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 static 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 zero;
  84 static unsigned int max_padding = PAGE_SIZE;
  85 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
  86 static unsigned int max_memreg = RPCRDMA_LAST - 1;
  87 static unsigned int dummy;
  88
  89 static struct ctl_table_header *sunrpc_table_header;
  90
  91 static struct ctl_table xr_tunables_table[] = {
  92         {
  93                 .procname       = "rdma_slot_table_entries",
  94                 .data           = &xprt_rdma_slot_table_entries,
  95                 .maxlen         = sizeof(unsigned int),
  96                 .mode           = 0644,
  97                 .proc_handler   = proc_dointvec_minmax,
  98                 .extra1         = &min_slot_table_size,
  99                 .extra2         = &max_slot_table_size
 100         },
 101         {
 102                 .procname       = "rdma_max_inline_read",
 103                 .data           = &xprt_rdma_max_inline_read,
 104                 .maxlen         = sizeof(unsigned int),
 105                 .mode           = 0644,
 106                 .proc_handler   = proc_dointvec_minmax,
 107                 .extra1         = &min_inline_size,
 108                 .extra2         = &max_inline_size,
 109         },
 110         {
 111                 .procname       = "rdma_max_inline_write",
 112                 .data           = &xprt_rdma_max_inline_write,
 113                 .maxlen         = sizeof(unsigned int),
 114                 .mode           = 0644,
 115                 .proc_handler   = proc_dointvec_minmax,
 116                 .extra1         = &min_inline_size,
 117                 .extra2         = &max_inline_size,
 118         },
 119         {
 120                 .procname       = "rdma_inline_write_padding",
 121                 .data           = &dummy,
 122                 .maxlen         = sizeof(unsigned int),
 123                 .mode           = 0644,
 124                 .proc_handler   = proc_dointvec_minmax,
 125                 .extra1         = &zero,
 126                 .extra2         = &max_padding,
 127         },
 128         {
 129                 .procname       = "rdma_memreg_strategy",
 130                 .data           = &xprt_rdma_memreg_strategy,
 131                 .maxlen         = sizeof(unsigned int),
 132                 .mode           = 0644,
 133                 .proc_handler   = proc_dointvec_minmax,
 134                 .extra1         = &min_memreg,
 135                 .extra2         = &max_memreg,
 136         },
 137         {
 138                 .procname       = "rdma_pad_optimize",
 139                 .data           = &xprt_rdma_pad_optimize,
 140                 .maxlen         = sizeof(unsigned int),
 141                 .mode           = 0644,
 142                 .proc_handler   = proc_dointvec,
 143         },
 144         { },
 145 };
 146
 147 static struct ctl_table sunrpc_table[] = {
 148         {
 149                 .procname       = "sunrpc",
 150                 .mode           = 0555,
 151                 .child          = xr_tunables_table
 152         },
 153         { },
 154 };
 155
 156 #endif
 157
 158 static const struct rpc_xprt_ops xprt_rdma_procs;
 159
 160 static void
 161 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
 162 {
 163         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
 164         char buf[20];
 165
 166         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
 167         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 168
 169         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
 170 }
 171
 172 static void
 173 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
 174 {
 175         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
 176         char buf[40];
 177
 178         snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
 179         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 180
 181         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
 182 }
 183
 184 void
 185 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
 186 {
 187         char buf[128];
 188
 189         switch (sap->sa_family) {
 190         case AF_INET:
 191                 xprt_rdma_format_addresses4(xprt, sap);
 192                 break;
 193         case AF_INET6:
 194                 xprt_rdma_format_addresses6(xprt, sap);
 195                 break;
 196         default:
 197                 pr_err("rpcrdma: Unrecognized address family\n");
 198                 return;
 199         }
 200
 201         (void)rpc_ntop(sap, buf, sizeof(buf));
 202         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
 203
 204         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
 205         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 206
 207         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
 208         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 209
 210         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
 211 }
 212
 213 void
 214 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
 215 {
 216         unsigned int i;
 217
 218         for (i = 0; i < RPC_DISPLAY_MAX; i++)
 219                 switch (i) {
 220                 case RPC_DISPLAY_PROTO:
 221                 case RPC_DISPLAY_NETID:
 222                         continue;
 223                 default:
 224                         kfree(xprt->address_strings[i]);
 225                 }
 226 }
 227
 228 /**
 229  * xprt_rdma_connect_worker - establish connection in the background
 230  * @work: worker thread context
 231  *
 232  * Requester holds the xprt's send lock to prevent activity on this
 233  * transport while a fresh connection is being established. RPC tasks
 234  * sleep on the xprt's pending queue waiting for connect to complete.
 235  */
 236 static void
 237 xprt_rdma_connect_worker(struct work_struct *work)
 238 {
 239         struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
 240                                                    rx_connect_worker.work);
 241         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 242         int rc;
 243
 244         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 245         xprt_clear_connecting(xprt);
 246         if (r_xprt->rx_ep.rep_connected > 0) {
 247                 if (!xprt_test_and_set_connected(xprt)) {
 248                         xprt->stat.connect_count++;
 249                         xprt->stat.connect_time += (long)jiffies -
 250                                                    xprt->stat.connect_start;
 251                         xprt_wake_pending_tasks(xprt, -EAGAIN);
 252                 }
 253         } else {
 254                 if (xprt_test_and_clear_connected(xprt))
 255                         xprt_wake_pending_tasks(xprt, rc);
 256         }
 257 }
 258
 259 /**
 260  * xprt_rdma_inject_disconnect - inject a connection fault
 261  * @xprt: transport context
 262  *
 263  * If @xprt is connected, disconnect it to simulate spurious connection
 264  * loss.
 265  */
 266 static void
 267 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
 268 {
 269         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 270
 271         trace_xprtrdma_inject_dsc(r_xprt);
 272         rdma_disconnect(r_xprt->rx_ia.ri_id);
 273 }
 274
 275 /**
 276  * xprt_rdma_destroy - Full tear down of transport
 277  * @xprt: doomed transport context
 278  *
 279  * Caller guarantees there will be no more calls to us with
 280  * this @xprt.
 281  */
 282 static void
 283 xprt_rdma_destroy(struct rpc_xprt *xprt)
 284 {
 285         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 286
 287         trace_xprtrdma_destroy(r_xprt);
 288
 289         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
 290
 291         rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
 292         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 293         rpcrdma_ia_close(&r_xprt->rx_ia);
 294
 295         xprt_rdma_free_addresses(xprt);
 296         xprt_free(xprt);
 297
 298         module_put(THIS_MODULE);
 299 }
 300
 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 rpcrdma_create_data_internal cdata;
 315         struct rpc_xprt *xprt;
 316         struct rpcrdma_xprt *new_xprt;
 317         struct rpcrdma_ep *new_ep;
 318         struct sockaddr *sap;
 319         int rc;
 320
 321         if (args->addrlen > sizeof(xprt->addr)) {
 322                 dprintk("RPC:       %s: address too large\n", __func__);
 323                 return ERR_PTR(-EBADF);
 324         }
 325
 326         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0, 0);
 327         if (xprt == NULL) {
 328                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
 329                         __func__);
 330                 return ERR_PTR(-ENOMEM);
 331         }
 332
 333         /* 60 second timeout, no retries */
 334         xprt->timeout = &xprt_rdma_default_timeout;
 335         xprt->bind_timeout = RPCRDMA_BIND_TO;
 336         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 337         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
 338
 339         xprt->resvport = 0;             /* privileged port not needed */
 340         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
 341         xprt->ops = &xprt_rdma_procs;
 342
 343         /*
 344          * Set up RDMA-specific connect data.
 345          */
 346         sap = args->dstaddr;
 347
 348         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 349          * address, for any side protocols which peek at it */
 350         xprt->prot = IPPROTO_TCP;
 351         xprt->addrlen = args->addrlen;
 352         memcpy(&xprt->addr, sap, xprt->addrlen);
 353
 354         if (rpc_get_port(sap))
 355                 xprt_set_bound(xprt);
 356         xprt_rdma_format_addresses(xprt, sap);
 357
 358         cdata.max_requests = xprt_rdma_slot_table_entries;
 359
 360         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
 361         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
 362
 363         cdata.inline_wsize = xprt_rdma_max_inline_write;
 364         if (cdata.inline_wsize > cdata.wsize)
 365                 cdata.inline_wsize = cdata.wsize;
 366
 367         cdata.inline_rsize = xprt_rdma_max_inline_read;
 368         if (cdata.inline_rsize > cdata.rsize)
 369                 cdata.inline_rsize = cdata.rsize;
 370
 371         /*
 372          * Create new transport instance, which includes initialized
 373          *  o ia
 374          *  o endpoint
 375          *  o buffers
 376          */
 377
 378         new_xprt = rpcx_to_rdmax(xprt);
 379
 380         rc = rpcrdma_ia_open(new_xprt);
 381         if (rc)
 382                 goto out1;
 383
 384         /*
 385          * initialize and create ep
 386          */
 387         new_xprt->rx_data = cdata;
 388         new_ep = &new_xprt->rx_ep;
 389
 390         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
 391                                 &new_xprt->rx_ia, &new_xprt->rx_data);
 392         if (rc)
 393                 goto out2;
 394
 395         rc = rpcrdma_buffer_create(new_xprt);
 396         if (rc)
 397                 goto out3;
 398
 399         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
 400                           xprt_rdma_connect_worker);
 401
 402         xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
 403         if (xprt->max_payload == 0)
 404                 goto out4;
 405         xprt->max_payload <<= PAGE_SHIFT;
 406         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
 407                 __func__, xprt->max_payload);
 408
 409         if (!try_module_get(THIS_MODULE))
 410                 goto out4;
 411
 412         dprintk("RPC:       %s: %s:%s\n", __func__,
 413                 xprt->address_strings[RPC_DISPLAY_ADDR],
 414                 xprt->address_strings[RPC_DISPLAY_PORT]);
 415         trace_xprtrdma_create(new_xprt);
 416         return xprt;
 417
 418 out4:
 419         rpcrdma_buffer_destroy(&new_xprt->rx_buf);
 420         rc = -ENODEV;
 421 out3:
 422         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
 423 out2:
 424         rpcrdma_ia_close(&new_xprt->rx_ia);
 425 out1:
 426         trace_xprtrdma_destroy(new_xprt);
 427         xprt_rdma_free_addresses(xprt);
 428         xprt_free(xprt);
 429         return ERR_PTR(rc);
 430 }
 431
 432 /**
 433  * xprt_rdma_close - close a transport connection
 434  * @xprt: transport context
 435  *
 436  * Called during transport shutdown, reconnect, or device removal.
 437  * Caller holds @xprt's send lock to prevent activity on this
 438  * transport while the connection is torn down.
 439  */
 440 void xprt_rdma_close(struct rpc_xprt *xprt)
 441 {
 442         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 443         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 444         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 445
 446         might_sleep();
 447
 448         dprintk("RPC:       %s: closing xprt %p\n", __func__, xprt);
 449
 450         /* Prevent marshaling and sending of new requests */
 451         xprt_clear_connected(xprt);
 452
 453         if (test_and_clear_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags)) {
 454                 rpcrdma_ia_remove(ia);
 455                 goto out;
 456         }
 457
 458         if (ep->rep_connected == -ENODEV)
 459                 return;
 460         if (ep->rep_connected > 0)
 461                 xprt->reestablish_timeout = 0;
 462         rpcrdma_ep_disconnect(ep, ia);
 463
 464         /* Prepare @xprt for the next connection by reinitializing
 465          * its credit grant to one (see RFC 8166, Section 3.3.3).
 466          */
 467         r_xprt->rx_buf.rb_credits = 1;
 468         xprt->cwnd = RPC_CWNDSHIFT;
 469
 470 out:
 471         ++xprt->connect_cookie;
 472         xprt_disconnect_done(xprt);
 473 }
 474
 475 /**
 476  * xprt_rdma_set_port - update server port with rpcbind result
 477  * @xprt: controlling RPC transport
 478  * @port: new port value
 479  *
 480  * Transport connect status is unchanged.
 481  */
 482 static void
 483 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 484 {
 485         struct sockaddr *sap = (struct sockaddr *)&xprt->addr;
 486         char buf[8];
 487
 488         dprintk("RPC:       %s: setting port for xprt %p (%s:%s) to %u\n",
 489                 __func__, xprt,
 490                 xprt->address_strings[RPC_DISPLAY_ADDR],
 491                 xprt->address_strings[RPC_DISPLAY_PORT],
 492                 port);
 493
 494         rpc_set_port(sap, port);
 495
 496         kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
 497         snprintf(buf, sizeof(buf), "%u", port);
 498         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 499
 500         kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
 501         snprintf(buf, sizeof(buf), "%4hx", port);
 502         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 503 }
 504
 505 /**
 506  * xprt_rdma_timer - invoked when an RPC times out
 507  * @xprt: controlling RPC transport
 508  * @task: RPC task that timed out
 509  *
 510  * Invoked when the transport is still connected, but an RPC
 511  * retransmit timeout occurs.
 512  *
 513  * Since RDMA connections don't have a keep-alive, forcibly
 514  * disconnect and retry to connect. This drives full
 515  * detection of the network path, and retransmissions of
 516  * all pending RPCs.
 517  */
 518 static void
 519 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
 520 {
 521         xprt_force_disconnect(xprt);
 522 }
 523
 524 /**
 525  * xprt_rdma_connect - try to establish a transport connection
 526  * @xprt: transport state
 527  * @task: RPC scheduler context
 528  *
 529  */
 530 static void
 531 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
 532 {
 533         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 534
 535         if (r_xprt->rx_ep.rep_connected != 0) {
 536                 /* Reconnect */
 537                 schedule_delayed_work(&r_xprt->rx_connect_worker,
 538                                       xprt->reestablish_timeout);
 539                 xprt->reestablish_timeout <<= 1;
 540                 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
 541                         xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
 542                 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
 543                         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 544         } else {
 545                 schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
 546                 if (!RPC_IS_ASYNC(task))
 547                         flush_delayed_work(&r_xprt->rx_connect_worker);
 548         }
 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         rpc_sleep_on(&xprt->backlog, task, NULL);
 575         task->tk_status = -EAGAIN;
 576 }
 577
 578 /**
 579  * xprt_rdma_free_slot - release an rpc_rqst
 580  * @xprt: controlling RPC transport
 581  * @rqst: rpc_rqst to release
 582  *
 583  */
 584 static void
 585 xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst)
 586 {
 587         memset(rqst, 0, sizeof(*rqst));
 588         rpcrdma_buffer_put(rpcr_to_rdmar(rqst));
 589         rpc_wake_up_next(&xprt->backlog);
 590 }
 591
 592 static bool
 593 rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 594                     size_t size, gfp_t flags)
 595 {
 596         struct rpcrdma_regbuf *rb;
 597
 598         if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
 599                 return true;
 600
 601         rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
 602         if (IS_ERR(rb))
 603                 return false;
 604
 605         rpcrdma_free_regbuf(req->rl_sendbuf);
 606         r_xprt->rx_stats.hardway_register_count += size;
 607         req->rl_sendbuf = rb;
 608         return true;
 609 }
 610
 611 /* The rq_rcv_buf is used only if a Reply chunk is necessary.
 612  * The decision to use a Reply chunk is made later in
 613  * rpcrdma_marshal_req. This buffer is registered at that time.
 614  *
 615  * Otherwise, the associated RPC Reply arrives in a separate
 616  * Receive buffer, arbitrarily chosen by the HCA. The buffer
 617  * allocated here for the RPC Reply is not utilized in that
 618  * case. See rpcrdma_inline_fixup.
 619  *
 620  * A regbuf is used here to remember the buffer size.
 621  */
 622 static bool
 623 rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 624                     size_t size, gfp_t flags)
 625 {
 626         struct rpcrdma_regbuf *rb;
 627
 628         if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
 629                 return true;
 630
 631         rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
 632         if (IS_ERR(rb))
 633                 return false;
 634
 635         rpcrdma_free_regbuf(req->rl_recvbuf);
 636         r_xprt->rx_stats.hardway_register_count += size;
 637         req->rl_recvbuf = rb;
 638         return true;
 639 }
 640
 641 /**
 642  * xprt_rdma_allocate - allocate transport resources for an RPC
 643  * @task: RPC task
 644  *
 645  * Return values:
 646  *        0:    Success; rq_buffer points to RPC buffer to use
 647  *   ENOMEM:    Out of memory, call again later
 648  *      EIO:    A permanent error occurred, do not retry
 649  */
 650 static int
 651 xprt_rdma_allocate(struct rpc_task *task)
 652 {
 653         struct rpc_rqst *rqst = task->tk_rqstp;
 654         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 655         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 656         gfp_t flags;
 657
 658         flags = RPCRDMA_DEF_GFP;
 659         if (RPC_IS_SWAPPER(task))
 660                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
 661
 662         if (!rpcrdma_get_sendbuf(r_xprt, req, rqst->rq_callsize, flags))
 663                 goto out_fail;
 664         if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
 665                 goto out_fail;
 666
 667         rqst->rq_buffer = req->rl_sendbuf->rg_base;
 668         rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
 669         trace_xprtrdma_allocate(task, req);
 670         return 0;
 671
 672 out_fail:
 673         trace_xprtrdma_allocate(task, NULL);
 674         return -ENOMEM;
 675 }
 676
 677 /**
 678  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
 679  * @task: RPC task
 680  *
 681  * Caller guarantees rqst->rq_buffer is non-NULL.
 682  */
 683 static void
 684 xprt_rdma_free(struct rpc_task *task)
 685 {
 686         struct rpc_rqst *rqst = task->tk_rqstp;
 687         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 688         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 689
 690         if (test_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags))
 691                 rpcrdma_release_rqst(r_xprt, req);
 692         trace_xprtrdma_rpc_done(task, req);
 693 }
 694
 695 /**
 696  * xprt_rdma_send_request - marshal and send an RPC request
 697  * @rqst: RPC message in rq_snd_buf
 698  *
 699  * Caller holds the transport's write lock.
 700  *
 701  * Returns:
 702  *      %0 if the RPC message has been sent
 703  *      %-ENOTCONN if the caller should reconnect and call again
 704  *      %-EAGAIN if the caller should call again
 705  *      %-ENOBUFS if the caller should call again after a delay
 706  *      %-EIO if a permanent error occurred and the request was not
 707  *              sent. Do not try to send this message again.
 708  */
 709 static int
 710 xprt_rdma_send_request(struct rpc_rqst *rqst)
 711 {
 712         struct rpc_xprt *xprt = rqst->rq_xprt;
 713         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 714         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 715         int rc = 0;
 716
 717 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
 718         if (unlikely(!rqst->rq_buffer))
 719                 return xprt_rdma_bc_send_reply(rqst);
 720 #endif  /* CONFIG_SUNRPC_BACKCHANNEL */
 721
 722         if (!xprt_connected(xprt))
 723                 return -ENOTCONN;
 724
 725         if (!xprt_request_get_cong(xprt, rqst))
 726                 return -EBADSLT;
 727
 728         rc = rpcrdma_marshal_req(r_xprt, rqst);
 729         if (rc < 0)
 730                 goto failed_marshal;
 731
 732         /* Must suppress retransmit to maintain credits */
 733         if (rqst->rq_connect_cookie == xprt->connect_cookie)
 734                 goto drop_connection;
 735         rqst->rq_xtime = ktime_get();
 736
 737         __set_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags);
 738         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 739                 goto drop_connection;
 740
 741         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
 742         rqst->rq_bytes_sent = 0;
 743
 744         /* An RPC with no reply will throw off credit accounting,
 745          * so drop the connection to reset the credit grant.
 746          */
 747         if (!rpc_reply_expected(rqst->rq_task))
 748                 goto drop_connection;
 749         return 0;
 750
 751 failed_marshal:
 752         if (rc != -ENOTCONN)
 753                 return rc;
 754 drop_connection:
 755         xprt_rdma_close(xprt);
 756         return -ENOTCONN;
 757 }
 758
 759 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 760 {
 761         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 762         long idle_time = 0;
 763
 764         if (xprt_connected(xprt))
 765                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 766
 767         seq_puts(seq, "\txprt:\trdma ");
 768         seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
 769                    0,   /* need a local port? */
 770                    xprt->stat.bind_count,
 771                    xprt->stat.connect_count,
 772                    xprt->stat.connect_time / HZ,
 773                    idle_time,
 774                    xprt->stat.sends,
 775                    xprt->stat.recvs,
 776                    xprt->stat.bad_xids,
 777                    xprt->stat.req_u,
 778                    xprt->stat.bklog_u);
 779         seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
 780                    r_xprt->rx_stats.read_chunk_count,
 781                    r_xprt->rx_stats.write_chunk_count,
 782                    r_xprt->rx_stats.reply_chunk_count,
 783                    r_xprt->rx_stats.total_rdma_request,
 784                    r_xprt->rx_stats.total_rdma_reply,
 785                    r_xprt->rx_stats.pullup_copy_count,
 786                    r_xprt->rx_stats.fixup_copy_count,
 787                    r_xprt->rx_stats.hardway_register_count,
 788                    r_xprt->rx_stats.failed_marshal_count,
 789                    r_xprt->rx_stats.bad_reply_count,
 790                    r_xprt->rx_stats.nomsg_call_count);
 791         seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
 792                    r_xprt->rx_stats.mrs_recycled,
 793                    r_xprt->rx_stats.mrs_orphaned,
 794                    r_xprt->rx_stats.mrs_allocated,
 795                    r_xprt->rx_stats.local_inv_needed,
 796                    r_xprt->rx_stats.empty_sendctx_q,
 797                    r_xprt->rx_stats.reply_waits_for_send);
 798 }
 799
 800 static int
 801 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
 802 {
 803         return 0;
 804 }
 805
 806 static void
 807 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
 808 {
 809 }
 810
 811 /*
 812  * Plumbing for rpc transport switch and kernel module
 813  */
 814
 815 static const struct rpc_xprt_ops xprt_rdma_procs = {
 816         .reserve_xprt           = xprt_reserve_xprt_cong,
 817         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 818         .alloc_slot             = xprt_rdma_alloc_slot,
 819         .free_slot              = xprt_rdma_free_slot,
 820         .release_request        = xprt_release_rqst_cong,       /* ditto */
 821         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
 822         .timer                  = xprt_rdma_timer,
 823         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 824         .set_port               = xprt_rdma_set_port,
 825         .connect                = xprt_rdma_connect,
 826         .buf_alloc              = xprt_rdma_allocate,
 827         .buf_free               = xprt_rdma_free,
 828         .send_request           = xprt_rdma_send_request,
 829         .close                  = xprt_rdma_close,
 830         .destroy                = xprt_rdma_destroy,
 831         .print_stats            = xprt_rdma_print_stats,
 832         .enable_swap            = xprt_rdma_enable_swap,
 833         .disable_swap           = xprt_rdma_disable_swap,
 834         .inject_disconnect      = xprt_rdma_inject_disconnect,
 835 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
 836         .bc_setup               = xprt_rdma_bc_setup,
 837         .bc_up                  = xprt_rdma_bc_up,
 838         .bc_maxpayload          = xprt_rdma_bc_maxpayload,
 839         .bc_free_rqst           = xprt_rdma_bc_free_rqst,
 840         .bc_destroy             = xprt_rdma_bc_destroy,
 841 #endif
 842 };
 843
 844 static struct xprt_class xprt_rdma = {
 845         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 846         .name                   = "rdma",
 847         .owner                  = THIS_MODULE,
 848         .ident                  = XPRT_TRANSPORT_RDMA,
 849         .setup                  = xprt_setup_rdma,
 850 };
 851
 852 void xprt_rdma_cleanup(void)
 853 {
 854         int rc;
 855
 856         dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
 857 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 858         if (sunrpc_table_header) {
 859                 unregister_sysctl_table(sunrpc_table_header);
 860                 sunrpc_table_header = NULL;
 861         }
 862 #endif
 863         rc = xprt_unregister_transport(&xprt_rdma);
 864         if (rc)
 865                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
 866                         __func__, rc);
 867
 868         rc = xprt_unregister_transport(&xprt_rdma_bc);
 869         if (rc)
 870                 dprintk("RPC:       %s: xprt_unregister(bc) returned %i\n",
 871                         __func__, rc);
 872 }
 873
 874 int xprt_rdma_init(void)
 875 {
 876         int rc;
 877
 878         rc = xprt_register_transport(&xprt_rdma);
 879         if (rc)
 880                 return rc;
 881
 882         rc = xprt_register_transport(&xprt_rdma_bc);
 883         if (rc) {
 884                 xprt_unregister_transport(&xprt_rdma);
 885                 return rc;
 886         }
 887
 888         dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
 889
 890         dprintk("Defaults:\n");
 891         dprintk("\tSlots %d\n"
 892                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
 893                 xprt_rdma_slot_table_entries,
 894                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
 895         dprintk("\tPadding 0\n\tMemreg %d\n", xprt_rdma_memreg_strategy);
 896
 897 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 898         if (!sunrpc_table_header)
 899                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 900 #endif
 901         return 0;
 902 }