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 void
 229 rpcrdma_conn_func(struct rpcrdma_ep *ep)
 230 {
 231         schedule_delayed_work(&ep->rep_connect_worker, 0);
 232 }
 233
 234 void
 235 rpcrdma_connect_worker(struct work_struct *work)
 236 {
 237         struct rpcrdma_ep *ep =
 238                 container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
 239         struct rpcrdma_xprt *r_xprt =
 240                 container_of(ep, struct rpcrdma_xprt, rx_ep);
 241         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 242
 243         spin_lock_bh(&xprt->transport_lock);
 244         if (ep->rep_connected > 0) {
 245                 if (!xprt_test_and_set_connected(xprt))
 246                         xprt_wake_pending_tasks(xprt, 0);
 247         } else {
 248                 if (xprt_test_and_clear_connected(xprt))
 249                         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 250         }
 251         spin_unlock_bh(&xprt->transport_lock);
 252 }
 253
 254 static void
 255 xprt_rdma_connect_worker(struct work_struct *work)
 256 {
 257         struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
 258                                                    rx_connect_worker.work);
 259         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 260         int rc = 0;
 261
 262         xprt_clear_connected(xprt);
 263
 264         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 265         if (rc)
 266                 xprt_wake_pending_tasks(xprt, rc);
 267
 268         xprt_clear_connecting(xprt);
 269 }
 270
 271 static void
 272 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
 273 {
 274         struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
 275                                                    rx_xprt);
 276
 277         trace_xprtrdma_inject_dsc(r_xprt);
 278         rdma_disconnect(r_xprt->rx_ia.ri_id);
 279 }
 280
 281 /*
 282  * xprt_rdma_destroy
 283  *
 284  * Destroy the xprt.
 285  * Free all memory associated with the object, including its own.
 286  * NOTE: none of the *destroy methods free memory for their top-level
 287  * objects, even though they may have allocated it (they do free
 288  * private memory). It's up to the caller to handle it. In this
 289  * case (RDMA transport), all structure memory is inlined with the
 290  * struct rpcrdma_xprt.
 291  */
 292 static void
 293 xprt_rdma_destroy(struct rpc_xprt *xprt)
 294 {
 295         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 296
 297         trace_xprtrdma_destroy(r_xprt);
 298
 299         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
 300
 301         xprt_clear_connected(xprt);
 302
 303         rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
 304         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 305         rpcrdma_ia_close(&r_xprt->rx_ia);
 306
 307         xprt_rdma_free_addresses(xprt);
 308         xprt_free(xprt);
 309
 310         module_put(THIS_MODULE);
 311 }
 312
 313 static const struct rpc_timeout xprt_rdma_default_timeout = {
 314         .to_initval = 60 * HZ,
 315         .to_maxval = 60 * HZ,
 316 };
 317
 318 /**
 319  * xprt_setup_rdma - Set up transport to use RDMA
 320  *
 321  * @args: rpc transport arguments
 322  */
 323 static struct rpc_xprt *
 324 xprt_setup_rdma(struct xprt_create *args)
 325 {
 326         struct rpcrdma_create_data_internal cdata;
 327         struct rpc_xprt *xprt;
 328         struct rpcrdma_xprt *new_xprt;
 329         struct rpcrdma_ep *new_ep;
 330         struct sockaddr *sap;
 331         int rc;
 332
 333         if (args->addrlen > sizeof(xprt->addr)) {
 334                 dprintk("RPC:       %s: address too large\n", __func__);
 335                 return ERR_PTR(-EBADF);
 336         }
 337
 338         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0, 0);
 339         if (xprt == NULL) {
 340                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
 341                         __func__);
 342                 return ERR_PTR(-ENOMEM);
 343         }
 344
 345         /* 60 second timeout, no retries */
 346         xprt->timeout = &xprt_rdma_default_timeout;
 347         xprt->bind_timeout = RPCRDMA_BIND_TO;
 348         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 349         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
 350
 351         xprt->resvport = 0;             /* privileged port not needed */
 352         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
 353         xprt->ops = &xprt_rdma_procs;
 354
 355         /*
 356          * Set up RDMA-specific connect data.
 357          */
 358         sap = args->dstaddr;
 359
 360         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 361          * address, for any side protocols which peek at it */
 362         xprt->prot = IPPROTO_TCP;
 363         xprt->addrlen = args->addrlen;
 364         memcpy(&xprt->addr, sap, xprt->addrlen);
 365
 366         if (rpc_get_port(sap))
 367                 xprt_set_bound(xprt);
 368         xprt_rdma_format_addresses(xprt, sap);
 369
 370         cdata.max_requests = xprt_rdma_slot_table_entries;
 371
 372         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
 373         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
 374
 375         cdata.inline_wsize = xprt_rdma_max_inline_write;
 376         if (cdata.inline_wsize > cdata.wsize)
 377                 cdata.inline_wsize = cdata.wsize;
 378
 379         cdata.inline_rsize = xprt_rdma_max_inline_read;
 380         if (cdata.inline_rsize > cdata.rsize)
 381                 cdata.inline_rsize = cdata.rsize;
 382
 383         /*
 384          * Create new transport instance, which includes initialized
 385          *  o ia
 386          *  o endpoint
 387          *  o buffers
 388          */
 389
 390         new_xprt = rpcx_to_rdmax(xprt);
 391
 392         rc = rpcrdma_ia_open(new_xprt);
 393         if (rc)
 394                 goto out1;
 395
 396         /*
 397          * initialize and create ep
 398          */
 399         new_xprt->rx_data = cdata;
 400         new_ep = &new_xprt->rx_ep;
 401
 402         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
 403                                 &new_xprt->rx_ia, &new_xprt->rx_data);
 404         if (rc)
 405                 goto out2;
 406
 407         rc = rpcrdma_buffer_create(new_xprt);
 408         if (rc)
 409                 goto out3;
 410
 411         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
 412                           xprt_rdma_connect_worker);
 413
 414         xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
 415         if (xprt->max_payload == 0)
 416                 goto out4;
 417         xprt->max_payload <<= PAGE_SHIFT;
 418         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
 419                 __func__, xprt->max_payload);
 420
 421         if (!try_module_get(THIS_MODULE))
 422                 goto out4;
 423
 424         dprintk("RPC:       %s: %s:%s\n", __func__,
 425                 xprt->address_strings[RPC_DISPLAY_ADDR],
 426                 xprt->address_strings[RPC_DISPLAY_PORT]);
 427         trace_xprtrdma_create(new_xprt);
 428         return xprt;
 429
 430 out4:
 431         rpcrdma_buffer_destroy(&new_xprt->rx_buf);
 432         rc = -ENODEV;
 433 out3:
 434         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
 435 out2:
 436         rpcrdma_ia_close(&new_xprt->rx_ia);
 437 out1:
 438         trace_xprtrdma_destroy(new_xprt);
 439         xprt_rdma_free_addresses(xprt);
 440         xprt_free(xprt);
 441         return ERR_PTR(rc);
 442 }
 443
 444 /**
 445  * xprt_rdma_close - Close down RDMA connection
 446  * @xprt: generic transport to be closed
 447  *
 448  * Called during transport shutdown reconnect, or device
 449  * removal. Caller holds the transport's write lock.
 450  */
 451 static void
 452 xprt_rdma_close(struct rpc_xprt *xprt)
 453 {
 454         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 455         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 456         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 457
 458         dprintk("RPC:       %s: closing xprt %p\n", __func__, xprt);
 459
 460         if (test_and_clear_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags)) {
 461                 xprt_clear_connected(xprt);
 462                 rpcrdma_ia_remove(ia);
 463                 return;
 464         }
 465         if (ep->rep_connected == -ENODEV)
 466                 return;
 467         if (ep->rep_connected > 0)
 468                 xprt->reestablish_timeout = 0;
 469         xprt_disconnect_done(xprt);
 470         rpcrdma_ep_disconnect(ep, ia);
 471 }
 472
 473 /**
 474  * xprt_rdma_set_port - update server port with rpcbind result
 475  * @xprt: controlling RPC transport
 476  * @port: new port value
 477  *
 478  * Transport connect status is unchanged.
 479  */
 480 static void
 481 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 482 {
 483         struct sockaddr *sap = (struct sockaddr *)&xprt->addr;
 484         char buf[8];
 485
 486         dprintk("RPC:       %s: setting port for xprt %p (%s:%s) to %u\n",
 487                 __func__, xprt,
 488                 xprt->address_strings[RPC_DISPLAY_ADDR],
 489                 xprt->address_strings[RPC_DISPLAY_PORT],
 490                 port);
 491
 492         rpc_set_port(sap, port);
 493
 494         kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
 495         snprintf(buf, sizeof(buf), "%u", port);
 496         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 497
 498         kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]);
 499         snprintf(buf, sizeof(buf), "%4hx", port);
 500         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 501 }
 502
 503 /**
 504  * xprt_rdma_timer - invoked when an RPC times out
 505  * @xprt: controlling RPC transport
 506  * @task: RPC task that timed out
 507  *
 508  * Invoked when the transport is still connected, but an RPC
 509  * retransmit timeout occurs.
 510  *
 511  * Since RDMA connections don't have a keep-alive, forcibly
 512  * disconnect and retry to connect. This drives full
 513  * detection of the network path, and retransmissions of
 514  * all pending RPCs.
 515  */
 516 static void
 517 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
 518 {
 519         xprt_force_disconnect(xprt);
 520 }
 521
 522 static void
 523 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
 524 {
 525         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 526
 527         if (r_xprt->rx_ep.rep_connected != 0) {
 528                 /* Reconnect */
 529                 schedule_delayed_work(&r_xprt->rx_connect_worker,
 530                                       xprt->reestablish_timeout);
 531                 xprt->reestablish_timeout <<= 1;
 532                 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
 533                         xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
 534                 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
 535                         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 536         } else {
 537                 schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
 538                 if (!RPC_IS_ASYNC(task))
 539                         flush_delayed_work(&r_xprt->rx_connect_worker);
 540         }
 541 }
 542
 543 /**
 544  * xprt_rdma_alloc_slot - allocate an rpc_rqst
 545  * @xprt: controlling RPC transport
 546  * @task: RPC task requesting a fresh rpc_rqst
 547  *
 548  * tk_status values:
 549  *      %0 if task->tk_rqstp points to a fresh rpc_rqst
 550  *      %-EAGAIN if no rpc_rqst is available; queued on backlog
 551  */
 552 static void
 553 xprt_rdma_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
 554 {
 555         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 556         struct rpcrdma_req *req;
 557
 558         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
 559         if (!req)
 560                 goto out_sleep;
 561         task->tk_rqstp = &req->rl_slot;
 562         task->tk_status = 0;
 563         return;
 564
 565 out_sleep:
 566         rpc_sleep_on(&xprt->backlog, task, NULL);
 567         task->tk_status = -EAGAIN;
 568 }
 569
 570 /**
 571  * xprt_rdma_free_slot - release an rpc_rqst
 572  * @xprt: controlling RPC transport
 573  * @rqst: rpc_rqst to release
 574  *
 575  */
 576 static void
 577 xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst)
 578 {
 579         memset(rqst, 0, sizeof(*rqst));
 580         rpcrdma_buffer_put(rpcr_to_rdmar(rqst));
 581         rpc_wake_up_next(&xprt->backlog);
 582 }
 583
 584 static bool
 585 rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 586                     size_t size, gfp_t flags)
 587 {
 588         struct rpcrdma_regbuf *rb;
 589
 590         if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
 591                 return true;
 592
 593         rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
 594         if (IS_ERR(rb))
 595                 return false;
 596
 597         rpcrdma_free_regbuf(req->rl_sendbuf);
 598         r_xprt->rx_stats.hardway_register_count += size;
 599         req->rl_sendbuf = rb;
 600         return true;
 601 }
 602
 603 /* The rq_rcv_buf is used only if a Reply chunk is necessary.
 604  * The decision to use a Reply chunk is made later in
 605  * rpcrdma_marshal_req. This buffer is registered at that time.
 606  *
 607  * Otherwise, the associated RPC Reply arrives in a separate
 608  * Receive buffer, arbitrarily chosen by the HCA. The buffer
 609  * allocated here for the RPC Reply is not utilized in that
 610  * case. See rpcrdma_inline_fixup.
 611  *
 612  * A regbuf is used here to remember the buffer size.
 613  */
 614 static bool
 615 rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 616                     size_t size, gfp_t flags)
 617 {
 618         struct rpcrdma_regbuf *rb;
 619
 620         if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
 621                 return true;
 622
 623         rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
 624         if (IS_ERR(rb))
 625                 return false;
 626
 627         rpcrdma_free_regbuf(req->rl_recvbuf);
 628         r_xprt->rx_stats.hardway_register_count += size;
 629         req->rl_recvbuf = rb;
 630         return true;
 631 }
 632
 633 /**
 634  * xprt_rdma_allocate - allocate transport resources for an RPC
 635  * @task: RPC task
 636  *
 637  * Return values:
 638  *        0:    Success; rq_buffer points to RPC buffer to use
 639  *   ENOMEM:    Out of memory, call again later
 640  *      EIO:    A permanent error occurred, do not retry
 641  *
 642  * The RDMA allocate/free functions need the task structure as a place
 643  * to hide the struct rpcrdma_req, which is necessary for the actual
 644  * send/recv sequence.
 645  *
 646  * xprt_rdma_allocate provides buffers that are already mapped for
 647  * DMA, and a local DMA lkey is provided for each.
 648  */
 649 static int
 650 xprt_rdma_allocate(struct rpc_task *task)
 651 {
 652         struct rpc_rqst *rqst = task->tk_rqstp;
 653         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 654         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 655         gfp_t flags;
 656
 657         flags = RPCRDMA_DEF_GFP;
 658         if (RPC_IS_SWAPPER(task))
 659                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
 660
 661         if (!rpcrdma_get_sendbuf(r_xprt, req, rqst->rq_callsize, flags))
 662                 goto out_fail;
 663         if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
 664                 goto out_fail;
 665
 666         rqst->rq_buffer = req->rl_sendbuf->rg_base;
 667         rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
 668         trace_xprtrdma_allocate(task, req);
 669         return 0;
 670
 671 out_fail:
 672         trace_xprtrdma_allocate(task, NULL);
 673         return -ENOMEM;
 674 }
 675
 676 /**
 677  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
 678  * @task: RPC task
 679  *
 680  * Caller guarantees rqst->rq_buffer is non-NULL.
 681  */
 682 static void
 683 xprt_rdma_free(struct rpc_task *task)
 684 {
 685         struct rpc_rqst *rqst = task->tk_rqstp;
 686         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 687         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 688
 689         if (test_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags))
 690                 rpcrdma_release_rqst(r_xprt, req);
 691         trace_xprtrdma_rpc_done(task, req);
 692 }
 693
 694 /**
 695  * xprt_rdma_send_request - marshal and send an RPC request
 696  * @task: RPC task with an RPC message in rq_snd_buf
 697  *
 698  * Caller holds the transport's write lock.
 699  *
 700  * Returns:
 701  *      %0 if the RPC message has been sent
 702  *      %-ENOTCONN if the caller should reconnect and call again
 703  *      %-EAGAIN if the caller should call again
 704  *      %-ENOBUFS if the caller should call again after a delay
 705  *      %-EIO if a permanent error occurred and the request was not
 706  *              sent. Do not try to send this message again.
 707  */
 708 static int
 709 xprt_rdma_send_request(struct rpc_task *task)
 710 {
 711         struct rpc_rqst *rqst = task->tk_rqstp;
 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                 goto drop_connection;
 724
 725         rc = rpcrdma_marshal_req(r_xprt, rqst);
 726         if (rc < 0)
 727                 goto failed_marshal;
 728
 729         /* Must suppress retransmit to maintain credits */
 730         if (rqst->rq_connect_cookie == xprt->connect_cookie)
 731                 goto drop_connection;
 732         rqst->rq_xtime = ktime_get();
 733
 734         __set_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags);
 735         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 736                 goto drop_connection;
 737
 738         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
 739         rqst->rq_bytes_sent = 0;
 740
 741         /* An RPC with no reply will throw off credit accounting,
 742          * so drop the connection to reset the credit grant.
 743          */
 744         if (!rpc_reply_expected(task))
 745                 goto drop_connection;
 746         return 0;
 747
 748 failed_marshal:
 749         if (rc != -ENOTCONN)
 750                 return rc;
 751 drop_connection:
 752         xprt_disconnect_done(xprt);
 753         return -ENOTCONN;       /* implies disconnect */
 754 }
 755
 756 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 757 {
 758         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 759         long idle_time = 0;
 760
 761         if (xprt_connected(xprt))
 762                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 763
 764         seq_puts(seq, "\txprt:\trdma ");
 765         seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
 766                    0,   /* need a local port? */
 767                    xprt->stat.bind_count,
 768                    xprt->stat.connect_count,
 769                    xprt->stat.connect_time,
 770                    idle_time,
 771                    xprt->stat.sends,
 772                    xprt->stat.recvs,
 773                    xprt->stat.bad_xids,
 774                    xprt->stat.req_u,
 775                    xprt->stat.bklog_u);
 776         seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
 777                    r_xprt->rx_stats.read_chunk_count,
 778                    r_xprt->rx_stats.write_chunk_count,
 779                    r_xprt->rx_stats.reply_chunk_count,
 780                    r_xprt->rx_stats.total_rdma_request,
 781                    r_xprt->rx_stats.total_rdma_reply,
 782                    r_xprt->rx_stats.pullup_copy_count,
 783                    r_xprt->rx_stats.fixup_copy_count,
 784                    r_xprt->rx_stats.hardway_register_count,
 785                    r_xprt->rx_stats.failed_marshal_count,
 786                    r_xprt->rx_stats.bad_reply_count,
 787                    r_xprt->rx_stats.nomsg_call_count);
 788         seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
 789                    r_xprt->rx_stats.mrs_recovered,
 790                    r_xprt->rx_stats.mrs_orphaned,
 791                    r_xprt->rx_stats.mrs_allocated,
 792                    r_xprt->rx_stats.local_inv_needed,
 793                    r_xprt->rx_stats.empty_sendctx_q,
 794                    r_xprt->rx_stats.reply_waits_for_send);
 795 }
 796
 797 static int
 798 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
 799 {
 800         return 0;
 801 }
 802
 803 static void
 804 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
 805 {
 806 }
 807
 808 /*
 809  * Plumbing for rpc transport switch and kernel module
 810  */
 811
 812 static const struct rpc_xprt_ops xprt_rdma_procs = {
 813         .reserve_xprt           = xprt_reserve_xprt_cong,
 814         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 815         .alloc_slot             = xprt_rdma_alloc_slot,
 816         .free_slot              = xprt_rdma_free_slot,
 817         .release_request        = xprt_release_rqst_cong,       /* ditto */
 818         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
 819         .timer                  = xprt_rdma_timer,
 820         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 821         .set_port               = xprt_rdma_set_port,
 822         .connect                = xprt_rdma_connect,
 823         .buf_alloc              = xprt_rdma_allocate,
 824         .buf_free               = xprt_rdma_free,
 825         .send_request           = xprt_rdma_send_request,
 826         .close                  = xprt_rdma_close,
 827         .destroy                = xprt_rdma_destroy,
 828         .print_stats            = xprt_rdma_print_stats,
 829         .enable_swap            = xprt_rdma_enable_swap,
 830         .disable_swap           = xprt_rdma_disable_swap,
 831         .inject_disconnect      = xprt_rdma_inject_disconnect,
 832 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
 833         .bc_setup               = xprt_rdma_bc_setup,
 834         .bc_up                  = xprt_rdma_bc_up,
 835         .bc_maxpayload          = xprt_rdma_bc_maxpayload,
 836         .bc_free_rqst           = xprt_rdma_bc_free_rqst,
 837         .bc_destroy             = xprt_rdma_bc_destroy,
 838 #endif
 839 };
 840
 841 static struct xprt_class xprt_rdma = {
 842         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 843         .name                   = "rdma",
 844         .owner                  = THIS_MODULE,
 845         .ident                  = XPRT_TRANSPORT_RDMA,
 846         .setup                  = xprt_setup_rdma,
 847 };
 848
 849 void xprt_rdma_cleanup(void)
 850 {
 851         int rc;
 852
 853         dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
 854 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 855         if (sunrpc_table_header) {
 856                 unregister_sysctl_table(sunrpc_table_header);
 857                 sunrpc_table_header = NULL;
 858         }
 859 #endif
 860         rc = xprt_unregister_transport(&xprt_rdma);
 861         if (rc)
 862                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
 863                         __func__, rc);
 864
 865         rpcrdma_destroy_wq();
 866
 867         rc = xprt_unregister_transport(&xprt_rdma_bc);
 868         if (rc)
 869                 dprintk("RPC:       %s: xprt_unregister(bc) returned %i\n",
 870                         __func__, rc);
 871 }
 872
 873 int xprt_rdma_init(void)
 874 {
 875         int rc;
 876
 877         rc = rpcrdma_alloc_wq();
 878         if (rc)
 879                 return rc;
 880
 881         rc = xprt_register_transport(&xprt_rdma);
 882         if (rc) {
 883                 rpcrdma_destroy_wq();
 884                 return rc;
 885         }
 886
 887         rc = xprt_register_transport(&xprt_rdma_bc);
 888         if (rc) {
 889                 xprt_unregister_transport(&xprt_rdma);
 890                 rpcrdma_destroy_wq();
 891                 return rc;
 892         }
 893
 894         dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
 895
 896         dprintk("Defaults:\n");
 897         dprintk("\tSlots %d\n"
 898                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
 899                 xprt_rdma_slot_table_entries,
 900                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
 901         dprintk("\tPadding 0\n\tMemreg %d\n", xprt_rdma_memreg_strategy);
 902
 903 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 904         if (!sunrpc_table_header)
 905                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 906 #endif
 907         return 0;
 908 }