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