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         if (!xprt->shutdown) {
 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                         goto out;
 210         }
 211         goto out_clear;
 212
 213 out:
 214         xprt_wake_pending_tasks(xprt, rc);
 215
 216 out_clear:
 217         dprintk("RPC:       %s: exit\n", __func__);
 218         xprt_clear_connecting(xprt);
 219 }
 220
 221 /*
 222  * xprt_rdma_destroy
 223  *
 224  * Destroy the xprt.
 225  * Free all memory associated with the object, including its own.
 226  * NOTE: none of the *destroy methods free memory for their top-level
 227  * objects, even though they may have allocated it (they do free
 228  * private memory). It's up to the caller to handle it. In this
 229  * case (RDMA transport), all structure memory is inlined with the
 230  * struct rpcrdma_xprt.
 231  */
 232 static void
 233 xprt_rdma_destroy(struct rpc_xprt *xprt)
 234 {
 235         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 236         int rc;
 237
 238         dprintk("RPC:       %s: called\n", __func__);
 239
 240         cancel_delayed_work(&r_xprt->rdma_connect);
 241         flush_scheduled_work();
 242
 243         xprt_clear_connected(xprt);
 244
 245         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 246         rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
 247         if (rc)
 248                 dprintk("RPC:       %s: rpcrdma_ep_destroy returned %i\n",
 249                         __func__, rc);
 250         rpcrdma_ia_close(&r_xprt->rx_ia);
 251
 252         xprt_rdma_free_addresses(xprt);
 253
 254         kfree(xprt->slot);
 255         xprt->slot = NULL;
 256         kfree(xprt);
 257
 258         dprintk("RPC:       %s: returning\n", __func__);
 259
 260         module_put(THIS_MODULE);
 261 }
 262
 263 static const struct rpc_timeout xprt_rdma_default_timeout = {
 264         .to_initval = 60 * HZ,
 265         .to_maxval = 60 * HZ,
 266 };
 267
 268 /**
 269  * xprt_setup_rdma - Set up transport to use RDMA
 270  *
 271  * @args: rpc transport arguments
 272  */
 273 static struct rpc_xprt *
 274 xprt_setup_rdma(struct xprt_create *args)
 275 {
 276         struct rpcrdma_create_data_internal cdata;
 277         struct rpc_xprt *xprt;
 278         struct rpcrdma_xprt *new_xprt;
 279         struct rpcrdma_ep *new_ep;
 280         struct sockaddr_in *sin;
 281         int rc;
 282
 283         if (args->addrlen > sizeof(xprt->addr)) {
 284                 dprintk("RPC:       %s: address too large\n", __func__);
 285                 return ERR_PTR(-EBADF);
 286         }
 287
 288         xprt = kzalloc(sizeof(struct rpcrdma_xprt), GFP_KERNEL);
 289         if (xprt == NULL) {
 290                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
 291                         __func__);
 292                 return ERR_PTR(-ENOMEM);
 293         }
 294
 295         xprt->max_reqs = xprt_rdma_slot_table_entries;
 296         xprt->slot = kcalloc(xprt->max_reqs,
 297                                 sizeof(struct rpc_rqst), GFP_KERNEL);
 298         if (xprt->slot == NULL) {
 299                 dprintk("RPC:       %s: couldn't allocate %d slots\n",
 300                         __func__, xprt->max_reqs);
 301                 kfree(xprt);
 302                 return ERR_PTR(-ENOMEM);
 303         }
 304
 305         /* 60 second timeout, no retries */
 306         xprt->timeout = &xprt_rdma_default_timeout;
 307         xprt->bind_timeout = (60U * HZ);
 308         xprt->reestablish_timeout = (5U * HZ);
 309         xprt->idle_timeout = (5U * 60 * HZ);
 310
 311         xprt->resvport = 0;             /* privileged port not needed */
 312         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
 313         xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE;
 314         xprt->ops = &xprt_rdma_procs;
 315
 316         /*
 317          * Set up RDMA-specific connect data.
 318          */
 319
 320         /* Put server RDMA address in local cdata */
 321         memcpy(&cdata.addr, args->dstaddr, args->addrlen);
 322
 323         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 324          * address, for any side protocols which peek at it */
 325         xprt->prot = IPPROTO_TCP;
 326         xprt->addrlen = args->addrlen;
 327         memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
 328
 329         sin = (struct sockaddr_in *)&cdata.addr;
 330         if (ntohs(sin->sin_port) != 0)
 331                 xprt_set_bound(xprt);
 332
 333         dprintk("RPC:       %s: %pI4:%u\n",
 334                 __func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
 335
 336         /* Set max requests */
 337         cdata.max_requests = xprt->max_reqs;
 338
 339         /* Set some length limits */
 340         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
 341         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
 342
 343         cdata.inline_wsize = xprt_rdma_max_inline_write;
 344         if (cdata.inline_wsize > cdata.wsize)
 345                 cdata.inline_wsize = cdata.wsize;
 346
 347         cdata.inline_rsize = xprt_rdma_max_inline_read;
 348         if (cdata.inline_rsize > cdata.rsize)
 349                 cdata.inline_rsize = cdata.rsize;
 350
 351         cdata.padding = xprt_rdma_inline_write_padding;
 352
 353         /*
 354          * Create new transport instance, which includes initialized
 355          *  o ia
 356          *  o endpoint
 357          *  o buffers
 358          */
 359
 360         new_xprt = rpcx_to_rdmax(xprt);
 361
 362         rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
 363                                 xprt_rdma_memreg_strategy);
 364         if (rc)
 365                 goto out1;
 366
 367         /*
 368          * initialize and create ep
 369          */
 370         new_xprt->rx_data = cdata;
 371         new_ep = &new_xprt->rx_ep;
 372         new_ep->rep_remote_addr = cdata.addr;
 373
 374         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
 375                                 &new_xprt->rx_ia, &new_xprt->rx_data);
 376         if (rc)
 377                 goto out2;
 378
 379         /*
 380          * Allocate pre-registered send and receive buffers for headers and
 381          * any inline data. Also specify any padding which will be provided
 382          * from a preregistered zero buffer.
 383          */
 384         rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia,
 385                                 &new_xprt->rx_data);
 386         if (rc)
 387                 goto out3;
 388
 389         /*
 390          * Register a callback for connection events. This is necessary because
 391          * connection loss notification is async. We also catch connection loss
 392          * when reaping receives.
 393          */
 394         INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker);
 395         new_ep->rep_func = rpcrdma_conn_func;
 396         new_ep->rep_xprt = xprt;
 397
 398         xprt_rdma_format_addresses(xprt);
 399
 400         if (!try_module_get(THIS_MODULE))
 401                 goto out4;
 402
 403         return xprt;
 404
 405 out4:
 406         xprt_rdma_free_addresses(xprt);
 407         rc = -EINVAL;
 408 out3:
 409         (void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
 410 out2:
 411         rpcrdma_ia_close(&new_xprt->rx_ia);
 412 out1:
 413         kfree(xprt->slot);
 414         kfree(xprt);
 415         return ERR_PTR(rc);
 416 }
 417
 418 /*
 419  * Close a connection, during shutdown or timeout/reconnect
 420  */
 421 static void
 422 xprt_rdma_close(struct rpc_xprt *xprt)
 423 {
 424         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 425
 426         dprintk("RPC:       %s: closing\n", __func__);
 427         if (r_xprt->rx_ep.rep_connected > 0)
 428                 xprt->reestablish_timeout = 0;
 429         xprt_disconnect_done(xprt);
 430         (void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 431 }
 432
 433 static void
 434 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 435 {
 436         struct sockaddr_in *sap;
 437
 438         sap = (struct sockaddr_in *)&xprt->addr;
 439         sap->sin_port = htons(port);
 440         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
 441         sap->sin_port = htons(port);
 442         dprintk("RPC:       %s: %u\n", __func__, port);
 443 }
 444
 445 static void
 446 xprt_rdma_connect(struct rpc_task *task)
 447 {
 448         struct rpc_xprt *xprt = (struct rpc_xprt *)task->tk_xprt;
 449         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 450
 451         if (r_xprt->rx_ep.rep_connected != 0) {
 452                 /* Reconnect */
 453                 schedule_delayed_work(&r_xprt->rdma_connect,
 454                         xprt->reestablish_timeout);
 455                 xprt->reestablish_timeout <<= 1;
 456                 if (xprt->reestablish_timeout > (30 * HZ))
 457                         xprt->reestablish_timeout = (30 * HZ);
 458                 else if (xprt->reestablish_timeout < (5 * HZ))
 459                         xprt->reestablish_timeout = (5 * HZ);
 460         } else {
 461                 schedule_delayed_work(&r_xprt->rdma_connect, 0);
 462                 if (!RPC_IS_ASYNC(task))
 463                         flush_scheduled_work();
 464         }
 465 }
 466
 467 static int
 468 xprt_rdma_reserve_xprt(struct rpc_task *task)
 469 {
 470         struct rpc_xprt *xprt = task->tk_xprt;
 471         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 472         int credits = atomic_read(&r_xprt->rx_buf.rb_credits);
 473
 474         /* == RPC_CWNDSCALE @ init, but *after* setup */
 475         if (r_xprt->rx_buf.rb_cwndscale == 0UL) {
 476                 r_xprt->rx_buf.rb_cwndscale = xprt->cwnd;
 477                 dprintk("RPC:       %s: cwndscale %lu\n", __func__,
 478                         r_xprt->rx_buf.rb_cwndscale);
 479                 BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0);
 480         }
 481         xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale;
 482         return xprt_reserve_xprt_cong(task);
 483 }
 484
 485 /*
 486  * The RDMA allocate/free functions need the task structure as a place
 487  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
 488  * sequence. For this reason, the recv buffers are attached to send
 489  * buffers for portions of the RPC. Note that the RPC layer allocates
 490  * both send and receive buffers in the same call. We may register
 491  * the receive buffer portion when using reply chunks.
 492  */
 493 static void *
 494 xprt_rdma_allocate(struct rpc_task *task, size_t size)
 495 {
 496         struct rpc_xprt *xprt = task->tk_xprt;
 497         struct rpcrdma_req *req, *nreq;
 498
 499         req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf);
 500         BUG_ON(NULL == req);
 501
 502         if (size > req->rl_size) {
 503                 dprintk("RPC:       %s: size %zd too large for buffer[%zd]: "
 504                         "prog %d vers %d proc %d\n",
 505                         __func__, size, req->rl_size,
 506                         task->tk_client->cl_prog, task->tk_client->cl_vers,
 507                         task->tk_msg.rpc_proc->p_proc);
 508                 /*
 509                  * Outgoing length shortage. Our inline write max must have
 510                  * been configured to perform direct i/o.
 511                  *
 512                  * This is therefore a large metadata operation, and the
 513                  * allocate call was made on the maximum possible message,
 514                  * e.g. containing long filename(s) or symlink data. In
 515                  * fact, while these metadata operations *might* carry
 516                  * large outgoing payloads, they rarely *do*. However, we
 517                  * have to commit to the request here, so reallocate and
 518                  * register it now. The data path will never require this
 519                  * reallocation.
 520                  *
 521                  * If the allocation or registration fails, the RPC framework
 522                  * will (doggedly) retry.
 523                  */
 524                 if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy ==
 525                                 RPCRDMA_BOUNCEBUFFERS) {
 526                         /* forced to "pure inline" */
 527                         dprintk("RPC:       %s: too much data (%zd) for inline "
 528                                         "(r/w max %d/%d)\n", __func__, size,
 529                                         rpcx_to_rdmad(xprt).inline_rsize,
 530                                         rpcx_to_rdmad(xprt).inline_wsize);
 531                         size = req->rl_size;
 532                         rpc_exit(task, -EIO);           /* fail the operation */
 533                         rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
 534                         goto out;
 535                 }
 536                 if (task->tk_flags & RPC_TASK_SWAPPER)
 537                         nreq = kmalloc(sizeof *req + size, GFP_ATOMIC);
 538                 else
 539                         nreq = kmalloc(sizeof *req + size, GFP_NOFS);
 540                 if (nreq == NULL)
 541                         goto outfail;
 542
 543                 if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia,
 544                                 nreq->rl_base, size + sizeof(struct rpcrdma_req)
 545                                 - offsetof(struct rpcrdma_req, rl_base),
 546                                 &nreq->rl_handle, &nreq->rl_iov)) {
 547                         kfree(nreq);
 548                         goto outfail;
 549                 }
 550                 rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size;
 551                 nreq->rl_size = size;
 552                 nreq->rl_niovs = 0;
 553                 nreq->rl_nchunks = 0;
 554                 nreq->rl_buffer = (struct rpcrdma_buffer *)req;
 555                 nreq->rl_reply = req->rl_reply;
 556                 memcpy(nreq->rl_segments,
 557                         req->rl_segments, sizeof nreq->rl_segments);
 558                 /* flag the swap with an unused field */
 559                 nreq->rl_iov.length = 0;
 560                 req->rl_reply = NULL;
 561                 req = nreq;
 562         }
 563         dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
 564 out:
 565         req->rl_connect_cookie = 0;     /* our reserved value */
 566         return req->rl_xdr_buf;
 567
 568 outfail:
 569         rpcrdma_buffer_put(req);
 570         rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
 571         return NULL;
 572 }
 573
 574 /*
 575  * This function returns all RDMA resources to the pool.
 576  */
 577 static void
 578 xprt_rdma_free(void *buffer)
 579 {
 580         struct rpcrdma_req *req;
 581         struct rpcrdma_xprt *r_xprt;
 582         struct rpcrdma_rep *rep;
 583         int i;
 584
 585         if (buffer == NULL)
 586                 return;
 587
 588         req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]);
 589         if (req->rl_iov.length == 0) {  /* see allocate above */
 590                 r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer,
 591                                       struct rpcrdma_xprt, rx_buf);
 592         } else
 593                 r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
 594         rep = req->rl_reply;
 595
 596         dprintk("RPC:       %s: called on 0x%p%s\n",
 597                 __func__, rep, (rep && rep->rr_func) ? " (with waiter)" : "");
 598
 599         /*
 600          * Finish the deregistration. When using mw bind, this was
 601          * begun in rpcrdma_reply_handler(). In all other modes, we
 602          * do it here, in thread context. The process is considered
 603          * complete when the rr_func vector becomes NULL - this
 604          * was put in place during rpcrdma_reply_handler() - the wait
 605          * call below will not block if the dereg is "done". If
 606          * interrupted, our framework will clean up.
 607          */
 608         for (i = 0; req->rl_nchunks;) {
 609                 --req->rl_nchunks;
 610                 i += rpcrdma_deregister_external(
 611                         &req->rl_segments[i], r_xprt, NULL);
 612         }
 613
 614         if (rep && wait_event_interruptible(rep->rr_unbind, !rep->rr_func)) {
 615                 rep->rr_func = NULL;    /* abandon the callback */
 616                 req->rl_reply = NULL;
 617         }
 618
 619         if (req->rl_iov.length == 0) {  /* see allocate above */
 620                 struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer;
 621                 oreq->rl_reply = req->rl_reply;
 622                 (void) rpcrdma_deregister_internal(&r_xprt->rx_ia,
 623                                                    req->rl_handle,
 624                                                    &req->rl_iov);
 625                 kfree(req);
 626                 req = oreq;
 627         }
 628
 629         /* Put back request+reply buffers */
 630         rpcrdma_buffer_put(req);
 631 }
 632
 633 /*
 634  * send_request invokes the meat of RPC RDMA. It must do the following:
 635  *  1.  Marshal the RPC request into an RPC RDMA request, which means
 636  *      putting a header in front of data, and creating IOVs for RDMA
 637  *      from those in the request.
 638  *  2.  In marshaling, detect opportunities for RDMA, and use them.
 639  *  3.  Post a recv message to set up asynch completion, then send
 640  *      the request (rpcrdma_ep_post).
 641  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
 642  */
 643
 644 static int
 645 xprt_rdma_send_request(struct rpc_task *task)
 646 {
 647         struct rpc_rqst *rqst = task->tk_rqstp;
 648         struct rpc_xprt *xprt = task->tk_xprt;
 649         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 650         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 651
 652         /* marshal the send itself */
 653         if (req->rl_niovs == 0 && rpcrdma_marshal_req(rqst) != 0) {
 654                 r_xprt->rx_stats.failed_marshal_count++;
 655                 dprintk("RPC:       %s: rpcrdma_marshal_req failed\n",
 656                         __func__);
 657                 return -EIO;
 658         }
 659
 660         if (req->rl_reply == NULL)              /* e.g. reconnection */
 661                 rpcrdma_recv_buffer_get(req);
 662
 663         if (req->rl_reply) {
 664                 req->rl_reply->rr_func = rpcrdma_reply_handler;
 665                 /* this need only be done once, but... */
 666                 req->rl_reply->rr_xprt = xprt;
 667         }
 668
 669         /* Must suppress retransmit to maintain credits */
 670         if (req->rl_connect_cookie == xprt->connect_cookie)
 671                 goto drop_connection;
 672         req->rl_connect_cookie = xprt->connect_cookie;
 673
 674         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 675                 goto drop_connection;
 676
 677         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
 678         rqst->rq_bytes_sent = 0;
 679         return 0;
 680
 681 drop_connection:
 682         xprt_disconnect_done(xprt);
 683         return -ENOTCONN;       /* implies disconnect */
 684 }
 685
 686 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 687 {
 688         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 689         long idle_time = 0;
 690
 691         if (xprt_connected(xprt))
 692                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 693
 694         seq_printf(seq,
 695           "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
 696           "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
 697
 698            0,   /* need a local port? */
 699            xprt->stat.bind_count,
 700            xprt->stat.connect_count,
 701            xprt->stat.connect_time,
 702            idle_time,
 703            xprt->stat.sends,
 704            xprt->stat.recvs,
 705            xprt->stat.bad_xids,
 706            xprt->stat.req_u,
 707            xprt->stat.bklog_u,
 708
 709            r_xprt->rx_stats.read_chunk_count,
 710            r_xprt->rx_stats.write_chunk_count,
 711            r_xprt->rx_stats.reply_chunk_count,
 712            r_xprt->rx_stats.total_rdma_request,
 713            r_xprt->rx_stats.total_rdma_reply,
 714            r_xprt->rx_stats.pullup_copy_count,
 715            r_xprt->rx_stats.fixup_copy_count,
 716            r_xprt->rx_stats.hardway_register_count,
 717            r_xprt->rx_stats.failed_marshal_count,
 718            r_xprt->rx_stats.bad_reply_count);
 719 }
 720
 721 /*
 722  * Plumbing for rpc transport switch and kernel module
 723  */
 724
 725 static struct rpc_xprt_ops xprt_rdma_procs = {
 726         .reserve_xprt           = xprt_rdma_reserve_xprt,
 727         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 728         .release_request        = xprt_release_rqst_cong,       /* ditto */
 729         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
 730         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 731         .set_port               = xprt_rdma_set_port,
 732         .connect                = xprt_rdma_connect,
 733         .buf_alloc              = xprt_rdma_allocate,
 734         .buf_free               = xprt_rdma_free,
 735         .send_request           = xprt_rdma_send_request,
 736         .close                  = xprt_rdma_close,
 737         .destroy                = xprt_rdma_destroy,
 738         .print_stats            = xprt_rdma_print_stats
 739 };
 740
 741 static struct xprt_class xprt_rdma = {
 742         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 743         .name                   = "rdma",
 744         .owner                  = THIS_MODULE,
 745         .ident                  = XPRT_TRANSPORT_RDMA,
 746         .setup                  = xprt_setup_rdma,
 747 };
 748
 749 static void __exit xprt_rdma_cleanup(void)
 750 {
 751         int rc;
 752
 753         dprintk(KERN_INFO "RPCRDMA Module Removed, deregister RPC RDMA transport\n");
 754 #ifdef RPC_DEBUG
 755         if (sunrpc_table_header) {
 756                 unregister_sysctl_table(sunrpc_table_header);
 757                 sunrpc_table_header = NULL;
 758         }
 759 #endif
 760         rc = xprt_unregister_transport(&xprt_rdma);
 761         if (rc)
 762                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
 763                         __func__, rc);
 764 }
 765
 766 static int __init xprt_rdma_init(void)
 767 {
 768         int rc;
 769
 770         rc = xprt_register_transport(&xprt_rdma);
 771
 772         if (rc)
 773                 return rc;
 774
 775         dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n");
 776
 777         dprintk(KERN_INFO "Defaults:\n");
 778         dprintk(KERN_INFO "\tSlots %d\n"
 779                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
 780                 xprt_rdma_slot_table_entries,
 781                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
 782         dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n",
 783                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
 784
 785 #ifdef RPC_DEBUG
 786         if (!sunrpc_table_header)
 787                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 788 #endif
 789         return 0;
 790 }
 791
 792 module_init(xprt_rdma_init);
 793 module_exit(xprt_rdma_cleanup);