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
  56 #include "xprt_rdma.h"
  57
  58 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  59 # define RPCDBG_FACILITY        RPCDBG_TRANS
  60 #endif
  61
  62 /*
  63  * tunables
  64  */
  65
  66 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
  67 unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
  68 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
  69 static unsigned int xprt_rdma_inline_write_padding;
  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
  84 static struct ctl_table_header *sunrpc_table_header;
  85
  86 static struct ctl_table xr_tunables_table[] = {
  87         {
  88                 .procname       = "rdma_slot_table_entries",
  89                 .data           = &xprt_rdma_slot_table_entries,
  90                 .maxlen         = sizeof(unsigned int),
  91                 .mode           = 0644,
  92                 .proc_handler   = proc_dointvec_minmax,
  93                 .extra1         = &min_slot_table_size,
  94                 .extra2         = &max_slot_table_size
  95         },
  96         {
  97                 .procname       = "rdma_max_inline_read",
  98                 .data           = &xprt_rdma_max_inline_read,
  99                 .maxlen         = sizeof(unsigned int),
 100                 .mode           = 0644,
 101                 .proc_handler   = proc_dointvec_minmax,
 102                 .extra1         = &min_inline_size,
 103                 .extra2         = &max_inline_size,
 104         },
 105         {
 106                 .procname       = "rdma_max_inline_write",
 107                 .data           = &xprt_rdma_max_inline_write,
 108                 .maxlen         = sizeof(unsigned int),
 109                 .mode           = 0644,
 110                 .proc_handler   = proc_dointvec_minmax,
 111                 .extra1         = &min_inline_size,
 112                 .extra2         = &max_inline_size,
 113         },
 114         {
 115                 .procname       = "rdma_inline_write_padding",
 116                 .data           = &xprt_rdma_inline_write_padding,
 117                 .maxlen         = sizeof(unsigned int),
 118                 .mode           = 0644,
 119                 .proc_handler   = proc_dointvec_minmax,
 120                 .extra1         = &zero,
 121                 .extra2         = &max_padding,
 122         },
 123         {
 124                 .procname       = "rdma_memreg_strategy",
 125                 .data           = &xprt_rdma_memreg_strategy,
 126                 .maxlen         = sizeof(unsigned int),
 127                 .mode           = 0644,
 128                 .proc_handler   = proc_dointvec_minmax,
 129                 .extra1         = &min_memreg,
 130                 .extra2         = &max_memreg,
 131         },
 132         {
 133                 .procname       = "rdma_pad_optimize",
 134                 .data           = &xprt_rdma_pad_optimize,
 135                 .maxlen         = sizeof(unsigned int),
 136                 .mode           = 0644,
 137                 .proc_handler   = proc_dointvec,
 138         },
 139         { },
 140 };
 141
 142 static struct ctl_table sunrpc_table[] = {
 143         {
 144                 .procname       = "sunrpc",
 145                 .mode           = 0555,
 146                 .child          = xr_tunables_table
 147         },
 148         { },
 149 };
 150
 151 #endif
 152
 153 static const struct rpc_xprt_ops xprt_rdma_procs;
 154
 155 static void
 156 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
 157 {
 158         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
 159         char buf[20];
 160
 161         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
 162         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 163
 164         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
 165 }
 166
 167 static void
 168 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
 169 {
 170         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
 171         char buf[40];
 172
 173         snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
 174         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 175
 176         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
 177 }
 178
 179 void
 180 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
 181 {
 182         char buf[128];
 183
 184         switch (sap->sa_family) {
 185         case AF_INET:
 186                 xprt_rdma_format_addresses4(xprt, sap);
 187                 break;
 188         case AF_INET6:
 189                 xprt_rdma_format_addresses6(xprt, sap);
 190                 break;
 191         default:
 192                 pr_err("rpcrdma: Unrecognized address family\n");
 193                 return;
 194         }
 195
 196         (void)rpc_ntop(sap, buf, sizeof(buf));
 197         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
 198
 199         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
 200         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 201
 202         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
 203         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 204
 205         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
 206 }
 207
 208 void
 209 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
 210 {
 211         unsigned int i;
 212
 213         for (i = 0; i < RPC_DISPLAY_MAX; i++)
 214                 switch (i) {
 215                 case RPC_DISPLAY_PROTO:
 216                 case RPC_DISPLAY_NETID:
 217                         continue;
 218                 default:
 219                         kfree(xprt->address_strings[i]);
 220                 }
 221 }
 222
 223 void
 224 rpcrdma_conn_func(struct rpcrdma_ep *ep)
 225 {
 226         schedule_delayed_work(&ep->rep_connect_worker, 0);
 227 }
 228
 229 void
 230 rpcrdma_connect_worker(struct work_struct *work)
 231 {
 232         struct rpcrdma_ep *ep =
 233                 container_of(work, struct rpcrdma_ep, rep_connect_worker.work);
 234         struct rpcrdma_xprt *r_xprt =
 235                 container_of(ep, struct rpcrdma_xprt, rx_ep);
 236         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 237
 238         spin_lock_bh(&xprt->transport_lock);
 239         if (++xprt->connect_cookie == 0)        /* maintain a reserved value */
 240                 ++xprt->connect_cookie;
 241         if (ep->rep_connected > 0) {
 242                 if (!xprt_test_and_set_connected(xprt))
 243                         xprt_wake_pending_tasks(xprt, 0);
 244         } else {
 245                 if (xprt_test_and_clear_connected(xprt))
 246                         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 247         }
 248         spin_unlock_bh(&xprt->transport_lock);
 249 }
 250
 251 static void
 252 xprt_rdma_connect_worker(struct work_struct *work)
 253 {
 254         struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
 255                                                    rx_connect_worker.work);
 256         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 257         int rc = 0;
 258
 259         xprt_clear_connected(xprt);
 260
 261         dprintk("RPC:       %s: %sconnect\n", __func__,
 262                         r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
 263         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 264         if (rc)
 265                 xprt_wake_pending_tasks(xprt, rc);
 266
 267         dprintk("RPC:       %s: exit\n", __func__);
 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         pr_info("rpcrdma: injecting transport disconnect on xprt=%p\n", 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         dprintk("RPC:       %s: called\n", __func__);
 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
 309         xprt_free(xprt);
 310
 311         dprintk("RPC:       %s: returning\n", __func__);
 312
 313         module_put(THIS_MODULE);
 314 }
 315
 316 static const struct rpc_timeout xprt_rdma_default_timeout = {
 317         .to_initval = 60 * HZ,
 318         .to_maxval = 60 * HZ,
 319 };
 320
 321 /**
 322  * xprt_setup_rdma - Set up transport to use RDMA
 323  *
 324  * @args: rpc transport arguments
 325  */
 326 static struct rpc_xprt *
 327 xprt_setup_rdma(struct xprt_create *args)
 328 {
 329         struct rpcrdma_create_data_internal cdata;
 330         struct rpc_xprt *xprt;
 331         struct rpcrdma_xprt *new_xprt;
 332         struct rpcrdma_ep *new_ep;
 333         struct sockaddr *sap;
 334         int rc;
 335
 336         if (args->addrlen > sizeof(xprt->addr)) {
 337                 dprintk("RPC:       %s: address too large\n", __func__);
 338                 return ERR_PTR(-EBADF);
 339         }
 340
 341         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
 342                         xprt_rdma_slot_table_entries,
 343                         xprt_rdma_slot_table_entries);
 344         if (xprt == NULL) {
 345                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
 346                         __func__);
 347                 return ERR_PTR(-ENOMEM);
 348         }
 349
 350         /* 60 second timeout, no retries */
 351         xprt->timeout = &xprt_rdma_default_timeout;
 352         xprt->bind_timeout = RPCRDMA_BIND_TO;
 353         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 354         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
 355
 356         xprt->resvport = 0;             /* privileged port not needed */
 357         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
 358         xprt->ops = &xprt_rdma_procs;
 359
 360         /*
 361          * Set up RDMA-specific connect data.
 362          */
 363
 364         sap = (struct sockaddr *)&cdata.addr;
 365         memcpy(sap, args->dstaddr, args->addrlen);
 366
 367         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 368          * address, for any side protocols which peek at it */
 369         xprt->prot = IPPROTO_TCP;
 370         xprt->addrlen = args->addrlen;
 371         memcpy(&xprt->addr, sap, xprt->addrlen);
 372
 373         if (rpc_get_port(sap))
 374                 xprt_set_bound(xprt);
 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         cdata.padding = xprt_rdma_inline_write_padding;
 390
 391         /*
 392          * Create new transport instance, which includes initialized
 393          *  o ia
 394          *  o endpoint
 395          *  o buffers
 396          */
 397
 398         new_xprt = rpcx_to_rdmax(xprt);
 399
 400         rc = rpcrdma_ia_open(new_xprt, sap);
 401         if (rc)
 402                 goto out1;
 403
 404         /*
 405          * initialize and create ep
 406          */
 407         new_xprt->rx_data = cdata;
 408         new_ep = &new_xprt->rx_ep;
 409         new_ep->rep_remote_addr = cdata.addr;
 410
 411         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
 412                                 &new_xprt->rx_ia, &new_xprt->rx_data);
 413         if (rc)
 414                 goto out2;
 415
 416         /*
 417          * Allocate pre-registered send and receive buffers for headers and
 418          * any inline data. Also specify any padding which will be provided
 419          * from a preregistered zero buffer.
 420          */
 421         rc = rpcrdma_buffer_create(new_xprt);
 422         if (rc)
 423                 goto out3;
 424
 425         /*
 426          * Register a callback for connection events. This is necessary because
 427          * connection loss notification is async. We also catch connection loss
 428          * when reaping receives.
 429          */
 430         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
 431                           xprt_rdma_connect_worker);
 432
 433         xprt_rdma_format_addresses(xprt, sap);
 434         xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
 435         if (xprt->max_payload == 0)
 436                 goto out4;
 437         xprt->max_payload <<= PAGE_SHIFT;
 438         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
 439                 __func__, xprt->max_payload);
 440
 441         if (!try_module_get(THIS_MODULE))
 442                 goto out4;
 443
 444         dprintk("RPC:       %s: %s:%s\n", __func__,
 445                 xprt->address_strings[RPC_DISPLAY_ADDR],
 446                 xprt->address_strings[RPC_DISPLAY_PORT]);
 447         return xprt;
 448
 449 out4:
 450         xprt_rdma_free_addresses(xprt);
 451         rc = -EINVAL;
 452 out3:
 453         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
 454 out2:
 455         rpcrdma_ia_close(&new_xprt->rx_ia);
 456 out1:
 457         xprt_free(xprt);
 458         return ERR_PTR(rc);
 459 }
 460
 461 /**
 462  * xprt_rdma_close - Close down RDMA connection
 463  * @xprt: generic transport to be closed
 464  *
 465  * Called during transport shutdown reconnect, or device
 466  * removal. Caller holds the transport's write lock.
 467  */
 468 static void
 469 xprt_rdma_close(struct rpc_xprt *xprt)
 470 {
 471         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 472         struct rpcrdma_ep *ep = &r_xprt->rx_ep;
 473         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 474
 475         dprintk("RPC:       %s: closing xprt %p\n", __func__, xprt);
 476
 477         if (test_and_clear_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags)) {
 478                 xprt_clear_connected(xprt);
 479                 rpcrdma_ia_remove(ia);
 480                 return;
 481         }
 482         if (ep->rep_connected == -ENODEV)
 483                 return;
 484         if (ep->rep_connected > 0)
 485                 xprt->reestablish_timeout = 0;
 486         xprt_disconnect_done(xprt);
 487         rpcrdma_ep_disconnect(ep, ia);
 488 }
 489
 490 static void
 491 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 492 {
 493         struct sockaddr_in *sap;
 494
 495         sap = (struct sockaddr_in *)&xprt->addr;
 496         sap->sin_port = htons(port);
 497         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
 498         sap->sin_port = htons(port);
 499         dprintk("RPC:       %s: %u\n", __func__, port);
 500 }
 501
 502 /**
 503  * xprt_rdma_timer - invoked when an RPC times out
 504  * @xprt: controlling RPC transport
 505  * @task: RPC task that timed out
 506  *
 507  * Invoked when the transport is still connected, but an RPC
 508  * retransmit timeout occurs.
 509  *
 510  * Since RDMA connections don't have a keep-alive, forcibly
 511  * disconnect and retry to connect. This drives full
 512  * detection of the network path, and retransmissions of
 513  * all pending RPCs.
 514  */
 515 static void
 516 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task)
 517 {
 518         dprintk("RPC: %5u %s: xprt = %p\n", task->tk_pid, __func__, xprt);
 519
 520         xprt_force_disconnect(xprt);
 521 }
 522
 523 static void
 524 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
 525 {
 526         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 527
 528         if (r_xprt->rx_ep.rep_connected != 0) {
 529                 /* Reconnect */
 530                 schedule_delayed_work(&r_xprt->rx_connect_worker,
 531                                       xprt->reestablish_timeout);
 532                 xprt->reestablish_timeout <<= 1;
 533                 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
 534                         xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
 535                 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
 536                         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 537         } else {
 538                 schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
 539                 if (!RPC_IS_ASYNC(task))
 540                         flush_delayed_work(&r_xprt->rx_connect_worker);
 541         }
 542 }
 543
 544 /* Allocate a fixed-size buffer in which to construct and send the
 545  * RPC-over-RDMA header for this request.
 546  */
 547 static bool
 548 rpcrdma_get_rdmabuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 549                     gfp_t flags)
 550 {
 551         size_t size = RPCRDMA_HDRBUF_SIZE;
 552         struct rpcrdma_regbuf *rb;
 553
 554         if (req->rl_rdmabuf)
 555                 return true;
 556
 557         rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
 558         if (IS_ERR(rb))
 559                 return false;
 560
 561         r_xprt->rx_stats.hardway_register_count += size;
 562         req->rl_rdmabuf = rb;
 563         xdr_buf_init(&req->rl_hdrbuf, rb->rg_base, rdmab_length(rb));
 564         return true;
 565 }
 566
 567 static bool
 568 rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 569                     size_t size, gfp_t flags)
 570 {
 571         struct rpcrdma_regbuf *rb;
 572
 573         if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
 574                 return true;
 575
 576         rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
 577         if (IS_ERR(rb))
 578                 return false;
 579
 580         rpcrdma_free_regbuf(req->rl_sendbuf);
 581         r_xprt->rx_stats.hardway_register_count += size;
 582         req->rl_sendbuf = rb;
 583         return true;
 584 }
 585
 586 /* The rq_rcv_buf is used only if a Reply chunk is necessary.
 587  * The decision to use a Reply chunk is made later in
 588  * rpcrdma_marshal_req. This buffer is registered at that time.
 589  *
 590  * Otherwise, the associated RPC Reply arrives in a separate
 591  * Receive buffer, arbitrarily chosen by the HCA. The buffer
 592  * allocated here for the RPC Reply is not utilized in that
 593  * case. See rpcrdma_inline_fixup.
 594  *
 595  * A regbuf is used here to remember the buffer size.
 596  */
 597 static bool
 598 rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 599                     size_t size, gfp_t flags)
 600 {
 601         struct rpcrdma_regbuf *rb;
 602
 603         if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
 604                 return true;
 605
 606         rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
 607         if (IS_ERR(rb))
 608                 return false;
 609
 610         rpcrdma_free_regbuf(req->rl_recvbuf);
 611         r_xprt->rx_stats.hardway_register_count += size;
 612         req->rl_recvbuf = rb;
 613         return true;
 614 }
 615
 616 /**
 617  * xprt_rdma_allocate - allocate transport resources for an RPC
 618  * @task: RPC task
 619  *
 620  * Return values:
 621  *        0:    Success; rq_buffer points to RPC buffer to use
 622  *   ENOMEM:    Out of memory, call again later
 623  *      EIO:    A permanent error occurred, do not retry
 624  *
 625  * The RDMA allocate/free functions need the task structure as a place
 626  * to hide the struct rpcrdma_req, which is necessary for the actual
 627  * send/recv sequence.
 628  *
 629  * xprt_rdma_allocate provides buffers that are already mapped for
 630  * DMA, and a local DMA lkey is provided for each.
 631  */
 632 static int
 633 xprt_rdma_allocate(struct rpc_task *task)
 634 {
 635         struct rpc_rqst *rqst = task->tk_rqstp;
 636         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 637         struct rpcrdma_req *req;
 638         gfp_t flags;
 639
 640         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
 641         if (req == NULL)
 642                 return -ENOMEM;
 643
 644         flags = RPCRDMA_DEF_GFP;
 645         if (RPC_IS_SWAPPER(task))
 646                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
 647
 648         if (!rpcrdma_get_rdmabuf(r_xprt, req, flags))
 649                 goto out_fail;
 650         if (!rpcrdma_get_sendbuf(r_xprt, req, rqst->rq_callsize, flags))
 651                 goto out_fail;
 652         if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
 653                 goto out_fail;
 654
 655         dprintk("RPC: %5u %s: send size = %zd, recv size = %zd, req = %p\n",
 656                 task->tk_pid, __func__, rqst->rq_callsize,
 657                 rqst->rq_rcvsize, req);
 658
 659         req->rl_connect_cookie = 0;     /* our reserved value */
 660         rpcrdma_set_xprtdata(rqst, req);
 661         rqst->rq_buffer = req->rl_sendbuf->rg_base;
 662         rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
 663         return 0;
 664
 665 out_fail:
 666         rpcrdma_buffer_put(req);
 667         return -ENOMEM;
 668 }
 669
 670 /**
 671  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
 672  * @task: RPC task
 673  *
 674  * Caller guarantees rqst->rq_buffer is non-NULL.
 675  */
 676 static void
 677 xprt_rdma_free(struct rpc_task *task)
 678 {
 679         struct rpc_rqst *rqst = task->tk_rqstp;
 680         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 681         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 682
 683         if (test_bit(RPCRDMA_REQ_F_BACKCHANNEL, &req->rl_flags))
 684                 return;
 685
 686         dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
 687
 688         if (test_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags))
 689                 rpcrdma_release_rqst(r_xprt, req);
 690         rpcrdma_buffer_put(req);
 691 }
 692
 693 /**
 694  * xprt_rdma_send_request - marshal and send an RPC request
 695  * @task: RPC task with an RPC message in rq_snd_buf
 696  *
 697  * Caller holds the transport's write lock.
 698  *
 699  * Return values:
 700  *        0:    The request has been sent
 701  * ENOTCONN:    Caller needs to invoke connect logic then call again
 702  *  ENOBUFS:    Call again later to send the request
 703  *      EIO:    A permanent error occurred. The request was not sent,
 704  *              and don't try it again
 705  *
 706  * send_request invokes the meat of RPC RDMA. It must do the following:
 707  *
 708  *  1.  Marshal the RPC request into an RPC RDMA request, which means
 709  *      putting a header in front of data, and creating IOVs for RDMA
 710  *      from those in the request.
 711  *  2.  In marshaling, detect opportunities for RDMA, and use them.
 712  *  3.  Post a recv message to set up asynch completion, then send
 713  *      the request (rpcrdma_ep_post).
 714  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
 715  */
 716 static int
 717 xprt_rdma_send_request(struct rpc_task *task)
 718 {
 719         struct rpc_rqst *rqst = task->tk_rqstp;
 720         struct rpc_xprt *xprt = rqst->rq_xprt;
 721         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 722         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 723         int rc = 0;
 724
 725         if (!xprt_connected(xprt))
 726                 goto drop_connection;
 727
 728         /* On retransmit, remove any previously registered chunks */
 729         if (unlikely(!list_empty(&req->rl_registered)))
 730                 r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt,
 731                                                     &req->rl_registered);
 732
 733         rc = rpcrdma_marshal_req(r_xprt, rqst);
 734         if (rc < 0)
 735                 goto failed_marshal;
 736
 737         if (req->rl_reply == NULL)              /* e.g. reconnection */
 738                 rpcrdma_recv_buffer_get(req);
 739
 740         /* Must suppress retransmit to maintain credits */
 741         if (req->rl_connect_cookie == xprt->connect_cookie)
 742                 goto drop_connection;
 743         req->rl_connect_cookie = xprt->connect_cookie;
 744
 745         set_bit(RPCRDMA_REQ_F_PENDING, &req->rl_flags);
 746         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 747                 goto drop_connection;
 748
 749         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
 750         rqst->rq_bytes_sent = 0;
 751         return 0;
 752
 753 failed_marshal:
 754         if (rc != -ENOTCONN)
 755                 return rc;
 756 drop_connection:
 757         xprt_disconnect_done(xprt);
 758         return -ENOTCONN;       /* implies disconnect */
 759 }
 760
 761 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 762 {
 763         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 764         long idle_time = 0;
 765
 766         if (xprt_connected(xprt))
 767                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 768
 769         seq_puts(seq, "\txprt:\trdma ");
 770         seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
 771                    0,   /* need a local port? */
 772                    xprt->stat.bind_count,
 773                    xprt->stat.connect_count,
 774                    xprt->stat.connect_time,
 775                    idle_time,
 776                    xprt->stat.sends,
 777                    xprt->stat.recvs,
 778                    xprt->stat.bad_xids,
 779                    xprt->stat.req_u,
 780                    xprt->stat.bklog_u);
 781         seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
 782                    r_xprt->rx_stats.read_chunk_count,
 783                    r_xprt->rx_stats.write_chunk_count,
 784                    r_xprt->rx_stats.reply_chunk_count,
 785                    r_xprt->rx_stats.total_rdma_request,
 786                    r_xprt->rx_stats.total_rdma_reply,
 787                    r_xprt->rx_stats.pullup_copy_count,
 788                    r_xprt->rx_stats.fixup_copy_count,
 789                    r_xprt->rx_stats.hardway_register_count,
 790                    r_xprt->rx_stats.failed_marshal_count,
 791                    r_xprt->rx_stats.bad_reply_count,
 792                    r_xprt->rx_stats.nomsg_call_count);
 793         seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n",
 794                    r_xprt->rx_stats.mrs_recovered,
 795                    r_xprt->rx_stats.mrs_orphaned,
 796                    r_xprt->rx_stats.mrs_allocated,
 797                    r_xprt->rx_stats.local_inv_needed,
 798                    r_xprt->rx_stats.empty_sendctx_q,
 799                    r_xprt->rx_stats.reply_waits_for_send);
 800 }
 801
 802 static int
 803 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
 804 {
 805         return 0;
 806 }
 807
 808 static void
 809 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
 810 {
 811 }
 812
 813 /*
 814  * Plumbing for rpc transport switch and kernel module
 815  */
 816
 817 static const struct rpc_xprt_ops xprt_rdma_procs = {
 818         .reserve_xprt           = xprt_reserve_xprt_cong,
 819         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 820         .alloc_slot             = xprt_alloc_slot,
 821         .release_request        = xprt_release_rqst_cong,       /* ditto */
 822         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
 823         .timer                  = xprt_rdma_timer,
 824         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 825         .set_port               = xprt_rdma_set_port,
 826         .connect                = xprt_rdma_connect,
 827         .buf_alloc              = xprt_rdma_allocate,
 828         .buf_free               = xprt_rdma_free,
 829         .send_request           = xprt_rdma_send_request,
 830         .close                  = xprt_rdma_close,
 831         .destroy                = xprt_rdma_destroy,
 832         .print_stats            = xprt_rdma_print_stats,
 833         .enable_swap            = xprt_rdma_enable_swap,
 834         .disable_swap           = xprt_rdma_disable_swap,
 835         .inject_disconnect      = xprt_rdma_inject_disconnect,
 836 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
 837         .bc_setup               = xprt_rdma_bc_setup,
 838         .bc_up                  = xprt_rdma_bc_up,
 839         .bc_maxpayload          = xprt_rdma_bc_maxpayload,
 840         .bc_free_rqst           = xprt_rdma_bc_free_rqst,
 841         .bc_destroy             = xprt_rdma_bc_destroy,
 842 #endif
 843 };
 844
 845 static struct xprt_class xprt_rdma = {
 846         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 847         .name                   = "rdma",
 848         .owner                  = THIS_MODULE,
 849         .ident                  = XPRT_TRANSPORT_RDMA,
 850         .setup                  = xprt_setup_rdma,
 851 };
 852
 853 void xprt_rdma_cleanup(void)
 854 {
 855         int rc;
 856
 857         dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
 858 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 859         if (sunrpc_table_header) {
 860                 unregister_sysctl_table(sunrpc_table_header);
 861                 sunrpc_table_header = NULL;
 862         }
 863 #endif
 864         rc = xprt_unregister_transport(&xprt_rdma);
 865         if (rc)
 866                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
 867                         __func__, rc);
 868
 869         rpcrdma_destroy_wq();
 870
 871         rc = xprt_unregister_transport(&xprt_rdma_bc);
 872         if (rc)
 873                 dprintk("RPC:       %s: xprt_unregister(bc) returned %i\n",
 874                         __func__, rc);
 875 }
 876
 877 int xprt_rdma_init(void)
 878 {
 879         int rc;
 880
 881         rc = rpcrdma_alloc_wq();
 882         if (rc)
 883                 return rc;
 884
 885         rc = xprt_register_transport(&xprt_rdma);
 886         if (rc) {
 887                 rpcrdma_destroy_wq();
 888                 return rc;
 889         }
 890
 891         rc = xprt_register_transport(&xprt_rdma_bc);
 892         if (rc) {
 893                 xprt_unregister_transport(&xprt_rdma);
 894                 rpcrdma_destroy_wq();
 895                 return rc;
 896         }
 897
 898         dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
 899
 900         dprintk("Defaults:\n");
 901         dprintk("\tSlots %d\n"
 902                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
 903                 xprt_rdma_slot_table_entries,
 904                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
 905         dprintk("\tPadding %d\n\tMemreg %d\n",
 906                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
 907
 908 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 909         if (!sunrpc_table_header)
 910                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 911 #endif
 912         return 0;
 913 }