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 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_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 244         xprt_clear_connecting(xprt);
 245         if (r_xprt->rx_ep.rep_connected > 0) {
 246                 if (!xprt_test_and_set_connected(xprt)) {
 247                         xprt->stat.connect_count++;
 248                         xprt->stat.connect_time += (long)jiffies -
 249                                                    xprt->stat.connect_start;
 250                         xprt_wake_pending_tasks(xprt, -EAGAIN);
 251                 }
 252         } else {
 253                 if (xprt_test_and_clear_connected(xprt))
 254                         xprt_wake_pending_tasks(xprt, rc);
 255         }
 256 }
 257
 258 /**
 259  * xprt_rdma_inject_disconnect - inject a connection fault
 260  * @xprt: transport context
 261  *
 262  * If @xprt is connected, disconnect it to simulate spurious connection
 263  * loss.
 264  */
 265 static void
 266 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
 267 {
 268         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 269
 270         trace_xprtrdma_op_inject_dsc(r_xprt);
 271         rdma_disconnect(r_xprt->rx_ia.ri_id);
 272 }
 273
 274 /**
 275  * xprt_rdma_destroy - Full tear down of transport
 276  * @xprt: doomed transport context
 277  *
 278  * Caller guarantees there will be no more calls to us with
 279  * this @xprt.
 280  */
 281 static void
 282 xprt_rdma_destroy(struct rpc_xprt *xprt)
 283 {
 284         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 285
 286         trace_xprtrdma_op_destroy(r_xprt);
 287
 288         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
 289
 290         rpcrdma_ep_destroy(r_xprt);
 291         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 292         rpcrdma_ia_close(&r_xprt->rx_ia);
 293
 294         xprt_rdma_free_addresses(xprt);
 295         xprt_free(xprt);
 296
 297         module_put(THIS_MODULE);
 298 }
 299
 300 /* 60 second timeout, no retries */
 301 static const struct rpc_timeout xprt_rdma_default_timeout = {
 302         .to_initval = 60 * HZ,
 303         .to_maxval = 60 * HZ,
 304 };
 305
 306 /**
 307  * xprt_setup_rdma - Set up transport to use RDMA
 308  *
 309  * @args: rpc transport arguments
 310  */
 311 static struct rpc_xprt *
 312 xprt_setup_rdma(struct xprt_create *args)
 313 {
 314         struct rpc_xprt *xprt;
 315         struct rpcrdma_xprt *new_xprt;
 316         struct sockaddr *sap;
 317         int rc;
 318
 319         if (args->addrlen > sizeof(xprt->addr))
 320                 return ERR_PTR(-EBADF);
 321
 322         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0, 0);
 323         if (!xprt)
 324                 return ERR_PTR(-ENOMEM);
 325
 326         xprt->timeout = &xprt_rdma_default_timeout;
 327         xprt->connect_timeout = xprt->timeout->to_initval;
 328         xprt->max_reconnect_timeout = xprt->timeout->to_maxval;
 329         xprt->bind_timeout = RPCRDMA_BIND_TO;
 330         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 331         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
 332
 333         xprt->resvport = 0;             /* privileged port not needed */
 334         xprt->ops = &xprt_rdma_procs;
 335
 336         /*
 337          * Set up RDMA-specific connect data.
 338          */
 339         sap = args->dstaddr;
 340
 341         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 342          * address, for any side protocols which peek at it */
 343         xprt->prot = IPPROTO_TCP;
 344         xprt->addrlen = args->addrlen;
 345         memcpy(&xprt->addr, sap, xprt->addrlen);
 346
 347         if (rpc_get_port(sap))
 348                 xprt_set_bound(xprt);
 349         xprt_rdma_format_addresses(xprt, sap);
 350
 351         new_xprt = rpcx_to_rdmax(xprt);
 352         rc = rpcrdma_ia_open(new_xprt);
 353         if (rc)
 354                 goto out1;
 355
 356         rc = rpcrdma_ep_create(new_xprt);
 357         if (rc)
 358                 goto out2;
 359
 360         rc = rpcrdma_buffer_create(new_xprt);
 361         if (rc)
 362                 goto out3;
 363
 364         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
 365                           xprt_rdma_connect_worker);
 366
 367         xprt->max_payload = frwr_maxpages(new_xprt);
 368         if (xprt->max_payload == 0)
 369                 goto out4;
 370         xprt->max_payload <<= PAGE_SHIFT;
 371         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
 372                 __func__, xprt->max_payload);
 373
 374         if (!try_module_get(THIS_MODULE))
 375                 goto out4;
 376
 377         dprintk("RPC:       %s: %s:%s\n", __func__,
 378                 xprt->address_strings[RPC_DISPLAY_ADDR],
 379                 xprt->address_strings[RPC_DISPLAY_PORT]);
 380         trace_xprtrdma_create(new_xprt);
 381         return xprt;
 382
 383 out4:
 384         rpcrdma_buffer_destroy(&new_xprt->rx_buf);
 385         rc = -ENODEV;
 386 out3:
 387         rpcrdma_ep_destroy(new_xprt);
 388 out2:
 389         rpcrdma_ia_close(&new_xprt->rx_ia);
 390 out1:
 391         trace_xprtrdma_op_destroy(new_xprt);
 392         xprt_rdma_free_addresses(xprt);
 393         xprt_free(xprt);
 394         return ERR_PTR(rc);
 395 }
 396
 397 /**
 398  * xprt_rdma_close - close a transport connection
 399  * @xprt: transport context
 400  *
 401  * Called during autoclose or device removal.
 402  *
 403  * Caller holds @xprt's send lock to prevent activity on this
 404  * transport while the connection is torn down.
 405  */
 406 void xprt_rdma_close(struct rpc_xprt *xprt)
 407 {
 408         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 409         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 410         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 411
 412         might_sleep();
 413
 414         trace_xprtrdma_op_close(r_xprt);
 415
 416         /* Prevent marshaling and sending of new requests */
 417         xprt_clear_connected(xprt);
 418
 419         if (test_and_clear_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags)) {
 420                 rpcrdma_ia_remove(ia);
 421                 goto out;
 422         }
 423
 424         if (ep->rep_connected == -ENODEV)
 425                 return;
 426         if (ep->rep_connected > 0)
 427                 xprt->reestablish_timeout = 0;
 428         rpcrdma_ep_disconnect(ep, ia);
 429
 430         /* Prepare @xprt for the next connection by reinitializing
 431          * its credit grant to one (see RFC 8166, Section 3.3.3).
 432          */
 433         r_xprt->rx_buf.rb_credits = 1;
 434         xprt->cwnd = RPC_CWNDSHIFT;
 435
 436 out:
 437         ++xprt->connect_cookie;
 438         xprt_disconnect_done(xprt);
 439 }
 440
 441 /**
 442  * xprt_rdma_set_port - update server port with rpcbind result
 443  * @xprt: controlling RPC transport
 444  * @port: new port value
 445  *
 446  * Transport connect status is unchanged.
 447  */
 448 static void
 449 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 450 {
 451         struct sockaddr *sap = (struct sockaddr *)&xprt->addr;
 452         char buf[8];
 453
 454         dprintk("RPC:       %s: setting port for xprt %p (%s:%s) to %u\n",
 455                 __func__, xprt,
 456                 xprt->address_strings[RPC_DISPLAY_ADDR],
 457                 xprt->address_strings[RPC_DISPLAY_PORT],
 458                 port);
 459
 460         rpc_set_port(sap, port);
 461
 462         kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
 463         snprintf(buf, sizeof(buf), "%u", port);
 464         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 465
 466         kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
 467         snprintf(buf, sizeof(buf), "%4hx", port);
 468         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 469 }
 470
 471 /**
 472  * xprt_rdma_timer - invoked when an RPC times out
 473  * @xprt: controlling RPC transport
 474  * @task: RPC task that timed out
 475  *
 476  * Invoked when the transport is still connected, but an RPC
 477  * retransmit timeout occurs.
 478  *
 479  * Since RDMA connections don't have a keep-alive, forcibly
 480  * disconnect and retry to connect. This drives full
 481  * detection of the network path, and retransmissions of
 482  * all pending RPCs.
 483  */
 484 static void
 485 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
 486 {
 487         xprt_force_disconnect(xprt);
 488 }
 489
 490 /**
 491  * xprt_rdma_set_connect_timeout - set timeouts for establishing a connection
 492  * @xprt: controlling transport instance
 493  * @connect_timeout: reconnect timeout after client disconnects
 494  * @reconnect_timeout: reconnect timeout after server disconnects
 495  *
 496  */
 497 static void xprt_rdma_tcp_set_connect_timeout(struct rpc_xprt *xprt,
 498                                               unsigned long connect_timeout,
 499                                               unsigned long reconnect_timeout)
 500 {
 501         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 502
 503         trace_xprtrdma_op_set_cto(r_xprt, connect_timeout, reconnect_timeout);
 504
 505         spin_lock(&xprt->transport_lock);
 506
 507         if (connect_timeout < xprt->connect_timeout) {
 508                 struct rpc_timeout to;
 509                 unsigned long initval;
 510
 511                 to = *xprt->timeout;
 512                 initval = connect_timeout;
 513                 if (initval < RPCRDMA_INIT_REEST_TO << 1)
 514                         initval = RPCRDMA_INIT_REEST_TO << 1;
 515                 to.to_initval = initval;
 516                 to.to_maxval = initval;
 517                 r_xprt->rx_timeout = to;
 518                 xprt->timeout = &r_xprt->rx_timeout;
 519                 xprt->connect_timeout = connect_timeout;
 520         }
 521
 522         if (reconnect_timeout < xprt->max_reconnect_timeout)
 523                 xprt->max_reconnect_timeout = reconnect_timeout;
 524
 525         spin_unlock(&xprt->transport_lock);
 526 }
 527
 528 /**
 529  * xprt_rdma_connect - schedule an attempt to reconnect
 530  * @xprt: transport state
 531  * @task: RPC scheduler context (unused)
 532  *
 533  */
 534 static void
 535 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
 536 {
 537         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 538         unsigned long delay;
 539
 540         trace_xprtrdma_op_connect(r_xprt);
 541
 542         delay = 0;
 543         if (r_xprt->rx_ep.rep_connected != 0) {
 544                 delay = xprt_reconnect_delay(xprt);
 545                 xprt_reconnect_backoff(xprt, RPCRDMA_INIT_REEST_TO);
 546         }
 547         queue_delayed_work(xprtiod_workqueue, &r_xprt->rx_connect_worker,
 548                            delay);
 549 }
 550
 551 /**
 552  * xprt_rdma_alloc_slot - allocate an rpc_rqst
 553  * @xprt: controlling RPC transport
 554  * @task: RPC task requesting a fresh rpc_rqst
 555  *
 556  * tk_status values:
 557  *      %0 if task->tk_rqstp points to a fresh rpc_rqst
 558  *      %-EAGAIN if no rpc_rqst is available; queued on backlog
 559  */
 560 static void
 561 xprt_rdma_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
 562 {
 563         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 564         struct rpcrdma_req *req;
 565
 566         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
 567         if (!req)
 568                 goto out_sleep;
 569         task->tk_rqstp = &req->rl_slot;
 570         task->tk_status = 0;
 571         return;
 572
 573 out_sleep:
 574         set_bit(XPRT_CONGESTED, &xprt->state);
 575         rpc_sleep_on(&xprt->backlog, task, NULL);
 576         task->tk_status = -EAGAIN;
 577 }
 578
 579 /**
 580  * xprt_rdma_free_slot - release an rpc_rqst
 581  * @xprt: controlling RPC transport
 582  * @rqst: rpc_rqst to release
 583  *
 584  */
 585 static void
 586 xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst)
 587 {
 588         struct rpcrdma_xprt *r_xprt =
 589                 container_of(xprt, struct rpcrdma_xprt, rx_xprt);
 590
 591         memset(rqst, 0, sizeof(*rqst));
 592         rpcrdma_buffer_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst));
 593         if (unlikely(!rpc_wake_up_next(&xprt->backlog)))
 594                 clear_bit(XPRT_CONGESTED, &xprt->state);
 595 }
 596
 597 static bool rpcrdma_check_regbuf(struct rpcrdma_xprt *r_xprt,
 598                                  struct rpcrdma_regbuf *rb, size_t size,
 599                                  gfp_t flags)
 600 {
 601         if (unlikely(rdmab_length(rb) < size)) {
 602                 if (!rpcrdma_regbuf_realloc(rb, size, flags))
 603                         return false;
 604                 r_xprt->rx_stats.hardway_register_count += size;
 605         }
 606         return true;
 607 }
 608
 609 /**
 610  * xprt_rdma_allocate - allocate transport resources for an RPC
 611  * @task: RPC task
 612  *
 613  * Return values:
 614  *        0:    Success; rq_buffer points to RPC buffer to use
 615  *   ENOMEM:    Out of memory, call again later
 616  *      EIO:    A permanent error occurred, do not retry
 617  */
 618 static int
 619 xprt_rdma_allocate(struct rpc_task *task)
 620 {
 621         struct rpc_rqst *rqst = task->tk_rqstp;
 622         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 623         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 624         gfp_t flags;
 625
 626         flags = RPCRDMA_DEF_GFP;
 627         if (RPC_IS_SWAPPER(task))
 628                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
 629
 630         if (!rpcrdma_check_regbuf(r_xprt, req->rl_sendbuf, rqst->rq_callsize,
 631                                   flags))
 632                 goto out_fail;
 633         if (!rpcrdma_check_regbuf(r_xprt, req->rl_recvbuf, rqst->rq_rcvsize,
 634                                   flags))
 635                 goto out_fail;
 636
 637         rqst->rq_buffer = rdmab_data(req->rl_sendbuf);
 638         rqst->rq_rbuffer = rdmab_data(req->rl_recvbuf);
 639         trace_xprtrdma_op_allocate(task, req);
 640         return 0;
 641
 642 out_fail:
 643         trace_xprtrdma_op_allocate(task, NULL);
 644         return -ENOMEM;
 645 }
 646
 647 /**
 648  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
 649  * @task: RPC task
 650  *
 651  * Caller guarantees rqst->rq_buffer is non-NULL.
 652  */
 653 static void
 654 xprt_rdma_free(struct rpc_task *task)
 655 {
 656         struct rpc_rqst *rqst = task->tk_rqstp;
 657         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 658         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 659
 660         trace_xprtrdma_op_free(task, req);
 661
 662         if (!list_empty(&req->rl_registered))
 663                 frwr_unmap_sync(r_xprt, req);
 664
 665         /* XXX: If the RPC is completing because of a signal and
 666          * not because a reply was received, we ought to ensure
 667          * that the Send completion has fired, so that memory
 668          * involved with the Send is not still visible to the NIC.
 669          */
 670 }
 671
 672 /**
 673  * xprt_rdma_send_request - marshal and send an RPC request
 674  * @rqst: RPC message in rq_snd_buf
 675  *
 676  * Caller holds the transport's write lock.
 677  *
 678  * Returns:
 679  *      %0 if the RPC message has been sent
 680  *      %-ENOTCONN if the caller should reconnect and call again
 681  *      %-EAGAIN if the caller should call again
 682  *      %-ENOBUFS if the caller should call again after a delay
 683  *      %-EMSGSIZE if encoding ran out of buffer space. The request
 684  *              was not sent. Do not try to send this message again.
 685  *      %-EIO if an I/O error occurred. The request was not sent.
 686  *              Do not try to send this message again.
 687  */
 688 static int
 689 xprt_rdma_send_request(struct rpc_rqst *rqst)
 690 {
 691         struct rpc_xprt *xprt = rqst->rq_xprt;
 692         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 693         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 694         int rc = 0;
 695
 696 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
 697         if (unlikely(!rqst->rq_buffer))
 698                 return xprt_rdma_bc_send_reply(rqst);
 699 #endif  /* CONFIG_SUNRPC_BACKCHANNEL */
 700
 701         if (!xprt_connected(xprt))
 702                 return -ENOTCONN;
 703
 704         if (!xprt_request_get_cong(xprt, rqst))
 705                 return -EBADSLT;
 706
 707         rc = rpcrdma_marshal_req(r_xprt, rqst);
 708         if (rc < 0)
 709                 goto failed_marshal;
 710
 711         /* Must suppress retransmit to maintain credits */
 712         if (rqst->rq_connect_cookie == xprt->connect_cookie)
 713                 goto drop_connection;
 714         rqst->rq_xtime = ktime_get();
 715
 716         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 717                 goto drop_connection;
 718
 719         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
 720
 721         /* An RPC with no reply will throw off credit accounting,
 722          * so drop the connection to reset the credit grant.
 723          */
 724         if (!rpc_reply_expected(rqst->rq_task))
 725                 goto drop_connection;
 726         return 0;
 727
 728 failed_marshal:
 729         if (rc != -ENOTCONN)
 730                 return rc;
 731 drop_connection:
 732         xprt_rdma_close(xprt);
 733         return -ENOTCONN;
 734 }
 735
 736 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 737 {
 738         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 739         long idle_time = 0;
 740
 741         if (xprt_connected(xprt))
 742                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 743
 744         seq_puts(seq, "\txprt:\trdma ");
 745         seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
 746                    0,   /* need a local port? */
 747                    xprt->stat.bind_count,
 748                    xprt->stat.connect_count,
 749                    xprt->stat.connect_time / HZ,
 750                    idle_time,
 751                    xprt->stat.sends,
 752                    xprt->stat.recvs,
 753                    xprt->stat.bad_xids,
 754                    xprt->stat.req_u,
 755                    xprt->stat.bklog_u);
 756         seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
 757                    r_xprt->rx_stats.read_chunk_count,
 758                    r_xprt->rx_stats.write_chunk_count,
 759                    r_xprt->rx_stats.reply_chunk_count,
 760                    r_xprt->rx_stats.total_rdma_request,
 761                    r_xprt->rx_stats.total_rdma_reply,
 762                    r_xprt->rx_stats.pullup_copy_count,
 763                    r_xprt->rx_stats.fixup_copy_count,
 764                    r_xprt->rx_stats.hardway_register_count,
 765                    r_xprt->rx_stats.failed_marshal_count,
 766                    r_xprt->rx_stats.bad_reply_count,
 767                    r_xprt->rx_stats.nomsg_call_count);
 768         seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
 769                    r_xprt->rx_stats.mrs_recycled,
 770                    r_xprt->rx_stats.mrs_orphaned,
 771                    r_xprt->rx_stats.mrs_allocated,
 772                    r_xprt->rx_stats.local_inv_needed,
 773                    r_xprt->rx_stats.empty_sendctx_q,
 774                    r_xprt->rx_stats.reply_waits_for_send);
 775 }
 776
 777 static int
 778 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
 779 {
 780         return 0;
 781 }
 782
 783 static void
 784 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
 785 {
 786 }
 787
 788 /*
 789  * Plumbing for rpc transport switch and kernel module
 790  */
 791
 792 static const struct rpc_xprt_ops xprt_rdma_procs = {
 793         .reserve_xprt           = xprt_reserve_xprt_cong,
 794         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 795         .alloc_slot             = xprt_rdma_alloc_slot,
 796         .free_slot              = xprt_rdma_free_slot,
 797         .release_request        = xprt_release_rqst_cong,       /* ditto */
 798         .wait_for_reply_request = xprt_wait_for_reply_request_def, /* ditto */
 799         .timer                  = xprt_rdma_timer,
 800         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 801         .set_port               = xprt_rdma_set_port,
 802         .connect                = xprt_rdma_connect,
 803         .buf_alloc              = xprt_rdma_allocate,
 804         .buf_free               = xprt_rdma_free,
 805         .send_request           = xprt_rdma_send_request,
 806         .close                  = xprt_rdma_close,
 807         .destroy                = xprt_rdma_destroy,
 808         .set_connect_timeout    = xprt_rdma_tcp_set_connect_timeout,
 809         .print_stats            = xprt_rdma_print_stats,
 810         .enable_swap            = xprt_rdma_enable_swap,
 811         .disable_swap           = xprt_rdma_disable_swap,
 812         .inject_disconnect      = xprt_rdma_inject_disconnect,
 813 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
 814         .bc_setup               = xprt_rdma_bc_setup,
 815         .bc_maxpayload          = xprt_rdma_bc_maxpayload,
 816         .bc_num_slots           = xprt_rdma_bc_max_slots,
 817         .bc_free_rqst           = xprt_rdma_bc_free_rqst,
 818         .bc_destroy             = xprt_rdma_bc_destroy,
 819 #endif
 820 };
 821
 822 static struct xprt_class xprt_rdma = {
 823         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 824         .name                   = "rdma",
 825         .owner                  = THIS_MODULE,
 826         .ident                  = XPRT_TRANSPORT_RDMA,
 827         .setup                  = xprt_setup_rdma,
 828 };
 829
 830 void xprt_rdma_cleanup(void)
 831 {
 832 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 833         if (sunrpc_table_header) {
 834                 unregister_sysctl_table(sunrpc_table_header);
 835                 sunrpc_table_header = NULL;
 836         }
 837 #endif
 838
 839         xprt_unregister_transport(&xprt_rdma);
 840         xprt_unregister_transport(&xprt_rdma_bc);
 841 }
 842
 843 int xprt_rdma_init(void)
 844 {
 845         int rc;
 846
 847         rc = xprt_register_transport(&xprt_rdma);
 848         if (rc)
 849                 return rc;
 850
 851         rc = xprt_register_transport(&xprt_rdma_bc);
 852         if (rc) {
 853                 xprt_unregister_transport(&xprt_rdma);
 854                 return rc;
 855         }
 856
 857 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 858         if (!sunrpc_table_header)
 859                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 860 #endif
 861         return 0;
 862 }