net/sunrpc/xprtrdma/transport.c

   1 /*
   2  * Copyright (c) 2014-2017 Oracle.  All rights reserved.
   3  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the BSD-type
   9  * license below:
  10  *
  11  * Redistribution and use in source and binary forms, with or without
  12  * modification, are permitted provided that the following conditions
  13  * are met:
  14  *
  15  *      Redistributions of source code must retain the above copyright
  16  *      notice, this list of conditions and the following disclaimer.
  17  *
  18  *      Redistributions in binary form must reproduce the above
  19  *      copyright notice, this list of conditions and the following
  20  *      disclaimer in the documentation and/or other materials provided
  21  *      with the distribution.
  22  *
  23  *      Neither the name of the Network Appliance, Inc. nor the names of
  24  *      its contributors may be used to endorse or promote products
  25  *      derived from this software without specific prior written
  26  *      permission.
  27  *
  28  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  29  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  30  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  31  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  32  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  33  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  34  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  35  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  36  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  37  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  38  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  39  */
  40
  41 /*
  42  * transport.c
  43  *
  44  * This file contains the top-level implementation of an RPC RDMA
  45  * transport.
  46  *
  47  * Naming convention: functions beginning with xprt_ are part of the
  48  * transport switch. All others are RPC RDMA internal.
  49  */
  50
  51 #include <linux/module.h>
  52 #include <linux/slab.h>
  53 #include <linux/seq_file.h>
  54 #include <linux/sunrpc/addr.h>
  55 #include <linux/smp.h>
  56
  57 #include "xprt_rdma.h"
  58
  59 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  60 # define RPCDBG_FACILITY        RPCDBG_TRANS
  61 #endif
  62
  63 /*
  64  * tunables
  65  */
  66
  67 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
  68 unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
  69 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
  70 unsigned int xprt_rdma_memreg_strategy          = RPCRDMA_FRMR;
  71 int xprt_rdma_pad_optimize;
  72
  73 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  74
  75 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
  76 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
  77 static unsigned int min_inline_size = RPCRDMA_MIN_INLINE;
  78 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
  79 static unsigned int zero;
  80 static unsigned int max_padding = PAGE_SIZE;
  81 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
  82 static unsigned int max_memreg = RPCRDMA_LAST - 1;
  83 static unsigned int dummy;
  84
  85 static struct ctl_table_header *sunrpc_table_header;
  86
  87 static struct ctl_table xr_tunables_table[] = {
  88         {
  89                 .procname       = "rdma_slot_table_entries",
  90                 .data           = &xprt_rdma_slot_table_entries,
  91                 .maxlen         = sizeof(unsigned int),
  92                 .mode           = 0644,
  93                 .proc_handler   = proc_dointvec_minmax,
  94                 .extra1         = &min_slot_table_size,
  95                 .extra2         = &max_slot_table_size
  96         },
  97         {
  98                 .procname       = "rdma_max_inline_read",
  99                 .data           = &xprt_rdma_max_inline_read,
 100                 .maxlen         = sizeof(unsigned int),
 101                 .mode           = 0644,
 102                 .proc_handler   = proc_dointvec_minmax,
 103                 .extra1         = &min_inline_size,
 104                 .extra2         = &max_inline_size,
 105         },
 106         {
 107                 .procname       = "rdma_max_inline_write",
 108                 .data           = &xprt_rdma_max_inline_write,
 109                 .maxlen         = sizeof(unsigned int),
 110                 .mode           = 0644,
 111                 .proc_handler   = proc_dointvec_minmax,
 112                 .extra1         = &min_inline_size,
 113                 .extra2         = &max_inline_size,
 114         },
 115         {
 116                 .procname       = "rdma_inline_write_padding",
 117                 .data           = &dummy,
 118                 .maxlen         = sizeof(unsigned int),
 119                 .mode           = 0644,
 120                 .proc_handler   = proc_dointvec_minmax,
 121                 .extra1         = &zero,
 122                 .extra2         = &max_padding,
 123         },
 124         {
 125                 .procname       = "rdma_memreg_strategy",
 126                 .data           = &xprt_rdma_memreg_strategy,
 127                 .maxlen         = sizeof(unsigned int),
 128                 .mode           = 0644,
 129                 .proc_handler   = proc_dointvec_minmax,
 130                 .extra1         = &min_memreg,
 131                 .extra2         = &max_memreg,
 132         },
 133         {
 134                 .procname       = "rdma_pad_optimize",
 135                 .data           = &xprt_rdma_pad_optimize,
 136                 .maxlen         = sizeof(unsigned int),
 137                 .mode           = 0644,
 138                 .proc_handler   = proc_dointvec,
 139         },
 140         { },
 141 };
 142
 143 static struct ctl_table sunrpc_table[] = {
 144         {
 145                 .procname       = "sunrpc",
 146                 .mode           = 0555,
 147                 .child          = xr_tunables_table
 148         },
 149         { },
 150 };
 151
 152 #endif
 153
 154 static const struct rpc_xprt_ops xprt_rdma_procs;
 155
 156 static void
 157 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
 158 {
 159         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
 160         char buf[20];
 161
 162         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
 163         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 164
 165         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
 166 }
 167
 168 static void
 169 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
 170 {
 171         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
 172         char buf[40];
 173
 174         snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
 175         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 176
 177         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
 178 }
 179
 180 void
 181 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
 182 {
 183         char buf[128];
 184
 185         switch (sap->sa_family) {
 186         case AF_INET:
 187                 xprt_rdma_format_addresses4(xprt, sap);
 188                 break;
 189         case AF_INET6:
 190                 xprt_rdma_format_addresses6(xprt, sap);
 191                 break;
 192         default:
 193                 pr_err("rpcrdma: Unrecognized address family\n");
 194                 return;
 195         }
 196
 197         (void)rpc_ntop(sap, buf, sizeof(buf));
 198         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
 199
 200         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
 201         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 202
 203         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
 204         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 205
 206         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
 207 }
 208
 209 void
 210 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
 211 {
 212         unsigned int i;
 213
 214         for (i = 0; i < RPC_DISPLAY_MAX; i++)
 215                 switch (i) {
 216                 case RPC_DISPLAY_PROTO:
 217                 case RPC_DISPLAY_NETID:
 218                         continue;
 219                 default:
 220                         kfree(xprt->address_strings[i]);
 221                 }
 222 }
 223
 224 void
 225 rpcrdma_conn_func(struct rpcrdma_ep *ep)
 226 {
 227         schedule_delayed_work(&ep->rep_connect_worker, 0);
 228 }
 229
 230 void
 231 rpcrdma_connect_worker(struct work_struct *work)
 232 {
 233         struct rpcrdma_ep *ep =
 234                 container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
 235         struct rpcrdma_xprt *r_xprt =
 236                 container_of(ep, struct rpcrdma_xprt, rx_ep);
 237         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 238
 239         spin_lock_bh(&xprt->transport_lock);
 240         if (++xprt->connect_cookie == 0)        /* maintain a reserved value */
 241                 ++xprt->connect_cookie;
 242         if (ep->rep_connected > 0) {
 243                 if (!xprt_test_and_set_connected(xprt))
 244                         xprt_wake_pending_tasks(xprt, 0);
 245         } else {
 246                 if (xprt_test_and_clear_connected(xprt))
 247                         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 248         }
 249         spin_unlock_bh(&xprt->transport_lock);
 250 }
 251
 252 static void
 253 xprt_rdma_connect_worker(struct work_struct *work)
 254 {
 255         struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
 256                                                    rx_connect_worker.work);
 257         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 258         int rc = 0;
 259
 260         xprt_clear_connected(xprt);
 261
 262         dprintk("RPC:       %s: %sconnect\n", __func__,
 263                         r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
 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         dprintk("RPC:       %s: exit\n", __func__);
 269         xprt_clear_connecting(xprt);
 270 }
 271
 272 static void
 273 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
 274 {
 275         struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
 276                                                    rx_xprt);
 277
 278         pr_info("rpcrdma: injecting transport disconnect on xprt=%p\n", xprt);
 279         rdma_disconnect(r_xprt->rx_ia.ri_id);
 280 }
 281
 282 /*
 283  * xprt_rdma_destroy
 284  *
 285  * Destroy the xprt.
 286  * Free all memory associated with the object, including its own.
 287  * NOTE: none of the *destroy methods free memory for their top-level
 288  * objects, even though they may have allocated it (they do free
 289  * private memory). It's up to the caller to handle it. In this
 290  * case (RDMA transport), all structure memory is inlined with the
 291  * struct rpcrdma_xprt.
 292  */
 293 static void
 294 xprt_rdma_destroy(struct rpc_xprt *xprt)
 295 {
 296         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 297
 298         dprintk("RPC:       %s: called\n", __func__);
 299
 300         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
 301
 302         xprt_clear_connected(xprt);
 303
 304         rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
 305         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 306         rpcrdma_ia_close(&r_xprt->rx_ia);
 307
 308         xprt_rdma_free_addresses(xprt);
 309
 310         xprt_free(xprt);
 311
 312         dprintk("RPC:       %s: returning\n", __func__);
 313
 314         module_put(THIS_MODULE);
 315 }
 316
 317 static const struct rpc_timeout xprt_rdma_default_timeout = {
 318         .to_initval = 60 * HZ,
 319         .to_maxval = 60 * HZ,
 320 };
 321
 322 /**
 323  * xprt_setup_rdma - Set up transport to use RDMA
 324  *
 325  * @args: rpc transport arguments
 326  */
 327 static struct rpc_xprt *
 328 xprt_setup_rdma(struct xprt_create *args)
 329 {
 330         struct rpcrdma_create_data_internal cdata;
 331         struct rpc_xprt *xprt;
 332         struct rpcrdma_xprt *new_xprt;
 333         struct rpcrdma_ep *new_ep;
 334         struct sockaddr *sap;
 335         int rc;
 336
 337         if (args->addrlen > sizeof(xprt->addr)) {
 338                 dprintk("RPC:       %s: address too large\n", __func__);
 339                 return ERR_PTR(-EBADF);
 340         }
 341
 342         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
 343                         xprt_rdma_slot_table_entries,
 344                         xprt_rdma_slot_table_entries);
 345         if (xprt == NULL) {
 346                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
 347                         __func__);
 348                 return ERR_PTR(-ENOMEM);
 349         }
 350
 351         /* 60 second timeout, no retries */
 352         xprt->timeout = &xprt_rdma_default_timeout;
 353         xprt->bind_timeout = RPCRDMA_BIND_TO;
 354         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 355         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
 356
 357         xprt->resvport = 0;             /* privileged port not needed */
 358         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
 359         xprt->ops = &xprt_rdma_procs;
 360
 361         /*
 362          * Set up RDMA-specific connect data.
 363          */
 364         sap = args->dstaddr;
 365
 366         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 367          * address, for any side protocols which peek at it */
 368         xprt->prot = IPPROTO_TCP;
 369         xprt->addrlen = args->addrlen;
 370         memcpy(&xprt->addr, sap, xprt->addrlen);
 371
 372         if (rpc_get_port(sap))
 373                 xprt_set_bound(xprt);
 374         xprt_rdma_format_addresses(xprt, sap);
 375
 376         cdata.max_requests = xprt->max_reqs;
 377
 378         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
 379         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
 380
 381         cdata.inline_wsize = xprt_rdma_max_inline_write;
 382         if (cdata.inline_wsize > cdata.wsize)
 383                 cdata.inline_wsize = cdata.wsize;
 384
 385         cdata.inline_rsize = xprt_rdma_max_inline_read;
 386         if (cdata.inline_rsize > cdata.rsize)
 387                 cdata.inline_rsize = cdata.rsize;
 388
 389         /*
 390          * Create new transport instance, which includes initialized
 391          *  o ia
 392          *  o endpoint
 393          *  o buffers
 394          */
 395
 396         new_xprt = rpcx_to_rdmax(xprt);
 397
 398         rc = rpcrdma_ia_open(new_xprt);
 399         if (rc)
 400                 goto out1;
 401
 402         /*
 403          * initialize and create ep
 404          */
 405         new_xprt->rx_data = cdata;
 406         new_ep = &new_xprt->rx_ep;
 407
 408         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
 409                                 &new_xprt->rx_ia, &new_xprt->rx_data);
 410         if (rc)
 411                 goto out2;
 412
 413         rc = rpcrdma_buffer_create(new_xprt);
 414         if (rc)
 415                 goto out3;
 416
 417         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
 418                           xprt_rdma_connect_worker);
 419
 420         xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
 421         if (xprt->max_payload == 0)
 422                 goto out4;
 423         xprt->max_payload <<= PAGE_SHIFT;
 424         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
 425                 __func__, xprt->max_payload);
 426
 427         if (!try_module_get(THIS_MODULE))
 428                 goto out4;
 429
 430         dprintk("RPC:       %s: %s:%s\n", __func__,
 431                 xprt->address_strings[RPC_DISPLAY_ADDR],
 432                 xprt->address_strings[RPC_DISPLAY_PORT]);
 433         return xprt;
 434
 435 out4:
 436         rpcrdma_buffer_destroy(&new_xprt->rx_buf);
 437         rc = -ENODEV;
 438 out3:
 439         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
 440 out2:
 441         rpcrdma_ia_close(&new_xprt->rx_ia);
 442 out1:
 443         xprt_rdma_free_addresses(xprt);
 444         xprt_free(xprt);
 445         return ERR_PTR(rc);
 446 }
 447
 448 /**
 449  * xprt_rdma_close - Close down RDMA connection
 450  * @xprt: generic transport to be closed
 451  *
 452  * Called during transport shutdown reconnect, or device
 453  * removal. Caller holds the transport's write lock.
 454  */
 455 static void
 456 xprt_rdma_close(struct rpc_xprt *xprt)
 457 {
 458         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 459         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 460         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 461
 462         dprintk("RPC:       %s: closing xprt %p\n", __func__, xprt);
 463
 464         if (test_and_clear_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags)) {
 465                 xprt_clear_connected(xprt);
 466                 rpcrdma_ia_remove(ia);
 467                 return;
 468         }
 469         if (ep->rep_connected == -ENODEV)
 470                 return;
 471         if (ep->rep_connected > 0)
 472                 xprt->reestablish_timeout = 0;
 473         xprt_disconnect_done(xprt);
 474         rpcrdma_ep_disconnect(ep, ia);
 475 }
 476
 477 static void
 478 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 479 {
 480         struct sockaddr_in *sap;
 481
 482         sap = (struct sockaddr_in *)&xprt->addr;
 483         sap->sin_port = htons(port);
 484         dprintk("RPC:       %s: %u\n", __func__, port);
 485 }
 486
 487 /**
 488  * xprt_rdma_timer - invoked when an RPC times out
 489  * @xprt: controlling RPC transport
 490  * @task: RPC task that timed out
 491  *
 492  * Invoked when the transport is still connected, but an RPC
 493  * retransmit timeout occurs.
 494  *
 495  * Since RDMA connections don't have a keep-alive, forcibly
 496  * disconnect and retry to connect. This drives full
 497  * detection of the network path, and retransmissions of
 498  * all pending RPCs.
 499  */
 500 static void
 501 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
 502 {
 503         dprintk("RPC: %5u %s: xprt = %p\n", task->tk_pid, __func__, xprt);
 504
 505         xprt_force_disconnect(xprt);
 506 }
 507
 508 static void
 509 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
 510 {
 511         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 512
 513         if (r_xprt->rx_ep.rep_connected != 0) {
 514                 /* Reconnect */
 515                 schedule_delayed_work(&r_xprt->rx_connect_worker,
 516                                       xprt->reestablish_timeout);
 517                 xprt->reestablish_timeout <<= 1;
 518                 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
 519                         xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
 520                 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
 521                         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 522         } else {
 523                 schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
 524                 if (!RPC_IS_ASYNC(task))
 525                         flush_delayed_work(&r_xprt->rx_connect_worker);
 526         }
 527 }
 528
 529 /* Allocate a fixed-size buffer in which to construct and send the
 530  * RPC-over-RDMA header for this request.
 531  */
 532 static bool
 533 rpcrdma_get_rdmabuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 534                     gfp_t flags)
 535 {
 536         size_t size = RPCRDMA_HDRBUF_SIZE;
 537         struct rpcrdma_regbuf *rb;
 538
 539         if (req->rl_rdmabuf)
 540                 return true;
 541
 542         rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
 543         if (IS_ERR(rb))
 544                 return false;
 545
 546         r_xprt->rx_stats.hardway_register_count += size;
 547         req->rl_rdmabuf = rb;
 548         xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
 549         return true;
 550 }
 551
 552 static bool
 553 rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 554                     size_t size, gfp_t flags)
 555 {
 556         struct rpcrdma_regbuf *rb;
 557
 558         if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
 559                 return true;
 560
 561         rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
 562         if (IS_ERR(rb))
 563                 return false;
 564
 565         rpcrdma_free_regbuf(req->rl_sendbuf);
 566         r_xprt->rx_stats.hardway_register_count += size;
 567         req->rl_sendbuf = rb;
 568         return true;
 569 }
 570
 571 /* The rq_rcv_buf is used only if a Reply chunk is necessary.
 572  * The decision to use a Reply chunk is made later in
 573  * rpcrdma_marshal_req. This buffer is registered at that time.
 574  *
 575  * Otherwise, the associated RPC Reply arrives in a separate
 576  * Receive buffer, arbitrarily chosen by the HCA. The buffer
 577  * allocated here for the RPC Reply is not utilized in that
 578  * case. See rpcrdma_inline_fixup.
 579  *
 580  * A regbuf is used here to remember the buffer size.
 581  */
 582 static bool
 583 rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 584                     size_t size, gfp_t flags)
 585 {
 586         struct rpcrdma_regbuf *rb;
 587
 588         if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
 589                 return true;
 590
 591         rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
 592         if (IS_ERR(rb))
 593                 return false;
 594
 595         rpcrdma_free_regbuf(req->rl_recvbuf);
 596         r_xprt->rx_stats.hardway_register_count += size;
 597         req->rl_recvbuf = rb;
 598         return true;
 599 }
 600
 601 /**
 602  * xprt_rdma_allocate - allocate transport resources for an RPC
 603  * @task: RPC task
 604  *
 605  * Return values:
 606  *        0:    Success; rq_buffer points to RPC buffer to use
 607  *   ENOMEM:    Out of memory, call again later
 608  *      EIO:    A permanent error occurred, do not retry
 609  *
 610  * The RDMA allocate/free functions need the task structure as a place
 611  * to hide the struct rpcrdma_req, which is necessary for the actual
 612  * send/recv sequence.
 613  *
 614  * xprt_rdma_allocate provides buffers that are already mapped for
 615  * DMA, and a local DMA lkey is provided for each.
 616  */
 617 static int
 618 xprt_rdma_allocate(struct rpc_task *task)
 619 {
 620         struct rpc_rqst *rqst = task->tk_rqstp;
 621         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 622         struct rpcrdma_req *req;
 623         gfp_t flags;
 624
 625         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
 626         if (req == NULL)
 627                 return -ENOMEM;
 628
 629         flags = RPCRDMA_DEF_GFP;
 630         if (RPC_IS_SWAPPER(task))
 631                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
 632
 633         if (!rpcrdma_get_rdmabuf(r_xprt, req, flags))
 634                 goto out_fail;
 635         if (!rpcrdma_get_sendbuf(r_xprt, req, rqst->rq_callsize, flags))
 636                 goto out_fail;
 637         if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
 638                 goto out_fail;
 639
 640         dprintk("RPC: %5u %s: send size = %zd, recv size = %zd, req = %p\n",
 641                 task->tk_pid, __func__, rqst->rq_callsize,
 642                 rqst->rq_rcvsize, req);
 643
 644         req->rl_cpu = smp_processor_id();
 645         req->rl_connect_cookie = 0;     /* our reserved value */
 646         rpcrdma_set_xprtdata(rqst, req);
 647         rqst->rq_buffer = req->rl_sendbuf->rg_base;
 648         rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
 649         return 0;
 650
 651 out_fail:
 652         rpcrdma_buffer_put(req);
 653         return -ENOMEM;
 654 }
 655
 656 /**
 657  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
 658  * @task: RPC task
 659  *
 660  * Caller guarantees rqst->rq_buffer is non-NULL.
 661  */
 662 static void
 663 xprt_rdma_free(struct rpc_task *task)
 664 {
 665         struct rpc_rqst *rqst = task->tk_rqstp;
 666         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 667         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 668
 669         if (test_bit(RPCRDMA_REQ_F_BACKCHANNEL, &req->rl_flags))
 670                 return;
 671
 672         dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
 673
 674         if (test_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags))
 675                 rpcrdma_release_rqst(r_xprt, req);
 676         rpcrdma_buffer_put(req);
 677 }
 678
 679 /**
 680  * xprt_rdma_send_request - marshal and send an RPC request
 681  * @task: RPC task with an RPC message in rq_snd_buf
 682  *
 683  * Caller holds the transport's write lock.
 684  *
 685  * Return values:
 686  *        0:    The request has been sent
 687  * ENOTCONN:    Caller needs to invoke connect logic then call again
 688  *  ENOBUFS:    Call again later to send the request
 689  *      EIO:    A permanent error occurred. The request was not sent,
 690  *              and don't try it again
 691  *
 692  * send_request invokes the meat of RPC RDMA. It must do the following:
 693  *
 694  *  1.  Marshal the RPC request into an RPC RDMA request, which means
 695  *      putting a header in front of data, and creating IOVs for RDMA
 696  *      from those in the request.
 697  *  2.  In marshaling, detect opportunities for RDMA, and use them.
 698  *  3.  Post a recv message to set up asynch completion, then send
 699  *      the request (rpcrdma_ep_post).
 700  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
 701  */
 702 static int
 703 xprt_rdma_send_request(struct rpc_task *task)
 704 {
 705         struct rpc_rqst *rqst = task->tk_rqstp;
 706         struct rpc_xprt *xprt = rqst->rq_xprt;
 707         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 708         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 709         int rc = 0;
 710
 711         if (!xprt_connected(xprt))
 712                 goto drop_connection;
 713
 714         /* On retransmit, remove any previously registered chunks */
 715         if (unlikely(!list_empty(&req->rl_registered)))
 716                 r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt,
 717                                                     &req->rl_registered);
 718
 719         rc = rpcrdma_marshal_req(r_xprt, rqst);
 720         if (rc < 0)
 721                 goto failed_marshal;
 722
 723         if (req->rl_reply == NULL)              /* e.g. reconnection */
 724                 rpcrdma_recv_buffer_get(req);
 725
 726         /* Must suppress retransmit to maintain credits */
 727         if (req->rl_connect_cookie == xprt->connect_cookie)
 728                 goto drop_connection;
 729         req->rl_connect_cookie = xprt->connect_cookie;
 730
 731         __set_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags);
 732         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 733                 goto drop_connection;
 734
 735         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
 736         rqst->rq_bytes_sent = 0;
 737         return 0;
 738
 739 failed_marshal:
 740         if (rc != -ENOTCONN)
 741                 return rc;
 742 drop_connection:
 743         xprt_disconnect_done(xprt);
 744         return -ENOTCONN;       /* implies disconnect */
 745 }
 746
 747 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 748 {
 749         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 750         long idle_time = 0;
 751
 752         if (xprt_connected(xprt))
 753                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 754
 755         seq_puts(seq, "\txprt:\trdma ");
 756         seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
 757                    0,   /* need a local port? */
 758                    xprt->stat.bind_count,
 759                    xprt->stat.connect_count,
 760                    xprt->stat.connect_time,
 761                    idle_time,
 762                    xprt->stat.sends,
 763                    xprt->stat.recvs,
 764                    xprt->stat.bad_xids,
 765                    xprt->stat.req_u,
 766                    xprt->stat.bklog_u);
 767         seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
 768                    r_xprt->rx_stats.read_chunk_count,
 769                    r_xprt->rx_stats.write_chunk_count,
 770                    r_xprt->rx_stats.reply_chunk_count,
 771                    r_xprt->rx_stats.total_rdma_request,
 772                    r_xprt->rx_stats.total_rdma_reply,
 773                    r_xprt->rx_stats.pullup_copy_count,
 774                    r_xprt->rx_stats.fixup_copy_count,
 775                    r_xprt->rx_stats.hardway_register_count,
 776                    r_xprt->rx_stats.failed_marshal_count,
 777                    r_xprt->rx_stats.bad_reply_count,
 778                    r_xprt->rx_stats.nomsg_call_count);
 779         seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
 780                    r_xprt->rx_stats.mrs_recovered,
 781                    r_xprt->rx_stats.mrs_orphaned,
 782                    r_xprt->rx_stats.mrs_allocated,
 783                    r_xprt->rx_stats.local_inv_needed,
 784                    r_xprt->rx_stats.empty_sendctx_q,
 785                    r_xprt->rx_stats.reply_waits_for_send);
 786 }
 787
 788 static int
 789 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
 790 {
 791         return 0;
 792 }
 793
 794 static void
 795 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
 796 {
 797 }
 798
 799 /*
 800  * Plumbing for rpc transport switch and kernel module
 801  */
 802
 803 static const struct rpc_xprt_ops xprt_rdma_procs = {
 804         .reserve_xprt           = xprt_reserve_xprt_cong,
 805         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 806         .alloc_slot             = xprt_alloc_slot,
 807         .release_request        = xprt_release_rqst_cong,       /* ditto */
 808         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
 809         .timer                  = xprt_rdma_timer,
 810         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 811         .set_port               = xprt_rdma_set_port,
 812         .connect                = xprt_rdma_connect,
 813         .buf_alloc              = xprt_rdma_allocate,
 814         .buf_free               = xprt_rdma_free,
 815         .send_request           = xprt_rdma_send_request,
 816         .close                  = xprt_rdma_close,
 817         .destroy                = xprt_rdma_destroy,
 818         .print_stats            = xprt_rdma_print_stats,
 819         .enable_swap            = xprt_rdma_enable_swap,
 820         .disable_swap           = xprt_rdma_disable_swap,
 821         .inject_disconnect      = xprt_rdma_inject_disconnect,
 822 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
 823         .bc_setup               = xprt_rdma_bc_setup,
 824         .bc_up                  = xprt_rdma_bc_up,
 825         .bc_maxpayload          = xprt_rdma_bc_maxpayload,
 826         .bc_free_rqst           = xprt_rdma_bc_free_rqst,
 827         .bc_destroy             = xprt_rdma_bc_destroy,
 828 #endif
 829 };
 830
 831 static struct xprt_class xprt_rdma = {
 832         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 833         .name                   = "rdma",
 834         .owner                  = THIS_MODULE,
 835         .ident                  = XPRT_TRANSPORT_RDMA,
 836         .setup                  = xprt_setup_rdma,
 837 };
 838
 839 void xprt_rdma_cleanup(void)
 840 {
 841         int rc;
 842
 843         dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
 844 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 845         if (sunrpc_table_header) {
 846                 unregister_sysctl_table(sunrpc_table_header);
 847                 sunrpc_table_header = NULL;
 848         }
 849 #endif
 850         rc = xprt_unregister_transport(&xprt_rdma);
 851         if (rc)
 852                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
 853                         __func__, rc);
 854
 855         rpcrdma_destroy_wq();
 856
 857         rc = xprt_unregister_transport(&xprt_rdma_bc);
 858         if (rc)
 859                 dprintk("RPC:       %s: xprt_unregister(bc) returned %i\n",
 860                         __func__, rc);
 861 }
 862
 863 int xprt_rdma_init(void)
 864 {
 865         int rc;
 866
 867         rc = rpcrdma_alloc_wq();
 868         if (rc)
 869                 return rc;
 870
 871         rc = xprt_register_transport(&xprt_rdma);
 872         if (rc) {
 873                 rpcrdma_destroy_wq();
 874                 return rc;
 875         }
 876
 877         rc = xprt_register_transport(&xprt_rdma_bc);
 878         if (rc) {
 879                 xprt_unregister_transport(&xprt_rdma);
 880                 rpcrdma_destroy_wq();
 881                 return rc;
 882         }
 883
 884         dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
 885
 886         dprintk("Defaults:\n");
 887         dprintk("\tSlots %d\n"
 888                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
 889                 xprt_rdma_slot_table_entries,
 890                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
 891         dprintk("\tPadding 0\n\tMemreg %d\n", xprt_rdma_memreg_strategy);
 892
 893 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 894         if (!sunrpc_table_header)
 895                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 896 #endif
 897         return 0;
 898 }