net/sunrpc/xprtrdma/transport.c

   1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
   2 /*
   3  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
   4  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the BSD-type
  10  * license below:
  11  *
  12  * Redistribution and use in source and binary forms, with or without
  13  * modification, are permitted provided that the following conditions
  14  * are met:
  15  *
  16  *      Redistributions of source code must retain the above copyright
  17  *      notice, this list of conditions and the following disclaimer.
  18  *
  19  *      Redistributions in binary form must reproduce the above
  20  *      copyright notice, this list of conditions and the following
  21  *      disclaimer in the documentation and/or other materials provided
  22  *      with the distribution.
  23  *
  24  *      Neither the name of the Network Appliance, Inc. nor the names of
  25  *      its contributors may be used to endorse or promote products
  26  *      derived from this software without specific prior written
  27  *      permission.
  28  *
  29  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  30  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  31  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  32  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  33  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  34  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  35  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  36  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  37  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  38  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  39  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  40  */
  41
  42 /*
  43  * transport.c
  44  *
  45  * This file contains the top-level implementation of an RPC RDMA
  46  * transport.
  47  *
  48  * Naming convention: functions beginning with xprt_ are part of the
  49  * transport switch. All others are RPC RDMA internal.
  50  */
  51
  52 #include <linux/module.h>
  53 #include <linux/slab.h>
  54 #include <linux/seq_file.h>
  55 #include <linux/smp.h>
  56
  57 #include <linux/sunrpc/addr.h>
  58 #include <linux/sunrpc/svc_rdma.h>
  59
  60 #include "xprt_rdma.h"
  61 #include <trace/events/rpcrdma.h>
  62
  63 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  64 # define RPCDBG_FACILITY        RPCDBG_TRANS
  65 #endif
  66
  67 /*
  68  * tunables
  69  */
  70
  71 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
  72 unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
  73 unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
  74 unsigned int xprt_rdma_memreg_strategy          = RPCRDMA_FRWR;
  75 int xprt_rdma_pad_optimize;
  76
  77 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  78
  79 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
  80 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
  81 static unsigned int min_inline_size = RPCRDMA_MIN_INLINE;
  82 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
  83 static unsigned int max_padding = PAGE_SIZE;
  84 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
  85 static unsigned int max_memreg = RPCRDMA_LAST - 1;
  86 static unsigned int dummy;
  87
  88 static struct ctl_table_header *sunrpc_table_header;
  89
  90 static struct ctl_table xr_tunables_table[] = {
  91         {
  92                 .procname       = "rdma_slot_table_entries",
  93                 .data           = &xprt_rdma_slot_table_entries,
  94                 .maxlen         = sizeof(unsigned int),
  95                 .mode           = 0644,
  96                 .proc_handler   = proc_dointvec_minmax,
  97                 .extra1         = &min_slot_table_size,
  98                 .extra2         = &max_slot_table_size
  99         },
 100         {
 101                 .procname       = "rdma_max_inline_read",
 102                 .data           = &xprt_rdma_max_inline_read,
 103                 .maxlen         = sizeof(unsigned int),
 104                 .mode           = 0644,
 105                 .proc_handler   = proc_dointvec_minmax,
 106                 .extra1         = &min_inline_size,
 107                 .extra2         = &max_inline_size,
 108         },
 109         {
 110                 .procname       = "rdma_max_inline_write",
 111                 .data           = &xprt_rdma_max_inline_write,
 112                 .maxlen         = sizeof(unsigned int),
 113                 .mode           = 0644,
 114                 .proc_handler   = proc_dointvec_minmax,
 115                 .extra1         = &min_inline_size,
 116                 .extra2         = &max_inline_size,
 117         },
 118         {
 119                 .procname       = "rdma_inline_write_padding",
 120                 .data           = &dummy,
 121                 .maxlen         = sizeof(unsigned int),
 122                 .mode           = 0644,
 123                 .proc_handler   = proc_dointvec_minmax,
 124                 .extra1         = SYSCTL_ZERO,
 125                 .extra2         = &max_padding,
 126         },
 127         {
 128                 .procname       = "rdma_memreg_strategy",
 129                 .data           = &xprt_rdma_memreg_strategy,
 130                 .maxlen         = sizeof(unsigned int),
 131                 .mode           = 0644,
 132                 .proc_handler   = proc_dointvec_minmax,
 133                 .extra1         = &min_memreg,
 134                 .extra2         = &max_memreg,
 135         },
 136         {
 137                 .procname       = "rdma_pad_optimize",
 138                 .data           = &xprt_rdma_pad_optimize,
 139                 .maxlen         = sizeof(unsigned int),
 140                 .mode           = 0644,
 141                 .proc_handler   = proc_dointvec,
 142         },
 143         { },
 144 };
 145
 146 static struct ctl_table sunrpc_table[] = {
 147         {
 148                 .procname       = "sunrpc",
 149                 .mode           = 0555,
 150                 .child          = xr_tunables_table
 151         },
 152         { },
 153 };
 154
 155 #endif
 156
 157 static const struct rpc_xprt_ops xprt_rdma_procs;
 158
 159 static void
 160 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
 161 {
 162         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
 163         char buf[20];
 164
 165         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
 166         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 167
 168         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
 169 }
 170
 171 static void
 172 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
 173 {
 174         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
 175         char buf[40];
 176
 177         snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
 178         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 179
 180         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
 181 }
 182
 183 void
 184 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
 185 {
 186         char buf[128];
 187
 188         switch (sap->sa_family) {
 189         case AF_INET:
 190                 xprt_rdma_format_addresses4(xprt, sap);
 191                 break;
 192         case AF_INET6:
 193                 xprt_rdma_format_addresses6(xprt, sap);
 194                 break;
 195         default:
 196                 pr_err("rpcrdma: Unrecognized address family\n");
 197                 return;
 198         }
 199
 200         (void)rpc_ntop(sap, buf, sizeof(buf));
 201         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
 202
 203         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
 204         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 205
 206         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
 207         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 208
 209         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
 210 }
 211
 212 void
 213 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
 214 {
 215         unsigned int i;
 216
 217         for (i = 0; i < RPC_DISPLAY_MAX; i++)
 218                 switch (i) {
 219                 case RPC_DISPLAY_PROTO:
 220                 case RPC_DISPLAY_NETID:
 221                         continue;
 222                 default:
 223                         kfree(xprt->address_strings[i]);
 224                 }
 225 }
 226
 227 /**
 228  * xprt_rdma_connect_worker - establish connection in the background
 229  * @work: worker thread context
 230  *
 231  * Requester holds the xprt's send lock to prevent activity on this
 232  * transport while a fresh connection is being established. RPC tasks
 233  * sleep on the xprt's pending queue waiting for connect to complete.
 234  */
 235 static void
 236 xprt_rdma_connect_worker(struct work_struct *work)
 237 {
 238         struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
 239                                                    rx_connect_worker.work);
 240         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 241         int rc;
 242
 243         rc = rpcrdma_xprt_connect(r_xprt);
 244         xprt_clear_connecting(xprt);
 245         if (!rc) {
 246                 xprt->connect_cookie++;
 247                 xprt->stat.connect_count++;
 248                 xprt->stat.connect_time += (long)jiffies -
 249                                            xprt->stat.connect_start;
 250                 xprt_set_connected(xprt);
 251                 rc = -EAGAIN;
 252         }
 253         xprt_wake_pending_tasks(xprt, rc);
 254 }
 255
 256 /**
 257  * xprt_rdma_inject_disconnect - inject a connection fault
 258  * @xprt: transport context
 259  *
 260  * If @xprt is connected, disconnect it to simulate spurious connection
 261  * loss.
 262  */
 263 static void
 264 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
 265 {
 266         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 267
 268         trace_xprtrdma_op_inject_dsc(r_xprt);
 269         rdma_disconnect(r_xprt->rx_ep->re_id);
 270 }
 271
 272 /**
 273  * xprt_rdma_destroy - Full tear down of transport
 274  * @xprt: doomed transport context
 275  *
 276  * Caller guarantees there will be no more calls to us with
 277  * this @xprt.
 278  */
 279 static void
 280 xprt_rdma_destroy(struct rpc_xprt *xprt)
 281 {
 282         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 283
 284         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
 285
 286         rpcrdma_xprt_disconnect(r_xprt);
 287         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 288
 289         xprt_rdma_free_addresses(xprt);
 290         xprt_free(xprt);
 291
 292         module_put(THIS_MODULE);
 293 }
 294
 295 /* 60 second timeout, no retries */
 296 static const struct rpc_timeout xprt_rdma_default_timeout = {
 297         .to_initval = 60 * HZ,
 298         .to_maxval = 60 * HZ,
 299 };
 300
 301 /**
 302  * xprt_setup_rdma - Set up transport to use RDMA
 303  *
 304  * @args: rpc transport arguments
 305  */
 306 static struct rpc_xprt *
 307 xprt_setup_rdma(struct xprt_create *args)
 308 {
 309         struct rpc_xprt *xprt;
 310         struct rpcrdma_xprt *new_xprt;
 311         struct sockaddr *sap;
 312         int rc;
 313
 314         if (args->addrlen > sizeof(xprt->addr))
 315                 return ERR_PTR(-EBADF);
 316
 317         if (!try_module_get(THIS_MODULE))
 318                 return ERR_PTR(-EIO);
 319
 320         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0,
 321                           xprt_rdma_slot_table_entries);
 322         if (!xprt) {
 323                 module_put(THIS_MODULE);
 324                 return ERR_PTR(-ENOMEM);
 325         }
 326
 327         xprt->timeout = &xprt_rdma_default_timeout;
 328         xprt->connect_timeout = xprt->timeout->to_initval;
 329         xprt->max_reconnect_timeout = xprt->timeout->to_maxval;
 330         xprt->bind_timeout = RPCRDMA_BIND_TO;
 331         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 332         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
 333
 334         xprt->resvport = 0;             /* privileged port not needed */
 335         xprt->ops = &xprt_rdma_procs;
 336
 337         /*
 338          * Set up RDMA-specific connect data.
 339          */
 340         sap = args->dstaddr;
 341
 342         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 343          * address, for any side protocols which peek at it */
 344         xprt->prot = IPPROTO_TCP;
 345         xprt->addrlen = args->addrlen;
 346         memcpy(&xprt->addr, sap, xprt->addrlen);
 347
 348         if (rpc_get_port(sap))
 349                 xprt_set_bound(xprt);
 350         xprt_rdma_format_addresses(xprt, sap);
 351
 352         new_xprt = rpcx_to_rdmax(xprt);
 353         rc = rpcrdma_buffer_create(new_xprt);
 354         if (rc) {
 355                 xprt_rdma_free_addresses(xprt);
 356                 xprt_free(xprt);
 357                 module_put(THIS_MODULE);
 358                 return ERR_PTR(rc);
 359         }
 360
 361         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
 362                           xprt_rdma_connect_worker);
 363
 364         xprt->max_payload = RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT;
 365
 366         return xprt;
 367 }
 368
 369 /**
 370  * xprt_rdma_close - close a transport connection
 371  * @xprt: transport context
 372  *
 373  * Called during autoclose or device removal.
 374  *
 375  * Caller holds @xprt's send lock to prevent activity on this
 376  * transport while the connection is torn down.
 377  */
 378 void xprt_rdma_close(struct rpc_xprt *xprt)
 379 {
 380         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 381
 382         rpcrdma_xprt_disconnect(r_xprt);
 383
 384         xprt->reestablish_timeout = 0;
 385         ++xprt->connect_cookie;
 386         xprt_disconnect_done(xprt);
 387 }
 388
 389 /**
 390  * xprt_rdma_set_port - update server port with rpcbind result
 391  * @xprt: controlling RPC transport
 392  * @port: new port value
 393  *
 394  * Transport connect status is unchanged.
 395  */
 396 static void
 397 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 398 {
 399         struct sockaddr *sap = (struct sockaddr *)&xprt->addr;
 400         char buf[8];
 401
 402         rpc_set_port(sap, port);
 403
 404         kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
 405         snprintf(buf, sizeof(buf), "%u", port);
 406         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 407
 408         kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
 409         snprintf(buf, sizeof(buf), "%4hx", port);
 410         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 411
 412         trace_xprtrdma_op_setport(container_of(xprt, struct rpcrdma_xprt,
 413                                                rx_xprt));
 414 }
 415
 416 /**
 417  * xprt_rdma_timer - invoked when an RPC times out
 418  * @xprt: controlling RPC transport
 419  * @task: RPC task that timed out
 420  *
 421  * Invoked when the transport is still connected, but an RPC
 422  * retransmit timeout occurs.
 423  *
 424  * Since RDMA connections don't have a keep-alive, forcibly
 425  * disconnect and retry to connect. This drives full
 426  * detection of the network path, and retransmissions of
 427  * all pending RPCs.
 428  */
 429 static void
 430 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
 431 {
 432         xprt_force_disconnect(xprt);
 433 }
 434
 435 /**
 436  * xprt_rdma_set_connect_timeout - set timeouts for establishing a connection
 437  * @xprt: controlling transport instance
 438  * @connect_timeout: reconnect timeout after client disconnects
 439  * @reconnect_timeout: reconnect timeout after server disconnects
 440  *
 441  */
 442 static void xprt_rdma_set_connect_timeout(struct rpc_xprt *xprt,
 443                                           unsigned long connect_timeout,
 444                                           unsigned long reconnect_timeout)
 445 {
 446         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 447
 448         trace_xprtrdma_op_set_cto(r_xprt, connect_timeout, reconnect_timeout);
 449
 450         spin_lock(&xprt->transport_lock);
 451
 452         if (connect_timeout < xprt->connect_timeout) {
 453                 struct rpc_timeout to;
 454                 unsigned long initval;
 455
 456                 to = *xprt->timeout;
 457                 initval = connect_timeout;
 458                 if (initval < RPCRDMA_INIT_REEST_TO << 1)
 459                         initval = RPCRDMA_INIT_REEST_TO << 1;
 460                 to.to_initval = initval;
 461                 to.to_maxval = initval;
 462                 r_xprt->rx_timeout = to;
 463                 xprt->timeout = &r_xprt->rx_timeout;
 464                 xprt->connect_timeout = connect_timeout;
 465         }
 466
 467         if (reconnect_timeout < xprt->max_reconnect_timeout)
 468                 xprt->max_reconnect_timeout = reconnect_timeout;
 469
 470         spin_unlock(&xprt->transport_lock);
 471 }
 472
 473 /**
 474  * xprt_rdma_connect - schedule an attempt to reconnect
 475  * @xprt: transport state
 476  * @task: RPC scheduler context (unused)
 477  *
 478  */
 479 static void
 480 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
 481 {
 482         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 483         struct rpcrdma_ep *ep = r_xprt->rx_ep;
 484         unsigned long delay;
 485
 486         delay = 0;
 487         if (ep && ep->re_connect_status != 0) {
 488                 delay = xprt_reconnect_delay(xprt);
 489                 xprt_reconnect_backoff(xprt, RPCRDMA_INIT_REEST_TO);
 490         }
 491         trace_xprtrdma_op_connect(r_xprt, delay);
 492         queue_delayed_work(xprtiod_workqueue, &r_xprt->rx_connect_worker,
 493                            delay);
 494 }
 495
 496 /**
 497  * xprt_rdma_alloc_slot - allocate an rpc_rqst
 498  * @xprt: controlling RPC transport
 499  * @task: RPC task requesting a fresh rpc_rqst
 500  *
 501  * tk_status values:
 502  *      %0 if task->tk_rqstp points to a fresh rpc_rqst
 503  *      %-EAGAIN if no rpc_rqst is available; queued on backlog
 504  */
 505 static void
 506 xprt_rdma_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
 507 {
 508         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 509         struct rpcrdma_req *req;
 510
 511         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
 512         if (!req)
 513                 goto out_sleep;
 514         task->tk_rqstp = &req->rl_slot;
 515         task->tk_status = 0;
 516         return;
 517
 518 out_sleep:
 519         set_bit(XPRT_CONGESTED, &xprt->state);
 520         rpc_sleep_on(&xprt->backlog, task, NULL);
 521         task->tk_status = -EAGAIN;
 522 }
 523
 524 /**
 525  * xprt_rdma_free_slot - release an rpc_rqst
 526  * @xprt: controlling RPC transport
 527  * @rqst: rpc_rqst to release
 528  *
 529  */
 530 static void
 531 xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst)
 532 {
 533         struct rpcrdma_xprt *r_xprt =
 534                 container_of(xprt, struct rpcrdma_xprt, rx_xprt);
 535
 536         memset(rqst, 0, sizeof(*rqst));
 537         rpcrdma_buffer_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst));
 538         if (unlikely(!rpc_wake_up_next(&xprt->backlog)))
 539                 clear_bit(XPRT_CONGESTED, &xprt->state);
 540 }
 541
 542 static bool rpcrdma_check_regbuf(struct rpcrdma_xprt *r_xprt,
 543                                  struct rpcrdma_regbuf *rb, size_t size,
 544                                  gfp_t flags)
 545 {
 546         if (unlikely(rdmab_length(rb) < size)) {
 547                 if (!rpcrdma_regbuf_realloc(rb, size, flags))
 548                         return false;
 549                 r_xprt->rx_stats.hardway_register_count += size;
 550         }
 551         return true;
 552 }
 553
 554 /**
 555  * xprt_rdma_allocate - allocate transport resources for an RPC
 556  * @task: RPC task
 557  *
 558  * Return values:
 559  *        0:    Success; rq_buffer points to RPC buffer to use
 560  *   ENOMEM:    Out of memory, call again later
 561  *      EIO:    A permanent error occurred, do not retry
 562  */
 563 static int
 564 xprt_rdma_allocate(struct rpc_task *task)
 565 {
 566         struct rpc_rqst *rqst = task->tk_rqstp;
 567         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 568         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 569         gfp_t flags;
 570
 571         flags = RPCRDMA_DEF_GFP;
 572         if (RPC_IS_SWAPPER(task))
 573                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
 574
 575         if (!rpcrdma_check_regbuf(r_xprt, req->rl_sendbuf, rqst->rq_callsize,
 576                                   flags))
 577                 goto out_fail;
 578         if (!rpcrdma_check_regbuf(r_xprt, req->rl_recvbuf, rqst->rq_rcvsize,
 579                                   flags))
 580                 goto out_fail;
 581
 582         rqst->rq_buffer = rdmab_data(req->rl_sendbuf);
 583         rqst->rq_rbuffer = rdmab_data(req->rl_recvbuf);
 584         trace_xprtrdma_op_allocate(task, req);
 585         return 0;
 586
 587 out_fail:
 588         trace_xprtrdma_op_allocate(task, NULL);
 589         return -ENOMEM;
 590 }
 591
 592 /**
 593  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
 594  * @task: RPC task
 595  *
 596  * Caller guarantees rqst->rq_buffer is non-NULL.
 597  */
 598 static void
 599 xprt_rdma_free(struct rpc_task *task)
 600 {
 601         struct rpc_rqst *rqst = task->tk_rqstp;
 602         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 603         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 604
 605         trace_xprtrdma_op_free(task, req);
 606
 607         if (!list_empty(&req->rl_registered))
 608                 frwr_unmap_sync(r_xprt, req);
 609
 610         /* XXX: If the RPC is completing because of a signal and
 611          * not because a reply was received, we ought to ensure
 612          * that the Send completion has fired, so that memory
 613          * involved with the Send is not still visible to the NIC.
 614          */
 615 }
 616
 617 /**
 618  * xprt_rdma_send_request - marshal and send an RPC request
 619  * @rqst: RPC message in rq_snd_buf
 620  *
 621  * Caller holds the transport's write lock.
 622  *
 623  * Returns:
 624  *      %0 if the RPC message has been sent
 625  *      %-ENOTCONN if the caller should reconnect and call again
 626  *      %-EAGAIN if the caller should call again
 627  *      %-ENOBUFS if the caller should call again after a delay
 628  *      %-EMSGSIZE if encoding ran out of buffer space. The request
 629  *              was not sent. Do not try to send this message again.
 630  *      %-EIO if an I/O error occurred. The request was not sent.
 631  *              Do not try to send this message again.
 632  */
 633 static int
 634 xprt_rdma_send_request(struct rpc_rqst *rqst)
 635 {
 636         struct rpc_xprt *xprt = rqst->rq_xprt;
 637         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 638         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 639         int rc = 0;
 640
 641 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
 642         if (unlikely(!rqst->rq_buffer))
 643                 return xprt_rdma_bc_send_reply(rqst);
 644 #endif  /* CONFIG_SUNRPC_BACKCHANNEL */
 645
 646         if (!xprt_connected(xprt))
 647                 return -ENOTCONN;
 648
 649         if (!xprt_request_get_cong(xprt, rqst))
 650                 return -EBADSLT;
 651
 652         rc = rpcrdma_marshal_req(r_xprt, rqst);
 653         if (rc < 0)
 654                 goto failed_marshal;
 655
 656         /* Must suppress retransmit to maintain credits */
 657         if (rqst->rq_connect_cookie == xprt->connect_cookie)
 658                 goto drop_connection;
 659         rqst->rq_xtime = ktime_get();
 660
 661         if (rpcrdma_post_sends(r_xprt, req))
 662                 goto drop_connection;
 663
 664         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
 665
 666         /* An RPC with no reply will throw off credit accounting,
 667          * so drop the connection to reset the credit grant.
 668          */
 669         if (!rpc_reply_expected(rqst->rq_task))
 670                 goto drop_connection;
 671         return 0;
 672
 673 failed_marshal:
 674         if (rc != -ENOTCONN)
 675                 return rc;
 676 drop_connection:
 677         xprt_rdma_close(xprt);
 678         return -ENOTCONN;
 679 }
 680
 681 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 682 {
 683         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 684         long idle_time = 0;
 685
 686         if (xprt_connected(xprt))
 687                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 688
 689         seq_puts(seq, "\txprt:\trdma ");
 690         seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
 691                    0,   /* need a local port? */
 692                    xprt->stat.bind_count,
 693                    xprt->stat.connect_count,
 694                    xprt->stat.connect_time / HZ,
 695                    idle_time,
 696                    xprt->stat.sends,
 697                    xprt->stat.recvs,
 698                    xprt->stat.bad_xids,
 699                    xprt->stat.req_u,
 700                    xprt->stat.bklog_u);
 701         seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
 702                    r_xprt->rx_stats.read_chunk_count,
 703                    r_xprt->rx_stats.write_chunk_count,
 704                    r_xprt->rx_stats.reply_chunk_count,
 705                    r_xprt->rx_stats.total_rdma_request,
 706                    r_xprt->rx_stats.total_rdma_reply,
 707                    r_xprt->rx_stats.pullup_copy_count,
 708                    r_xprt->rx_stats.fixup_copy_count,
 709                    r_xprt->rx_stats.hardway_register_count,
 710                    r_xprt->rx_stats.failed_marshal_count,
 711                    r_xprt->rx_stats.bad_reply_count,
 712                    r_xprt->rx_stats.nomsg_call_count);
 713         seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
 714                    r_xprt->rx_stats.mrs_recycled,
 715                    r_xprt->rx_stats.mrs_orphaned,
 716                    r_xprt->rx_stats.mrs_allocated,
 717                    r_xprt->rx_stats.local_inv_needed,
 718                    r_xprt->rx_stats.empty_sendctx_q,
 719                    r_xprt->rx_stats.reply_waits_for_send);
 720 }
 721
 722 static int
 723 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
 724 {
 725         return 0;
 726 }
 727
 728 static void
 729 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
 730 {
 731 }
 732
 733 /*
 734  * Plumbing for rpc transport switch and kernel module
 735  */
 736
 737 static const struct rpc_xprt_ops xprt_rdma_procs = {
 738         .reserve_xprt           = xprt_reserve_xprt_cong,
 739         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 740         .alloc_slot             = xprt_rdma_alloc_slot,
 741         .free_slot              = xprt_rdma_free_slot,
 742         .release_request        = xprt_release_rqst_cong,       /* ditto */
 743         .wait_for_reply_request = xprt_wait_for_reply_request_def, /* ditto */
 744         .timer                  = xprt_rdma_timer,
 745         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 746         .set_port               = xprt_rdma_set_port,
 747         .connect                = xprt_rdma_connect,
 748         .buf_alloc              = xprt_rdma_allocate,
 749         .buf_free               = xprt_rdma_free,
 750         .send_request           = xprt_rdma_send_request,
 751         .close                  = xprt_rdma_close,
 752         .destroy                = xprt_rdma_destroy,
 753         .set_connect_timeout    = xprt_rdma_set_connect_timeout,
 754         .print_stats            = xprt_rdma_print_stats,
 755         .enable_swap            = xprt_rdma_enable_swap,
 756         .disable_swap           = xprt_rdma_disable_swap,
 757         .inject_disconnect      = xprt_rdma_inject_disconnect,
 758 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
 759         .bc_setup               = xprt_rdma_bc_setup,
 760         .bc_maxpayload          = xprt_rdma_bc_maxpayload,
 761         .bc_num_slots           = xprt_rdma_bc_max_slots,
 762         .bc_free_rqst           = xprt_rdma_bc_free_rqst,
 763         .bc_destroy             = xprt_rdma_bc_destroy,
 764 #endif
 765 };
 766
 767 static struct xprt_class xprt_rdma = {
 768         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 769         .name                   = "rdma",
 770         .owner                  = THIS_MODULE,
 771         .ident                  = XPRT_TRANSPORT_RDMA,
 772         .setup                  = xprt_setup_rdma,
 773 };
 774
 775 void xprt_rdma_cleanup(void)
 776 {
 777 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 778         if (sunrpc_table_header) {
 779                 unregister_sysctl_table(sunrpc_table_header);
 780                 sunrpc_table_header = NULL;
 781         }
 782 #endif
 783
 784         xprt_unregister_transport(&xprt_rdma);
 785         xprt_unregister_transport(&xprt_rdma_bc);
 786 }
 787
 788 int xprt_rdma_init(void)
 789 {
 790         int rc;
 791
 792         rc = xprt_register_transport(&xprt_rdma);
 793         if (rc)
 794                 return rc;
 795
 796         rc = xprt_register_transport(&xprt_rdma_bc);
 797         if (rc) {
 798                 xprt_unregister_transport(&xprt_rdma);
 799                 return rc;
 800         }
 801
 802 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 803         if (!sunrpc_table_header)
 804                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 805 #endif
 806         return 0;
 807 }