Pileus Git - ~andy/linux/blob - 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/slab.h>
  53 #include <linux/seq_file.h>
  54
  55 #include "xprt_rdma.h"
  56
  57 #ifdef RPC_DEBUG
  58 # define RPCDBG_FACILITY        RPCDBG_TRANS
  59 #endif
  60
  61 MODULE_LICENSE("Dual BSD/GPL");
  62
  63 MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS");
  64 MODULE_AUTHOR("Network Appliance, Inc.");
  65
  66 /*
  67  * tunables
  68  */
  69
  70 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
  71 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
  72 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
  73 static unsigned int xprt_rdma_inline_write_padding;
  74 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
  75                 int xprt_rdma_pad_optimize = 0;
  76
  77 #ifdef RPC_DEBUG
  78
  79 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
  80 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
  81 static unsigned int zero;
  82 static unsigned int max_padding = PAGE_SIZE;
  83 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
  84 static unsigned int max_memreg = RPCRDMA_LAST - 1;
  85
  86 static struct ctl_table_header *sunrpc_table_header;
  87
  88 static ctl_table xr_tunables_table[] = {
  89         {
  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                 .extra1         = &min_slot_table_size,
  96                 .extra2         = &max_slot_table_size
  97         },
  98         {
  99                 .procname       = "rdma_max_inline_read",
 100                 .data           = &xprt_rdma_max_inline_read,
 101                 .maxlen         = sizeof(unsigned int),
 102                 .mode           = 0644,
 103                 .proc_handler   = proc_dointvec,
 104         },
 105         {
 106                 .procname       = "rdma_max_inline_write",
 107                 .data           = &xprt_rdma_max_inline_write,
 108                 .maxlen         = sizeof(unsigned int),
 109                 .mode           = 0644,
 110                 .proc_handler   = proc_dointvec,
 111         },
 112         {
 113                 .procname       = "rdma_inline_write_padding",
 114                 .data           = &xprt_rdma_inline_write_padding,
 115                 .maxlen         = sizeof(unsigned int),
 116                 .mode           = 0644,
 117                 .proc_handler   = proc_dointvec_minmax,
 118                 .extra1         = &zero,
 119                 .extra2         = &max_padding,
 120         },
 121         {
 122                 .procname       = "rdma_memreg_strategy",
 123                 .data           = &xprt_rdma_memreg_strategy,
 124                 .maxlen         = sizeof(unsigned int),
 125                 .mode           = 0644,
 126                 .proc_handler   = proc_dointvec_minmax,
 127                 .extra1         = &min_memreg,
 128                 .extra2         = &max_memreg,
 129         },
 130         {
 131                 .procname       = "rdma_pad_optimize",
 132                 .data           = &xprt_rdma_pad_optimize,
 133                 .maxlen         = sizeof(unsigned int),
 134                 .mode           = 0644,
 135                 .proc_handler   = proc_dointvec,
 136         },
 137         { },
 138 };
 139
 140 static ctl_table sunrpc_table[] = {
 141         {
 142                 .procname       = "sunrpc",
 143                 .mode           = 0555,
 144                 .child          = xr_tunables_table
 145         },
 146         { },
 147 };
 148
 149 #endif
 150
 151 static struct rpc_xprt_ops xprt_rdma_procs;     /* forward reference */
 152
 153 static void
 154 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
 155 {
 156         struct sockaddr *sap = (struct sockaddr *)
 157                                         &rpcx_to_rdmad(xprt).addr;
 158         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
 159         char buf[64];
 160
 161         (void)rpc_ntop(sap, buf, sizeof(buf));
 162         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
 163
 164         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
 165         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 166
 167         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
 168
 169         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
 170         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 171
 172         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
 173         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 174
 175         /* netid */
 176         xprt->address_strings[RPC_DISPLAY_NETID] = "rdma";
 177 }
 178
 179 static void
 180 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
 181 {
 182         unsigned int i;
 183
 184         for (i = 0; i < RPC_DISPLAY_MAX; i++)
 185                 switch (i) {
 186                 case RPC_DISPLAY_PROTO:
 187                 case RPC_DISPLAY_NETID:
 188                         continue;
 189                 default:
 190                         kfree(xprt->address_strings[i]);
 191                 }
 192 }
 193
 194 static void
 195 xprt_rdma_connect_worker(struct work_struct *work)
 196 {
 197         struct rpcrdma_xprt *r_xprt =
 198                 container_of(work, struct rpcrdma_xprt, rdma_connect.work);
 199         struct rpc_xprt *xprt = &r_xprt->xprt;
 200         int rc = 0;
 201
 202         current->flags |= PF_FSTRANS;
 203         xprt_clear_connected(xprt);
 204
 205         dprintk("RPC:       %s: %sconnect\n", __func__,
 206                         r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
 207         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 208         if (rc)
 209                 xprt_wake_pending_tasks(xprt, rc);
 210
 211         dprintk("RPC:       %s: exit\n", __func__);
 212         xprt_clear_connecting(xprt);
 213         current->flags &= ~PF_FSTRANS;
 214 }
 215
 216 /*
 217  * xprt_rdma_destroy
 218  *
 219  * Destroy the xprt.
 220  * Free all memory associated with the object, including its own.
 221  * NOTE: none of the *destroy methods free memory for their top-level
 222  * objects, even though they may have allocated it (they do free
 223  * private memory). It's up to the caller to handle it. In this
 224  * case (RDMA transport), all structure memory is inlined with the
 225  * struct rpcrdma_xprt.
 226  */
 227 static void
 228 xprt_rdma_destroy(struct rpc_xprt *xprt)
 229 {
 230         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 231         int rc;
 232
 233         dprintk("RPC:       %s: called\n", __func__);
 234
 235         cancel_delayed_work_sync(&r_xprt->rdma_connect);
 236
 237         xprt_clear_connected(xprt);
 238
 239         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 240         rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
 241         if (rc)
 242                 dprintk("RPC:       %s: rpcrdma_ep_destroy returned %i\n",
 243                         __func__, rc);
 244         rpcrdma_ia_close(&r_xprt->rx_ia);
 245
 246         xprt_rdma_free_addresses(xprt);
 247
 248         xprt_free(xprt);
 249
 250         dprintk("RPC:       %s: returning\n", __func__);
 251
 252         module_put(THIS_MODULE);
 253 }
 254
 255 static const struct rpc_timeout xprt_rdma_default_timeout = {
 256         .to_initval = 60 * HZ,
 257         .to_maxval = 60 * HZ,
 258 };
 259
 260 /**
 261  * xprt_setup_rdma - Set up transport to use RDMA
 262  *
 263  * @args: rpc transport arguments
 264  */
 265 static struct rpc_xprt *
 266 xprt_setup_rdma(struct xprt_create *args)
 267 {
 268         struct rpcrdma_create_data_internal cdata;
 269         struct rpc_xprt *xprt;
 270         struct rpcrdma_xprt *new_xprt;
 271         struct rpcrdma_ep *new_ep;
 272         struct sockaddr_in *sin;
 273         int rc;
 274
 275         if (args->addrlen > sizeof(xprt->addr)) {
 276                 dprintk("RPC:       %s: address too large\n", __func__);
 277                 return ERR_PTR(-EBADF);
 278         }
 279
 280         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
 281                         xprt_rdma_slot_table_entries,
 282                         xprt_rdma_slot_table_entries);
 283         if (xprt == NULL) {
 284                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
 285                         __func__);
 286                 return ERR_PTR(-ENOMEM);
 287         }
 288
 289         /* 60 second timeout, no retries */
 290         xprt->timeout = &xprt_rdma_default_timeout;
 291         xprt->bind_timeout = (60U * HZ);
 292         xprt->reestablish_timeout = (5U * HZ);
 293         xprt->idle_timeout = (5U * 60 * HZ);
 294
 295         xprt->resvport = 0;             /* privileged port not needed */
 296         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
 297         xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE;
 298         xprt->ops = &xprt_rdma_procs;
 299
 300         /*
 301          * Set up RDMA-specific connect data.
 302          */
 303
 304         /* Put server RDMA address in local cdata */
 305         memcpy(&cdata.addr, args->dstaddr, args->addrlen);
 306
 307         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 308          * address, for any side protocols which peek at it */
 309         xprt->prot = IPPROTO_TCP;
 310         xprt->addrlen = args->addrlen;
 311         memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
 312
 313         sin = (struct sockaddr_in *)&cdata.addr;
 314         if (ntohs(sin->sin_port) != 0)
 315                 xprt_set_bound(xprt);
 316
 317         dprintk("RPC:       %s: %pI4:%u\n",
 318                 __func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
 319
 320         /* Set max requests */
 321         cdata.max_requests = xprt->max_reqs;
 322
 323         /* Set some length limits */
 324         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
 325         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
 326
 327         cdata.inline_wsize = xprt_rdma_max_inline_write;
 328         if (cdata.inline_wsize > cdata.wsize)
 329                 cdata.inline_wsize = cdata.wsize;
 330
 331         cdata.inline_rsize = xprt_rdma_max_inline_read;
 332         if (cdata.inline_rsize > cdata.rsize)
 333                 cdata.inline_rsize = cdata.rsize;
 334
 335         cdata.padding = xprt_rdma_inline_write_padding;
 336
 337         /*
 338          * Create new transport instance, which includes initialized
 339          *  o ia
 340          *  o endpoint
 341          *  o buffers
 342          */
 343
 344         new_xprt = rpcx_to_rdmax(xprt);
 345
 346         rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
 347                                 xprt_rdma_memreg_strategy);
 348         if (rc)
 349                 goto out1;
 350
 351         /*
 352          * initialize and create ep
 353          */
 354         new_xprt->rx_data = cdata;
 355         new_ep = &new_xprt->rx_ep;
 356         new_ep->rep_remote_addr = cdata.addr;
 357
 358         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
 359                                 &new_xprt->rx_ia, &new_xprt->rx_data);
 360         if (rc)
 361                 goto out2;
 362
 363         /*
 364          * Allocate pre-registered send and receive buffers for headers and
 365          * any inline data. Also specify any padding which will be provided
 366          * from a preregistered zero buffer.
 367          */
 368         rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia,
 369                                 &new_xprt->rx_data);
 370         if (rc)
 371                 goto out3;
 372
 373         /*
 374          * Register a callback for connection events. This is necessary because
 375          * connection loss notification is async. We also catch connection loss
 376          * when reaping receives.
 377          */
 378         INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker);
 379         new_ep->rep_func = rpcrdma_conn_func;
 380         new_ep->rep_xprt = xprt;
 381
 382         xprt_rdma_format_addresses(xprt);
 383
 384         if (!try_module_get(THIS_MODULE))
 385                 goto out4;
 386
 387         return xprt;
 388
 389 out4:
 390         xprt_rdma_free_addresses(xprt);
 391         rc = -EINVAL;
 392 out3:
 393         (void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
 394 out2:
 395         rpcrdma_ia_close(&new_xprt->rx_ia);
 396 out1:
 397         xprt_free(xprt);
 398         return ERR_PTR(rc);
 399 }
 400
 401 /*
 402  * Close a connection, during shutdown or timeout/reconnect
 403  */
 404 static void
 405 xprt_rdma_close(struct rpc_xprt *xprt)
 406 {
 407         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 408
 409         dprintk("RPC:       %s: closing\n", __func__);
 410         if (r_xprt->rx_ep.rep_connected > 0)
 411                 xprt->reestablish_timeout = 0;
 412         xprt_disconnect_done(xprt);
 413         (void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 414 }
 415
 416 static void
 417 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 418 {
 419         struct sockaddr_in *sap;
 420
 421         sap = (struct sockaddr_in *)&xprt->addr;
 422         sap->sin_port = htons(port);
 423         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
 424         sap->sin_port = htons(port);
 425         dprintk("RPC:       %s: %u\n", __func__, port);
 426 }
 427
 428 static void
 429 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
 430 {
 431         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 432
 433         if (r_xprt->rx_ep.rep_connected != 0) {
 434                 /* Reconnect */
 435                 schedule_delayed_work(&r_xprt->rdma_connect,
 436                         xprt->reestablish_timeout);
 437                 xprt->reestablish_timeout <<= 1;
 438                 if (xprt->reestablish_timeout > (30 * HZ))
 439                         xprt->reestablish_timeout = (30 * HZ);
 440                 else if (xprt->reestablish_timeout < (5 * HZ))
 441                         xprt->reestablish_timeout = (5 * HZ);
 442         } else {
 443                 schedule_delayed_work(&r_xprt->rdma_connect, 0);
 444                 if (!RPC_IS_ASYNC(task))
 445                         flush_delayed_work(&r_xprt->rdma_connect);
 446         }
 447 }
 448
 449 static int
 450 xprt_rdma_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 451 {
 452         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 453         int credits = atomic_read(&r_xprt->rx_buf.rb_credits);
 454
 455         /* == RPC_CWNDSCALE @ init, but *after* setup */
 456         if (r_xprt->rx_buf.rb_cwndscale == 0UL) {
 457                 r_xprt->rx_buf.rb_cwndscale = xprt->cwnd;
 458                 dprintk("RPC:       %s: cwndscale %lu\n", __func__,
 459                         r_xprt->rx_buf.rb_cwndscale);
 460                 BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0);
 461         }
 462         xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale;
 463         return xprt_reserve_xprt_cong(xprt, task);
 464 }
 465
 466 /*
 467  * The RDMA allocate/free functions need the task structure as a place
 468  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
 469  * sequence. For this reason, the recv buffers are attached to send
 470  * buffers for portions of the RPC. Note that the RPC layer allocates
 471  * both send and receive buffers in the same call. We may register
 472  * the receive buffer portion when using reply chunks.
 473  */
 474 static void *
 475 xprt_rdma_allocate(struct rpc_task *task, size_t size)
 476 {
 477         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 478         struct rpcrdma_req *req, *nreq;
 479
 480         req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf);
 481         BUG_ON(NULL == req);
 482
 483         if (size > req->rl_size) {
 484                 dprintk("RPC:       %s: size %zd too large for buffer[%zd]: "
 485                         "prog %d vers %d proc %d\n",
 486                         __func__, size, req->rl_size,
 487                         task->tk_client->cl_prog, task->tk_client->cl_vers,
 488                         task->tk_msg.rpc_proc->p_proc);
 489                 /*
 490                  * Outgoing length shortage. Our inline write max must have
 491                  * been configured to perform direct i/o.
 492                  *
 493                  * This is therefore a large metadata operation, and the
 494                  * allocate call was made on the maximum possible message,
 495                  * e.g. containing long filename(s) or symlink data. In
 496                  * fact, while these metadata operations *might* carry
 497                  * large outgoing payloads, they rarely *do*. However, we
 498                  * have to commit to the request here, so reallocate and
 499                  * register it now. The data path will never require this
 500                  * reallocation.
 501                  *
 502                  * If the allocation or registration fails, the RPC framework
 503                  * will (doggedly) retry.
 504                  */
 505                 if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy ==
 506                                 RPCRDMA_BOUNCEBUFFERS) {
 507                         /* forced to "pure inline" */
 508                         dprintk("RPC:       %s: too much data (%zd) for inline "
 509                                         "(r/w max %d/%d)\n", __func__, size,
 510                                         rpcx_to_rdmad(xprt).inline_rsize,
 511                                         rpcx_to_rdmad(xprt).inline_wsize);
 512                         size = req->rl_size;
 513                         rpc_exit(task, -EIO);           /* fail the operation */
 514                         rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
 515                         goto out;
 516                 }
 517                 if (task->tk_flags & RPC_TASK_SWAPPER)
 518                         nreq = kmalloc(sizeof *req + size, GFP_ATOMIC);
 519                 else
 520                         nreq = kmalloc(sizeof *req + size, GFP_NOFS);
 521                 if (nreq == NULL)
 522                         goto outfail;
 523
 524                 if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia,
 525                                 nreq->rl_base, size + sizeof(struct rpcrdma_req)
 526                                 - offsetof(struct rpcrdma_req, rl_base),
 527                                 &nreq->rl_handle, &nreq->rl_iov)) {
 528                         kfree(nreq);
 529                         goto outfail;
 530                 }
 531                 rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size;
 532                 nreq->rl_size = size;
 533                 nreq->rl_niovs = 0;
 534                 nreq->rl_nchunks = 0;
 535                 nreq->rl_buffer = (struct rpcrdma_buffer *)req;
 536                 nreq->rl_reply = req->rl_reply;
 537                 memcpy(nreq->rl_segments,
 538                         req->rl_segments, sizeof nreq->rl_segments);
 539                 /* flag the swap with an unused field */
 540                 nreq->rl_iov.length = 0;
 541                 req->rl_reply = NULL;
 542                 req = nreq;
 543         }
 544         dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
 545 out:
 546         req->rl_connect_cookie = 0;     /* our reserved value */
 547         return req->rl_xdr_buf;
 548
 549 outfail:
 550         rpcrdma_buffer_put(req);
 551         rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
 552         return NULL;
 553 }
 554
 555 /*
 556  * This function returns all RDMA resources to the pool.
 557  */
 558 static void
 559 xprt_rdma_free(void *buffer)
 560 {
 561         struct rpcrdma_req *req;
 562         struct rpcrdma_xprt *r_xprt;
 563         struct rpcrdma_rep *rep;
 564         int i;
 565
 566         if (buffer == NULL)
 567                 return;
 568
 569         req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]);
 570         if (req->rl_iov.length == 0) {  /* see allocate above */
 571                 r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer,
 572                                       struct rpcrdma_xprt, rx_buf);
 573         } else
 574                 r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
 575         rep = req->rl_reply;
 576
 577         dprintk("RPC:       %s: called on 0x%p%s\n",
 578                 __func__, rep, (rep && rep->rr_func) ? " (with waiter)" : "");
 579
 580         /*
 581          * Finish the deregistration. When using mw bind, this was
 582          * begun in rpcrdma_reply_handler(). In all other modes, we
 583          * do it here, in thread context. The process is considered
 584          * complete when the rr_func vector becomes NULL - this
 585          * was put in place during rpcrdma_reply_handler() - the wait
 586          * call below will not block if the dereg is "done". If
 587          * interrupted, our framework will clean up.
 588          */
 589         for (i = 0; req->rl_nchunks;) {
 590                 --req->rl_nchunks;
 591                 i += rpcrdma_deregister_external(
 592                         &req->rl_segments[i], r_xprt, NULL);
 593         }
 594
 595         if (rep && wait_event_interruptible(rep->rr_unbind, !rep->rr_func)) {
 596                 rep->rr_func = NULL;    /* abandon the callback */
 597                 req->rl_reply = NULL;
 598         }
 599
 600         if (req->rl_iov.length == 0) {  /* see allocate above */
 601                 struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer;
 602                 oreq->rl_reply = req->rl_reply;
 603                 (void) rpcrdma_deregister_internal(&r_xprt->rx_ia,
 604                                                    req->rl_handle,
 605                                                    &req->rl_iov);
 606                 kfree(req);
 607                 req = oreq;
 608         }
 609
 610         /* Put back request+reply buffers */
 611         rpcrdma_buffer_put(req);
 612 }
 613
 614 /*
 615  * send_request invokes the meat of RPC RDMA. It must do the following:
 616  *  1.  Marshal the RPC request into an RPC RDMA request, which means
 617  *      putting a header in front of data, and creating IOVs for RDMA
 618  *      from those in the request.
 619  *  2.  In marshaling, detect opportunities for RDMA, and use them.
 620  *  3.  Post a recv message to set up asynch completion, then send
 621  *      the request (rpcrdma_ep_post).
 622  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
 623  */
 624
 625 static int
 626 xprt_rdma_send_request(struct rpc_task *task)
 627 {
 628         struct rpc_rqst *rqst = task->tk_rqstp;
 629         struct rpc_xprt *xprt = rqst->rq_xprt;
 630         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 631         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 632
 633         /* marshal the send itself */
 634         if (req->rl_niovs == 0 && rpcrdma_marshal_req(rqst) != 0) {
 635                 r_xprt->rx_stats.failed_marshal_count++;
 636                 dprintk("RPC:       %s: rpcrdma_marshal_req failed\n",
 637                         __func__);
 638                 return -EIO;
 639         }
 640
 641         if (req->rl_reply == NULL)              /* e.g. reconnection */
 642                 rpcrdma_recv_buffer_get(req);
 643
 644         if (req->rl_reply) {
 645                 req->rl_reply->rr_func = rpcrdma_reply_handler;
 646                 /* this need only be done once, but... */
 647                 req->rl_reply->rr_xprt = xprt;
 648         }
 649
 650         /* Must suppress retransmit to maintain credits */
 651         if (req->rl_connect_cookie == xprt->connect_cookie)
 652                 goto drop_connection;
 653         req->rl_connect_cookie = xprt->connect_cookie;
 654
 655         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 656                 goto drop_connection;
 657
 658         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
 659         rqst->rq_bytes_sent = 0;
 660         return 0;
 661
 662 drop_connection:
 663         xprt_disconnect_done(xprt);
 664         return -ENOTCONN;       /* implies disconnect */
 665 }
 666
 667 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 668 {
 669         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 670         long idle_time = 0;
 671
 672         if (xprt_connected(xprt))
 673                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 674
 675         seq_printf(seq,
 676           "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
 677           "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
 678
 679            0,   /* need a local port? */
 680            xprt->stat.bind_count,
 681            xprt->stat.connect_count,
 682            xprt->stat.connect_time,
 683            idle_time,
 684            xprt->stat.sends,
 685            xprt->stat.recvs,
 686            xprt->stat.bad_xids,
 687            xprt->stat.req_u,
 688            xprt->stat.bklog_u,
 689
 690            r_xprt->rx_stats.read_chunk_count,
 691            r_xprt->rx_stats.write_chunk_count,
 692            r_xprt->rx_stats.reply_chunk_count,
 693            r_xprt->rx_stats.total_rdma_request,
 694            r_xprt->rx_stats.total_rdma_reply,
 695            r_xprt->rx_stats.pullup_copy_count,
 696            r_xprt->rx_stats.fixup_copy_count,
 697            r_xprt->rx_stats.hardway_register_count,
 698            r_xprt->rx_stats.failed_marshal_count,
 699            r_xprt->rx_stats.bad_reply_count);
 700 }
 701
 702 /*
 703  * Plumbing for rpc transport switch and kernel module
 704  */
 705
 706 static struct rpc_xprt_ops xprt_rdma_procs = {
 707         .reserve_xprt           = xprt_rdma_reserve_xprt,
 708         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 709         .alloc_slot             = xprt_alloc_slot,
 710         .release_request        = xprt_release_rqst_cong,       /* ditto */
 711         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
 712         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 713         .set_port               = xprt_rdma_set_port,
 714         .connect                = xprt_rdma_connect,
 715         .buf_alloc              = xprt_rdma_allocate,
 716         .buf_free               = xprt_rdma_free,
 717         .send_request           = xprt_rdma_send_request,
 718         .close                  = xprt_rdma_close,
 719         .destroy                = xprt_rdma_destroy,
 720         .print_stats            = xprt_rdma_print_stats
 721 };
 722
 723 static struct xprt_class xprt_rdma = {
 724         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 725         .name                   = "rdma",
 726         .owner                  = THIS_MODULE,
 727         .ident                  = XPRT_TRANSPORT_RDMA,
 728         .setup                  = xprt_setup_rdma,
 729 };
 730
 731 static void __exit xprt_rdma_cleanup(void)
 732 {
 733         int rc;
 734
 735         dprintk(KERN_INFO "RPCRDMA Module Removed, deregister RPC RDMA transport\n");
 736 #ifdef RPC_DEBUG
 737         if (sunrpc_table_header) {
 738                 unregister_sysctl_table(sunrpc_table_header);
 739                 sunrpc_table_header = NULL;
 740         }
 741 #endif
 742         rc = xprt_unregister_transport(&xprt_rdma);
 743         if (rc)
 744                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
 745                         __func__, rc);
 746 }
 747
 748 static int __init xprt_rdma_init(void)
 749 {
 750         int rc;
 751
 752         rc = xprt_register_transport(&xprt_rdma);
 753
 754         if (rc)
 755                 return rc;
 756
 757         dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n");
 758
 759         dprintk(KERN_INFO "Defaults:\n");
 760         dprintk(KERN_INFO "\tSlots %d\n"
 761                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
 762                 xprt_rdma_slot_table_entries,
 763                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
 764         dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n",
 765                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
 766
 767 #ifdef RPC_DEBUG
 768         if (!sunrpc_table_header)
 769                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 770 #endif
 771         return 0;
 772 }
 773
 774 module_init(xprt_rdma_init);
 775 module_exit(xprt_rdma_cleanup);