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 #if IS_ENABLED(CONFIG_SUNRPC_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 = 1;
  77
  78 #if IS_ENABLED(CONFIG_SUNRPC_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_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
 161 {
 162         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
 163         char buf[20];
 164
 165         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
 166         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 167
 168         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
 169 }
 170
 171 static void
 172 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
 173 {
 174         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
 175         char buf[40];
 176
 177         snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
 178         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 179
 180         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
 181 }
 182
 183 static void
 184 xprt_rdma_format_addresses(struct rpc_xprt *xprt)
 185 {
 186         struct sockaddr *sap = (struct sockaddr *)
 187                                         &rpcx_to_rdmad(xprt).addr;
 188         char buf[128];
 189
 190         switch (sap->sa_family) {
 191         case AF_INET:
 192                 xprt_rdma_format_addresses4(xprt, sap);
 193                 break;
 194         case AF_INET6:
 195                 xprt_rdma_format_addresses6(xprt, sap);
 196                 break;
 197         default:
 198                 pr_err("rpcrdma: Unrecognized address family\n");
 199                 return;
 200         }
 201
 202         (void)rpc_ntop(sap, buf, sizeof(buf));
 203         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
 204
 205         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
 206         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 207
 208         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
 209         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 210
 211         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
 212 }
 213
 214 static void
 215 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
 216 {
 217         unsigned int i;
 218
 219         for (i = 0; i < RPC_DISPLAY_MAX; i++)
 220                 switch (i) {
 221                 case RPC_DISPLAY_PROTO:
 222                 case RPC_DISPLAY_NETID:
 223                         continue;
 224                 default:
 225                         kfree(xprt->address_strings[i]);
 226                 }
 227 }
 228
 229 static void
 230 xprt_rdma_connect_worker(struct work_struct *work)
 231 {
 232         struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
 233                                                    rx_connect_worker.work);
 234         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 235         int rc = 0;
 236
 237         xprt_clear_connected(xprt);
 238
 239         dprintk("RPC:       %s: %sconnect\n", __func__,
 240                         r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
 241         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 242         if (rc)
 243                 xprt_wake_pending_tasks(xprt, rc);
 244
 245         dprintk("RPC:       %s: exit\n", __func__);
 246         xprt_clear_connecting(xprt);
 247 }
 248
 249 static void
 250 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
 251 {
 252         struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
 253                                                    rx_xprt);
 254
 255         pr_info("rpcrdma: injecting transport disconnect on xprt=%p\n", xprt);
 256         rdma_disconnect(r_xprt->rx_ia.ri_id);
 257 }
 258
 259 /*
 260  * xprt_rdma_destroy
 261  *
 262  * Destroy the xprt.
 263  * Free all memory associated with the object, including its own.
 264  * NOTE: none of the *destroy methods free memory for their top-level
 265  * objects, even though they may have allocated it (they do free
 266  * private memory). It's up to the caller to handle it. In this
 267  * case (RDMA transport), all structure memory is inlined with the
 268  * struct rpcrdma_xprt.
 269  */
 270 static void
 271 xprt_rdma_destroy(struct rpc_xprt *xprt)
 272 {
 273         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 274
 275         dprintk("RPC:       %s: called\n", __func__);
 276
 277         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
 278
 279         xprt_clear_connected(xprt);
 280
 281         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 282         rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
 283         rpcrdma_ia_close(&r_xprt->rx_ia);
 284
 285         xprt_rdma_free_addresses(xprt);
 286
 287         xprt_free(xprt);
 288
 289         dprintk("RPC:       %s: returning\n", __func__);
 290
 291         module_put(THIS_MODULE);
 292 }
 293
 294 static const struct rpc_timeout xprt_rdma_default_timeout = {
 295         .to_initval = 60 * HZ,
 296         .to_maxval = 60 * HZ,
 297 };
 298
 299 /**
 300  * xprt_setup_rdma - Set up transport to use RDMA
 301  *
 302  * @args: rpc transport arguments
 303  */
 304 static struct rpc_xprt *
 305 xprt_setup_rdma(struct xprt_create *args)
 306 {
 307         struct rpcrdma_create_data_internal cdata;
 308         struct rpc_xprt *xprt;
 309         struct rpcrdma_xprt *new_xprt;
 310         struct rpcrdma_ep *new_ep;
 311         struct sockaddr_in *sin;
 312         int rc;
 313
 314         if (args->addrlen > sizeof(xprt->addr)) {
 315                 dprintk("RPC:       %s: address too large\n", __func__);
 316                 return ERR_PTR(-EBADF);
 317         }
 318
 319         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
 320                         xprt_rdma_slot_table_entries,
 321                         xprt_rdma_slot_table_entries);
 322         if (xprt == NULL) {
 323                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
 324                         __func__);
 325                 return ERR_PTR(-ENOMEM);
 326         }
 327
 328         /* 60 second timeout, no retries */
 329         xprt->timeout = &xprt_rdma_default_timeout;
 330         xprt->bind_timeout = RPCRDMA_BIND_TO;
 331         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 332         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
 333
 334         xprt->resvport = 0;             /* privileged port not needed */
 335         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
 336         xprt->ops = &xprt_rdma_procs;
 337
 338         /*
 339          * Set up RDMA-specific connect data.
 340          */
 341
 342         /* Put server RDMA address in local cdata */
 343         memcpy(&cdata.addr, args->dstaddr, args->addrlen);
 344
 345         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 346          * address, for any side protocols which peek at it */
 347         xprt->prot = IPPROTO_TCP;
 348         xprt->addrlen = args->addrlen;
 349         memcpy(&xprt->addr, &cdata.addr, xprt->addrlen);
 350
 351         sin = (struct sockaddr_in *)&cdata.addr;
 352         if (ntohs(sin->sin_port) != 0)
 353                 xprt_set_bound(xprt);
 354
 355         dprintk("RPC:       %s: %pI4:%u\n",
 356                 __func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port));
 357
 358         /* Set max requests */
 359         cdata.max_requests = xprt->max_reqs;
 360
 361         /* Set some length limits */
 362         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
 363         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
 364
 365         cdata.inline_wsize = xprt_rdma_max_inline_write;
 366         if (cdata.inline_wsize > cdata.wsize)
 367                 cdata.inline_wsize = cdata.wsize;
 368
 369         cdata.inline_rsize = xprt_rdma_max_inline_read;
 370         if (cdata.inline_rsize > cdata.rsize)
 371                 cdata.inline_rsize = cdata.rsize;
 372
 373         cdata.padding = xprt_rdma_inline_write_padding;
 374
 375         /*
 376          * Create new transport instance, which includes initialized
 377          *  o ia
 378          *  o endpoint
 379          *  o buffers
 380          */
 381
 382         new_xprt = rpcx_to_rdmax(xprt);
 383
 384         rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr,
 385                                 xprt_rdma_memreg_strategy);
 386         if (rc)
 387                 goto out1;
 388
 389         /*
 390          * initialize and create ep
 391          */
 392         new_xprt->rx_data = cdata;
 393         new_ep = &new_xprt->rx_ep;
 394         new_ep->rep_remote_addr = cdata.addr;
 395
 396         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
 397                                 &new_xprt->rx_ia, &new_xprt->rx_data);
 398         if (rc)
 399                 goto out2;
 400
 401         /*
 402          * Allocate pre-registered send and receive buffers for headers and
 403          * any inline data. Also specify any padding which will be provided
 404          * from a preregistered zero buffer.
 405          */
 406         rc = rpcrdma_buffer_create(new_xprt);
 407         if (rc)
 408                 goto out3;
 409
 410         /*
 411          * Register a callback for connection events. This is necessary because
 412          * connection loss notification is async. We also catch connection loss
 413          * when reaping receives.
 414          */
 415         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
 416                           xprt_rdma_connect_worker);
 417
 418         xprt_rdma_format_addresses(xprt);
 419         xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
 420         if (xprt->max_payload == 0)
 421                 goto out4;
 422         xprt->max_payload <<= PAGE_SHIFT;
 423         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
 424                 __func__, xprt->max_payload);
 425
 426         if (!try_module_get(THIS_MODULE))
 427                 goto out4;
 428
 429         return xprt;
 430
 431 out4:
 432         xprt_rdma_free_addresses(xprt);
 433         rc = -EINVAL;
 434 out3:
 435         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
 436 out2:
 437         rpcrdma_ia_close(&new_xprt->rx_ia);
 438 out1:
 439         xprt_free(xprt);
 440         return ERR_PTR(rc);
 441 }
 442
 443 /*
 444  * Close a connection, during shutdown or timeout/reconnect
 445  */
 446 static void
 447 xprt_rdma_close(struct rpc_xprt *xprt)
 448 {
 449         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 450
 451         dprintk("RPC:       %s: closing\n", __func__);
 452         if (r_xprt->rx_ep.rep_connected > 0)
 453                 xprt->reestablish_timeout = 0;
 454         xprt_disconnect_done(xprt);
 455         rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 456 }
 457
 458 static void
 459 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 460 {
 461         struct sockaddr_in *sap;
 462
 463         sap = (struct sockaddr_in *)&xprt->addr;
 464         sap->sin_port = htons(port);
 465         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
 466         sap->sin_port = htons(port);
 467         dprintk("RPC:       %s: %u\n", __func__, port);
 468 }
 469
 470 static void
 471 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
 472 {
 473         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 474
 475         if (r_xprt->rx_ep.rep_connected != 0) {
 476                 /* Reconnect */
 477                 schedule_delayed_work(&r_xprt->rx_connect_worker,
 478                                       xprt->reestablish_timeout);
 479                 xprt->reestablish_timeout <<= 1;
 480                 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
 481                         xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
 482                 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
 483                         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 484         } else {
 485                 schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
 486                 if (!RPC_IS_ASYNC(task))
 487                         flush_delayed_work(&r_xprt->rx_connect_worker);
 488         }
 489 }
 490
 491 /*
 492  * The RDMA allocate/free functions need the task structure as a place
 493  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
 494  * sequence.
 495  *
 496  * The RPC layer allocates both send and receive buffers in the same call
 497  * (rq_send_buf and rq_rcv_buf are both part of a single contiguous buffer).
 498  * We may register rq_rcv_buf when using reply chunks.
 499  */
 500 static void *
 501 xprt_rdma_allocate(struct rpc_task *task, size_t size)
 502 {
 503         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 504         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 505         struct rpcrdma_regbuf *rb;
 506         struct rpcrdma_req *req;
 507         size_t min_size;
 508         gfp_t flags;
 509
 510         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
 511         if (req == NULL)
 512                 return NULL;
 513
 514         flags = GFP_NOIO | __GFP_NOWARN;
 515         if (RPC_IS_SWAPPER(task))
 516                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
 517
 518         if (req->rl_rdmabuf == NULL)
 519                 goto out_rdmabuf;
 520         if (req->rl_sendbuf == NULL)
 521                 goto out_sendbuf;
 522         if (size > req->rl_sendbuf->rg_size)
 523                 goto out_sendbuf;
 524
 525 out:
 526         dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
 527         req->rl_connect_cookie = 0;     /* our reserved value */
 528         return req->rl_sendbuf->rg_base;
 529
 530 out_rdmabuf:
 531         min_size = RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
 532         rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, min_size, flags);
 533         if (IS_ERR(rb))
 534                 goto out_fail;
 535         req->rl_rdmabuf = rb;
 536
 537 out_sendbuf:
 538         /* XDR encoding and RPC/RDMA marshaling of this request has not
 539          * yet occurred. Thus a lower bound is needed to prevent buffer
 540          * overrun during marshaling.
 541          *
 542          * RPC/RDMA marshaling may choose to send payload bearing ops
 543          * inline, if the result is smaller than the inline threshold.
 544          * The value of the "size" argument accounts for header
 545          * requirements but not for the payload in these cases.
 546          *
 547          * Likewise, allocate enough space to receive a reply up to the
 548          * size of the inline threshold.
 549          *
 550          * It's unlikely that both the send header and the received
 551          * reply will be large, but slush is provided here to allow
 552          * flexibility when marshaling.
 553          */
 554         min_size = RPCRDMA_INLINE_READ_THRESHOLD(task->tk_rqstp);
 555         min_size += RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
 556         if (size < min_size)
 557                 size = min_size;
 558
 559         rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, size, flags);
 560         if (IS_ERR(rb))
 561                 goto out_fail;
 562         rb->rg_owner = req;
 563
 564         r_xprt->rx_stats.hardway_register_count += size;
 565         rpcrdma_free_regbuf(&r_xprt->rx_ia, req->rl_sendbuf);
 566         req->rl_sendbuf = rb;
 567         goto out;
 568
 569 out_fail:
 570         rpcrdma_buffer_put(req);
 571         r_xprt->rx_stats.failed_marshal_count++;
 572         return NULL;
 573 }
 574
 575 /*
 576  * This function returns all RDMA resources to the pool.
 577  */
 578 static void
 579 xprt_rdma_free(void *buffer)
 580 {
 581         struct rpcrdma_req *req;
 582         struct rpcrdma_xprt *r_xprt;
 583         struct rpcrdma_regbuf *rb;
 584         int i;
 585
 586         if (buffer == NULL)
 587                 return;
 588
 589         rb = container_of(buffer, struct rpcrdma_regbuf, rg_base[0]);
 590         req = rb->rg_owner;
 591         r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
 592
 593         dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
 594
 595         for (i = 0; req->rl_nchunks;) {
 596                 --req->rl_nchunks;
 597                 i += r_xprt->rx_ia.ri_ops->ro_unmap(r_xprt,
 598                                                     &req->rl_segments[i]);
 599         }
 600
 601         rpcrdma_buffer_put(req);
 602 }
 603
 604 /*
 605  * send_request invokes the meat of RPC RDMA. It must do the following:
 606  *  1.  Marshal the RPC request into an RPC RDMA request, which means
 607  *      putting a header in front of data, and creating IOVs for RDMA
 608  *      from those in the request.
 609  *  2.  In marshaling, detect opportunities for RDMA, and use them.
 610  *  3.  Post a recv message to set up asynch completion, then send
 611  *      the request (rpcrdma_ep_post).
 612  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
 613  */
 614
 615 static int
 616 xprt_rdma_send_request(struct rpc_task *task)
 617 {
 618         struct rpc_rqst *rqst = task->tk_rqstp;
 619         struct rpc_xprt *xprt = rqst->rq_xprt;
 620         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 621         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 622         int rc = 0;
 623
 624         rc = rpcrdma_marshal_req(rqst);
 625         if (rc < 0)
 626                 goto failed_marshal;
 627
 628         if (req->rl_reply == NULL)              /* e.g. reconnection */
 629                 rpcrdma_recv_buffer_get(req);
 630
 631         /* Must suppress retransmit to maintain credits */
 632         if (req->rl_connect_cookie == xprt->connect_cookie)
 633                 goto drop_connection;
 634         req->rl_connect_cookie = xprt->connect_cookie;
 635
 636         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 637                 goto drop_connection;
 638
 639         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
 640         rqst->rq_bytes_sent = 0;
 641         return 0;
 642
 643 failed_marshal:
 644         r_xprt->rx_stats.failed_marshal_count++;
 645         dprintk("RPC:       %s: rpcrdma_marshal_req failed, status %i\n",
 646                 __func__, rc);
 647         if (rc == -EIO)
 648                 return -EIO;
 649 drop_connection:
 650         xprt_disconnect_done(xprt);
 651         return -ENOTCONN;       /* implies disconnect */
 652 }
 653
 654 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 655 {
 656         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 657         long idle_time = 0;
 658
 659         if (xprt_connected(xprt))
 660                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 661
 662         seq_printf(seq,
 663           "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu "
 664           "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n",
 665
 666            0,   /* need a local port? */
 667            xprt->stat.bind_count,
 668            xprt->stat.connect_count,
 669            xprt->stat.connect_time,
 670            idle_time,
 671            xprt->stat.sends,
 672            xprt->stat.recvs,
 673            xprt->stat.bad_xids,
 674            xprt->stat.req_u,
 675            xprt->stat.bklog_u,
 676
 677            r_xprt->rx_stats.read_chunk_count,
 678            r_xprt->rx_stats.write_chunk_count,
 679            r_xprt->rx_stats.reply_chunk_count,
 680            r_xprt->rx_stats.total_rdma_request,
 681            r_xprt->rx_stats.total_rdma_reply,
 682            r_xprt->rx_stats.pullup_copy_count,
 683            r_xprt->rx_stats.fixup_copy_count,
 684            r_xprt->rx_stats.hardway_register_count,
 685            r_xprt->rx_stats.failed_marshal_count,
 686            r_xprt->rx_stats.bad_reply_count);
 687 }
 688
 689 static int
 690 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
 691 {
 692         return -EINVAL;
 693 }
 694
 695 static void
 696 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
 697 {
 698 }
 699
 700 /*
 701  * Plumbing for rpc transport switch and kernel module
 702  */
 703
 704 static struct rpc_xprt_ops xprt_rdma_procs = {
 705         .reserve_xprt           = xprt_reserve_xprt_cong,
 706         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 707         .alloc_slot             = xprt_alloc_slot,
 708         .release_request        = xprt_release_rqst_cong,       /* ditto */
 709         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
 710         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 711         .set_port               = xprt_rdma_set_port,
 712         .connect                = xprt_rdma_connect,
 713         .buf_alloc              = xprt_rdma_allocate,
 714         .buf_free               = xprt_rdma_free,
 715         .send_request           = xprt_rdma_send_request,
 716         .close                  = xprt_rdma_close,
 717         .destroy                = xprt_rdma_destroy,
 718         .print_stats            = xprt_rdma_print_stats,
 719         .enable_swap            = xprt_rdma_enable_swap,
 720         .disable_swap           = xprt_rdma_disable_swap,
 721         .inject_disconnect      = xprt_rdma_inject_disconnect
 722 };
 723
 724 static struct xprt_class xprt_rdma = {
 725         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 726         .name                   = "rdma",
 727         .owner                  = THIS_MODULE,
 728         .ident                  = XPRT_TRANSPORT_RDMA,
 729         .setup                  = xprt_setup_rdma,
 730 };
 731
 732 static void __exit xprt_rdma_cleanup(void)
 733 {
 734         int rc;
 735
 736         dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
 737 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 738         if (sunrpc_table_header) {
 739                 unregister_sysctl_table(sunrpc_table_header);
 740                 sunrpc_table_header = NULL;
 741         }
 742 #endif
 743         rc = xprt_unregister_transport(&xprt_rdma);
 744         if (rc)
 745                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
 746                         __func__, rc);
 747
 748         frwr_destroy_recovery_wq();
 749 }
 750
 751 static int __init xprt_rdma_init(void)
 752 {
 753         int rc;
 754
 755         rc = frwr_alloc_recovery_wq();
 756         if (rc)
 757                 return rc;
 758
 759         rc = xprt_register_transport(&xprt_rdma);
 760         if (rc) {
 761                 frwr_destroy_recovery_wq();
 762                 return rc;
 763         }
 764
 765         dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
 766
 767         dprintk("Defaults:\n");
 768         dprintk("\tSlots %d\n"
 769                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
 770                 xprt_rdma_slot_table_entries,
 771                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
 772         dprintk("\tPadding %d\n\tMemreg %d\n",
 773                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
 774
 775 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 776         if (!sunrpc_table_header)
 777                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 778 #endif
 779         return 0;
 780 }
 781
 782 module_init(xprt_rdma_init);
 783 module_exit(xprt_rdma_cleanup);