net/sunrpc/xprtrdma/verbs.c

   1 /*
   2  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the BSD-type
   8  * license below:
   9  *
  10  * Redistribution and use in source and binary forms, with or without
  11  * modification, are permitted provided that the following conditions
  12  * are met:
  13  *
  14  *      Redistributions of source code must retain the above copyright
  15  *      notice, this list of conditions and the following disclaimer.
  16  *
  17  *      Redistributions in binary form must reproduce the above
  18  *      copyright notice, this list of conditions and the following
  19  *      disclaimer in the documentation and/or other materials provided
  20  *      with the distribution.
  21  *
  22  *      Neither the name of the Network Appliance, Inc. nor the names of
  23  *      its contributors may be used to endorse or promote products
  24  *      derived from this software without specific prior written
  25  *      permission.
  26  *
  27  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  28  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  29  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  30  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  31  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  32  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  33  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  34  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  35  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  36  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  37  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  38  */
  39
  40 /*
  41  * verbs.c
  42  *
  43  * Encapsulates the major functions managing:
  44  *  o adapters
  45  *  o endpoints
  46  *  o connections
  47  *  o buffer memory
  48  */
  49
  50 #include <linux/interrupt.h>
  51 #include <linux/slab.h>
  52 #include <linux/prefetch.h>
  53 #include <linux/sunrpc/addr.h>
  54 #include <linux/sunrpc/svc_rdma.h>
  55 #include <asm/bitops.h>
  56
  57 #include <rdma/ib_cm.h>
  58
  59 #include "xprt_rdma.h"
  60
  61 /*
  62  * Globals/Macros
  63  */
  64
  65 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  66 # define RPCDBG_FACILITY        RPCDBG_TRANS
  67 #endif
  68
  69 /*
  70  * internal functions
  71  */
  72 static void rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt);
  73 static void rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf);
  74 static void rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb);
  75
  76 struct workqueue_struct *rpcrdma_receive_wq __read_mostly;
  77
  78 int
  79 rpcrdma_alloc_wq(void)
  80 {
  81         struct workqueue_struct *recv_wq;
  82
  83         recv_wq = alloc_workqueue("xprtrdma_receive",
  84                                   WQ_MEM_RECLAIM | WQ_UNBOUND | WQ_HIGHPRI,
  85                                   0);
  86         if (!recv_wq)
  87                 return -ENOMEM;
  88
  89         rpcrdma_receive_wq = recv_wq;
  90         return 0;
  91 }
  92
  93 void
  94 rpcrdma_destroy_wq(void)
  95 {
  96         struct workqueue_struct *wq;
  97
  98         if (rpcrdma_receive_wq) {
  99                 wq = rpcrdma_receive_wq;
 100                 rpcrdma_receive_wq = NULL;
 101                 destroy_workqueue(wq);
 102         }
 103 }
 104
 105 static void
 106 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
 107 {
 108         struct rpcrdma_ep *ep = context;
 109
 110         pr_err("rpcrdma: %s on device %s ep %p\n",
 111                ib_event_msg(event->event), event->device->name, context);
 112
 113         if (ep->rep_connected == 1) {
 114                 ep->rep_connected = -EIO;
 115                 rpcrdma_conn_func(ep);
 116                 wake_up_all(&ep->rep_connect_wait);
 117         }
 118 }
 119
 120 /**
 121  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
 122  * @cq: completion queue (ignored)
 123  * @wc: completed WR
 124  *
 125  */
 126 static void
 127 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
 128 {
 129         /* WARNING: Only wr_cqe and status are reliable at this point */
 130         if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
 131                 pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
 132                        ib_wc_status_msg(wc->status),
 133                        wc->status, wc->vendor_err);
 134 }
 135
 136 /* Perform basic sanity checking to avoid using garbage
 137  * to update the credit grant value.
 138  */
 139 static void
 140 rpcrdma_update_granted_credits(struct rpcrdma_rep *rep)
 141 {
 142         struct rpcrdma_buffer *buffer = &rep->rr_rxprt->rx_buf;
 143         __be32 *p = rep->rr_rdmabuf->rg_base;
 144         u32 credits;
 145
 146         credits = be32_to_cpup(p + 2);
 147         if (credits == 0)
 148                 credits = 1;    /* don't deadlock */
 149         else if (credits > buffer->rb_max_requests)
 150                 credits = buffer->rb_max_requests;
 151
 152         atomic_set(&buffer->rb_credits, credits);
 153 }
 154
 155 /**
 156  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
 157  * @cq: completion queue (ignored)
 158  * @wc: completed WR
 159  *
 160  */
 161 static void
 162 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
 163 {
 164         struct ib_cqe *cqe = wc->wr_cqe;
 165         struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
 166                                                rr_cqe);
 167
 168         /* WARNING: Only wr_id and status are reliable at this point */
 169         if (wc->status != IB_WC_SUCCESS)
 170                 goto out_fail;
 171
 172         /* status == SUCCESS means all fields in wc are trustworthy */
 173         dprintk("RPC:       %s: rep %p opcode 'recv', length %u: success\n",
 174                 __func__, rep, wc->byte_len);
 175
 176         rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
 177         rep->rr_wc_flags = wc->wc_flags;
 178         rep->rr_inv_rkey = wc->ex.invalidate_rkey;
 179
 180         ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
 181                                    rdmab_addr(rep->rr_rdmabuf),
 182                                    wc->byte_len, DMA_FROM_DEVICE);
 183
 184         if (wc->byte_len >= RPCRDMA_HDRLEN_ERR)
 185                 rpcrdma_update_granted_credits(rep);
 186
 187 out_schedule:
 188         rpcrdma_reply_handler(rep);
 189         return;
 190
 191 out_fail:
 192         if (wc->status != IB_WC_WR_FLUSH_ERR)
 193                 pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
 194                        ib_wc_status_msg(wc->status),
 195                        wc->status, wc->vendor_err);
 196         rpcrdma_set_xdrlen(&rep->rr_hdrbuf, 0);
 197         goto out_schedule;
 198 }
 199
 200 static void
 201 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
 202                                struct rdma_conn_param *param)
 203 {
 204         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
 205         const struct rpcrdma_connect_private *pmsg = param->private_data;
 206         unsigned int rsize, wsize;
 207
 208         /* Default settings for RPC-over-RDMA Version One */
 209         r_xprt->rx_ia.ri_reminv_expected = false;
 210         r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
 211         rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
 212         wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
 213
 214         if (pmsg &&
 215             pmsg->cp_magic == rpcrdma_cmp_magic &&
 216             pmsg->cp_version == RPCRDMA_CMP_VERSION) {
 217                 r_xprt->rx_ia.ri_reminv_expected = true;
 218                 r_xprt->rx_ia.ri_implicit_roundup = true;
 219                 rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
 220                 wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
 221         }
 222
 223         if (rsize < cdata->inline_rsize)
 224                 cdata->inline_rsize = rsize;
 225         if (wsize < cdata->inline_wsize)
 226                 cdata->inline_wsize = wsize;
 227         dprintk("RPC:       %s: max send %u, max recv %u\n",
 228                 __func__, cdata->inline_wsize, cdata->inline_rsize);
 229         rpcrdma_set_max_header_sizes(r_xprt);
 230 }
 231
 232 static int
 233 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
 234 {
 235         struct rpcrdma_xprt *xprt = id->context;
 236         struct rpcrdma_ia *ia = &xprt->rx_ia;
 237         struct rpcrdma_ep *ep = &xprt->rx_ep;
 238 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 239         struct sockaddr *sap = (struct sockaddr *)&ep->rep_remote_addr;
 240 #endif
 241         int connstate = 0;
 242
 243         switch (event->event) {
 244         case RDMA_CM_EVENT_ADDR_RESOLVED:
 245         case RDMA_CM_EVENT_ROUTE_RESOLVED:
 246                 ia->ri_async_rc = 0;
 247                 complete(&ia->ri_done);
 248                 break;
 249         case RDMA_CM_EVENT_ADDR_ERROR:
 250                 ia->ri_async_rc = -EHOSTUNREACH;
 251                 dprintk("RPC:       %s: CM address resolution error, ep 0x%p\n",
 252                         __func__, ep);
 253                 complete(&ia->ri_done);
 254                 break;
 255         case RDMA_CM_EVENT_ROUTE_ERROR:
 256                 ia->ri_async_rc = -ENETUNREACH;
 257                 dprintk("RPC:       %s: CM route resolution error, ep 0x%p\n",
 258                         __func__, ep);
 259                 complete(&ia->ri_done);
 260                 break;
 261         case RDMA_CM_EVENT_DEVICE_REMOVAL:
 262 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 263                 pr_info("rpcrdma: removing device %s for %pIS:%u\n",
 264                         ia->ri_device->name,
 265                         sap, rpc_get_port(sap));
 266 #endif
 267                 set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
 268                 ep->rep_connected = -ENODEV;
 269                 xprt_force_disconnect(&xprt->rx_xprt);
 270                 wait_for_completion(&ia->ri_remove_done);
 271
 272                 ia->ri_id = NULL;
 273                 ia->ri_pd = NULL;
 274                 ia->ri_device = NULL;
 275                 /* Return 1 to ensure the core destroys the id. */
 276                 return 1;
 277         case RDMA_CM_EVENT_ESTABLISHED:
 278                 connstate = 1;
 279                 rpcrdma_update_connect_private(xprt, &event->param.conn);
 280                 goto connected;
 281         case RDMA_CM_EVENT_CONNECT_ERROR:
 282                 connstate = -ENOTCONN;
 283                 goto connected;
 284         case RDMA_CM_EVENT_UNREACHABLE:
 285                 connstate = -ENETDOWN;
 286                 goto connected;
 287         case RDMA_CM_EVENT_REJECTED:
 288                 dprintk("rpcrdma: connection to %pIS:%u rejected: %s\n",
 289                         sap, rpc_get_port(sap),
 290                         rdma_reject_msg(id, event->status));
 291                 connstate = -ECONNREFUSED;
 292                 if (event->status == IB_CM_REJ_STALE_CONN)
 293                         connstate = -EAGAIN;
 294                 goto connected;
 295         case RDMA_CM_EVENT_DISCONNECTED:
 296                 connstate = -ECONNABORTED;
 297 connected:
 298                 atomic_set(&xprt->rx_buf.rb_credits, 1);
 299                 ep->rep_connected = connstate;
 300                 rpcrdma_conn_func(ep);
 301                 wake_up_all(&ep->rep_connect_wait);
 302                 /*FALLTHROUGH*/
 303         default:
 304                 dprintk("RPC:       %s: %pIS:%u on %s/%s (ep 0x%p): %s\n",
 305                         __func__, sap, rpc_get_port(sap),
 306                         ia->ri_device->name, ia->ri_ops->ro_displayname,
 307                         ep, rdma_event_msg(event->event));
 308                 break;
 309         }
 310
 311         return 0;
 312 }
 313
 314 static struct rdma_cm_id *
 315 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
 316                         struct rpcrdma_ia *ia, struct sockaddr *addr)
 317 {
 318         unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
 319         struct rdma_cm_id *id;
 320         int rc;
 321
 322         init_completion(&ia->ri_done);
 323         init_completion(&ia->ri_remove_done);
 324
 325         id = rdma_create_id(&init_net, rpcrdma_conn_upcall, xprt, RDMA_PS_TCP,
 326                             IB_QPT_RC);
 327         if (IS_ERR(id)) {
 328                 rc = PTR_ERR(id);
 329                 dprintk("RPC:       %s: rdma_create_id() failed %i\n",
 330                         __func__, rc);
 331                 return id;
 332         }
 333
 334         ia->ri_async_rc = -ETIMEDOUT;
 335         rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
 336         if (rc) {
 337                 dprintk("RPC:       %s: rdma_resolve_addr() failed %i\n",
 338                         __func__, rc);
 339                 goto out;
 340         }
 341         rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
 342         if (rc < 0) {
 343                 dprintk("RPC:       %s: wait() exited: %i\n",
 344                         __func__, rc);
 345                 goto out;
 346         }
 347
 348         rc = ia->ri_async_rc;
 349         if (rc)
 350                 goto out;
 351
 352         ia->ri_async_rc = -ETIMEDOUT;
 353         rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
 354         if (rc) {
 355                 dprintk("RPC:       %s: rdma_resolve_route() failed %i\n",
 356                         __func__, rc);
 357                 goto out;
 358         }
 359         rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
 360         if (rc < 0) {
 361                 dprintk("RPC:       %s: wait() exited: %i\n",
 362                         __func__, rc);
 363                 goto out;
 364         }
 365         rc = ia->ri_async_rc;
 366         if (rc)
 367                 goto out;
 368
 369         return id;
 370
 371 out:
 372         rdma_destroy_id(id);
 373         return ERR_PTR(rc);
 374 }
 375
 376 /*
 377  * Exported functions.
 378  */
 379
 380 /**
 381  * rpcrdma_ia_open - Open and initialize an Interface Adapter.
 382  * @xprt: controlling transport
 383  * @addr: IP address of remote peer
 384  *
 385  * Returns 0 on success, negative errno if an appropriate
 386  * Interface Adapter could not be found and opened.
 387  */
 388 int
 389 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr)
 390 {
 391         struct rpcrdma_ia *ia = &xprt->rx_ia;
 392         int rc;
 393
 394         ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
 395         if (IS_ERR(ia->ri_id)) {
 396                 rc = PTR_ERR(ia->ri_id);
 397                 goto out_err;
 398         }
 399         ia->ri_device = ia->ri_id->device;
 400
 401         ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
 402         if (IS_ERR(ia->ri_pd)) {
 403                 rc = PTR_ERR(ia->ri_pd);
 404                 pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
 405                 goto out_err;
 406         }
 407
 408         switch (xprt_rdma_memreg_strategy) {
 409         case RPCRDMA_FRMR:
 410                 if (frwr_is_supported(ia)) {
 411                         ia->ri_ops = &rpcrdma_frwr_memreg_ops;
 412                         break;
 413                 }
 414                 /*FALLTHROUGH*/
 415         case RPCRDMA_MTHCAFMR:
 416                 if (fmr_is_supported(ia)) {
 417                         ia->ri_ops = &rpcrdma_fmr_memreg_ops;
 418                         break;
 419                 }
 420                 /*FALLTHROUGH*/
 421         default:
 422                 pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
 423                        ia->ri_device->name, xprt_rdma_memreg_strategy);
 424                 rc = -EINVAL;
 425                 goto out_err;
 426         }
 427
 428         return 0;
 429
 430 out_err:
 431         rpcrdma_ia_close(ia);
 432         return rc;
 433 }
 434
 435 /**
 436  * rpcrdma_ia_remove - Handle device driver unload
 437  * @ia: interface adapter being removed
 438  *
 439  * Divest transport H/W resources associated with this adapter,
 440  * but allow it to be restored later.
 441  */
 442 void
 443 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
 444 {
 445         struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
 446                                                    rx_ia);
 447         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 448         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 449         struct rpcrdma_req *req;
 450         struct rpcrdma_rep *rep;
 451
 452         cancel_delayed_work_sync(&buf->rb_refresh_worker);
 453
 454         /* This is similar to rpcrdma_ep_destroy, but:
 455          * - Don't cancel the connect worker.
 456          * - Don't call rpcrdma_ep_disconnect, which waits
 457          *   for another conn upcall, which will deadlock.
 458          * - rdma_disconnect is unneeded, the underlying
 459          *   connection is already gone.
 460          */
 461         if (ia->ri_id->qp) {
 462                 ib_drain_qp(ia->ri_id->qp);
 463                 rdma_destroy_qp(ia->ri_id);
 464                 ia->ri_id->qp = NULL;
 465         }
 466         ib_free_cq(ep->rep_attr.recv_cq);
 467         ib_free_cq(ep->rep_attr.send_cq);
 468
 469         /* The ULP is responsible for ensuring all DMA
 470          * mappings and MRs are gone.
 471          */
 472         list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
 473                 rpcrdma_dma_unmap_regbuf(rep->rr_rdmabuf);
 474         list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
 475                 rpcrdma_dma_unmap_regbuf(req->rl_rdmabuf);
 476                 rpcrdma_dma_unmap_regbuf(req->rl_sendbuf);
 477                 rpcrdma_dma_unmap_regbuf(req->rl_recvbuf);
 478         }
 479         rpcrdma_destroy_mrs(buf);
 480
 481         /* Allow waiters to continue */
 482         complete(&ia->ri_remove_done);
 483 }
 484
 485 /**
 486  * rpcrdma_ia_close - Clean up/close an IA.
 487  * @ia: interface adapter to close
 488  *
 489  */
 490 void
 491 rpcrdma_ia_close(struct rpcrdma_ia *ia)
 492 {
 493         dprintk("RPC:       %s: entering\n", __func__);
 494         if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
 495                 if (ia->ri_id->qp)
 496                         rdma_destroy_qp(ia->ri_id);
 497                 rdma_destroy_id(ia->ri_id);
 498         }
 499         ia->ri_id = NULL;
 500         ia->ri_device = NULL;
 501
 502         /* If the pd is still busy, xprtrdma missed freeing a resource */
 503         if (ia->ri_pd && !IS_ERR(ia->ri_pd))
 504                 ib_dealloc_pd(ia->ri_pd);
 505         ia->ri_pd = NULL;
 506 }
 507
 508 /*
 509  * Create unconnected endpoint.
 510  */
 511 int
 512 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
 513                   struct rpcrdma_create_data_internal *cdata)
 514 {
 515         struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
 516         unsigned int max_qp_wr, max_sge;
 517         struct ib_cq *sendcq, *recvcq;
 518         int rc;
 519
 520         max_sge = min_t(unsigned int, ia->ri_device->attrs.max_sge,
 521                         RPCRDMA_MAX_SEND_SGES);
 522         if (max_sge < RPCRDMA_MIN_SEND_SGES) {
 523                 pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
 524                 return -ENOMEM;
 525         }
 526         ia->ri_max_send_sges = max_sge - RPCRDMA_MIN_SEND_SGES;
 527
 528         if (ia->ri_device->attrs.max_qp_wr <= RPCRDMA_BACKWARD_WRS) {
 529                 dprintk("RPC:       %s: insufficient wqe's available\n",
 530                         __func__);
 531                 return -ENOMEM;
 532         }
 533         max_qp_wr = ia->ri_device->attrs.max_qp_wr - RPCRDMA_BACKWARD_WRS - 1;
 534
 535         /* check provider's send/recv wr limits */
 536         if (cdata->max_requests > max_qp_wr)
 537                 cdata->max_requests = max_qp_wr;
 538
 539         ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
 540         ep->rep_attr.qp_context = ep;
 541         ep->rep_attr.srq = NULL;
 542         ep->rep_attr.cap.max_send_wr = cdata->max_requests;
 543         ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
 544         ep->rep_attr.cap.max_send_wr += 1;      /* drain cqe */
 545         rc = ia->ri_ops->ro_open(ia, ep, cdata);
 546         if (rc)
 547                 return rc;
 548         ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
 549         ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
 550         ep->rep_attr.cap.max_recv_wr += 1;      /* drain cqe */
 551         ep->rep_attr.cap.max_send_sge = max_sge;
 552         ep->rep_attr.cap.max_recv_sge = 1;
 553         ep->rep_attr.cap.max_inline_data = 0;
 554         ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 555         ep->rep_attr.qp_type = IB_QPT_RC;
 556         ep->rep_attr.port_num = ~0;
 557
 558         dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
 559                 "iovs: send %d recv %d\n",
 560                 __func__,
 561                 ep->rep_attr.cap.max_send_wr,
 562                 ep->rep_attr.cap.max_recv_wr,
 563                 ep->rep_attr.cap.max_send_sge,
 564                 ep->rep_attr.cap.max_recv_sge);
 565
 566         /* set trigger for requesting send completion */
 567         ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 - 1;
 568         if (ep->rep_cqinit <= 2)
 569                 ep->rep_cqinit = 0;     /* always signal? */
 570         rpcrdma_init_cqcount(ep, 0);
 571         init_waitqueue_head(&ep->rep_connect_wait);
 572         INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
 573
 574         sendcq = ib_alloc_cq(ia->ri_device, NULL,
 575                              ep->rep_attr.cap.max_send_wr + 1,
 576                              0, IB_POLL_SOFTIRQ);
 577         if (IS_ERR(sendcq)) {
 578                 rc = PTR_ERR(sendcq);
 579                 dprintk("RPC:       %s: failed to create send CQ: %i\n",
 580                         __func__, rc);
 581                 goto out1;
 582         }
 583
 584         recvcq = ib_alloc_cq(ia->ri_device, NULL,
 585                              ep->rep_attr.cap.max_recv_wr + 1,
 586                              0, IB_POLL_WORKQUEUE);
 587         if (IS_ERR(recvcq)) {
 588                 rc = PTR_ERR(recvcq);
 589                 dprintk("RPC:       %s: failed to create recv CQ: %i\n",
 590                         __func__, rc);
 591                 goto out2;
 592         }
 593
 594         ep->rep_attr.send_cq = sendcq;
 595         ep->rep_attr.recv_cq = recvcq;
 596
 597         /* Initialize cma parameters */
 598         memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
 599
 600         /* Prepare RDMA-CM private message */
 601         pmsg->cp_magic = rpcrdma_cmp_magic;
 602         pmsg->cp_version = RPCRDMA_CMP_VERSION;
 603         pmsg->cp_flags |= ia->ri_ops->ro_send_w_inv_ok;
 604         pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize);
 605         pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize);
 606         ep->rep_remote_cma.private_data = pmsg;
 607         ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
 608
 609         /* Client offers RDMA Read but does not initiate */
 610         ep->rep_remote_cma.initiator_depth = 0;
 611         if (ia->ri_device->attrs.max_qp_rd_atom > 32)   /* arbitrary but <= 255 */
 612                 ep->rep_remote_cma.responder_resources = 32;
 613         else
 614                 ep->rep_remote_cma.responder_resources =
 615                                                 ia->ri_device->attrs.max_qp_rd_atom;
 616
 617         /* Limit transport retries so client can detect server
 618          * GID changes quickly. RPC layer handles re-establishing
 619          * transport connection and retransmission.
 620          */
 621         ep->rep_remote_cma.retry_count = 6;
 622
 623         /* RPC-over-RDMA handles its own flow control. In addition,
 624          * make all RNR NAKs visible so we know that RPC-over-RDMA
 625          * flow control is working correctly (no NAKs should be seen).
 626          */
 627         ep->rep_remote_cma.flow_control = 0;
 628         ep->rep_remote_cma.rnr_retry_count = 0;
 629
 630         return 0;
 631
 632 out2:
 633         ib_free_cq(sendcq);
 634 out1:
 635         return rc;
 636 }
 637
 638 /*
 639  * rpcrdma_ep_destroy
 640  *
 641  * Disconnect and destroy endpoint. After this, the only
 642  * valid operations on the ep are to free it (if dynamically
 643  * allocated) or re-create it.
 644  */
 645 void
 646 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 647 {
 648         dprintk("RPC:       %s: entering, connected is %d\n",
 649                 __func__, ep->rep_connected);
 650
 651         cancel_delayed_work_sync(&ep->rep_connect_worker);
 652
 653         if (ia->ri_id->qp) {
 654                 rpcrdma_ep_disconnect(ep, ia);
 655                 rdma_destroy_qp(ia->ri_id);
 656                 ia->ri_id->qp = NULL;
 657         }
 658
 659         ib_free_cq(ep->rep_attr.recv_cq);
 660         ib_free_cq(ep->rep_attr.send_cq);
 661 }
 662
 663 /* Re-establish a connection after a device removal event.
 664  * Unlike a normal reconnection, a fresh PD and a new set
 665  * of MRs and buffers is needed.
 666  */
 667 static int
 668 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
 669                          struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 670 {
 671         struct sockaddr *sap = (struct sockaddr *)&r_xprt->rx_data.addr;
 672         int rc, err;
 673
 674         pr_info("%s: r_xprt = %p\n", __func__, r_xprt);
 675
 676         rc = -EHOSTUNREACH;
 677         if (rpcrdma_ia_open(r_xprt, sap))
 678                 goto out1;
 679
 680         rc = -ENOMEM;
 681         err = rpcrdma_ep_create(ep, ia, &r_xprt->rx_data);
 682         if (err) {
 683                 pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
 684                 goto out2;
 685         }
 686
 687         rc = -ENETUNREACH;
 688         err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
 689         if (err) {
 690                 pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
 691                 goto out3;
 692         }
 693
 694         rpcrdma_create_mrs(r_xprt);
 695         return 0;
 696
 697 out3:
 698         rpcrdma_ep_destroy(ep, ia);
 699 out2:
 700         rpcrdma_ia_close(ia);
 701 out1:
 702         return rc;
 703 }
 704
 705 static int
 706 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
 707                      struct rpcrdma_ia *ia)
 708 {
 709         struct sockaddr *sap = (struct sockaddr *)&r_xprt->rx_data.addr;
 710         struct rdma_cm_id *id, *old;
 711         int err, rc;
 712
 713         dprintk("RPC:       %s: reconnecting...\n", __func__);
 714
 715         rpcrdma_ep_disconnect(ep, ia);
 716
 717         rc = -EHOSTUNREACH;
 718         id = rpcrdma_create_id(r_xprt, ia, sap);
 719         if (IS_ERR(id))
 720                 goto out;
 721
 722         /* As long as the new ID points to the same device as the
 723          * old ID, we can reuse the transport's existing PD and all
 724          * previously allocated MRs. Also, the same device means
 725          * the transport's previous DMA mappings are still valid.
 726          *
 727          * This is a sanity check only. There should be no way these
 728          * point to two different devices here.
 729          */
 730         old = id;
 731         rc = -ENETUNREACH;
 732         if (ia->ri_device != id->device) {
 733                 pr_err("rpcrdma: can't reconnect on different device!\n");
 734                 goto out_destroy;
 735         }
 736
 737         err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
 738         if (err) {
 739                 dprintk("RPC:       %s: rdma_create_qp returned %d\n",
 740                         __func__, err);
 741                 goto out_destroy;
 742         }
 743
 744         /* Atomically replace the transport's ID and QP. */
 745         rc = 0;
 746         old = ia->ri_id;
 747         ia->ri_id = id;
 748         rdma_destroy_qp(old);
 749
 750 out_destroy:
 751         rdma_destroy_id(old);
 752 out:
 753         return rc;
 754 }
 755
 756 /*
 757  * Connect unconnected endpoint.
 758  */
 759 int
 760 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 761 {
 762         struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
 763                                                    rx_ia);
 764         unsigned int extras;
 765         int rc;
 766
 767 retry:
 768         switch (ep->rep_connected) {
 769         case 0:
 770                 dprintk("RPC:       %s: connecting...\n", __func__);
 771                 rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
 772                 if (rc) {
 773                         dprintk("RPC:       %s: rdma_create_qp failed %i\n",
 774                                 __func__, rc);
 775                         rc = -ENETUNREACH;
 776                         goto out_noupdate;
 777                 }
 778                 break;
 779         case -ENODEV:
 780                 rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
 781                 if (rc)
 782                         goto out_noupdate;
 783                 break;
 784         default:
 785                 rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
 786                 if (rc)
 787                         goto out;
 788         }
 789
 790         ep->rep_connected = 0;
 791
 792         rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
 793         if (rc) {
 794                 dprintk("RPC:       %s: rdma_connect() failed with %i\n",
 795                                 __func__, rc);
 796                 goto out;
 797         }
 798
 799         wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
 800         if (ep->rep_connected <= 0) {
 801                 if (ep->rep_connected == -EAGAIN)
 802                         goto retry;
 803                 rc = ep->rep_connected;
 804                 goto out;
 805         }
 806
 807         dprintk("RPC:       %s: connected\n", __func__);
 808         extras = r_xprt->rx_buf.rb_bc_srv_max_requests;
 809         if (extras)
 810                 rpcrdma_ep_post_extra_recv(r_xprt, extras);
 811
 812 out:
 813         if (rc)
 814                 ep->rep_connected = rc;
 815
 816 out_noupdate:
 817         return rc;
 818 }
 819
 820 /*
 821  * rpcrdma_ep_disconnect
 822  *
 823  * This is separate from destroy to facilitate the ability
 824  * to reconnect without recreating the endpoint.
 825  *
 826  * This call is not reentrant, and must not be made in parallel
 827  * on the same endpoint.
 828  */
 829 void
 830 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
 831 {
 832         int rc;
 833
 834         rc = rdma_disconnect(ia->ri_id);
 835         if (!rc) {
 836                 /* returns without wait if not connected */
 837                 wait_event_interruptible(ep->rep_connect_wait,
 838                                                         ep->rep_connected != 1);
 839                 dprintk("RPC:       %s: after wait, %sconnected\n", __func__,
 840                         (ep->rep_connected == 1) ? "still " : "dis");
 841         } else {
 842                 dprintk("RPC:       %s: rdma_disconnect %i\n", __func__, rc);
 843                 ep->rep_connected = rc;
 844         }
 845
 846         ib_drain_qp(ia->ri_id->qp);
 847 }
 848
 849 static void
 850 rpcrdma_mr_recovery_worker(struct work_struct *work)
 851 {
 852         struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
 853                                                   rb_recovery_worker.work);
 854         struct rpcrdma_mw *mw;
 855
 856         spin_lock(&buf->rb_recovery_lock);
 857         while (!list_empty(&buf->rb_stale_mrs)) {
 858                 mw = rpcrdma_pop_mw(&buf->rb_stale_mrs);
 859                 spin_unlock(&buf->rb_recovery_lock);
 860
 861                 dprintk("RPC:       %s: recovering MR %p\n", __func__, mw);
 862                 mw->mw_xprt->rx_ia.ri_ops->ro_recover_mr(mw);
 863
 864                 spin_lock(&buf->rb_recovery_lock);
 865         }
 866         spin_unlock(&buf->rb_recovery_lock);
 867 }
 868
 869 void
 870 rpcrdma_defer_mr_recovery(struct rpcrdma_mw *mw)
 871 {
 872         struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
 873         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 874
 875         spin_lock(&buf->rb_recovery_lock);
 876         rpcrdma_push_mw(mw, &buf->rb_stale_mrs);
 877         spin_unlock(&buf->rb_recovery_lock);
 878
 879         schedule_delayed_work(&buf->rb_recovery_worker, 0);
 880 }
 881
 882 static void
 883 rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt)
 884 {
 885         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 886         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 887         unsigned int count;
 888         LIST_HEAD(free);
 889         LIST_HEAD(all);
 890
 891         for (count = 0; count < 32; count++) {
 892                 struct rpcrdma_mw *mw;
 893                 int rc;
 894
 895                 mw = kzalloc(sizeof(*mw), GFP_KERNEL);
 896                 if (!mw)
 897                         break;
 898
 899                 rc = ia->ri_ops->ro_init_mr(ia, mw);
 900                 if (rc) {
 901                         kfree(mw);
 902                         break;
 903                 }
 904
 905                 mw->mw_xprt = r_xprt;
 906
 907                 list_add(&mw->mw_list, &free);
 908                 list_add(&mw->mw_all, &all);
 909         }
 910
 911         spin_lock(&buf->rb_mwlock);
 912         list_splice(&free, &buf->rb_mws);
 913         list_splice(&all, &buf->rb_all);
 914         r_xprt->rx_stats.mrs_allocated += count;
 915         spin_unlock(&buf->rb_mwlock);
 916
 917         dprintk("RPC:       %s: created %u MRs\n", __func__, count);
 918 }
 919
 920 static void
 921 rpcrdma_mr_refresh_worker(struct work_struct *work)
 922 {
 923         struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
 924                                                   rb_refresh_worker.work);
 925         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
 926                                                    rx_buf);
 927
 928         rpcrdma_create_mrs(r_xprt);
 929 }
 930
 931 struct rpcrdma_req *
 932 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
 933 {
 934         struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
 935         struct rpcrdma_req *req;
 936
 937         req = kzalloc(sizeof(*req), GFP_KERNEL);
 938         if (req == NULL)
 939                 return ERR_PTR(-ENOMEM);
 940
 941         spin_lock(&buffer->rb_reqslock);
 942         list_add(&req->rl_all, &buffer->rb_allreqs);
 943         spin_unlock(&buffer->rb_reqslock);
 944         req->rl_cqe.done = rpcrdma_wc_send;
 945         req->rl_buffer = &r_xprt->rx_buf;
 946         INIT_LIST_HEAD(&req->rl_registered);
 947         req->rl_send_wr.next = NULL;
 948         req->rl_send_wr.wr_cqe = &req->rl_cqe;
 949         req->rl_send_wr.sg_list = req->rl_send_sge;
 950         req->rl_send_wr.opcode = IB_WR_SEND;
 951         return req;
 952 }
 953
 954 struct rpcrdma_rep *
 955 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
 956 {
 957         struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
 958         struct rpcrdma_rep *rep;
 959         int rc;
 960
 961         rc = -ENOMEM;
 962         rep = kzalloc(sizeof(*rep), GFP_KERNEL);
 963         if (rep == NULL)
 964                 goto out;
 965
 966         rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
 967                                                DMA_FROM_DEVICE, GFP_KERNEL);
 968         if (IS_ERR(rep->rr_rdmabuf)) {
 969                 rc = PTR_ERR(rep->rr_rdmabuf);
 970                 goto out_free;
 971         }
 972         xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base,
 973                      rdmab_length(rep->rr_rdmabuf));
 974
 975         rep->rr_cqe.done = rpcrdma_wc_receive;
 976         rep->rr_rxprt = r_xprt;
 977         INIT_WORK(&rep->rr_work, rpcrdma_deferred_completion);
 978         rep->rr_recv_wr.next = NULL;
 979         rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
 980         rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
 981         rep->rr_recv_wr.num_sge = 1;
 982         return rep;
 983
 984 out_free:
 985         kfree(rep);
 986 out:
 987         return ERR_PTR(rc);
 988 }
 989
 990 int
 991 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
 992 {
 993         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 994         int i, rc;
 995
 996         buf->rb_max_requests = r_xprt->rx_data.max_requests;
 997         buf->rb_bc_srv_max_requests = 0;
 998         atomic_set(&buf->rb_credits, 1);
 999         spin_lock_init(&buf->rb_mwlock);
1000         spin_lock_init(&buf->rb_lock);
1001         spin_lock_init(&buf->rb_recovery_lock);
1002         INIT_LIST_HEAD(&buf->rb_mws);
1003         INIT_LIST_HEAD(&buf->rb_all);
1004         INIT_LIST_HEAD(&buf->rb_stale_mrs);
1005         INIT_DELAYED_WORK(&buf->rb_refresh_worker,
1006                           rpcrdma_mr_refresh_worker);
1007         INIT_DELAYED_WORK(&buf->rb_recovery_worker,
1008                           rpcrdma_mr_recovery_worker);
1009
1010         rpcrdma_create_mrs(r_xprt);
1011
1012         INIT_LIST_HEAD(&buf->rb_send_bufs);
1013         INIT_LIST_HEAD(&buf->rb_allreqs);
1014         spin_lock_init(&buf->rb_reqslock);
1015         for (i = 0; i < buf->rb_max_requests; i++) {
1016                 struct rpcrdma_req *req;
1017
1018                 req = rpcrdma_create_req(r_xprt);
1019                 if (IS_ERR(req)) {
1020                         dprintk("RPC:       %s: request buffer %d alloc"
1021                                 " failed\n", __func__, i);
1022                         rc = PTR_ERR(req);
1023                         goto out;
1024                 }
1025                 req->rl_backchannel = false;
1026                 list_add(&req->rl_list, &buf->rb_send_bufs);
1027         }
1028
1029         INIT_LIST_HEAD(&buf->rb_recv_bufs);
1030         for (i = 0; i < buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS; i++) {
1031                 struct rpcrdma_rep *rep;
1032
1033                 rep = rpcrdma_create_rep(r_xprt);
1034                 if (IS_ERR(rep)) {
1035                         dprintk("RPC:       %s: reply buffer %d alloc failed\n",
1036                                 __func__, i);
1037                         rc = PTR_ERR(rep);
1038                         goto out;
1039                 }
1040                 list_add(&rep->rr_list, &buf->rb_recv_bufs);
1041         }
1042
1043         return 0;
1044 out:
1045         rpcrdma_buffer_destroy(buf);
1046         return rc;
1047 }
1048
1049 static struct rpcrdma_req *
1050 rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer *buf)
1051 {
1052         struct rpcrdma_req *req;
1053
1054         req = list_first_entry(&buf->rb_send_bufs,
1055                                struct rpcrdma_req, rl_list);
1056         list_del_init(&req->rl_list);
1057         return req;
1058 }
1059
1060 static struct rpcrdma_rep *
1061 rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer *buf)
1062 {
1063         struct rpcrdma_rep *rep;
1064
1065         rep = list_first_entry(&buf->rb_recv_bufs,
1066                                struct rpcrdma_rep, rr_list);
1067         list_del(&rep->rr_list);
1068         return rep;
1069 }
1070
1071 static void
1072 rpcrdma_destroy_rep(struct rpcrdma_rep *rep)
1073 {
1074         rpcrdma_free_regbuf(rep->rr_rdmabuf);
1075         kfree(rep);
1076 }
1077
1078 void
1079 rpcrdma_destroy_req(struct rpcrdma_req *req)
1080 {
1081         rpcrdma_free_regbuf(req->rl_recvbuf);
1082         rpcrdma_free_regbuf(req->rl_sendbuf);
1083         rpcrdma_free_regbuf(req->rl_rdmabuf);
1084         kfree(req);
1085 }
1086
1087 static void
1088 rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf)
1089 {
1090         struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1091                                                    rx_buf);
1092         struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1093         struct rpcrdma_mw *mw;
1094         unsigned int count;
1095
1096         count = 0;
1097         spin_lock(&buf->rb_mwlock);
1098         while (!list_empty(&buf->rb_all)) {
1099                 mw = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
1100                 list_del(&mw->mw_all);
1101
1102                 spin_unlock(&buf->rb_mwlock);
1103                 ia->ri_ops->ro_release_mr(mw);
1104                 count++;
1105                 spin_lock(&buf->rb_mwlock);
1106         }
1107         spin_unlock(&buf->rb_mwlock);
1108         r_xprt->rx_stats.mrs_allocated = 0;
1109
1110         dprintk("RPC:       %s: released %u MRs\n", __func__, count);
1111 }
1112
1113 void
1114 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1115 {
1116         cancel_delayed_work_sync(&buf->rb_recovery_worker);
1117         cancel_delayed_work_sync(&buf->rb_refresh_worker);
1118
1119         while (!list_empty(&buf->rb_recv_bufs)) {
1120                 struct rpcrdma_rep *rep;
1121
1122                 rep = rpcrdma_buffer_get_rep_locked(buf);
1123                 rpcrdma_destroy_rep(rep);
1124         }
1125         buf->rb_send_count = 0;
1126
1127         spin_lock(&buf->rb_reqslock);
1128         while (!list_empty(&buf->rb_allreqs)) {
1129                 struct rpcrdma_req *req;
1130
1131                 req = list_first_entry(&buf->rb_allreqs,
1132                                        struct rpcrdma_req, rl_all);
1133                 list_del(&req->rl_all);
1134
1135                 spin_unlock(&buf->rb_reqslock);
1136                 rpcrdma_destroy_req(req);
1137                 spin_lock(&buf->rb_reqslock);
1138         }
1139         spin_unlock(&buf->rb_reqslock);
1140         buf->rb_recv_count = 0;
1141
1142         rpcrdma_destroy_mrs(buf);
1143 }
1144
1145 struct rpcrdma_mw *
1146 rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt)
1147 {
1148         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1149         struct rpcrdma_mw *mw = NULL;
1150
1151         spin_lock(&buf->rb_mwlock);
1152         if (!list_empty(&buf->rb_mws))
1153                 mw = rpcrdma_pop_mw(&buf->rb_mws);
1154         spin_unlock(&buf->rb_mwlock);
1155
1156         if (!mw)
1157                 goto out_nomws;
1158         mw->mw_flags = 0;
1159         return mw;
1160
1161 out_nomws:
1162         dprintk("RPC:       %s: no MWs available\n", __func__);
1163         if (r_xprt->rx_ep.rep_connected != -ENODEV)
1164                 schedule_delayed_work(&buf->rb_refresh_worker, 0);
1165
1166         /* Allow the reply handler and refresh worker to run */
1167         cond_resched();
1168
1169         return NULL;
1170 }
1171
1172 void
1173 rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw)
1174 {
1175         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1176
1177         spin_lock(&buf->rb_mwlock);
1178         rpcrdma_push_mw(mw, &buf->rb_mws);
1179         spin_unlock(&buf->rb_mwlock);
1180 }
1181
1182 static struct rpcrdma_rep *
1183 rpcrdma_buffer_get_rep(struct rpcrdma_buffer *buffers)
1184 {
1185         /* If an RPC previously completed without a reply (say, a
1186          * credential problem or a soft timeout occurs) then hold off
1187          * on supplying more Receive buffers until the number of new
1188          * pending RPCs catches up to the number of posted Receives.
1189          */
1190         if (unlikely(buffers->rb_send_count < buffers->rb_recv_count))
1191                 return NULL;
1192
1193         if (unlikely(list_empty(&buffers->rb_recv_bufs)))
1194                 return NULL;
1195         buffers->rb_recv_count++;
1196         return rpcrdma_buffer_get_rep_locked(buffers);
1197 }
1198
1199 /*
1200  * Get a set of request/reply buffers.
1201  *
1202  * Reply buffer (if available) is attached to send buffer upon return.
1203  */
1204 struct rpcrdma_req *
1205 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1206 {
1207         struct rpcrdma_req *req;
1208
1209         spin_lock(&buffers->rb_lock);
1210         if (list_empty(&buffers->rb_send_bufs))
1211                 goto out_reqbuf;
1212         buffers->rb_send_count++;
1213         req = rpcrdma_buffer_get_req_locked(buffers);
1214         req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1215         spin_unlock(&buffers->rb_lock);
1216         return req;
1217
1218 out_reqbuf:
1219         spin_unlock(&buffers->rb_lock);
1220         pr_warn("RPC:       %s: out of request buffers\n", __func__);
1221         return NULL;
1222 }
1223
1224 /*
1225  * Put request/reply buffers back into pool.
1226  * Pre-decrement counter/array index.
1227  */
1228 void
1229 rpcrdma_buffer_put(struct rpcrdma_req *req)
1230 {
1231         struct rpcrdma_buffer *buffers = req->rl_buffer;
1232         struct rpcrdma_rep *rep = req->rl_reply;
1233
1234         req->rl_send_wr.num_sge = 0;
1235         req->rl_reply = NULL;
1236
1237         spin_lock(&buffers->rb_lock);
1238         buffers->rb_send_count--;
1239         list_add_tail(&req->rl_list, &buffers->rb_send_bufs);
1240         if (rep) {
1241                 buffers->rb_recv_count--;
1242                 list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1243         }
1244         spin_unlock(&buffers->rb_lock);
1245 }
1246
1247 /*
1248  * Recover reply buffers from pool.
1249  * This happens when recovering from disconnect.
1250  */
1251 void
1252 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1253 {
1254         struct rpcrdma_buffer *buffers = req->rl_buffer;
1255
1256         spin_lock(&buffers->rb_lock);
1257         req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1258         spin_unlock(&buffers->rb_lock);
1259 }
1260
1261 /*
1262  * Put reply buffers back into pool when not attached to
1263  * request. This happens in error conditions.
1264  */
1265 void
1266 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1267 {
1268         struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1269
1270         spin_lock(&buffers->rb_lock);
1271         buffers->rb_recv_count--;
1272         list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1273         spin_unlock(&buffers->rb_lock);
1274 }
1275
1276 /**
1277  * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1278  * @size: size of buffer to be allocated, in bytes
1279  * @direction: direction of data movement
1280  * @flags: GFP flags
1281  *
1282  * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
1283  * can be persistently DMA-mapped for I/O.
1284  *
1285  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1286  * receiving the payload of RDMA RECV operations. During Long Calls
1287  * or Replies they may be registered externally via ro_map.
1288  */
1289 struct rpcrdma_regbuf *
1290 rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
1291                      gfp_t flags)
1292 {
1293         struct rpcrdma_regbuf *rb;
1294
1295         rb = kmalloc(sizeof(*rb) + size, flags);
1296         if (rb == NULL)
1297                 return ERR_PTR(-ENOMEM);
1298
1299         rb->rg_device = NULL;
1300         rb->rg_direction = direction;
1301         rb->rg_iov.length = size;
1302
1303         return rb;
1304 }
1305
1306 /**
1307  * __rpcrdma_map_regbuf - DMA-map a regbuf
1308  * @ia: controlling rpcrdma_ia
1309  * @rb: regbuf to be mapped
1310  */
1311 bool
1312 __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1313 {
1314         struct ib_device *device = ia->ri_device;
1315
1316         if (rb->rg_direction == DMA_NONE)
1317                 return false;
1318
1319         rb->rg_iov.addr = ib_dma_map_single(device,
1320                                             (void *)rb->rg_base,
1321                                             rdmab_length(rb),
1322                                             rb->rg_direction);
1323         if (ib_dma_mapping_error(device, rdmab_addr(rb)))
1324                 return false;
1325
1326         rb->rg_device = device;
1327         rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey;
1328         return true;
1329 }
1330
1331 static void
1332 rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb)
1333 {
1334         if (!rpcrdma_regbuf_is_mapped(rb))
1335                 return;
1336
1337         ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb),
1338                             rdmab_length(rb), rb->rg_direction);
1339         rb->rg_device = NULL;
1340 }
1341
1342 /**
1343  * rpcrdma_free_regbuf - deregister and free registered buffer
1344  * @rb: regbuf to be deregistered and freed
1345  */
1346 void
1347 rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
1348 {
1349         if (!rb)
1350                 return;
1351
1352         rpcrdma_dma_unmap_regbuf(rb);
1353         kfree(rb);
1354 }
1355
1356 /*
1357  * Prepost any receive buffer, then post send.
1358  *
1359  * Receive buffer is donated to hardware, reclaimed upon recv completion.
1360  */
1361 int
1362 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1363                 struct rpcrdma_ep *ep,
1364                 struct rpcrdma_req *req)
1365 {
1366         struct ib_send_wr *send_wr = &req->rl_send_wr;
1367         struct ib_send_wr *send_wr_fail;
1368         int rc;
1369
1370         if (req->rl_reply) {
1371                 rc = rpcrdma_ep_post_recv(ia, req->rl_reply);
1372                 if (rc)
1373                         return rc;
1374                 req->rl_reply = NULL;
1375         }
1376
1377         dprintk("RPC:       %s: posting %d s/g entries\n",
1378                 __func__, send_wr->num_sge);
1379
1380         rpcrdma_set_signaled(ep, send_wr);
1381         rc = ib_post_send(ia->ri_id->qp, send_wr, &send_wr_fail);
1382         if (rc)
1383                 goto out_postsend_err;
1384         return 0;
1385
1386 out_postsend_err:
1387         pr_err("rpcrdma: RDMA Send ib_post_send returned %i\n", rc);
1388         return -ENOTCONN;
1389 }
1390
1391 int
1392 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1393                      struct rpcrdma_rep *rep)
1394 {
1395         struct ib_recv_wr *recv_wr_fail;
1396         int rc;
1397
1398         if (!rpcrdma_dma_map_regbuf(ia, rep->rr_rdmabuf))
1399                 goto out_map;
1400         rc = ib_post_recv(ia->ri_id->qp, &rep->rr_recv_wr, &recv_wr_fail);
1401         if (rc)
1402                 goto out_postrecv;
1403         return 0;
1404
1405 out_map:
1406         pr_err("rpcrdma: failed to DMA map the Receive buffer\n");
1407         return -EIO;
1408
1409 out_postrecv:
1410         pr_err("rpcrdma: ib_post_recv returned %i\n", rc);
1411         return -ENOTCONN;
1412 }
1413
1414 /**
1415  * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests
1416  * @r_xprt: transport associated with these backchannel resources
1417  * @min_reqs: minimum number of incoming requests expected
1418  *
1419  * Returns zero if all requested buffers were posted, or a negative errno.
1420  */
1421 int
1422 rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *r_xprt, unsigned int count)
1423 {
1424         struct rpcrdma_buffer *buffers = &r_xprt->rx_buf;
1425         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1426         struct rpcrdma_rep *rep;
1427         int rc;
1428
1429         while (count--) {
1430                 spin_lock(&buffers->rb_lock);
1431                 if (list_empty(&buffers->rb_recv_bufs))
1432                         goto out_reqbuf;
1433                 rep = rpcrdma_buffer_get_rep_locked(buffers);
1434                 spin_unlock(&buffers->rb_lock);
1435
1436                 rc = rpcrdma_ep_post_recv(ia, rep);
1437                 if (rc)
1438                         goto out_rc;
1439         }
1440
1441         return 0;
1442
1443 out_reqbuf:
1444         spin_unlock(&buffers->rb_lock);
1445         pr_warn("%s: no extra receive buffers\n", __func__);
1446         return -ENOMEM;
1447
1448 out_rc:
1449         rpcrdma_recv_buffer_put(rep);
1450         return rc;
1451 }