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 #include <linux/sunrpc/addr.h>
  55
  56 #include "xprt_rdma.h"
  57
  58 #ifdef RPC_DEBUG
  59 # define RPCDBG_FACILITY        RPCDBG_TRANS
  60 #endif
  61
  62 MODULE_LICENSE("Dual BSD/GPL");
  63
  64 MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS");
  65 MODULE_AUTHOR("Network Appliance, Inc.");
  66
  67 /*
  68  * tunables
  69  */
  70
  71 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
  72 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
  73 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
  74 static unsigned int xprt_rdma_inline_write_padding;
  75 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
  76                 int xprt_rdma_pad_optimize = 0;
  77
  78 #ifdef RPC_DEBUG
  79
  80 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
  81 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
  82 static unsigned int zero;
  83 static unsigned int max_padding = PAGE_SIZE;
  84 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
  85 static unsigned int max_memreg = RPCRDMA_LAST - 1;
  86
  87 static struct ctl_table_header *sunrpc_table_header;
  88
  89 static struct ctl_table xr_tunables_table[] = {
  90         {
  91                 .procname       = "rdma_slot_table_entries",
  92                 .data           = &xprt_rdma_slot_table_entries,
  93                 .maxlen         = sizeof(unsigned int),
  94                 .mode           = 0644,
  95                 .proc_handler   = proc_dointvec_minmax,
  96                 .extra1         = &min_slot_table_size,
  97                 .extra2         = &max_slot_table_size
  98         },
  99         {
 100                 .procname       = "rdma_max_inline_read",
 101                 .data           = &xprt_rdma_max_inline_read,
 102                 .maxlen         = sizeof(unsigned int),
 103                 .mode           = 0644,
 104                 .proc_handler   = proc_dointvec,
 105         },
 106         {
 107                 .procname       = "rdma_max_inline_write",
 108                 .data           = &xprt_rdma_max_inline_write,
 109                 .maxlen         = sizeof(unsigned int),
 110                 .mode           = 0644,
 111                 .proc_handler   = proc_dointvec,
 112         },
 113         {
 114                 .procname       = "rdma_inline_write_padding",
 115                 .data           = &xprt_rdma_inline_write_padding,
 116                 .maxlen         = sizeof(unsigned int),
 117                 .mode           = 0644,
 118                 .proc_handler   = proc_dointvec_minmax,
 119                 .extra1         = &zero,
 120                 .extra2         = &max_padding,
 121         },
 122         {
 123                 .procname       = "rdma_memreg_strategy",
 124                 .data           = &xprt_rdma_memreg_strategy,
 125                 .maxlen         = sizeof(unsigned int),
 126                 .mode           = 0644,
 127                 .proc_handler   = proc_dointvec_minmax,
 128                 .extra1         = &min_memreg,
 129                 .extra2         = &max_memreg,
 130         },
 131         {
 132                 .procname       = "rdma_pad_optimize",
 133                 .data           = &xprt_rdma_pad_optimize,
 134                 .maxlen         = sizeof(unsigned int),
 135                 .mode           = 0644,
 136                 .proc_handler   = proc_dointvec,
 137         },
 138         { },
 139 };
 140
 141 static struct ctl_table sunrpc_table[] = {
 142         {
 143                 .procname       = "sunrpc",
 144                 .mode           = 0555,
 145                 .child          = xr_tunables_table
 146         },
 147         { },
 148 };
 149
 150 #endif
 151
 152 #define RPCRDMA_BIND_TO         (60U * HZ)
 153 #define RPCRDMA_INIT_REEST_TO   (5U * HZ)
 154 #define RPCRDMA_MAX_REEST_TO    (30U * HZ)
 155 #define RPCRDMA_IDLE_DISC_TO    (5U * 60 * HZ)
 156
 157 static struct rpc_xprt_ops xprt_rdma_procs;     /* forward reference */
 158
 159 static void
 160 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
 161 {
 162         struct sockaddr *sap = (struct sockaddr *)
 163                                         &rpcx_to_rdmad(xprt).addr;
 164         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
 165         char buf[64];
 166
 167         (void)rpc_ntop(sap, buf, sizeof(buf));
 168         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
 169
 170         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
 171         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 172
 173         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
 174
 175         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
 176         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 177
 178         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
 179         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 180
 181         /* netid */
 182         xprt->address_strings[RPC_DISPLAY_NETID] = "rdma";
 183 }
 184
 185 static void
 186 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
 187 {
 188         unsigned int i;
 189
 190         for (i = 0; i < RPC_DISPLAY_MAX; i++)
 191                 switch (i) {
 192                 case RPC_DISPLAY_PROTO:
 193                 case RPC_DISPLAY_NETID:
 194                         continue;
 195                 default:
 196                         kfree(xprt->address_strings[i]);
 197                 }
 198 }
 199
 200 static void
 201 xprt_rdma_connect_worker(struct work_struct *work)
 202 {
 203         struct rpcrdma_xprt *r_xprt =
 204                 container_of(work, struct rpcrdma_xprt, rdma_connect.work);
 205         struct rpc_xprt *xprt = &r_xprt->xprt;
 206         int rc = 0;
 207
 208         current->flags |= PF_FSTRANS;
 209         xprt_clear_connected(xprt);
 210
 211         dprintk("RPC:       %s: %sconnect\n", __func__,
 212                         r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
 213         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 214         if (rc)
 215                 xprt_wake_pending_tasks(xprt, rc);
 216
 217         dprintk("RPC:       %s: exit\n", __func__);
 218         xprt_clear_connecting(xprt);
 219         current->flags &= ~PF_FSTRANS;
 220 }
 221
 222 /*
 223  * xprt_rdma_destroy
 224  *
 225  * Destroy the xprt.
 226  * Free all memory associated with the object, including its own.
 227  * NOTE: none of the *destroy methods free memory for their top-level
 228  * objects, even though they may have allocated it (they do free
 229  * private memory). It's up to the caller to handle it. In this
 230  * case (RDMA transport), all structure memory is inlined with the
 231  * struct rpcrdma_xprt.
 232  */
 233 static void
 234 xprt_rdma_destroy(struct rpc_xprt *xprt)
 235 {
 236         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 237
 238         dprintk("RPC:       %s: called\n", __func__);
 239
 240         cancel_delayed_work_sync(&r_xprt->rdma_connect);
 241
 242         xprt_clear_connected(xprt);
 243
 244         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 245         rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
 246         rpcrdma_ia_close(&r_xprt->rx_ia);
 247
 248         xprt_rdma_free_addresses(xprt);
 249
 250         xprt_free(xprt);
 251
 252         dprintk("RPC:       %s: returning\n", __func__);
 253
 254         module_put(THIS_MODULE);
 255 }
 256
 257 static const struct rpc_timeout xprt_rdma_default_timeout = {
 258         .to_initval = 60 * HZ,
 259         .to_maxval = 60 * HZ,
 260 };
 261
 262 /**
 263  * xprt_setup_rdma - Set up transport to use RDMA
 264  *
 265  * @args: rpc transport arguments
 266  */
 267 static struct rpc_xprt *
 268 xprt_setup_rdma(struct xprt_create *args)
 269 {
 270         struct rpcrdma_create_data_internal cdata;
 271         struct rpc_xprt *xprt;
 272         struct rpcrdma_xprt *new_xprt;
 273         struct rpcrdma_ep *new_ep;
 274         struct sockaddr_in *sin;
 275         int rc;
 276
 277         if (args->addrlen > sizeof(xprt->addr)) {
 278                 dprintk("RPC:       %s: address too large\n", __func__);
 279                 return ERR_PTR(-EBADF);
 280         }
 281
 282         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
 283                         xprt_rdma_slot_table_entries,
 284                         xprt_rdma_slot_table_entries);
 285         if (xprt == NULL) {
 286                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
 287                         __func__);
 288                 return ERR_PTR(-ENOMEM);
 289         }
 290
 291         /* 60 second timeout, no retries */
 292         xprt->timeout = &xprt_rdma_default_timeout;
 293         xprt->bind_timeout = RPCRDMA_BIND_TO;
 294         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 295         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
 296
 297         xprt->resvport = 0;             /* privileged port not needed */
 298         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
 299         xprt->ops = &xprt_rdma_procs;
 300
 301         /*
 302          * Set up RDMA-specific connect data.
 303          */
 304
 305         /* Put server RDMA address in local cdata */
 306         memcpy(&cdata.addr, args->dstaddr, args->addrlen);
 307
 308         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 309          * address, for any side protocols which peek at it */
 310         xprt->prot = IPPROTO_TCP;
 311         xprt->addrlen = args->addrlen;
 312         memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
 313
 314         sin = (struct sockaddr_in *)&cdata.addr;
 315         if (ntohs(sin->sin_port) != 0)
 316                 xprt_set_bound(xprt);
 317
 318         dprintk("RPC:       %s: %pI4:%u\n",
 319                 __func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
 320
 321         /* Set max requests */
 322         cdata.max_requests = xprt->max_reqs;
 323
 324         /* Set some length limits */
 325         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
 326         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
 327
 328         cdata.inline_wsize = xprt_rdma_max_inline_write;
 329         if (cdata.inline_wsize > cdata.wsize)
 330                 cdata.inline_wsize = cdata.wsize;
 331
 332         cdata.inline_rsize = xprt_rdma_max_inline_read;
 333         if (cdata.inline_rsize > cdata.rsize)
 334                 cdata.inline_rsize = cdata.rsize;
 335
 336         cdata.padding = xprt_rdma_inline_write_padding;
 337
 338         /*
 339          * Create new transport instance, which includes initialized
 340          *  o ia
 341          *  o endpoint
 342          *  o buffers
 343          */
 344
 345         new_xprt = rpcx_to_rdmax(xprt);
 346
 347         rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
 348                                 xprt_rdma_memreg_strategy);
 349         if (rc)
 350                 goto out1;
 351
 352         /*
 353          * initialize and create ep
 354          */
 355         new_xprt->rx_data = cdata;
 356         new_ep = &new_xprt->rx_ep;
 357         new_ep->rep_remote_addr = cdata.addr;
 358
 359         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
 360                                 &new_xprt->rx_ia, &new_xprt->rx_data);
 361         if (rc)
 362                 goto out2;
 363
 364         /*
 365          * Allocate pre-registered send and receive buffers for headers and
 366          * any inline data. Also specify any padding which will be provided
 367          * from a preregistered zero buffer.
 368          */
 369         rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia,
 370                                 &new_xprt->rx_data);
 371         if (rc)
 372                 goto out3;
 373
 374         /*
 375          * Register a callback for connection events. This is necessary because
 376          * connection loss notification is async. We also catch connection loss
 377          * when reaping receives.
 378          */
 379         INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker);
 380         new_ep->rep_func = rpcrdma_conn_func;
 381         new_ep->rep_xprt = xprt;
 382
 383         xprt_rdma_format_addresses(xprt);
 384         xprt->max_payload = rpcrdma_max_payload(new_xprt);
 385         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
 386                 __func__, xprt->max_payload);
 387
 388         if (!try_module_get(THIS_MODULE))
 389                 goto out4;
 390
 391         return xprt;
 392
 393 out4:
 394         xprt_rdma_free_addresses(xprt);
 395         rc = -EINVAL;
 396 out3:
 397         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
 398 out2:
 399         rpcrdma_ia_close(&new_xprt->rx_ia);
 400 out1:
 401         xprt_free(xprt);
 402         return ERR_PTR(rc);
 403 }
 404
 405 /*
 406  * Close a connection, during shutdown or timeout/reconnect
 407  */
 408 static void
 409 xprt_rdma_close(struct rpc_xprt *xprt)
 410 {
 411         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 412
 413         dprintk("RPC:       %s: closing\n", __func__);
 414         if (r_xprt->rx_ep.rep_connected > 0)
 415                 xprt->reestablish_timeout = 0;
 416         xprt_disconnect_done(xprt);
 417         rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 418 }
 419
 420 static void
 421 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 422 {
 423         struct sockaddr_in *sap;
 424
 425         sap = (struct sockaddr_in *)&xprt->addr;
 426         sap->sin_port = htons(port);
 427         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
 428         sap->sin_port = htons(port);
 429         dprintk("RPC:       %s: %u\n", __func__, port);
 430 }
 431
 432 static void
 433 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
 434 {
 435         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 436
 437         if (r_xprt->rx_ep.rep_connected != 0) {
 438                 /* Reconnect */
 439                 schedule_delayed_work(&r_xprt->rdma_connect,
 440                         xprt->reestablish_timeout);
 441                 xprt->reestablish_timeout <<= 1;
 442                 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
 443                         xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
 444                 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
 445                         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 446         } else {
 447                 schedule_delayed_work(&r_xprt->rdma_connect, 0);
 448                 if (!RPC_IS_ASYNC(task))
 449                         flush_delayed_work(&r_xprt->rdma_connect);
 450         }
 451 }
 452
 453 /*
 454  * The RDMA allocate/free functions need the task structure as a place
 455  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
 456  * sequence. For this reason, the recv buffers are attached to send
 457  * buffers for portions of the RPC. Note that the RPC layer allocates
 458  * both send and receive buffers in the same call. We may register
 459  * the receive buffer portion when using reply chunks.
 460  */
 461 static void *
 462 xprt_rdma_allocate(struct rpc_task *task, size_t size)
 463 {
 464         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 465         struct rpcrdma_req *req, *nreq;
 466
 467         req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf);
 468         if (req == NULL)
 469                 return NULL;
 470
 471         if (size > req->rl_size) {
 472                 dprintk("RPC:       %s: size %zd too large for buffer[%zd]: "
 473                         "prog %d vers %d proc %d\n",
 474                         __func__, size, req->rl_size,
 475                         task->tk_client->cl_prog, task->tk_client->cl_vers,
 476                         task->tk_msg.rpc_proc->p_proc);
 477                 /*
 478                  * Outgoing length shortage. Our inline write max must have
 479                  * been configured to perform direct i/o.
 480                  *
 481                  * This is therefore a large metadata operation, and the
 482                  * allocate call was made on the maximum possible message,
 483                  * e.g. containing long filename(s) or symlink data. In
 484                  * fact, while these metadata operations *might* carry
 485                  * large outgoing payloads, they rarely *do*. However, we
 486                  * have to commit to the request here, so reallocate and
 487                  * register it now. The data path will never require this
 488                  * reallocation.
 489                  *
 490                  * If the allocation or registration fails, the RPC framework
 491                  * will (doggedly) retry.
 492                  */
 493                 if (task->tk_flags & RPC_TASK_SWAPPER)
 494                         nreq = kmalloc(sizeof *req + size, GFP_ATOMIC);
 495                 else
 496                         nreq = kmalloc(sizeof *req + size, GFP_NOFS);
 497                 if (nreq == NULL)
 498                         goto outfail;
 499
 500                 if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia,
 501                                 nreq->rl_base, size + sizeof(struct rpcrdma_req)
 502                                 - offsetof(struct rpcrdma_req, rl_base),
 503                                 &nreq->rl_handle, &nreq->rl_iov)) {
 504                         kfree(nreq);
 505                         goto outfail;
 506                 }
 507                 rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size;
 508                 nreq->rl_size = size;
 509                 nreq->rl_niovs = 0;
 510                 nreq->rl_nchunks = 0;
 511                 nreq->rl_buffer = (struct rpcrdma_buffer *)req;
 512                 nreq->rl_reply = req->rl_reply;
 513                 memcpy(nreq->rl_segments,
 514                         req->rl_segments, sizeof nreq->rl_segments);
 515                 /* flag the swap with an unused field */
 516                 nreq->rl_iov.length = 0;
 517                 req->rl_reply = NULL;
 518                 req = nreq;
 519         }
 520         dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
 521         req->rl_connect_cookie = 0;     /* our reserved value */
 522         return req->rl_xdr_buf;
 523
 524 outfail:
 525         rpcrdma_buffer_put(req);
 526         rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++;
 527         return NULL;
 528 }
 529
 530 /*
 531  * This function returns all RDMA resources to the pool.
 532  */
 533 static void
 534 xprt_rdma_free(void *buffer)
 535 {
 536         struct rpcrdma_req *req;
 537         struct rpcrdma_xprt *r_xprt;
 538         struct rpcrdma_rep *rep;
 539         int i;
 540
 541         if (buffer == NULL)
 542                 return;
 543
 544         req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]);
 545         if (req->rl_iov.length == 0) {  /* see allocate above */
 546                 r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer,
 547                                       struct rpcrdma_xprt, rx_buf);
 548         } else
 549                 r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
 550         rep = req->rl_reply;
 551
 552         dprintk("RPC:       %s: called on 0x%p%s\n",
 553                 __func__, rep, (rep && rep->rr_func) ? " (with waiter)" : "");
 554
 555         /*
 556          * Finish the deregistration.  The process is considered
 557          * complete when the rr_func vector becomes NULL - this
 558          * was put in place during rpcrdma_reply_handler() - the wait
 559          * call below will not block if the dereg is "done". If
 560          * interrupted, our framework will clean up.
 561          */
 562         for (i = 0; req->rl_nchunks;) {
 563                 --req->rl_nchunks;
 564                 i += rpcrdma_deregister_external(
 565                         &req->rl_segments[i], r_xprt);
 566         }
 567
 568         if (req->rl_iov.length == 0) {  /* see allocate above */
 569                 struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer;
 570                 oreq->rl_reply = req->rl_reply;
 571                 (void) rpcrdma_deregister_internal(&r_xprt->rx_ia,
 572                                                    req->rl_handle,
 573                                                    &req->rl_iov);
 574                 kfree(req);
 575                 req = oreq;
 576         }
 577
 578         /* Put back request+reply buffers */
 579         rpcrdma_buffer_put(req);
 580 }
 581
 582 /*
 583  * send_request invokes the meat of RPC RDMA. It must do the following:
 584  *  1.  Marshal the RPC request into an RPC RDMA request, which means
 585  *      putting a header in front of data, and creating IOVs for RDMA
 586  *      from those in the request.
 587  *  2.  In marshaling, detect opportunities for RDMA, and use them.
 588  *  3.  Post a recv message to set up asynch completion, then send
 589  *      the request (rpcrdma_ep_post).
 590  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
 591  */
 592
 593 static int
 594 xprt_rdma_send_request(struct rpc_task *task)
 595 {
 596         struct rpc_rqst *rqst = task->tk_rqstp;
 597         struct rpc_xprt *xprt = rqst->rq_xprt;
 598         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 599         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 600         int rc = 0;
 601
 602         if (req->rl_niovs == 0)
 603                 rc = rpcrdma_marshal_req(rqst);
 604         else if (r_xprt->rx_ia.ri_memreg_strategy == RPCRDMA_FRMR)
 605                 rc = rpcrdma_marshal_chunks(rqst, 0);
 606         if (rc < 0)
 607                 goto failed_marshal;
 608
 609         if (req->rl_reply == NULL)              /* e.g. reconnection */
 610                 rpcrdma_recv_buffer_get(req);
 611
 612         if (req->rl_reply) {
 613                 req->rl_reply->rr_func = rpcrdma_reply_handler;
 614                 /* this need only be done once, but... */
 615                 req->rl_reply->rr_xprt = xprt;
 616         }
 617
 618         /* Must suppress retransmit to maintain credits */
 619         if (req->rl_connect_cookie == xprt->connect_cookie)
 620                 goto drop_connection;
 621         req->rl_connect_cookie = xprt->connect_cookie;
 622
 623         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 624                 goto drop_connection;
 625
 626         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
 627         rqst->rq_bytes_sent = 0;
 628         return 0;
 629
 630 failed_marshal:
 631         r_xprt->rx_stats.failed_marshal_count++;
 632         dprintk("RPC:       %s: rpcrdma_marshal_req failed, status %i\n",
 633                 __func__, rc);
 634         if (rc == -EIO)
 635                 return -EIO;
 636 drop_connection:
 637         xprt_disconnect_done(xprt);
 638         return -ENOTCONN;       /* implies disconnect */
 639 }
 640
 641 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 642 {
 643         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 644         long idle_time = 0;
 645
 646         if (xprt_connected(xprt))
 647                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 648
 649         seq_printf(seq,
 650           "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
 651           "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
 652
 653            0,   /* need a local port? */
 654            xprt->stat.bind_count,
 655            xprt->stat.connect_count,
 656            xprt->stat.connect_time,
 657            idle_time,
 658            xprt->stat.sends,
 659            xprt->stat.recvs,
 660            xprt->stat.bad_xids,
 661            xprt->stat.req_u,
 662            xprt->stat.bklog_u,
 663
 664            r_xprt->rx_stats.read_chunk_count,
 665            r_xprt->rx_stats.write_chunk_count,
 666            r_xprt->rx_stats.reply_chunk_count,
 667            r_xprt->rx_stats.total_rdma_request,
 668            r_xprt->rx_stats.total_rdma_reply,
 669            r_xprt->rx_stats.pullup_copy_count,
 670            r_xprt->rx_stats.fixup_copy_count,
 671            r_xprt->rx_stats.hardway_register_count,
 672            r_xprt->rx_stats.failed_marshal_count,
 673            r_xprt->rx_stats.bad_reply_count);
 674 }
 675
 676 /*
 677  * Plumbing for rpc transport switch and kernel module
 678  */
 679
 680 static struct rpc_xprt_ops xprt_rdma_procs = {
 681         .reserve_xprt           = xprt_reserve_xprt_cong,
 682         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 683         .alloc_slot             = xprt_alloc_slot,
 684         .release_request        = xprt_release_rqst_cong,       /* ditto */
 685         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
 686         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 687         .set_port               = xprt_rdma_set_port,
 688         .connect                = xprt_rdma_connect,
 689         .buf_alloc              = xprt_rdma_allocate,
 690         .buf_free               = xprt_rdma_free,
 691         .send_request           = xprt_rdma_send_request,
 692         .close                  = xprt_rdma_close,
 693         .destroy                = xprt_rdma_destroy,
 694         .print_stats            = xprt_rdma_print_stats
 695 };
 696
 697 static struct xprt_class xprt_rdma = {
 698         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 699         .name                   = "rdma",
 700         .owner                  = THIS_MODULE,
 701         .ident                  = XPRT_TRANSPORT_RDMA,
 702         .setup                  = xprt_setup_rdma,
 703 };
 704
 705 static void __exit xprt_rdma_cleanup(void)
 706 {
 707         int rc;
 708
 709         dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
 710 #ifdef RPC_DEBUG
 711         if (sunrpc_table_header) {
 712                 unregister_sysctl_table(sunrpc_table_header);
 713                 sunrpc_table_header = NULL;
 714         }
 715 #endif
 716         rc = xprt_unregister_transport(&xprt_rdma);
 717         if (rc)
 718                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
 719                         __func__, rc);
 720 }
 721
 722 static int __init xprt_rdma_init(void)
 723 {
 724         int rc;
 725
 726         rc = xprt_register_transport(&xprt_rdma);
 727
 728         if (rc)
 729                 return rc;
 730
 731         dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
 732
 733         dprintk("Defaults:\n");
 734         dprintk("\tSlots %d\n"
 735                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
 736                 xprt_rdma_slot_table_entries,
 737                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
 738         dprintk("\tPadding %d\n\tMemreg %d\n",
 739                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
 740
 741 #ifdef RPC_DEBUG
 742         if (!sunrpc_table_header)
 743                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 744 #endif
 745         return 0;
 746 }
 747
 748 module_init(xprt_rdma_init);
 749 module_exit(xprt_rdma_cleanup);