net/sunrpc/xprtrdma/transport.c

   1 /*
   2  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the BSD-type
   8  * license below:
   9  *
  10  * Redistribution and use in source and binary forms, with or without
  11  * modification, are permitted provided that the following conditions
  12  * are met:
  13  *
  14  *      Redistributions of source code must retain the above copyright
  15  *      notice, this list of conditions and the following disclaimer.
  16  *
  17  *      Redistributions in binary form must reproduce the above
  18  *      copyright notice, this list of conditions and the following
  19  *      disclaimer in the documentation and/or other materials provided
  20  *      with the distribution.
  21  *
  22  *      Neither the name of the Network Appliance, Inc. nor the names of
  23  *      its contributors may be used to endorse or promote products
  24  *      derived from this software without specific prior written
  25  *      permission.
  26  *
  27  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  28  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  29  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  30  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  31  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  32  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  33  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  34  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  35  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  36  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  37  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  38  */
  39
  40 /*
  41  * transport.c
  42  *
  43  * This file contains the top-level implementation of an RPC RDMA
  44  * transport.
  45  *
  46  * Naming convention: functions beginning with xprt_ are part of the
  47  * transport switch. All others are RPC RDMA internal.
  48  */
  49
  50 #include <linux/module.h>
  51 #include <linux/slab.h>
  52 #include <linux/seq_file.h>
  53 #include <linux/sunrpc/addr.h>
  54
  55 #include "xprt_rdma.h"
  56
  57 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  58 # define RPCDBG_FACILITY        RPCDBG_TRANS
  59 #endif
  60
  61 /*
  62  * tunables
  63  */
  64
  65 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
  66 unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
  67 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
  68 static unsigned int xprt_rdma_inline_write_padding;
  69 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
  70                 int xprt_rdma_pad_optimize = 0;
  71
  72 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  73
  74 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
  75 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
  76 static unsigned int min_inline_size = RPCRDMA_MIN_INLINE;
  77 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE;
  78 static unsigned int zero;
  79 static unsigned int max_padding = PAGE_SIZE;
  80 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
  81 static unsigned int max_memreg = RPCRDMA_LAST - 1;
  82
  83 static struct ctl_table_header *sunrpc_table_header;
  84
  85 static struct ctl_table xr_tunables_table[] = {
  86         {
  87                 .procname       = "rdma_slot_table_entries",
  88                 .data           = &xprt_rdma_slot_table_entries,
  89                 .maxlen         = sizeof(unsigned int),
  90                 .mode           = 0644,
  91                 .proc_handler   = proc_dointvec_minmax,
  92                 .extra1         = &min_slot_table_size,
  93                 .extra2         = &max_slot_table_size
  94         },
  95         {
  96                 .procname       = "rdma_max_inline_read",
  97                 .data           = &xprt_rdma_max_inline_read,
  98                 .maxlen         = sizeof(unsigned int),
  99                 .mode           = 0644,
 100                 .proc_handler   = proc_dointvec_minmax,
 101                 .extra1         = &min_inline_size,
 102                 .extra2         = &max_inline_size,
 103         },
 104         {
 105                 .procname       = "rdma_max_inline_write",
 106                 .data           = &xprt_rdma_max_inline_write,
 107                 .maxlen         = sizeof(unsigned int),
 108                 .mode           = 0644,
 109                 .proc_handler   = proc_dointvec_minmax,
 110                 .extra1         = &min_inline_size,
 111                 .extra2         = &max_inline_size,
 112         },
 113         {
 114                 .procname       = "rdma_inline_write_padding",
 115                 .data           = &xprt_rdma_inline_write_padding,
 116                 .maxlen         = sizeof(unsigned int),
 117                 .mode           = 0644,
 118                 .proc_handler   = proc_dointvec_minmax,
 119                 .extra1         = &zero,
 120                 .extra2         = &max_padding,
 121         },
 122         {
 123                 .procname       = "rdma_memreg_strategy",
 124                 .data           = &xprt_rdma_memreg_strategy,
 125                 .maxlen         = sizeof(unsigned int),
 126                 .mode           = 0644,
 127                 .proc_handler   = proc_dointvec_minmax,
 128                 .extra1         = &min_memreg,
 129                 .extra2         = &max_memreg,
 130         },
 131         {
 132                 .procname       = "rdma_pad_optimize",
 133                 .data           = &xprt_rdma_pad_optimize,
 134                 .maxlen         = sizeof(unsigned int),
 135                 .mode           = 0644,
 136                 .proc_handler   = proc_dointvec,
 137         },
 138         { },
 139 };
 140
 141 static struct ctl_table sunrpc_table[] = {
 142         {
 143                 .procname       = "sunrpc",
 144                 .mode           = 0555,
 145                 .child          = xr_tunables_table
 146         },
 147         { },
 148 };
 149
 150 #endif
 151
 152 static struct rpc_xprt_ops xprt_rdma_procs;     /*forward reference */
 153
 154 static void
 155 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
 156 {
 157         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
 158         char buf[20];
 159
 160         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
 161         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 162
 163         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
 164 }
 165
 166 static void
 167 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
 168 {
 169         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
 170         char buf[40];
 171
 172         snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
 173         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 174
 175         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
 176 }
 177
 178 void
 179 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
 180 {
 181         char buf[128];
 182
 183         switch (sap->sa_family) {
 184         case AF_INET:
 185                 xprt_rdma_format_addresses4(xprt, sap);
 186                 break;
 187         case AF_INET6:
 188                 xprt_rdma_format_addresses6(xprt, sap);
 189                 break;
 190         default:
 191                 pr_err("rpcrdma: Unrecognized address family\n");
 192                 return;
 193         }
 194
 195         (void)rpc_ntop(sap, buf, sizeof(buf));
 196         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
 197
 198         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
 199         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 200
 201         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
 202         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 203
 204         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
 205 }
 206
 207 void
 208 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
 209 {
 210         unsigned int i;
 211
 212         for (i = 0; i < RPC_DISPLAY_MAX; i++)
 213                 switch (i) {
 214                 case RPC_DISPLAY_PROTO:
 215                 case RPC_DISPLAY_NETID:
 216                         continue;
 217                 default:
 218                         kfree(xprt->address_strings[i]);
 219                 }
 220 }
 221
 222 void
 223 rpcrdma_conn_func(struct rpcrdma_ep *ep)
 224 {
 225         schedule_delayed_work(&ep->rep_connect_worker, 0);
 226 }
 227
 228 void
 229 rpcrdma_connect_worker(struct work_struct *work)
 230 {
 231         struct rpcrdma_ep *ep =
 232                 container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
 233         struct rpcrdma_xprt *r_xprt =
 234                 container_of(ep, struct rpcrdma_xprt, rx_ep);
 235         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 236
 237         spin_lock_bh(&xprt->transport_lock);
 238         if (++xprt->connect_cookie == 0)        /* maintain a reserved value */
 239                 ++xprt->connect_cookie;
 240         if (ep->rep_connected > 0) {
 241                 if (!xprt_test_and_set_connected(xprt))
 242                         xprt_wake_pending_tasks(xprt, 0);
 243         } else {
 244                 if (xprt_test_and_clear_connected(xprt))
 245                         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 246         }
 247         spin_unlock_bh(&xprt->transport_lock);
 248 }
 249
 250 static void
 251 xprt_rdma_connect_worker(struct work_struct *work)
 252 {
 253         struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
 254                                                    rx_connect_worker.work);
 255         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 256         int rc = 0;
 257
 258         xprt_clear_connected(xprt);
 259
 260         dprintk("RPC:       %s: %sconnect\n", __func__,
 261                         r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
 262         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 263         if (rc)
 264                 xprt_wake_pending_tasks(xprt, rc);
 265
 266         dprintk("RPC:       %s: exit\n", __func__);
 267         xprt_clear_connecting(xprt);
 268 }
 269
 270 static void
 271 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
 272 {
 273         struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
 274                                                    rx_xprt);
 275
 276         pr_info("rpcrdma: injecting transport disconnect on xprt=%p\n", xprt);
 277         rdma_disconnect(r_xprt->rx_ia.ri_id);
 278 }
 279
 280 /*
 281  * xprt_rdma_destroy
 282  *
 283  * Destroy the xprt.
 284  * Free all memory associated with the object, including its own.
 285  * NOTE: none of the *destroy methods free memory for their top-level
 286  * objects, even though they may have allocated it (they do free
 287  * private memory). It's up to the caller to handle it. In this
 288  * case (RDMA transport), all structure memory is inlined with the
 289  * struct rpcrdma_xprt.
 290  */
 291 static void
 292 xprt_rdma_destroy(struct rpc_xprt *xprt)
 293 {
 294         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 295
 296         dprintk("RPC:       %s: called\n", __func__);
 297
 298         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
 299
 300         xprt_clear_connected(xprt);
 301
 302         rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
 303         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 304         rpcrdma_ia_close(&r_xprt->rx_ia);
 305
 306         xprt_rdma_free_addresses(xprt);
 307
 308         xprt_free(xprt);
 309
 310         dprintk("RPC:       %s: returning\n", __func__);
 311
 312         module_put(THIS_MODULE);
 313 }
 314
 315 static const struct rpc_timeout xprt_rdma_default_timeout = {
 316         .to_initval = 60 * HZ,
 317         .to_maxval = 60 * HZ,
 318 };
 319
 320 /**
 321  * xprt_setup_rdma - Set up transport to use RDMA
 322  *
 323  * @args: rpc transport arguments
 324  */
 325 static struct rpc_xprt *
 326 xprt_setup_rdma(struct xprt_create *args)
 327 {
 328         struct rpcrdma_create_data_internal cdata;
 329         struct rpc_xprt *xprt;
 330         struct rpcrdma_xprt *new_xprt;
 331         struct rpcrdma_ep *new_ep;
 332         struct sockaddr *sap;
 333         int rc;
 334
 335         if (args->addrlen > sizeof(xprt->addr)) {
 336                 dprintk("RPC:       %s: address too large\n", __func__);
 337                 return ERR_PTR(-EBADF);
 338         }
 339
 340         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
 341                         xprt_rdma_slot_table_entries,
 342                         xprt_rdma_slot_table_entries);
 343         if (xprt == NULL) {
 344                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
 345                         __func__);
 346                 return ERR_PTR(-ENOMEM);
 347         }
 348
 349         /* 60 second timeout, no retries */
 350         xprt->timeout = &xprt_rdma_default_timeout;
 351         xprt->bind_timeout = RPCRDMA_BIND_TO;
 352         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 353         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
 354
 355         xprt->resvport = 0;             /* privileged port not needed */
 356         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
 357         xprt->ops = &xprt_rdma_procs;
 358
 359         /*
 360          * Set up RDMA-specific connect data.
 361          */
 362
 363         sap = (struct sockaddr *)&cdata.addr;
 364         memcpy(sap, args->dstaddr, args->addrlen);
 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
 375         cdata.max_requests = xprt->max_reqs;
 376
 377         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
 378         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
 379
 380         cdata.inline_wsize = xprt_rdma_max_inline_write;
 381         if (cdata.inline_wsize > cdata.wsize)
 382                 cdata.inline_wsize = cdata.wsize;
 383
 384         cdata.inline_rsize = xprt_rdma_max_inline_read;
 385         if (cdata.inline_rsize > cdata.rsize)
 386                 cdata.inline_rsize = cdata.rsize;
 387
 388         cdata.padding = xprt_rdma_inline_write_padding;
 389
 390         /*
 391          * Create new transport instance, which includes initialized
 392          *  o ia
 393          *  o endpoint
 394          *  o buffers
 395          */
 396
 397         new_xprt = rpcx_to_rdmax(xprt);
 398
 399         rc = rpcrdma_ia_open(new_xprt, sap, xprt_rdma_memreg_strategy);
 400         if (rc)
 401                 goto out1;
 402
 403         /*
 404          * initialize and create ep
 405          */
 406         new_xprt->rx_data = cdata;
 407         new_ep = &new_xprt->rx_ep;
 408         new_ep->rep_remote_addr = cdata.addr;
 409
 410         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
 411                                 &new_xprt->rx_ia, &new_xprt->rx_data);
 412         if (rc)
 413                 goto out2;
 414
 415         /*
 416          * Allocate pre-registered send and receive buffers for headers and
 417          * any inline data. Also specify any padding which will be provided
 418          * from a preregistered zero buffer.
 419          */
 420         rc = rpcrdma_buffer_create(new_xprt);
 421         if (rc)
 422                 goto out3;
 423
 424         /*
 425          * Register a callback for connection events. This is necessary because
 426          * connection loss notification is async. We also catch connection loss
 427          * when reaping receives.
 428          */
 429         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
 430                           xprt_rdma_connect_worker);
 431
 432         xprt_rdma_format_addresses(xprt, sap);
 433         xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
 434         if (xprt->max_payload == 0)
 435                 goto out4;
 436         xprt->max_payload <<= PAGE_SHIFT;
 437         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
 438                 __func__, xprt->max_payload);
 439
 440         if (!try_module_get(THIS_MODULE))
 441                 goto out4;
 442
 443         dprintk("RPC:       %s: %s:%s\n", __func__,
 444                 xprt->address_strings[RPC_DISPLAY_ADDR],
 445                 xprt->address_strings[RPC_DISPLAY_PORT]);
 446         return xprt;
 447
 448 out4:
 449         xprt_rdma_free_addresses(xprt);
 450         rc = -EINVAL;
 451 out3:
 452         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
 453 out2:
 454         rpcrdma_ia_close(&new_xprt->rx_ia);
 455 out1:
 456         xprt_free(xprt);
 457         return ERR_PTR(rc);
 458 }
 459
 460 /*
 461  * Close a connection, during shutdown or timeout/reconnect
 462  */
 463 static void
 464 xprt_rdma_close(struct rpc_xprt *xprt)
 465 {
 466         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 467
 468         dprintk("RPC:       %s: closing\n", __func__);
 469         if (r_xprt->rx_ep.rep_connected > 0)
 470                 xprt->reestablish_timeout = 0;
 471         xprt_disconnect_done(xprt);
 472         rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 473 }
 474
 475 static void
 476 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 477 {
 478         struct sockaddr_in *sap;
 479
 480         sap = (struct sockaddr_in *)&xprt->addr;
 481         sap->sin_port = htons(port);
 482         sap = (struct sockaddr_in *)&rpcx_to_rdmad(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         return true;
 549 }
 550
 551 static bool
 552 rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 553                     size_t size, gfp_t flags)
 554 {
 555         struct rpcrdma_regbuf *rb;
 556
 557         if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
 558                 return true;
 559
 560         rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
 561         if (IS_ERR(rb))
 562                 return false;
 563
 564         rpcrdma_free_regbuf(req->rl_sendbuf);
 565         r_xprt->rx_stats.hardway_register_count += size;
 566         req->rl_sendbuf = rb;
 567         return true;
 568 }
 569
 570 /* The rq_rcv_buf is used only if a Reply chunk is necessary.
 571  * The decision to use a Reply chunk is made later in
 572  * rpcrdma_marshal_req. This buffer is registered at that time.
 573  *
 574  * Otherwise, the associated RPC Reply arrives in a separate
 575  * Receive buffer, arbitrarily chosen by the HCA. The buffer
 576  * allocated here for the RPC Reply is not utilized in that
 577  * case. See rpcrdma_inline_fixup.
 578  *
 579  * A regbuf is used here to remember the buffer size.
 580  */
 581 static bool
 582 rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 583                     size_t size, gfp_t flags)
 584 {
 585         struct rpcrdma_regbuf *rb;
 586
 587         if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
 588                 return true;
 589
 590         rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
 591         if (IS_ERR(rb))
 592                 return false;
 593
 594         rpcrdma_free_regbuf(req->rl_recvbuf);
 595         r_xprt->rx_stats.hardway_register_count += size;
 596         req->rl_recvbuf = rb;
 597         return true;
 598 }
 599
 600 /**
 601  * xprt_rdma_allocate - allocate transport resources for an RPC
 602  * @task: RPC task
 603  *
 604  * Return values:
 605  *        0:    Success; rq_buffer points to RPC buffer to use
 606  *   ENOMEM:    Out of memory, call again later
 607  *      EIO:    A permanent error occurred, do not retry
 608  *
 609  * The RDMA allocate/free functions need the task structure as a place
 610  * to hide the struct rpcrdma_req, which is necessary for the actual
 611  * send/recv sequence.
 612  *
 613  * xprt_rdma_allocate provides buffers that are already mapped for
 614  * DMA, and a local DMA lkey is provided for each.
 615  */
 616 static int
 617 xprt_rdma_allocate(struct rpc_task *task)
 618 {
 619         struct rpc_rqst *rqst = task->tk_rqstp;
 620         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 621         struct rpcrdma_req *req;
 622         gfp_t flags;
 623
 624         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
 625         if (req == NULL)
 626                 return -ENOMEM;
 627
 628         flags = RPCRDMA_DEF_GFP;
 629         if (RPC_IS_SWAPPER(task))
 630                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
 631
 632         if (!rpcrdma_get_rdmabuf(r_xprt, req, flags))
 633                 goto out_fail;
 634         if (!rpcrdma_get_sendbuf(r_xprt, req, rqst->rq_callsize, flags))
 635                 goto out_fail;
 636         if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
 637                 goto out_fail;
 638
 639         dprintk("RPC: %5u %s: send size = %zd, recv size = %zd, req = %p\n",
 640                 task->tk_pid, __func__, rqst->rq_callsize,
 641                 rqst->rq_rcvsize, req);
 642
 643         req->rl_connect_cookie = 0;     /* our reserved value */
 644         rpcrdma_set_xprtdata(rqst, req);
 645         rqst->rq_buffer = req->rl_sendbuf->rg_base;
 646         rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
 647         return 0;
 648
 649 out_fail:
 650         rpcrdma_buffer_put(req);
 651         return -ENOMEM;
 652 }
 653
 654 /**
 655  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
 656  * @task: RPC task
 657  *
 658  * Caller guarantees rqst->rq_buffer is non-NULL.
 659  */
 660 static void
 661 xprt_rdma_free(struct rpc_task *task)
 662 {
 663         struct rpc_rqst *rqst = task->tk_rqstp;
 664         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 665         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 666         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 667
 668         if (req->rl_backchannel)
 669                 return;
 670
 671         dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
 672
 673         if (unlikely(!list_empty(&req->rl_registered)))
 674                 ia->ri_ops->ro_unmap_safe(r_xprt, req, !RPC_IS_ASYNC(task));
 675         rpcrdma_unmap_sges(ia, 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  * Return values:
 684  *        0:    The request has been sent
 685  * ENOTCONN:    Caller needs to invoke connect logic then call again
 686  *  ENOBUFS:    Call again later to send the request
 687  *      EIO:    A permanent error occurred. The request was not sent,
 688  *              and don't try it again
 689  *
 690  * send_request invokes the meat of RPC RDMA. It must do the following:
 691  *
 692  *  1.  Marshal the RPC request into an RPC RDMA request, which means
 693  *      putting a header in front of data, and creating IOVs for RDMA
 694  *      from those in the request.
 695  *  2.  In marshaling, detect opportunities for RDMA, and use them.
 696  *  3.  Post a recv message to set up asynch completion, then send
 697  *      the request (rpcrdma_ep_post).
 698  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
 699  */
 700 static int
 701 xprt_rdma_send_request(struct rpc_task *task)
 702 {
 703         struct rpc_rqst *rqst = task->tk_rqstp;
 704         struct rpc_xprt *xprt = rqst->rq_xprt;
 705         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 706         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 707         int rc = 0;
 708
 709         /* On retransmit, remove any previously registered chunks */
 710         if (unlikely(!list_empty(&req->rl_registered)))
 711                 r_xprt->rx_ia.ri_ops->ro_unmap_safe(r_xprt, req, false);
 712
 713         rc = rpcrdma_marshal_req(rqst);
 714         if (rc < 0)
 715                 goto failed_marshal;
 716
 717         if (req->rl_reply == NULL)              /* e.g. reconnection */
 718                 rpcrdma_recv_buffer_get(req);
 719
 720         /* Must suppress retransmit to maintain credits */
 721         if (req->rl_connect_cookie == xprt->connect_cookie)
 722                 goto drop_connection;
 723         req->rl_connect_cookie = xprt->connect_cookie;
 724
 725         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 726                 goto drop_connection;
 727
 728         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
 729         rqst->rq_bytes_sent = 0;
 730         return 0;
 731
 732 failed_marshal:
 733         if (rc != -ENOTCONN)
 734                 return rc;
 735 drop_connection:
 736         xprt_disconnect_done(xprt);
 737         return -ENOTCONN;       /* implies disconnect */
 738 }
 739
 740 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 741 {
 742         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 743         long idle_time = 0;
 744
 745         if (xprt_connected(xprt))
 746                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 747
 748         seq_puts(seq, "\txprt:\trdma ");
 749         seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
 750                    0,   /* need a local port? */
 751                    xprt->stat.bind_count,
 752                    xprt->stat.connect_count,
 753                    xprt->stat.connect_time,
 754                    idle_time,
 755                    xprt->stat.sends,
 756                    xprt->stat.recvs,
 757                    xprt->stat.bad_xids,
 758                    xprt->stat.req_u,
 759                    xprt->stat.bklog_u);
 760         seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
 761                    r_xprt->rx_stats.read_chunk_count,
 762                    r_xprt->rx_stats.write_chunk_count,
 763                    r_xprt->rx_stats.reply_chunk_count,
 764                    r_xprt->rx_stats.total_rdma_request,
 765                    r_xprt->rx_stats.total_rdma_reply,
 766                    r_xprt->rx_stats.pullup_copy_count,
 767                    r_xprt->rx_stats.fixup_copy_count,
 768                    r_xprt->rx_stats.hardway_register_count,
 769                    r_xprt->rx_stats.failed_marshal_count,
 770                    r_xprt->rx_stats.bad_reply_count,
 771                    r_xprt->rx_stats.nomsg_call_count);
 772         seq_printf(seq, "%lu %lu %lu %lu\n",
 773                    r_xprt->rx_stats.mrs_recovered,
 774                    r_xprt->rx_stats.mrs_orphaned,
 775                    r_xprt->rx_stats.mrs_allocated,
 776                    r_xprt->rx_stats.local_inv_needed);
 777 }
 778
 779 static int
 780 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
 781 {
 782         return 0;
 783 }
 784
 785 static void
 786 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
 787 {
 788 }
 789
 790 /*
 791  * Plumbing for rpc transport switch and kernel module
 792  */
 793
 794 static struct rpc_xprt_ops xprt_rdma_procs = {
 795         .reserve_xprt           = xprt_reserve_xprt_cong,
 796         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 797         .alloc_slot             = xprt_alloc_slot,
 798         .release_request        = xprt_release_rqst_cong,       /* ditto */
 799         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
 800         .timer                  = xprt_rdma_timer,
 801         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 802         .set_port               = xprt_rdma_set_port,
 803         .connect                = xprt_rdma_connect,
 804         .buf_alloc              = xprt_rdma_allocate,
 805         .buf_free               = xprt_rdma_free,
 806         .send_request           = xprt_rdma_send_request,
 807         .close                  = xprt_rdma_close,
 808         .destroy                = xprt_rdma_destroy,
 809         .print_stats            = xprt_rdma_print_stats,
 810         .enable_swap            = xprt_rdma_enable_swap,
 811         .disable_swap           = xprt_rdma_disable_swap,
 812         .inject_disconnect      = xprt_rdma_inject_disconnect,
 813 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
 814         .bc_setup               = xprt_rdma_bc_setup,
 815         .bc_up                  = xprt_rdma_bc_up,
 816         .bc_maxpayload          = xprt_rdma_bc_maxpayload,
 817         .bc_free_rqst           = xprt_rdma_bc_free_rqst,
 818         .bc_destroy             = xprt_rdma_bc_destroy,
 819 #endif
 820 };
 821
 822 static struct xprt_class xprt_rdma = {
 823         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 824         .name                   = "rdma",
 825         .owner                  = THIS_MODULE,
 826         .ident                  = XPRT_TRANSPORT_RDMA,
 827         .setup                  = xprt_setup_rdma,
 828 };
 829
 830 void xprt_rdma_cleanup(void)
 831 {
 832         int rc;
 833
 834         dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
 835 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 836         if (sunrpc_table_header) {
 837                 unregister_sysctl_table(sunrpc_table_header);
 838                 sunrpc_table_header = NULL;
 839         }
 840 #endif
 841         rc = xprt_unregister_transport(&xprt_rdma);
 842         if (rc)
 843                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
 844                         __func__, rc);
 845
 846         rpcrdma_destroy_wq();
 847
 848         rc = xprt_unregister_transport(&xprt_rdma_bc);
 849         if (rc)
 850                 dprintk("RPC:       %s: xprt_unregister(bc) returned %i\n",
 851                         __func__, rc);
 852 }
 853
 854 int xprt_rdma_init(void)
 855 {
 856         int rc;
 857
 858         rc = rpcrdma_alloc_wq();
 859         if (rc)
 860                 return rc;
 861
 862         rc = xprt_register_transport(&xprt_rdma);
 863         if (rc) {
 864                 rpcrdma_destroy_wq();
 865                 return rc;
 866         }
 867
 868         rc = xprt_register_transport(&xprt_rdma_bc);
 869         if (rc) {
 870                 xprt_unregister_transport(&xprt_rdma);
 871                 rpcrdma_destroy_wq();
 872                 return rc;
 873         }
 874
 875         dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
 876
 877         dprintk("Defaults:\n");
 878         dprintk("\tSlots %d\n"
 879                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
 880                 xprt_rdma_slot_table_entries,
 881                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
 882         dprintk("\tPadding %d\n\tMemreg %d\n",
 883                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
 884
 885 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 886         if (!sunrpc_table_header)
 887                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 888 #endif
 889         return 0;
 890 }