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 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
  74                 int xprt_rdma_pad_optimize = 0;
  75
  76 #ifdef RPC_DEBUG
  77
  78 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
  79 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
  80 static unsigned int zero;
  81 static unsigned int max_padding = PAGE_SIZE;
  82 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
  83 static unsigned int max_memreg = RPCRDMA_LAST - 1;
  84
  85 static struct ctl_table_header *sunrpc_table_header;
  86
  87 static ctl_table xr_tunables_table[] = {
  88         {
  89                 .ctl_name       = CTL_UNNUMBERED,
  90                 .procname       = "rdma_slot_table_entries",
  91                 .data           = &xprt_rdma_slot_table_entries,
  92                 .maxlen         = sizeof(unsigned int),
  93                 .mode           = 0644,
  94                 .proc_handler   = &proc_dointvec_minmax,
  95                 .strategy       = &sysctl_intvec,
  96                 .extra1         = &min_slot_table_size,
  97                 .extra2         = &max_slot_table_size
  98         },
  99         {
 100                 .ctl_name       = CTL_UNNUMBERED,
 101                 .procname       = "rdma_max_inline_read",
 102                 .data           = &xprt_rdma_max_inline_read,
 103                 .maxlen         = sizeof(unsigned int),
 104                 .mode           = 0644,
 105                 .proc_handler   = &proc_dointvec,
 106                 .strategy       = &sysctl_intvec,
 107         },
 108         {
 109                 .ctl_name       = CTL_UNNUMBERED,
 110                 .procname       = "rdma_max_inline_write",
 111                 .data           = &xprt_rdma_max_inline_write,
 112                 .maxlen         = sizeof(unsigned int),
 113                 .mode           = 0644,
 114                 .proc_handler   = &proc_dointvec,
 115                 .strategy       = &sysctl_intvec,
 116         },
 117         {
 118                 .ctl_name       = CTL_UNNUMBERED,
 119                 .procname       = "rdma_inline_write_padding",
 120                 .data           = &xprt_rdma_inline_write_padding,
 121                 .maxlen         = sizeof(unsigned int),
 122                 .mode           = 0644,
 123                 .proc_handler   = &proc_dointvec_minmax,
 124                 .strategy       = &sysctl_intvec,
 125                 .extra1         = &zero,
 126                 .extra2         = &max_padding,
 127         },
 128         {
 129                 .ctl_name       = CTL_UNNUMBERED,
 130                 .procname       = "rdma_memreg_strategy",
 131                 .data           = &xprt_rdma_memreg_strategy,
 132                 .maxlen         = sizeof(unsigned int),
 133                 .mode           = 0644,
 134                 .proc_handler   = &proc_dointvec_minmax,
 135                 .strategy       = &sysctl_intvec,
 136                 .extra1         = &min_memreg,
 137                 .extra2         = &max_memreg,
 138         },
 139         {
 140                 .ctl_name       = CTL_UNNUMBERED,
 141                 .procname       = "rdma_pad_optimize",
 142                 .data           = &xprt_rdma_pad_optimize,
 143                 .maxlen         = sizeof(unsigned int),
 144                 .mode           = 0644,
 145                 .proc_handler   = &proc_dointvec,
 146         },
 147         {
 148                 .ctl_name = 0,
 149         },
 150 };
 151
 152 static ctl_table sunrpc_table[] = {
 153         {
 154                 .ctl_name       = CTL_SUNRPC,
 155                 .procname       = "sunrpc",
 156                 .mode           = 0555,
 157                 .child          = xr_tunables_table
 158         },
 159         {
 160                 .ctl_name = 0,
 161         },
 162 };
 163
 164 #endif
 165
 166 static struct rpc_xprt_ops xprt_rdma_procs;     /* forward reference */
 167
 168 static void
 169 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
 170 {
 171         struct sockaddr *sap = (struct sockaddr *)
 172                                         &rpcx_to_rdmad(xprt).addr;
 173         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
 174         char buf[64];
 175
 176         (void)rpc_ntop(sap, buf, sizeof(buf));
 177         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
 178
 179         (void)snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
 180         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 181
 182         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
 183
 184         (void)snprintf(buf, sizeof(buf), "addr=%s port=%s proto=rdma",
 185                         xprt->address_strings[RPC_DISPLAY_ADDR],
 186                         xprt->address_strings[RPC_DISPLAY_PORT]);
 187         xprt->address_strings[RPC_DISPLAY_ALL] = kstrdup(buf, GFP_KERNEL);
 188
 189         (void)snprintf(buf, sizeof(buf), "%02x%02x%02x%02x",
 190                                 NIPQUAD(sin->sin_addr.s_addr));
 191         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 192
 193         (void)snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
 194         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 195
 196         /* netid */
 197         xprt->address_strings[RPC_DISPLAY_NETID] = "rdma";
 198 }
 199
 200 static void
 201 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
 202 {
 203         unsigned int i;
 204
 205         for (i = 0; i < RPC_DISPLAY_MAX; i++)
 206                 switch (i) {
 207                 case RPC_DISPLAY_PROTO:
 208                 case RPC_DISPLAY_NETID:
 209                         continue;
 210                 default:
 211                         kfree(xprt->address_strings[i]);
 212                 }
 213 }
 214
 215 static void
 216 xprt_rdma_connect_worker(struct work_struct *work)
 217 {
 218         struct rpcrdma_xprt *r_xprt =
 219                 container_of(work, struct rpcrdma_xprt, rdma_connect.work);
 220         struct rpc_xprt *xprt = &r_xprt->xprt;
 221         int rc = 0;
 222
 223         if (!xprt->shutdown) {
 224                 xprt_clear_connected(xprt);
 225
 226                 dprintk("RPC:       %s: %sconnect\n", __func__,
 227                                 r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
 228                 rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 229                 if (rc)
 230                         goto out;
 231         }
 232         goto out_clear;
 233
 234 out:
 235         xprt_wake_pending_tasks(xprt, rc);
 236
 237 out_clear:
 238         dprintk("RPC:       %s: exit\n", __func__);
 239         xprt_clear_connecting(xprt);
 240 }
 241
 242 /*
 243  * xprt_rdma_destroy
 244  *
 245  * Destroy the xprt.
 246  * Free all memory associated with the object, including its own.
 247  * NOTE: none of the *destroy methods free memory for their top-level
 248  * objects, even though they may have allocated it (they do free
 249  * private memory). It's up to the caller to handle it. In this
 250  * case (RDMA transport), all structure memory is inlined with the
 251  * struct rpcrdma_xprt.
 252  */
 253 static void
 254 xprt_rdma_destroy(struct rpc_xprt *xprt)
 255 {
 256         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 257         int rc;
 258
 259         dprintk("RPC:       %s: called\n", __func__);
 260
 261         cancel_delayed_work(&r_xprt->rdma_connect);
 262         flush_scheduled_work();
 263
 264         xprt_clear_connected(xprt);
 265
 266         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 267         rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
 268         if (rc)
 269                 dprintk("RPC:       %s: rpcrdma_ep_destroy returned %i\n",
 270                         __func__, rc);
 271         rpcrdma_ia_close(&r_xprt->rx_ia);
 272
 273         xprt_rdma_free_addresses(xprt);
 274
 275         kfree(xprt->slot);
 276         xprt->slot = NULL;
 277         kfree(xprt);
 278
 279         dprintk("RPC:       %s: returning\n", __func__);
 280
 281         module_put(THIS_MODULE);
 282 }
 283
 284 static const struct rpc_timeout xprt_rdma_default_timeout = {
 285         .to_initval = 60 * HZ,
 286         .to_maxval = 60 * HZ,
 287 };
 288
 289 /**
 290  * xprt_setup_rdma - Set up transport to use RDMA
 291  *
 292  * @args: rpc transport arguments
 293  */
 294 static struct rpc_xprt *
 295 xprt_setup_rdma(struct xprt_create *args)
 296 {
 297         struct rpcrdma_create_data_internal cdata;
 298         struct rpc_xprt *xprt;
 299         struct rpcrdma_xprt *new_xprt;
 300         struct rpcrdma_ep *new_ep;
 301         struct sockaddr_in *sin;
 302         int rc;
 303
 304         if (args->addrlen > sizeof(xprt->addr)) {
 305                 dprintk("RPC:       %s: address too large\n", __func__);
 306                 return ERR_PTR(-EBADF);
 307         }
 308
 309         xprt = kzalloc(sizeof(struct rpcrdma_xprt), GFP_KERNEL);
 310         if (xprt == NULL) {
 311                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
 312                         __func__);
 313                 return ERR_PTR(-ENOMEM);
 314         }
 315
 316         xprt->max_reqs = xprt_rdma_slot_table_entries;
 317         xprt->slot = kcalloc(xprt->max_reqs,
 318                                 sizeof(struct rpc_rqst), GFP_KERNEL);
 319         if (xprt->slot == NULL) {
 320                 dprintk("RPC:       %s: couldn't allocate %d slots\n",
 321                         __func__, xprt->max_reqs);
 322                 kfree(xprt);
 323                 return ERR_PTR(-ENOMEM);
 324         }
 325
 326         /* 60 second timeout, no retries */
 327         xprt->timeout = &xprt_rdma_default_timeout;
 328         xprt->bind_timeout = (60U * HZ);
 329         xprt->connect_timeout = (60U * HZ);
 330         xprt->reestablish_timeout = (5U * HZ);
 331         xprt->idle_timeout = (5U * 60 * HZ);
 332
 333         xprt->resvport = 0;             /* privileged port not needed */
 334         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
 335         xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE;
 336         xprt->ops = &xprt_rdma_procs;
 337
 338         /*
 339          * Set up RDMA-specific connect data.
 340          */
 341
 342         /* Put server RDMA address in local cdata */
 343         memcpy(&cdata.addr, args->dstaddr, args->addrlen);
 344
 345         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 346          * address, for any side protocols which peek at it */
 347         xprt->prot = IPPROTO_TCP;
 348         xprt->addrlen = args->addrlen;
 349         memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
 350
 351         sin = (struct sockaddr_in *)&cdata.addr;
 352         if (ntohs(sin->sin_port) != 0)
 353                 xprt_set_bound(xprt);
 354
 355         dprintk("RPC:       %s: %pI4:%u\n",
 356                 __func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
 357
 358         /* Set max requests */
 359         cdata.max_requests = xprt->max_reqs;
 360
 361         /* Set some length limits */
 362         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
 363         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
 364
 365         cdata.inline_wsize = xprt_rdma_max_inline_write;
 366         if (cdata.inline_wsize > cdata.wsize)
 367                 cdata.inline_wsize = cdata.wsize;
 368
 369         cdata.inline_rsize = xprt_rdma_max_inline_read;
 370         if (cdata.inline_rsize > cdata.rsize)
 371                 cdata.inline_rsize = cdata.rsize;
 372
 373         cdata.padding = xprt_rdma_inline_write_padding;
 374
 375         /*
 376          * Create new transport instance, which includes initialized
 377          *  o ia
 378          *  o endpoint
 379          *  o buffers
 380          */
 381
 382         new_xprt = rpcx_to_rdmax(xprt);
 383
 384         rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
 385                                 xprt_rdma_memreg_strategy);
 386         if (rc)
 387                 goto out1;
 388
 389         /*
 390          * initialize and create ep
 391          */
 392         new_xprt->rx_data = cdata;
 393         new_ep = &new_xprt->rx_ep;
 394         new_ep->rep_remote_addr = cdata.addr;
 395
 396         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
 397                                 &new_xprt->rx_ia, &new_xprt->rx_data);
 398         if (rc)
 399                 goto out2;
 400
 401         /*
 402          * Allocate pre-registered send and receive buffers for headers and
 403          * any inline data. Also specify any padding which will be provided
 404          * from a preregistered zero buffer.
 405          */
 406         rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia,
 407                                 &new_xprt->rx_data);
 408         if (rc)
 409                 goto out3;
 410
 411         /*
 412          * Register a callback for connection events. This is necessary because
 413          * connection loss notification is async. We also catch connection loss
 414          * when reaping receives.
 415          */
 416         INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker);
 417         new_ep->rep_func = rpcrdma_conn_func;
 418         new_ep->rep_xprt = xprt;
 419
 420         xprt_rdma_format_addresses(xprt);
 421
 422         if (!try_module_get(THIS_MODULE))
 423                 goto out4;
 424
 425         return xprt;
 426
 427 out4:
 428         xprt_rdma_free_addresses(xprt);
 429         rc = -EINVAL;
 430 out3:
 431         (void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
 432 out2:
 433         rpcrdma_ia_close(&new_xprt->rx_ia);
 434 out1:
 435         kfree(xprt->slot);
 436         kfree(xprt);
 437         return ERR_PTR(rc);
 438 }
 439
 440 /*
 441  * Close a connection, during shutdown or timeout/reconnect
 442  */
 443 static void
 444 xprt_rdma_close(struct rpc_xprt *xprt)
 445 {
 446         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 447
 448         dprintk("RPC:       %s: closing\n", __func__);
 449         if (r_xprt->rx_ep.rep_connected > 0)
 450                 xprt->reestablish_timeout = 0;
 451         xprt_disconnect_done(xprt);
 452         (void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 453 }
 454
 455 static void
 456 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 457 {
 458         struct sockaddr_in *sap;
 459
 460         sap = (struct sockaddr_in *)&xprt->addr;
 461         sap->sin_port = htons(port);
 462         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
 463         sap->sin_port = htons(port);
 464         dprintk("RPC:       %s: %u\n", __func__, port);
 465 }
 466
 467 static void
 468 xprt_rdma_connect(struct rpc_task *task)
 469 {
 470         struct rpc_xprt *xprt = (struct rpc_xprt *)task->tk_xprt;
 471         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 472
 473         if (!xprt_test_and_set_connecting(xprt)) {
 474                 if (r_xprt->rx_ep.rep_connected != 0) {
 475                         /* Reconnect */
 476                         schedule_delayed_work(&r_xprt->rdma_connect,
 477                                 xprt->reestablish_timeout);
 478                         xprt->reestablish_timeout <<= 1;
 479                         if (xprt->reestablish_timeout > (30 * HZ))
 480                                 xprt->reestablish_timeout = (30 * HZ);
 481                         else if (xprt->reestablish_timeout < (5 * HZ))
 482                                 xprt->reestablish_timeout = (5 * HZ);
 483                 } else {
 484                         schedule_delayed_work(&r_xprt->rdma_connect, 0);
 485                         if (!RPC_IS_ASYNC(task))
 486                                 flush_scheduled_work();
 487                 }
 488         }
 489 }
 490
 491 static int
 492 xprt_rdma_reserve_xprt(struct rpc_task *task)
 493 {
 494         struct rpc_xprt *xprt = task->tk_xprt;
 495         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 496         int credits = atomic_read(&r_xprt->rx_buf.rb_credits);
 497
 498         /* == RPC_CWNDSCALE @ init, but *after* setup */
 499         if (r_xprt->rx_buf.rb_cwndscale == 0UL) {
 500                 r_xprt->rx_buf.rb_cwndscale = xprt->cwnd;
 501                 dprintk("RPC:       %s: cwndscale %lu\n", __func__,
 502                         r_xprt->rx_buf.rb_cwndscale);
 503                 BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0);
 504         }
 505         xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale;
 506         return xprt_reserve_xprt_cong(task);
 507 }
 508
 509 /*
 510  * The RDMA allocate/free functions need the task structure as a place
 511  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
 512  * sequence. For this reason, the recv buffers are attached to send
 513  * buffers for portions of the RPC. Note that the RPC layer allocates
 514  * both send and receive buffers in the same call. We may register
 515  * the receive buffer portion when using reply chunks.
 516  */
 517 static void *
 518 xprt_rdma_allocate(struct rpc_task *task, size_t size)
 519 {
 520         struct rpc_xprt *xprt = task->tk_xprt;
 521         struct rpcrdma_req *req, *nreq;
 522
 523         req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf);
 524         BUG_ON(NULL == req);
 525
 526         if (size > req->rl_size) {
 527                 dprintk("RPC:       %s: size %zd too large for buffer[%zd]: "
 528                         "prog %d vers %d proc %d\n",
 529                         __func__, size, req->rl_size,
 530                         task->tk_client->cl_prog, task->tk_client->cl_vers,
 531                         task->tk_msg.rpc_proc->p_proc);
 532                 /*
 533                  * Outgoing length shortage. Our inline write max must have
 534                  * been configured to perform direct i/o.
 535                  *
 536                  * This is therefore a large metadata operation, and the
 537                  * allocate call was made on the maximum possible message,
 538                  * e.g. containing long filename(s) or symlink data. In
 539                  * fact, while these metadata operations *might* carry
 540                  * large outgoing payloads, they rarely *do*. However, we
 541                  * have to commit to the request here, so reallocate and
 542                  * register it now. The data path will never require this
 543                  * reallocation.
 544                  *
 545                  * If the allocation or registration fails, the RPC framework
 546                  * will (doggedly) retry.
 547                  */
 548                 if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy ==
 549                                 RPCRDMA_BOUNCEBUFFERS) {
 550                         /* forced to "pure inline" */
 551                         dprintk("RPC:       %s: too much data (%zd) for inline "
 552                                         "(r/w max %d/%d)\n", __func__, size,
 553                                         rpcx_to_rdmad(xprt).inline_rsize,
 554                                         rpcx_to_rdmad(xprt).inline_wsize);
 555                         size = req->rl_size;
 556                         rpc_exit(task, -EIO);           /* fail the operation */
 557                         rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
 558                         goto out;
 559                 }
 560                 if (task->tk_flags & RPC_TASK_SWAPPER)
 561                         nreq = kmalloc(sizeof *req + size, GFP_ATOMIC);
 562                 else
 563                         nreq = kmalloc(sizeof *req + size, GFP_NOFS);
 564                 if (nreq == NULL)
 565                         goto outfail;
 566
 567                 if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia,
 568                                 nreq->rl_base, size + sizeof(struct rpcrdma_req)
 569                                 - offsetof(struct rpcrdma_req, rl_base),
 570                                 &nreq->rl_handle, &nreq->rl_iov)) {
 571                         kfree(nreq);
 572                         goto outfail;
 573                 }
 574                 rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size;
 575                 nreq->rl_size = size;
 576                 nreq->rl_niovs = 0;
 577                 nreq->rl_nchunks = 0;
 578                 nreq->rl_buffer = (struct rpcrdma_buffer *)req;
 579                 nreq->rl_reply = req->rl_reply;
 580                 memcpy(nreq->rl_segments,
 581                         req->rl_segments, sizeof nreq->rl_segments);
 582                 /* flag the swap with an unused field */
 583                 nreq->rl_iov.length = 0;
 584                 req->rl_reply = NULL;
 585                 req = nreq;
 586         }
 587         dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
 588 out:
 589         req->rl_connect_cookie = 0;     /* our reserved value */
 590         return req->rl_xdr_buf;
 591
 592 outfail:
 593         rpcrdma_buffer_put(req);
 594         rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
 595         return NULL;
 596 }
 597
 598 /*
 599  * This function returns all RDMA resources to the pool.
 600  */
 601 static void
 602 xprt_rdma_free(void *buffer)
 603 {
 604         struct rpcrdma_req *req;
 605         struct rpcrdma_xprt *r_xprt;
 606         struct rpcrdma_rep *rep;
 607         int i;
 608
 609         if (buffer == NULL)
 610                 return;
 611
 612         req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]);
 613         if (req->rl_iov.length == 0) {  /* see allocate above */
 614                 r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer,
 615                                       struct rpcrdma_xprt, rx_buf);
 616         } else
 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         /* Must suppress retransmit to maintain credits */
 694         if (req->rl_connect_cookie == xprt->connect_cookie)
 695                 goto drop_connection;
 696         req->rl_connect_cookie = xprt->connect_cookie;
 697
 698         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 699                 goto drop_connection;
 700
 701         task->tk_bytes_sent += rqst->rq_snd_buf.len;
 702         rqst->rq_bytes_sent = 0;
 703         return 0;
 704
 705 drop_connection:
 706         xprt_disconnect_done(xprt);
 707         return -ENOTCONN;       /* implies disconnect */
 708 }
 709
 710 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 711 {
 712         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 713         long idle_time = 0;
 714
 715         if (xprt_connected(xprt))
 716                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 717
 718         seq_printf(seq,
 719           "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
 720           "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
 721
 722            0,   /* need a local port? */
 723            xprt->stat.bind_count,
 724            xprt->stat.connect_count,
 725            xprt->stat.connect_time,
 726            idle_time,
 727            xprt->stat.sends,
 728            xprt->stat.recvs,
 729            xprt->stat.bad_xids,
 730            xprt->stat.req_u,
 731            xprt->stat.bklog_u,
 732
 733            r_xprt->rx_stats.read_chunk_count,
 734            r_xprt->rx_stats.write_chunk_count,
 735            r_xprt->rx_stats.reply_chunk_count,
 736            r_xprt->rx_stats.total_rdma_request,
 737            r_xprt->rx_stats.total_rdma_reply,
 738            r_xprt->rx_stats.pullup_copy_count,
 739            r_xprt->rx_stats.fixup_copy_count,
 740            r_xprt->rx_stats.hardway_register_count,
 741            r_xprt->rx_stats.failed_marshal_count,
 742            r_xprt->rx_stats.bad_reply_count);
 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_rdma_reserve_xprt,
 751         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 752         .release_request        = xprt_release_rqst_cong,       /* ditto */
 753         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
 754         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 755         .set_port               = xprt_rdma_set_port,
 756         .connect                = xprt_rdma_connect,
 757         .buf_alloc              = xprt_rdma_allocate,
 758         .buf_free               = xprt_rdma_free,
 759         .send_request           = xprt_rdma_send_request,
 760         .close                  = xprt_rdma_close,
 761         .destroy                = xprt_rdma_destroy,
 762         .print_stats            = xprt_rdma_print_stats
 763 };
 764
 765 static struct xprt_class xprt_rdma = {
 766         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 767         .name                   = "rdma",
 768         .owner                  = THIS_MODULE,
 769         .ident                  = XPRT_TRANSPORT_RDMA,
 770         .setup                  = xprt_setup_rdma,
 771 };
 772
 773 static void __exit xprt_rdma_cleanup(void)
 774 {
 775         int rc;
 776
 777         dprintk(KERN_INFO "RPCRDMA Module Removed, deregister RPC RDMA transport\n");
 778 #ifdef RPC_DEBUG
 779         if (sunrpc_table_header) {
 780                 unregister_sysctl_table(sunrpc_table_header);
 781                 sunrpc_table_header = NULL;
 782         }
 783 #endif
 784         rc = xprt_unregister_transport(&xprt_rdma);
 785         if (rc)
 786                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
 787                         __func__, rc);
 788 }
 789
 790 static int __init xprt_rdma_init(void)
 791 {
 792         int rc;
 793
 794         rc = xprt_register_transport(&xprt_rdma);
 795
 796         if (rc)
 797                 return rc;
 798
 799         dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n");
 800
 801         dprintk(KERN_INFO "Defaults:\n");
 802         dprintk(KERN_INFO "\tSlots %d\n"
 803                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
 804                 xprt_rdma_slot_table_entries,
 805                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
 806         dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n",
 807                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
 808
 809 #ifdef RPC_DEBUG
 810         if (!sunrpc_table_header)
 811                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 812 #endif
 813         return 0;
 814 }
 815
 816 module_init(xprt_rdma_init);
 817 module_exit(xprt_rdma_cleanup);