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 static void
 223 xprt_rdma_connect_worker(struct work_struct *work)
 224 {
 225         struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
 226                                                    rx_connect_worker.work);
 227         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 228         int rc = 0;
 229
 230         xprt_clear_connected(xprt);
 231
 232         dprintk("RPC:       %s: %sconnect\n", __func__,
 233                         r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
 234         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 235         if (rc)
 236                 xprt_wake_pending_tasks(xprt, rc);
 237
 238         dprintk("RPC:       %s: exit\n", __func__);
 239         xprt_clear_connecting(xprt);
 240 }
 241
 242 static void
 243 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
 244 {
 245         struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
 246                                                    rx_xprt);
 247
 248         pr_info("rpcrdma: injecting transport disconnect on xprt=%p\n", xprt);
 249         rdma_disconnect(r_xprt->rx_ia.ri_id);
 250 }
 251
 252 /*
 253  * xprt_rdma_destroy
 254  *
 255  * Destroy the xprt.
 256  * Free all memory associated with the object, including its own.
 257  * NOTE: none of the *destroy methods free memory for their top-level
 258  * objects, even though they may have allocated it (they do free
 259  * private memory). It's up to the caller to handle it. In this
 260  * case (RDMA transport), all structure memory is inlined with the
 261  * struct rpcrdma_xprt.
 262  */
 263 static void
 264 xprt_rdma_destroy(struct rpc_xprt *xprt)
 265 {
 266         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 267
 268         dprintk("RPC:       %s: called\n", __func__);
 269
 270         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
 271
 272         xprt_clear_connected(xprt);
 273
 274         rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
 275         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 276         rpcrdma_ia_close(&r_xprt->rx_ia);
 277
 278         xprt_rdma_free_addresses(xprt);
 279
 280         xprt_free(xprt);
 281
 282         dprintk("RPC:       %s: returning\n", __func__);
 283
 284         module_put(THIS_MODULE);
 285 }
 286
 287 static const struct rpc_timeout xprt_rdma_default_timeout = {
 288         .to_initval = 60 * HZ,
 289         .to_maxval = 60 * HZ,
 290 };
 291
 292 /**
 293  * xprt_setup_rdma - Set up transport to use RDMA
 294  *
 295  * @args: rpc transport arguments
 296  */
 297 static struct rpc_xprt *
 298 xprt_setup_rdma(struct xprt_create *args)
 299 {
 300         struct rpcrdma_create_data_internal cdata;
 301         struct rpc_xprt *xprt;
 302         struct rpcrdma_xprt *new_xprt;
 303         struct rpcrdma_ep *new_ep;
 304         struct sockaddr *sap;
 305         int rc;
 306
 307         if (args->addrlen > sizeof(xprt->addr)) {
 308                 dprintk("RPC:       %s: address too large\n", __func__);
 309                 return ERR_PTR(-EBADF);
 310         }
 311
 312         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
 313                         xprt_rdma_slot_table_entries,
 314                         xprt_rdma_slot_table_entries);
 315         if (xprt == NULL) {
 316                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
 317                         __func__);
 318                 return ERR_PTR(-ENOMEM);
 319         }
 320
 321         /* 60 second timeout, no retries */
 322         xprt->timeout = &xprt_rdma_default_timeout;
 323         xprt->bind_timeout = RPCRDMA_BIND_TO;
 324         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 325         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
 326
 327         xprt->resvport = 0;             /* privileged port not needed */
 328         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
 329         xprt->ops = &xprt_rdma_procs;
 330
 331         /*
 332          * Set up RDMA-specific connect data.
 333          */
 334
 335         sap = (struct sockaddr *)&cdata.addr;
 336         memcpy(sap, args->dstaddr, args->addrlen);
 337
 338         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 339          * address, for any side protocols which peek at it */
 340         xprt->prot = IPPROTO_TCP;
 341         xprt->addrlen = args->addrlen;
 342         memcpy(&xprt->addr, sap, xprt->addrlen);
 343
 344         if (rpc_get_port(sap))
 345                 xprt_set_bound(xprt);
 346
 347         cdata.max_requests = xprt->max_reqs;
 348
 349         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
 350         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
 351
 352         cdata.inline_wsize = xprt_rdma_max_inline_write;
 353         if (cdata.inline_wsize > cdata.wsize)
 354                 cdata.inline_wsize = cdata.wsize;
 355
 356         cdata.inline_rsize = xprt_rdma_max_inline_read;
 357         if (cdata.inline_rsize > cdata.rsize)
 358                 cdata.inline_rsize = cdata.rsize;
 359
 360         cdata.padding = xprt_rdma_inline_write_padding;
 361
 362         /*
 363          * Create new transport instance, which includes initialized
 364          *  o ia
 365          *  o endpoint
 366          *  o buffers
 367          */
 368
 369         new_xprt = rpcx_to_rdmax(xprt);
 370
 371         rc = rpcrdma_ia_open(new_xprt, sap, xprt_rdma_memreg_strategy);
 372         if (rc)
 373                 goto out1;
 374
 375         /*
 376          * initialize and create ep
 377          */
 378         new_xprt->rx_data = cdata;
 379         new_ep = &new_xprt->rx_ep;
 380         new_ep->rep_remote_addr = cdata.addr;
 381
 382         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
 383                                 &new_xprt->rx_ia, &new_xprt->rx_data);
 384         if (rc)
 385                 goto out2;
 386
 387         /*
 388          * Allocate pre-registered send and receive buffers for headers and
 389          * any inline data. Also specify any padding which will be provided
 390          * from a preregistered zero buffer.
 391          */
 392         rc = rpcrdma_buffer_create(new_xprt);
 393         if (rc)
 394                 goto out3;
 395
 396         /*
 397          * Register a callback for connection events. This is necessary because
 398          * connection loss notification is async. We also catch connection loss
 399          * when reaping receives.
 400          */
 401         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
 402                           xprt_rdma_connect_worker);
 403
 404         xprt_rdma_format_addresses(xprt, sap);
 405         xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
 406         if (xprt->max_payload == 0)
 407                 goto out4;
 408         xprt->max_payload <<= PAGE_SHIFT;
 409         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
 410                 __func__, xprt->max_payload);
 411
 412         if (!try_module_get(THIS_MODULE))
 413                 goto out4;
 414
 415         dprintk("RPC:       %s: %s:%s\n", __func__,
 416                 xprt->address_strings[RPC_DISPLAY_ADDR],
 417                 xprt->address_strings[RPC_DISPLAY_PORT]);
 418         return xprt;
 419
 420 out4:
 421         xprt_rdma_free_addresses(xprt);
 422         rc = -EINVAL;
 423 out3:
 424         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
 425 out2:
 426         rpcrdma_ia_close(&new_xprt->rx_ia);
 427 out1:
 428         xprt_free(xprt);
 429         return ERR_PTR(rc);
 430 }
 431
 432 /*
 433  * Close a connection, during shutdown or timeout/reconnect
 434  */
 435 static void
 436 xprt_rdma_close(struct rpc_xprt *xprt)
 437 {
 438         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 439
 440         dprintk("RPC:       %s: closing\n", __func__);
 441         if (r_xprt->rx_ep.rep_connected > 0)
 442                 xprt->reestablish_timeout = 0;
 443         xprt_disconnect_done(xprt);
 444         rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 445 }
 446
 447 static void
 448 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 449 {
 450         struct sockaddr_in *sap;
 451
 452         sap = (struct sockaddr_in *)&xprt->addr;
 453         sap->sin_port = htons(port);
 454         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
 455         sap->sin_port = htons(port);
 456         dprintk("RPC:       %s: %u\n", __func__, port);
 457 }
 458
 459 static void
 460 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
 461 {
 462         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 463
 464         if (r_xprt->rx_ep.rep_connected != 0) {
 465                 /* Reconnect */
 466                 schedule_delayed_work(&r_xprt->rx_connect_worker,
 467                                       xprt->reestablish_timeout);
 468                 xprt->reestablish_timeout <<= 1;
 469                 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
 470                         xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
 471                 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
 472                         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 473         } else {
 474                 schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
 475                 if (!RPC_IS_ASYNC(task))
 476                         flush_delayed_work(&r_xprt->rx_connect_worker);
 477         }
 478 }
 479
 480 /* Allocate a fixed-size buffer in which to construct and send the
 481  * RPC-over-RDMA header for this request.
 482  */
 483 static bool
 484 rpcrdma_get_rdmabuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 485                     gfp_t flags)
 486 {
 487         size_t size = RPCRDMA_HDRBUF_SIZE;
 488         struct rpcrdma_regbuf *rb;
 489
 490         if (req->rl_rdmabuf)
 491                 return true;
 492
 493         rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
 494         if (IS_ERR(rb))
 495                 return false;
 496
 497         r_xprt->rx_stats.hardway_register_count += size;
 498         req->rl_rdmabuf = rb;
 499         return true;
 500 }
 501
 502 static bool
 503 rpcrdma_get_sendbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 504                     size_t size, gfp_t flags)
 505 {
 506         struct rpcrdma_regbuf *rb;
 507
 508         if (req->rl_sendbuf && rdmab_length(req->rl_sendbuf) >= size)
 509                 return true;
 510
 511         rb = rpcrdma_alloc_regbuf(size, DMA_TO_DEVICE, flags);
 512         if (IS_ERR(rb))
 513                 return false;
 514
 515         rpcrdma_free_regbuf(req->rl_sendbuf);
 516         r_xprt->rx_stats.hardway_register_count += size;
 517         req->rl_sendbuf = rb;
 518         return true;
 519 }
 520
 521 /* The rq_rcv_buf is used only if a Reply chunk is necessary.
 522  * The decision to use a Reply chunk is made later in
 523  * rpcrdma_marshal_req. This buffer is registered at that time.
 524  *
 525  * Otherwise, the associated RPC Reply arrives in a separate
 526  * Receive buffer, arbitrarily chosen by the HCA. The buffer
 527  * allocated here for the RPC Reply is not utilized in that
 528  * case. See rpcrdma_inline_fixup.
 529  *
 530  * A regbuf is used here to remember the buffer size.
 531  */
 532 static bool
 533 rpcrdma_get_recvbuf(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 534                     size_t size, gfp_t flags)
 535 {
 536         struct rpcrdma_regbuf *rb;
 537
 538         if (req->rl_recvbuf && rdmab_length(req->rl_recvbuf) >= size)
 539                 return true;
 540
 541         rb = rpcrdma_alloc_regbuf(size, DMA_NONE, flags);
 542         if (IS_ERR(rb))
 543                 return false;
 544
 545         rpcrdma_free_regbuf(req->rl_recvbuf);
 546         r_xprt->rx_stats.hardway_register_count += size;
 547         req->rl_recvbuf = rb;
 548         return true;
 549 }
 550
 551 /**
 552  * xprt_rdma_allocate - allocate transport resources for an RPC
 553  * @task: RPC task
 554  *
 555  * Return values:
 556  *        0:    Success; rq_buffer points to RPC buffer to use
 557  *   ENOMEM:    Out of memory, call again later
 558  *      EIO:    A permanent error occurred, do not retry
 559  *
 560  * The RDMA allocate/free functions need the task structure as a place
 561  * to hide the struct rpcrdma_req, which is necessary for the actual
 562  * send/recv sequence.
 563  *
 564  * xprt_rdma_allocate provides buffers that are already mapped for
 565  * DMA, and a local DMA lkey is provided for each.
 566  */
 567 static int
 568 xprt_rdma_allocate(struct rpc_task *task)
 569 {
 570         struct rpc_rqst *rqst = task->tk_rqstp;
 571         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 572         struct rpcrdma_req *req;
 573         gfp_t flags;
 574
 575         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
 576         if (req == NULL)
 577                 return -ENOMEM;
 578
 579         flags = RPCRDMA_DEF_GFP;
 580         if (RPC_IS_ASYNC(task))
 581                 flags = GFP_NOWAIT | __GFP_NOWARN;
 582         if (RPC_IS_SWAPPER(task))
 583                 flags |= __GFP_MEMALLOC;
 584
 585         if (!rpcrdma_get_rdmabuf(r_xprt, req, flags))
 586                 goto out_fail;
 587         if (!rpcrdma_get_sendbuf(r_xprt, req, rqst->rq_callsize, flags))
 588                 goto out_fail;
 589         if (!rpcrdma_get_recvbuf(r_xprt, req, rqst->rq_rcvsize, flags))
 590                 goto out_fail;
 591
 592         dprintk("RPC: %5u %s: send size = %zd, recv size = %zd, req = %p\n",
 593                 task->tk_pid, __func__, rqst->rq_callsize,
 594                 rqst->rq_rcvsize, req);
 595
 596         req->rl_connect_cookie = 0;     /* our reserved value */
 597         rpcrdma_set_xprtdata(rqst, req);
 598         rqst->rq_buffer = req->rl_sendbuf->rg_base;
 599         rqst->rq_rbuffer = req->rl_recvbuf->rg_base;
 600         return 0;
 601
 602 out_fail:
 603         rpcrdma_buffer_put(req);
 604         return -ENOMEM;
 605 }
 606
 607 /**
 608  * xprt_rdma_free - release resources allocated by xprt_rdma_allocate
 609  * @task: RPC task
 610  *
 611  * Caller guarantees rqst->rq_buffer is non-NULL.
 612  */
 613 static void
 614 xprt_rdma_free(struct rpc_task *task)
 615 {
 616         struct rpc_rqst *rqst = task->tk_rqstp;
 617         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt);
 618         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 619         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 620
 621         if (req->rl_backchannel)
 622                 return;
 623
 624         dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
 625
 626         ia->ri_ops->ro_unmap_safe(r_xprt, req, !RPC_IS_ASYNC(task));
 627         rpcrdma_unmap_sges(ia, req);
 628         rpcrdma_buffer_put(req);
 629 }
 630
 631 /**
 632  * xprt_rdma_send_request - marshal and send an RPC request
 633  * @task: RPC task with an RPC message in rq_snd_buf
 634  *
 635  * Return values:
 636  *        0:    The request has been sent
 637  * ENOTCONN:    Caller needs to invoke connect logic then call again
 638  *  ENOBUFS:    Call again later to send the request
 639  *      EIO:    A permanent error occurred. The request was not sent,
 640  *              and don't try it again
 641  *
 642  * send_request invokes the meat of RPC RDMA. It must do the following:
 643  *
 644  *  1.  Marshal the RPC request into an RPC RDMA request, which means
 645  *      putting a header in front of data, and creating IOVs for RDMA
 646  *      from those in the request.
 647  *  2.  In marshaling, detect opportunities for RDMA, and use them.
 648  *  3.  Post a recv message to set up asynch completion, then send
 649  *      the request (rpcrdma_ep_post).
 650  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
 651  */
 652 static int
 653 xprt_rdma_send_request(struct rpc_task *task)
 654 {
 655         struct rpc_rqst *rqst = task->tk_rqstp;
 656         struct rpc_xprt *xprt = rqst->rq_xprt;
 657         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 658         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 659         int rc = 0;
 660
 661         /* On retransmit, remove any previously registered chunks */
 662         r_xprt->rx_ia.ri_ops->ro_unmap_safe(r_xprt, req, false);
 663
 664         rc = rpcrdma_marshal_req(rqst);
 665         if (rc < 0)
 666                 goto failed_marshal;
 667
 668         if (req->rl_reply == NULL)              /* e.g. reconnection */
 669                 rpcrdma_recv_buffer_get(req);
 670
 671         /* Must suppress retransmit to maintain credits */
 672         if (req->rl_connect_cookie == xprt->connect_cookie)
 673                 goto drop_connection;
 674         req->rl_connect_cookie = xprt->connect_cookie;
 675
 676         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 677                 goto drop_connection;
 678
 679         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
 680         rqst->rq_bytes_sent = 0;
 681         return 0;
 682
 683 failed_marshal:
 684         dprintk("RPC:       %s: rpcrdma_marshal_req failed, status %i\n",
 685                 __func__, rc);
 686         if (rc == -EIO)
 687                 r_xprt->rx_stats.failed_marshal_count++;
 688         if (rc != -ENOTCONN)
 689                 return rc;
 690 drop_connection:
 691         xprt_disconnect_done(xprt);
 692         return -ENOTCONN;       /* implies disconnect */
 693 }
 694
 695 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 696 {
 697         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 698         long idle_time = 0;
 699
 700         if (xprt_connected(xprt))
 701                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 702
 703         seq_puts(seq, "\txprt:\trdma ");
 704         seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
 705                    0,   /* need a local port? */
 706                    xprt->stat.bind_count,
 707                    xprt->stat.connect_count,
 708                    xprt->stat.connect_time,
 709                    idle_time,
 710                    xprt->stat.sends,
 711                    xprt->stat.recvs,
 712                    xprt->stat.bad_xids,
 713                    xprt->stat.req_u,
 714                    xprt->stat.bklog_u);
 715         seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ",
 716                    r_xprt->rx_stats.read_chunk_count,
 717                    r_xprt->rx_stats.write_chunk_count,
 718                    r_xprt->rx_stats.reply_chunk_count,
 719                    r_xprt->rx_stats.total_rdma_request,
 720                    r_xprt->rx_stats.total_rdma_reply,
 721                    r_xprt->rx_stats.pullup_copy_count,
 722                    r_xprt->rx_stats.fixup_copy_count,
 723                    r_xprt->rx_stats.hardway_register_count,
 724                    r_xprt->rx_stats.failed_marshal_count,
 725                    r_xprt->rx_stats.bad_reply_count,
 726                    r_xprt->rx_stats.nomsg_call_count);
 727         seq_printf(seq, "%lu %lu %lu %lu\n",
 728                    r_xprt->rx_stats.mrs_recovered,
 729                    r_xprt->rx_stats.mrs_orphaned,
 730                    r_xprt->rx_stats.mrs_allocated,
 731                    r_xprt->rx_stats.local_inv_needed);
 732 }
 733
 734 static int
 735 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
 736 {
 737         return 0;
 738 }
 739
 740 static void
 741 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
 742 {
 743 }
 744
 745 /*
 746  * Plumbing for rpc transport switch and kernel module
 747  */
 748
 749 static struct rpc_xprt_ops xprt_rdma_procs = {
 750         .reserve_xprt           = xprt_reserve_xprt_cong,
 751         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 752         .alloc_slot             = xprt_alloc_slot,
 753         .release_request        = xprt_release_rqst_cong,       /* ditto */
 754         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
 755         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 756         .set_port               = xprt_rdma_set_port,
 757         .connect                = xprt_rdma_connect,
 758         .buf_alloc              = xprt_rdma_allocate,
 759         .buf_free               = xprt_rdma_free,
 760         .send_request           = xprt_rdma_send_request,
 761         .close                  = xprt_rdma_close,
 762         .destroy                = xprt_rdma_destroy,
 763         .print_stats            = xprt_rdma_print_stats,
 764         .enable_swap            = xprt_rdma_enable_swap,
 765         .disable_swap           = xprt_rdma_disable_swap,
 766         .inject_disconnect      = xprt_rdma_inject_disconnect,
 767 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
 768         .bc_setup               = xprt_rdma_bc_setup,
 769         .bc_up                  = xprt_rdma_bc_up,
 770         .bc_maxpayload          = xprt_rdma_bc_maxpayload,
 771         .bc_free_rqst           = xprt_rdma_bc_free_rqst,
 772         .bc_destroy             = xprt_rdma_bc_destroy,
 773 #endif
 774 };
 775
 776 static struct xprt_class xprt_rdma = {
 777         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 778         .name                   = "rdma",
 779         .owner                  = THIS_MODULE,
 780         .ident                  = XPRT_TRANSPORT_RDMA,
 781         .setup                  = xprt_setup_rdma,
 782         .netid                  = { "rdma", "rdma6", "" },
 783 };
 784
 785 void xprt_rdma_cleanup(void)
 786 {
 787         int rc;
 788
 789         dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
 790 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 791         if (sunrpc_table_header) {
 792                 unregister_sysctl_table(sunrpc_table_header);
 793                 sunrpc_table_header = NULL;
 794         }
 795 #endif
 796         rc = xprt_unregister_transport(&xprt_rdma);
 797         if (rc)
 798                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
 799                         __func__, rc);
 800
 801         rpcrdma_destroy_wq();
 802
 803         rc = xprt_unregister_transport(&xprt_rdma_bc);
 804         if (rc)
 805                 dprintk("RPC:       %s: xprt_unregister(bc) returned %i\n",
 806                         __func__, rc);
 807 }
 808
 809 int xprt_rdma_init(void)
 810 {
 811         int rc;
 812
 813         rc = rpcrdma_alloc_wq();
 814         if (rc)
 815                 return rc;
 816
 817         rc = xprt_register_transport(&xprt_rdma);
 818         if (rc) {
 819                 rpcrdma_destroy_wq();
 820                 return rc;
 821         }
 822
 823         rc = xprt_register_transport(&xprt_rdma_bc);
 824         if (rc) {
 825                 xprt_unregister_transport(&xprt_rdma);
 826                 rpcrdma_destroy_wq();
 827                 return rc;
 828         }
 829
 830         dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
 831
 832         dprintk("Defaults:\n");
 833         dprintk("\tSlots %d\n"
 834                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
 835                 xprt_rdma_slot_table_entries,
 836                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
 837         dprintk("\tPadding %d\n\tMemreg %d\n",
 838                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
 839
 840 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 841         if (!sunrpc_table_header)
 842                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 843 #endif
 844         return 0;
 845 }