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/init.h>
  52 #include <linux/seq_file.h>
  53
  54 #include "xprt_rdma.h"
  55
  56 #ifdef RPC_DEBUG
  57 # define RPCDBG_FACILITY        RPCDBG_TRANS
  58 #endif
  59
  60 MODULE_LICENSE("Dual BSD/GPL");
  61
  62 MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS");
  63 MODULE_AUTHOR("Network Appliance, Inc.");
  64
  65 /*
  66  * tunables
  67  */
  68
  69 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
  70 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
  71 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
  72 static unsigned int xprt_rdma_inline_write_padding;
  73 #if !RPCRDMA_PERSISTENT_REGISTRATION
  74 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_REGISTER; /* FMR? */
  75 #else
  76 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_ALLPHYSICAL;
  77 #endif
  78
  79 #ifdef RPC_DEBUG
  80
  81 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
  82 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
  83 static unsigned int zero;
  84 static unsigned int max_padding = PAGE_SIZE;
  85 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
  86 static unsigned int max_memreg = RPCRDMA_LAST - 1;
  87
  88 static struct ctl_table_header *sunrpc_table_header;
  89
  90 static ctl_table xr_tunables_table[] = {
  91         {
  92                 .ctl_name       = CTL_UNNUMBERED,
  93                 .procname       = "rdma_slot_table_entries",
  94                 .data           = &xprt_rdma_slot_table_entries,
  95                 .maxlen         = sizeof(unsigned int),
  96                 .mode           = 0644,
  97                 .proc_handler   = &proc_dointvec_minmax,
  98                 .strategy       = &sysctl_intvec,
  99                 .extra1         = &min_slot_table_size,
 100                 .extra2         = &max_slot_table_size
 101         },
 102         {
 103                 .ctl_name       = CTL_UNNUMBERED,
 104                 .procname       = "rdma_max_inline_read",
 105                 .data           = &xprt_rdma_max_inline_read,
 106                 .maxlen         = sizeof(unsigned int),
 107                 .mode           = 0644,
 108                 .proc_handler   = &proc_dointvec,
 109                 .strategy       = &sysctl_intvec,
 110         },
 111         {
 112                 .ctl_name       = CTL_UNNUMBERED,
 113                 .procname       = "rdma_max_inline_write",
 114                 .data           = &xprt_rdma_max_inline_write,
 115                 .maxlen         = sizeof(unsigned int),
 116                 .mode           = 0644,
 117                 .proc_handler   = &proc_dointvec,
 118                 .strategy       = &sysctl_intvec,
 119         },
 120         {
 121                 .ctl_name       = CTL_UNNUMBERED,
 122                 .procname       = "rdma_inline_write_padding",
 123                 .data           = &xprt_rdma_inline_write_padding,
 124                 .maxlen         = sizeof(unsigned int),
 125                 .mode           = 0644,
 126                 .proc_handler   = &proc_dointvec_minmax,
 127                 .strategy       = &sysctl_intvec,
 128                 .extra1         = &zero,
 129                 .extra2         = &max_padding,
 130         },
 131         {
 132                 .ctl_name       = CTL_UNNUMBERED,
 133                 .procname       = "rdma_memreg_strategy",
 134                 .data           = &xprt_rdma_memreg_strategy,
 135                 .maxlen         = sizeof(unsigned int),
 136                 .mode           = 0644,
 137                 .proc_handler   = &proc_dointvec_minmax,
 138                 .strategy       = &sysctl_intvec,
 139                 .extra1         = &min_memreg,
 140                 .extra2         = &max_memreg,
 141         },
 142         {
 143                 .ctl_name = 0,
 144         },
 145 };
 146
 147 static ctl_table sunrpc_table[] = {
 148         {
 149                 .ctl_name       = CTL_SUNRPC,
 150                 .procname       = "sunrpc",
 151                 .mode           = 0555,
 152                 .child          = xr_tunables_table
 153         },
 154         {
 155                 .ctl_name = 0,
 156         },
 157 };
 158
 159 #endif
 160
 161 static struct rpc_xprt_ops xprt_rdma_procs;     /* forward reference */
 162
 163 static void
 164 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
 165 {
 166         struct sockaddr_in *addr = (struct sockaddr_in *)
 167                                         &rpcx_to_rdmad(xprt).addr;
 168         char *buf;
 169
 170         buf = kzalloc(20, GFP_KERNEL);
 171         if (buf)
 172                 snprintf(buf, 20, NIPQUAD_FMT, NIPQUAD(addr->sin_addr.s_addr));
 173         xprt->address_strings[RPC_DISPLAY_ADDR] = buf;
 174
 175         buf = kzalloc(8, GFP_KERNEL);
 176         if (buf)
 177                 snprintf(buf, 8, "%u", ntohs(addr->sin_port));
 178         xprt->address_strings[RPC_DISPLAY_PORT] = buf;
 179
 180         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
 181
 182         buf = kzalloc(48, GFP_KERNEL);
 183         if (buf)
 184                 snprintf(buf, 48, "addr="NIPQUAD_FMT" port=%u proto=%s",
 185                         NIPQUAD(addr->sin_addr.s_addr),
 186                         ntohs(addr->sin_port), "rdma");
 187         xprt->address_strings[RPC_DISPLAY_ALL] = buf;
 188
 189         buf = kzalloc(10, GFP_KERNEL);
 190         if (buf)
 191                 snprintf(buf, 10, "%02x%02x%02x%02x",
 192                         NIPQUAD(addr->sin_addr.s_addr));
 193         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = buf;
 194
 195         buf = kzalloc(8, GFP_KERNEL);
 196         if (buf)
 197                 snprintf(buf, 8, "%4hx", ntohs(addr->sin_port));
 198         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = buf;
 199
 200         buf = kzalloc(30, GFP_KERNEL);
 201         if (buf)
 202                 snprintf(buf, 30, NIPQUAD_FMT".%u.%u",
 203                         NIPQUAD(addr->sin_addr.s_addr),
 204                         ntohs(addr->sin_port) >> 8,
 205                         ntohs(addr->sin_port) & 0xff);
 206         xprt->address_strings[RPC_DISPLAY_UNIVERSAL_ADDR] = buf;
 207
 208         /* netid */
 209         xprt->address_strings[RPC_DISPLAY_NETID] = "rdma";
 210 }
 211
 212 static void
 213 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
 214 {
 215         unsigned int i;
 216
 217         for (i = 0; i < RPC_DISPLAY_MAX; i++)
 218                 switch (i) {
 219                 case RPC_DISPLAY_PROTO:
 220                 case RPC_DISPLAY_NETID:
 221                         continue;
 222                 default:
 223                         kfree(xprt->address_strings[i]);
 224                 }
 225 }
 226
 227 static void
 228 xprt_rdma_connect_worker(struct work_struct *work)
 229 {
 230         struct rpcrdma_xprt *r_xprt =
 231                 container_of(work, struct rpcrdma_xprt, rdma_connect.work);
 232         struct rpc_xprt *xprt = &r_xprt->xprt;
 233         int rc = 0;
 234
 235         if (!xprt->shutdown) {
 236                 xprt_clear_connected(xprt);
 237
 238                 dprintk("RPC:       %s: %sconnect\n", __func__,
 239                                 r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
 240                 rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 241                 if (rc)
 242                         goto out;
 243         }
 244         goto out_clear;
 245
 246 out:
 247         xprt_wake_pending_tasks(xprt, rc);
 248
 249 out_clear:
 250         dprintk("RPC:       %s: exit\n", __func__);
 251         xprt_clear_connecting(xprt);
 252 }
 253
 254 /*
 255  * xprt_rdma_destroy
 256  *
 257  * Destroy the xprt.
 258  * Free all memory associated with the object, including its own.
 259  * NOTE: none of the *destroy methods free memory for their top-level
 260  * objects, even though they may have allocated it (they do free
 261  * private memory). It's up to the caller to handle it. In this
 262  * case (RDMA transport), all structure memory is inlined with the
 263  * struct rpcrdma_xprt.
 264  */
 265 static void
 266 xprt_rdma_destroy(struct rpc_xprt *xprt)
 267 {
 268         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 269         int rc;
 270
 271         dprintk("RPC:       %s: called\n", __func__);
 272
 273         cancel_delayed_work(&r_xprt->rdma_connect);
 274         flush_scheduled_work();
 275
 276         xprt_clear_connected(xprt);
 277
 278         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 279         rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
 280         if (rc)
 281                 dprintk("RPC:       %s: rpcrdma_ep_destroy returned %i\n",
 282                         __func__, rc);
 283         rpcrdma_ia_close(&r_xprt->rx_ia);
 284
 285         xprt_rdma_free_addresses(xprt);
 286
 287         kfree(xprt->slot);
 288         xprt->slot = NULL;
 289         kfree(xprt);
 290
 291         dprintk("RPC:       %s: returning\n", __func__);
 292
 293         module_put(THIS_MODULE);
 294 }
 295
 296 static const struct rpc_timeout xprt_rdma_default_timeout = {
 297         .to_initval = 60 * HZ,
 298         .to_maxval = 60 * HZ,
 299 };
 300
 301 /**
 302  * xprt_setup_rdma - Set up transport to use RDMA
 303  *
 304  * @args: rpc transport arguments
 305  */
 306 static struct rpc_xprt *
 307 xprt_setup_rdma(struct xprt_create *args)
 308 {
 309         struct rpcrdma_create_data_internal cdata;
 310         struct rpc_xprt *xprt;
 311         struct rpcrdma_xprt *new_xprt;
 312         struct rpcrdma_ep *new_ep;
 313         struct sockaddr_in *sin;
 314         int rc;
 315
 316         if (args->addrlen > sizeof(xprt->addr)) {
 317                 dprintk("RPC:       %s: address too large\n", __func__);
 318                 return ERR_PTR(-EBADF);
 319         }
 320
 321         xprt = kzalloc(sizeof(struct rpcrdma_xprt), GFP_KERNEL);
 322         if (xprt == NULL) {
 323                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
 324                         __func__);
 325                 return ERR_PTR(-ENOMEM);
 326         }
 327
 328         xprt->max_reqs = xprt_rdma_slot_table_entries;
 329         xprt->slot = kcalloc(xprt->max_reqs,
 330                                 sizeof(struct rpc_rqst), GFP_KERNEL);
 331         if (xprt->slot == NULL) {
 332                 dprintk("RPC:       %s: couldn't allocate %d slots\n",
 333                         __func__, xprt->max_reqs);
 334                 kfree(xprt);
 335                 return ERR_PTR(-ENOMEM);
 336         }
 337
 338         /* 60 second timeout, no retries */
 339         xprt->timeout = &xprt_rdma_default_timeout;
 340         xprt->bind_timeout = (60U * HZ);
 341         xprt->connect_timeout = (60U * HZ);
 342         xprt->reestablish_timeout = (5U * HZ);
 343         xprt->idle_timeout = (5U * 60 * HZ);
 344
 345         xprt->resvport = 0;             /* privileged port not needed */
 346         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
 347         xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE;
 348         xprt->ops = &xprt_rdma_procs;
 349
 350         /*
 351          * Set up RDMA-specific connect data.
 352          */
 353
 354         /* Put server RDMA address in local cdata */
 355         memcpy(&cdata.addr, args->dstaddr, args->addrlen);
 356
 357         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 358          * address, for any side protocols which peek at it */
 359         xprt->prot = IPPROTO_TCP;
 360         xprt->addrlen = args->addrlen;
 361         memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
 362
 363         sin = (struct sockaddr_in *)&cdata.addr;
 364         if (ntohs(sin->sin_port) != 0)
 365                 xprt_set_bound(xprt);
 366
 367         dprintk("RPC:       %s: %u.%u.%u.%u:%u\n", __func__,
 368                         NIPQUAD(sin->sin_addr.s_addr), ntohs(sin->sin_port));
 369
 370         /* Set max requests */
 371         cdata.max_requests = xprt->max_reqs;
 372
 373         /* Set some length limits */
 374         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
 375         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
 376
 377         cdata.inline_wsize = xprt_rdma_max_inline_write;
 378         if (cdata.inline_wsize > cdata.wsize)
 379                 cdata.inline_wsize = cdata.wsize;
 380
 381         cdata.inline_rsize = xprt_rdma_max_inline_read;
 382         if (cdata.inline_rsize > cdata.rsize)
 383                 cdata.inline_rsize = cdata.rsize;
 384
 385         cdata.padding = xprt_rdma_inline_write_padding;
 386
 387         /*
 388          * Create new transport instance, which includes initialized
 389          *  o ia
 390          *  o endpoint
 391          *  o buffers
 392          */
 393
 394         new_xprt = rpcx_to_rdmax(xprt);
 395
 396         rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
 397                                 xprt_rdma_memreg_strategy);
 398         if (rc)
 399                 goto out1;
 400
 401         /*
 402          * initialize and create ep
 403          */
 404         new_xprt->rx_data = cdata;
 405         new_ep = &new_xprt->rx_ep;
 406         new_ep->rep_remote_addr = cdata.addr;
 407
 408         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
 409                                 &new_xprt->rx_ia, &new_xprt->rx_data);
 410         if (rc)
 411                 goto out2;
 412
 413         /*
 414          * Allocate pre-registered send and receive buffers for headers and
 415          * any inline data. Also specify any padding which will be provided
 416          * from a preregistered zero buffer.
 417          */
 418         rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia,
 419                                 &new_xprt->rx_data);
 420         if (rc)
 421                 goto out3;
 422
 423         /*
 424          * Register a callback for connection events. This is necessary because
 425          * connection loss notification is async. We also catch connection loss
 426          * when reaping receives.
 427          */
 428         INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker);
 429         new_ep->rep_func = rpcrdma_conn_func;
 430         new_ep->rep_xprt = xprt;
 431
 432         xprt_rdma_format_addresses(xprt);
 433
 434         if (!try_module_get(THIS_MODULE))
 435                 goto out4;
 436
 437         return xprt;
 438
 439 out4:
 440         xprt_rdma_free_addresses(xprt);
 441         rc = -EINVAL;
 442 out3:
 443         (void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
 444 out2:
 445         rpcrdma_ia_close(&new_xprt->rx_ia);
 446 out1:
 447         kfree(xprt->slot);
 448         kfree(xprt);
 449         return ERR_PTR(rc);
 450 }
 451
 452 /*
 453  * Close a connection, during shutdown or timeout/reconnect
 454  */
 455 static void
 456 xprt_rdma_close(struct rpc_xprt *xprt)
 457 {
 458         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 459
 460         dprintk("RPC:       %s: closing\n", __func__);
 461         xprt_disconnect_done(xprt);
 462         (void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 463 }
 464
 465 static void
 466 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 467 {
 468         struct sockaddr_in *sap;
 469
 470         sap = (struct sockaddr_in *)&xprt->addr;
 471         sap->sin_port = htons(port);
 472         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
 473         sap->sin_port = htons(port);
 474         dprintk("RPC:       %s: %u\n", __func__, port);
 475 }
 476
 477 static void
 478 xprt_rdma_connect(struct rpc_task *task)
 479 {
 480         struct rpc_xprt *xprt = (struct rpc_xprt *)task->tk_xprt;
 481         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 482
 483         if (!xprt_test_and_set_connecting(xprt)) {
 484                 if (r_xprt->rx_ep.rep_connected != 0) {
 485                         /* Reconnect */
 486                         schedule_delayed_work(&r_xprt->rdma_connect,
 487                                 xprt->reestablish_timeout);
 488                 } else {
 489                         schedule_delayed_work(&r_xprt->rdma_connect, 0);
 490                         if (!RPC_IS_ASYNC(task))
 491                                 flush_scheduled_work();
 492                 }
 493         }
 494 }
 495
 496 static int
 497 xprt_rdma_reserve_xprt(struct rpc_task *task)
 498 {
 499         struct rpc_xprt *xprt = task->tk_xprt;
 500         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 501         int credits = atomic_read(&r_xprt->rx_buf.rb_credits);
 502
 503         /* == RPC_CWNDSCALE @ init, but *after* setup */
 504         if (r_xprt->rx_buf.rb_cwndscale == 0UL) {
 505                 r_xprt->rx_buf.rb_cwndscale = xprt->cwnd;
 506                 dprintk("RPC:       %s: cwndscale %lu\n", __func__,
 507                         r_xprt->rx_buf.rb_cwndscale);
 508                 BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0);
 509         }
 510         xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale;
 511         return xprt_reserve_xprt_cong(task);
 512 }
 513
 514 /*
 515  * The RDMA allocate/free functions need the task structure as a place
 516  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
 517  * sequence. For this reason, the recv buffers are attached to send
 518  * buffers for portions of the RPC. Note that the RPC layer allocates
 519  * both send and receive buffers in the same call. We may register
 520  * the receive buffer portion when using reply chunks.
 521  */
 522 static void *
 523 xprt_rdma_allocate(struct rpc_task *task, size_t size)
 524 {
 525         struct rpc_xprt *xprt = task->tk_xprt;
 526         struct rpcrdma_req *req, *nreq;
 527
 528         req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf);
 529         BUG_ON(NULL == req);
 530
 531         if (size > req->rl_size) {
 532                 dprintk("RPC:       %s: size %zd too large for buffer[%zd]: "
 533                         "prog %d vers %d proc %d\n",
 534                         __func__, size, req->rl_size,
 535                         task->tk_client->cl_prog, task->tk_client->cl_vers,
 536                         task->tk_msg.rpc_proc->p_proc);
 537                 /*
 538                  * Outgoing length shortage. Our inline write max must have
 539                  * been configured to perform direct i/o.
 540                  *
 541                  * This is therefore a large metadata operation, and the
 542                  * allocate call was made on the maximum possible message,
 543                  * e.g. containing long filename(s) or symlink data. In
 544                  * fact, while these metadata operations *might* carry
 545                  * large outgoing payloads, they rarely *do*. However, we
 546                  * have to commit to the request here, so reallocate and
 547                  * register it now. The data path will never require this
 548                  * reallocation.
 549                  *
 550                  * If the allocation or registration fails, the RPC framework
 551                  * will (doggedly) retry.
 552                  */
 553                 if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy ==
 554                                 RPCRDMA_BOUNCEBUFFERS) {
 555                         /* forced to "pure inline" */
 556                         dprintk("RPC:       %s: too much data (%zd) for inline "
 557                                         "(r/w max %d/%d)\n", __func__, size,
 558                                         rpcx_to_rdmad(xprt).inline_rsize,
 559                                         rpcx_to_rdmad(xprt).inline_wsize);
 560                         size = req->rl_size;
 561                         rpc_exit(task, -EIO);           /* fail the operation */
 562                         rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
 563                         goto out;
 564                 }
 565                 if (task->tk_flags & RPC_TASK_SWAPPER)
 566                         nreq = kmalloc(sizeof *req + size, GFP_ATOMIC);
 567                 else
 568                         nreq = kmalloc(sizeof *req + size, GFP_NOFS);
 569                 if (nreq == NULL)
 570                         goto outfail;
 571
 572                 if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia,
 573                                 nreq->rl_base, size + sizeof(struct rpcrdma_req)
 574                                 - offsetof(struct rpcrdma_req, rl_base),
 575                                 &nreq->rl_handle, &nreq->rl_iov)) {
 576                         kfree(nreq);
 577                         goto outfail;
 578                 }
 579                 rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size;
 580                 nreq->rl_size = size;
 581                 nreq->rl_niovs = 0;
 582                 nreq->rl_nchunks = 0;
 583                 nreq->rl_buffer = (struct rpcrdma_buffer *)req;
 584                 nreq->rl_reply = req->rl_reply;
 585                 memcpy(nreq->rl_segments,
 586                         req->rl_segments, sizeof nreq->rl_segments);
 587                 /* flag the swap with an unused field */
 588                 nreq->rl_iov.length = 0;
 589                 req->rl_reply = NULL;
 590                 req = nreq;
 591         }
 592         dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
 593 out:
 594         return req->rl_xdr_buf;
 595
 596 outfail:
 597         rpcrdma_buffer_put(req);
 598         rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
 599         return NULL;
 600 }
 601
 602 /*
 603  * This function returns all RDMA resources to the pool.
 604  */
 605 static void
 606 xprt_rdma_free(void *buffer)
 607 {
 608         struct rpcrdma_req *req;
 609         struct rpcrdma_xprt *r_xprt;
 610         struct rpcrdma_rep *rep;
 611         int i;
 612
 613         if (buffer == NULL)
 614                 return;
 615
 616         req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]);
 617         r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
 618         rep = req->rl_reply;
 619
 620         dprintk("RPC:       %s: called on 0x%p%s\n",
 621                 __func__, rep, (rep && rep->rr_func) ? " (with waiter)" : "");
 622
 623         /*
 624          * Finish the deregistration. When using mw bind, this was
 625          * begun in rpcrdma_reply_handler(). In all other modes, we
 626          * do it here, in thread context. The process is considered
 627          * complete when the rr_func vector becomes NULL - this
 628          * was put in place during rpcrdma_reply_handler() - the wait
 629          * call below will not block if the dereg is "done". If
 630          * interrupted, our framework will clean up.
 631          */
 632         for (i = 0; req->rl_nchunks;) {
 633                 --req->rl_nchunks;
 634                 i += rpcrdma_deregister_external(
 635                         &req->rl_segments[i], r_xprt, NULL);
 636         }
 637
 638         if (rep && wait_event_interruptible(rep->rr_unbind, !rep->rr_func)) {
 639                 rep->rr_func = NULL;    /* abandon the callback */
 640                 req->rl_reply = NULL;
 641         }
 642
 643         if (req->rl_iov.length == 0) {  /* see allocate above */
 644                 struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer;
 645                 oreq->rl_reply = req->rl_reply;
 646                 (void) rpcrdma_deregister_internal(&r_xprt->rx_ia,
 647                                                    req->rl_handle,
 648                                                    &req->rl_iov);
 649                 kfree(req);
 650                 req = oreq;
 651         }
 652
 653         /* Put back request+reply buffers */
 654         rpcrdma_buffer_put(req);
 655 }
 656
 657 /*
 658  * send_request invokes the meat of RPC RDMA. It must do the following:
 659  *  1.  Marshal the RPC request into an RPC RDMA request, which means
 660  *      putting a header in front of data, and creating IOVs for RDMA
 661  *      from those in the request.
 662  *  2.  In marshaling, detect opportunities for RDMA, and use them.
 663  *  3.  Post a recv message to set up asynch completion, then send
 664  *      the request (rpcrdma_ep_post).
 665  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
 666  */
 667
 668 static int
 669 xprt_rdma_send_request(struct rpc_task *task)
 670 {
 671         struct rpc_rqst *rqst = task->tk_rqstp;
 672         struct rpc_xprt *xprt = task->tk_xprt;
 673         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 674         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 675
 676         /* marshal the send itself */
 677         if (req->rl_niovs == 0 && rpcrdma_marshal_req(rqst) != 0) {
 678                 r_xprt->rx_stats.failed_marshal_count++;
 679                 dprintk("RPC:       %s: rpcrdma_marshal_req failed\n",
 680                         __func__);
 681                 return -EIO;
 682         }
 683
 684         if (req->rl_reply == NULL)              /* e.g. reconnection */
 685                 rpcrdma_recv_buffer_get(req);
 686
 687         if (req->rl_reply) {
 688                 req->rl_reply->rr_func = rpcrdma_reply_handler;
 689                 /* this need only be done once, but... */
 690                 req->rl_reply->rr_xprt = xprt;
 691         }
 692
 693         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req)) {
 694                 xprt_disconnect_done(xprt);
 695                 return -ENOTCONN;       /* implies disconnect */
 696         }
 697
 698         rqst->rq_bytes_sent = 0;
 699         return 0;
 700 }
 701
 702 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 703 {
 704         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 705         long idle_time = 0;
 706
 707         if (xprt_connected(xprt))
 708                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 709
 710         seq_printf(seq,
 711           "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
 712           "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
 713
 714            0,   /* need a local port? */
 715            xprt->stat.bind_count,
 716            xprt->stat.connect_count,
 717            xprt->stat.connect_time,
 718            idle_time,
 719            xprt->stat.sends,
 720            xprt->stat.recvs,
 721            xprt->stat.bad_xids,
 722            xprt->stat.req_u,
 723            xprt->stat.bklog_u,
 724
 725            r_xprt->rx_stats.read_chunk_count,
 726            r_xprt->rx_stats.write_chunk_count,
 727            r_xprt->rx_stats.reply_chunk_count,
 728            r_xprt->rx_stats.total_rdma_request,
 729            r_xprt->rx_stats.total_rdma_reply,
 730            r_xprt->rx_stats.pullup_copy_count,
 731            r_xprt->rx_stats.fixup_copy_count,
 732            r_xprt->rx_stats.hardway_register_count,
 733            r_xprt->rx_stats.failed_marshal_count,
 734            r_xprt->rx_stats.bad_reply_count);
 735 }
 736
 737 /*
 738  * Plumbing for rpc transport switch and kernel module
 739  */
 740
 741 static struct rpc_xprt_ops xprt_rdma_procs = {
 742         .reserve_xprt           = xprt_rdma_reserve_xprt,
 743         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 744         .release_request        = xprt_release_rqst_cong,       /* ditto */
 745         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
 746         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 747         .set_port               = xprt_rdma_set_port,
 748         .connect                = xprt_rdma_connect,
 749         .buf_alloc              = xprt_rdma_allocate,
 750         .buf_free               = xprt_rdma_free,
 751         .send_request           = xprt_rdma_send_request,
 752         .close                  = xprt_rdma_close,
 753         .destroy                = xprt_rdma_destroy,
 754         .print_stats            = xprt_rdma_print_stats
 755 };
 756
 757 static struct xprt_class xprt_rdma = {
 758         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 759         .name                   = "rdma",
 760         .owner                  = THIS_MODULE,
 761         .ident                  = XPRT_TRANSPORT_RDMA,
 762         .setup                  = xprt_setup_rdma,
 763 };
 764
 765 static void __exit xprt_rdma_cleanup(void)
 766 {
 767         int rc;
 768
 769         dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
 770 #ifdef RPC_DEBUG
 771         if (sunrpc_table_header) {
 772                 unregister_sysctl_table(sunrpc_table_header);
 773                 sunrpc_table_header = NULL;
 774         }
 775 #endif
 776         rc = xprt_unregister_transport(&xprt_rdma);
 777         if (rc)
 778                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
 779                         __func__, rc);
 780 }
 781
 782 static int __init xprt_rdma_init(void)
 783 {
 784         int rc;
 785
 786         rc = xprt_register_transport(&xprt_rdma);
 787
 788         if (rc)
 789                 return rc;
 790
 791         dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n");
 792
 793         dprintk(KERN_INFO "Defaults:\n");
 794         dprintk(KERN_INFO "\tSlots %d\n"
 795                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
 796                 xprt_rdma_slot_table_entries,
 797                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
 798         dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n",
 799                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
 800
 801 #ifdef RPC_DEBUG
 802         if (!sunrpc_table_header)
 803                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 804 #endif
 805         return 0;
 806 }
 807
 808 module_init(xprt_rdma_init);
 809 module_exit(xprt_rdma_cleanup);