         dprintk("RPC:       %s: posting %d s/g entries\n",
1522                 __func__, send_wr->num_sge);
1523
1524         if (!ep->rep_send_count ||
1525             test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags)) {
1526                 send_wr->send_flags |= IB_SEND_SIGNALED;
1527                 ep->rep_send_count = ep->rep_send_batch;
1528         } else {
1529                 send_wr->send_flags &= ~IB_SEND_SIGNALED;
1530                 --ep->rep_send_count;
1531         }
1532         rc = ib_post_send(ia->ri_id->qp, send_wr, &send_wr_fail);
1533         if (rc)
1534                 goto out_postsend_err;
1535         return 0;
1536
1537 out_postsend_err:
1538         pr_err("rpcrdma: RDMA Send ib_post_send returned %i\n", rc);
1539         return -ENOTCONN;
1540 }
1541
1542 int
1543 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1544                      struct rpcrdma_rep *rep)
1545 {
1546         struct ib_recv_wr *recv_wr_fail;
1547         int rc;
1548
1549         if (!rpcrdma_dma_map_regbuf(ia, rep->rr_rdmabuf))
1550                 goto out_map;
1551         rc = ib_post_recv(ia->ri_id->qp, &rep->rr_recv_wr, &recv_wr_fail);
1552         if (rc)
1553                 goto out_postrecv;
1554         return 0;
1555
1556 out_map:
1557         pr_err("rpcrdma: failed to DMA map the Receive buffer\n");
1558         return -EIO;
1559
1560 out_postrecv:
1561         pr_err("rpcrdma: ib_post_recv returned %i\n", rc);
1562         return -ENOTCONN;
1563 }
1564
1565 /**
1566  * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests
1567  * @r_xprt: transport associated with these backchannel resources
1568  * @min_reqs: minimum number of incoming requests expected
1569  *
1570  * Returns zero if all requested buffers were posted, or a negative errno.
1571  */
1572 int
1573 rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *r_xprt, unsigned int count)
1574 {
1575         struct rpcrdma_buffer *buffers = &r_xprt->rx_buf;
1576         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1577         struct rpcrdma_rep *rep;
1578         int rc;
1579
1580         while (count--) {
1581                 spin_lock(&buffers->rb_lock);
1582                 if (list_empty(&buffers->rb_recv_bufs))
1583                         goto out_reqbuf;
1584                 rep = rpcrdma_buffer_get_rep_locked(buffers);
1585                 spin_unlock(&buffers->rb_lock);
1586
1587                 rc = rpcrdma_ep_post_recv(ia, rep);
1588                 if (rc)
1589                         goto out_rc;
1590         }
1591
1592         return 0;
1593
1594 out_reqbuf:
1595         spin_unlock(&buffers->rb_lock);
1596         pr_warn("%s: no extra receive buffers\n", __func__);
1597         return -ENOMEM;
1598
1599 out_rc:
1600         rpcrdma_recv_buffer_put(rep);
1601         return rc;
1602 }