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/slab.h>
  52 #include <linux/seq_file.h>
  53 #include <linux/sunrpc/addr.h>
  54
  55 #include "xprt_rdma.h"
  56
  57 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  58 # define RPCDBG_FACILITY        RPCDBG_TRANS
  59 #endif
  60
  61 /*
  62  * tunables
  63  */
  64
  65 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE;
  66 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE;
  67 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE;
  68 static unsigned int xprt_rdma_inline_write_padding;
  69 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR;
  70                 int xprt_rdma_pad_optimize = 1;
  71
  72 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  73
  74 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE;
  75 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE;
  76 static unsigned int zero;
  77 static unsigned int max_padding = PAGE_SIZE;
  78 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS;
  79 static unsigned int max_memreg = RPCRDMA_LAST - 1;
  80
  81 static struct ctl_table_header *sunrpc_table_header;
  82
  83 static struct ctl_table xr_tunables_table[] = {
  84         {
  85                 .procname       = "rdma_slot_table_entries",
  86                 .data           = &xprt_rdma_slot_table_entries,
  87                 .maxlen         = sizeof(unsigned int),
  88                 .mode           = 0644,
  89                 .proc_handler   = proc_dointvec_minmax,
  90                 .extra1         = &min_slot_table_size,
  91                 .extra2         = &max_slot_table_size
  92         },
  93         {
  94                 .procname       = "rdma_max_inline_read",
  95                 .data           = &xprt_rdma_max_inline_read,
  96                 .maxlen         = sizeof(unsigned int),
  97                 .mode           = 0644,
  98                 .proc_handler   = proc_dointvec,
  99         },
 100         {
 101                 .procname       = "rdma_max_inline_write",
 102                 .data           = &xprt_rdma_max_inline_write,
 103                 .maxlen         = sizeof(unsigned int),
 104                 .mode           = 0644,
 105                 .proc_handler   = proc_dointvec,
 106         },
 107         {
 108                 .procname       = "rdma_inline_write_padding",
 109                 .data           = &xprt_rdma_inline_write_padding,
 110                 .maxlen         = sizeof(unsigned int),
 111                 .mode           = 0644,
 112                 .proc_handler   = proc_dointvec_minmax,
 113                 .extra1         = &zero,
 114                 .extra2         = &max_padding,
 115         },
 116         {
 117                 .procname       = "rdma_memreg_strategy",
 118                 .data           = &xprt_rdma_memreg_strategy,
 119                 .maxlen         = sizeof(unsigned int),
 120                 .mode           = 0644,
 121                 .proc_handler   = proc_dointvec_minmax,
 122                 .extra1         = &min_memreg,
 123                 .extra2         = &max_memreg,
 124         },
 125         {
 126                 .procname       = "rdma_pad_optimize",
 127                 .data           = &xprt_rdma_pad_optimize,
 128                 .maxlen         = sizeof(unsigned int),
 129                 .mode           = 0644,
 130                 .proc_handler   = proc_dointvec,
 131         },
 132         { },
 133 };
 134
 135 static struct ctl_table sunrpc_table[] = {
 136         {
 137                 .procname       = "sunrpc",
 138                 .mode           = 0555,
 139                 .child          = xr_tunables_table
 140         },
 141         { },
 142 };
 143
 144 #endif
 145
 146 #define RPCRDMA_BIND_TO         (60U * HZ)
 147 #define RPCRDMA_INIT_REEST_TO   (5U * HZ)
 148 #define RPCRDMA_MAX_REEST_TO    (30U * HZ)
 149 #define RPCRDMA_IDLE_DISC_TO    (5U * 60 * HZ)
 150
 151 static struct rpc_xprt_ops xprt_rdma_procs;     /* forward reference */
 152
 153 static void
 154 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap)
 155 {
 156         struct sockaddr_in *sin = (struct sockaddr_in *)sap;
 157         char buf[20];
 158
 159         snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr));
 160         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 161
 162         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA;
 163 }
 164
 165 static void
 166 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap)
 167 {
 168         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
 169         char buf[40];
 170
 171         snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr);
 172         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL);
 173
 174         xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6;
 175 }
 176
 177 static void
 178 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap)
 179 {
 180         char buf[128];
 181
 182         switch (sap->sa_family) {
 183         case AF_INET:
 184                 xprt_rdma_format_addresses4(xprt, sap);
 185                 break;
 186         case AF_INET6:
 187                 xprt_rdma_format_addresses6(xprt, sap);
 188                 break;
 189         default:
 190                 pr_err("rpcrdma: Unrecognized address family\n");
 191                 return;
 192         }
 193
 194         (void)rpc_ntop(sap, buf, sizeof(buf));
 195         xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL);
 196
 197         snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap));
 198         xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL);
 199
 200         snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap));
 201         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL);
 202
 203         xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma";
 204 }
 205
 206 static void
 207 xprt_rdma_free_addresses(struct rpc_xprt *xprt)
 208 {
 209         unsigned int i;
 210
 211         for (i = 0; i < RPC_DISPLAY_MAX; i++)
 212                 switch (i) {
 213                 case RPC_DISPLAY_PROTO:
 214                 case RPC_DISPLAY_NETID:
 215                         continue;
 216                 default:
 217                         kfree(xprt->address_strings[i]);
 218                 }
 219 }
 220
 221 static void
 222 xprt_rdma_connect_worker(struct work_struct *work)
 223 {
 224         struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt,
 225                                                    rx_connect_worker.work);
 226         struct rpc_xprt *xprt = &r_xprt->rx_xprt;
 227         int rc = 0;
 228
 229         xprt_clear_connected(xprt);
 230
 231         dprintk("RPC:       %s: %sconnect\n", __func__,
 232                         r_xprt->rx_ep.rep_connected != 0 ? "re" : "");
 233         rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 234         if (rc)
 235                 xprt_wake_pending_tasks(xprt, rc);
 236
 237         dprintk("RPC:       %s: exit\n", __func__);
 238         xprt_clear_connecting(xprt);
 239 }
 240
 241 static void
 242 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt)
 243 {
 244         struct rpcrdma_xprt *r_xprt = container_of(xprt, struct rpcrdma_xprt,
 245                                                    rx_xprt);
 246
 247         pr_info("rpcrdma: injecting transport disconnect on xprt=%p\n", xprt);
 248         rdma_disconnect(r_xprt->rx_ia.ri_id);
 249 }
 250
 251 /*
 252  * xprt_rdma_destroy
 253  *
 254  * Destroy the xprt.
 255  * Free all memory associated with the object, including its own.
 256  * NOTE: none of the *destroy methods free memory for their top-level
 257  * objects, even though they may have allocated it (they do free
 258  * private memory). It's up to the caller to handle it. In this
 259  * case (RDMA transport), all structure memory is inlined with the
 260  * struct rpcrdma_xprt.
 261  */
 262 static void
 263 xprt_rdma_destroy(struct rpc_xprt *xprt)
 264 {
 265         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 266
 267         dprintk("RPC:       %s: called\n", __func__);
 268
 269         cancel_delayed_work_sync(&r_xprt->rx_connect_worker);
 270
 271         xprt_clear_connected(xprt);
 272
 273         rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia);
 274         rpcrdma_buffer_destroy(&r_xprt->rx_buf);
 275         rpcrdma_ia_close(&r_xprt->rx_ia);
 276
 277         xprt_rdma_free_addresses(xprt);
 278
 279         xprt_free(xprt);
 280
 281         dprintk("RPC:       %s: returning\n", __func__);
 282
 283         module_put(THIS_MODULE);
 284 }
 285
 286 static const struct rpc_timeout xprt_rdma_default_timeout = {
 287         .to_initval = 60 * HZ,
 288         .to_maxval = 60 * HZ,
 289 };
 290
 291 /**
 292  * xprt_setup_rdma - Set up transport to use RDMA
 293  *
 294  * @args: rpc transport arguments
 295  */
 296 static struct rpc_xprt *
 297 xprt_setup_rdma(struct xprt_create *args)
 298 {
 299         struct rpcrdma_create_data_internal cdata;
 300         struct rpc_xprt *xprt;
 301         struct rpcrdma_xprt *new_xprt;
 302         struct rpcrdma_ep *new_ep;
 303         struct sockaddr *sap;
 304         int rc;
 305
 306         if (args->addrlen > sizeof(xprt->addr)) {
 307                 dprintk("RPC:       %s: address too large\n", __func__);
 308                 return ERR_PTR(-EBADF);
 309         }
 310
 311         xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt),
 312                         xprt_rdma_slot_table_entries,
 313                         xprt_rdma_slot_table_entries);
 314         if (xprt == NULL) {
 315                 dprintk("RPC:       %s: couldn't allocate rpcrdma_xprt\n",
 316                         __func__);
 317                 return ERR_PTR(-ENOMEM);
 318         }
 319
 320         /* 60 second timeout, no retries */
 321         xprt->timeout = &xprt_rdma_default_timeout;
 322         xprt->bind_timeout = RPCRDMA_BIND_TO;
 323         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 324         xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO;
 325
 326         xprt->resvport = 0;             /* privileged port not needed */
 327         xprt->tsh_size = 0;             /* RPC-RDMA handles framing */
 328         xprt->ops = &xprt_rdma_procs;
 329
 330         /*
 331          * Set up RDMA-specific connect data.
 332          */
 333
 334         sap = (struct sockaddr *)&cdata.addr;
 335         memcpy(sap, args->dstaddr, args->addrlen);
 336
 337         /* Ensure xprt->addr holds valid server TCP (not RDMA)
 338          * address, for any side protocols which peek at it */
 339         xprt->prot = IPPROTO_TCP;
 340         xprt->addrlen = args->addrlen;
 341         memcpy(&xprt->addr, sap, xprt->addrlen);
 342
 343         if (rpc_get_port(sap))
 344                 xprt_set_bound(xprt);
 345
 346         cdata.max_requests = xprt->max_reqs;
 347
 348         cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */
 349         cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */
 350
 351         cdata.inline_wsize = xprt_rdma_max_inline_write;
 352         if (cdata.inline_wsize > cdata.wsize)
 353                 cdata.inline_wsize = cdata.wsize;
 354
 355         cdata.inline_rsize = xprt_rdma_max_inline_read;
 356         if (cdata.inline_rsize > cdata.rsize)
 357                 cdata.inline_rsize = cdata.rsize;
 358
 359         cdata.padding = xprt_rdma_inline_write_padding;
 360
 361         /*
 362          * Create new transport instance, which includes initialized
 363          *  o ia
 364          *  o endpoint
 365          *  o buffers
 366          */
 367
 368         new_xprt = rpcx_to_rdmax(xprt);
 369
 370         rc = rpcrdma_ia_open(new_xprt, sap, xprt_rdma_memreg_strategy);
 371         if (rc)
 372                 goto out1;
 373
 374         /*
 375          * initialize and create ep
 376          */
 377         new_xprt->rx_data = cdata;
 378         new_ep = &new_xprt->rx_ep;
 379         new_ep->rep_remote_addr = cdata.addr;
 380
 381         rc = rpcrdma_ep_create(&new_xprt->rx_ep,
 382                                 &new_xprt->rx_ia, &new_xprt->rx_data);
 383         if (rc)
 384                 goto out2;
 385
 386         /*
 387          * Allocate pre-registered send and receive buffers for headers and
 388          * any inline data. Also specify any padding which will be provided
 389          * from a preregistered zero buffer.
 390          */
 391         rc = rpcrdma_buffer_create(new_xprt);
 392         if (rc)
 393                 goto out3;
 394
 395         /*
 396          * Register a callback for connection events. This is necessary because
 397          * connection loss notification is async. We also catch connection loss
 398          * when reaping receives.
 399          */
 400         INIT_DELAYED_WORK(&new_xprt->rx_connect_worker,
 401                           xprt_rdma_connect_worker);
 402
 403         xprt_rdma_format_addresses(xprt, sap);
 404         xprt->max_payload = new_xprt->rx_ia.ri_ops->ro_maxpages(new_xprt);
 405         if (xprt->max_payload == 0)
 406                 goto out4;
 407         xprt->max_payload <<= PAGE_SHIFT;
 408         dprintk("RPC:       %s: transport data payload maximum: %zu bytes\n",
 409                 __func__, xprt->max_payload);
 410
 411         if (!try_module_get(THIS_MODULE))
 412                 goto out4;
 413
 414         dprintk("RPC:       %s: %s:%s\n", __func__,
 415                 xprt->address_strings[RPC_DISPLAY_ADDR],
 416                 xprt->address_strings[RPC_DISPLAY_PORT]);
 417         return xprt;
 418
 419 out4:
 420         xprt_rdma_free_addresses(xprt);
 421         rc = -EINVAL;
 422 out3:
 423         rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia);
 424 out2:
 425         rpcrdma_ia_close(&new_xprt->rx_ia);
 426 out1:
 427         xprt_free(xprt);
 428         return ERR_PTR(rc);
 429 }
 430
 431 /*
 432  * Close a connection, during shutdown or timeout/reconnect
 433  */
 434 static void
 435 xprt_rdma_close(struct rpc_xprt *xprt)
 436 {
 437         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 438
 439         dprintk("RPC:       %s: closing\n", __func__);
 440         if (r_xprt->rx_ep.rep_connected > 0)
 441                 xprt->reestablish_timeout = 0;
 442         xprt_disconnect_done(xprt);
 443         rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia);
 444 }
 445
 446 static void
 447 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port)
 448 {
 449         struct sockaddr_in *sap;
 450
 451         sap = (struct sockaddr_in *)&xprt->addr;
 452         sap->sin_port = htons(port);
 453         sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr;
 454         sap->sin_port = htons(port);
 455         dprintk("RPC:       %s: %u\n", __func__, port);
 456 }
 457
 458 static void
 459 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task)
 460 {
 461         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 462
 463         if (r_xprt->rx_ep.rep_connected != 0) {
 464                 /* Reconnect */
 465                 schedule_delayed_work(&r_xprt->rx_connect_worker,
 466                                       xprt->reestablish_timeout);
 467                 xprt->reestablish_timeout <<= 1;
 468                 if (xprt->reestablish_timeout > RPCRDMA_MAX_REEST_TO)
 469                         xprt->reestablish_timeout = RPCRDMA_MAX_REEST_TO;
 470                 else if (xprt->reestablish_timeout < RPCRDMA_INIT_REEST_TO)
 471                         xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO;
 472         } else {
 473                 schedule_delayed_work(&r_xprt->rx_connect_worker, 0);
 474                 if (!RPC_IS_ASYNC(task))
 475                         flush_delayed_work(&r_xprt->rx_connect_worker);
 476         }
 477 }
 478
 479 /*
 480  * The RDMA allocate/free functions need the task structure as a place
 481  * to hide the struct rpcrdma_req, which is necessary for the actual send/recv
 482  * sequence.
 483  *
 484  * The RPC layer allocates both send and receive buffers in the same call
 485  * (rq_send_buf and rq_rcv_buf are both part of a single contiguous buffer).
 486  * We may register rq_rcv_buf when using reply chunks.
 487  */
 488 static void *
 489 xprt_rdma_allocate(struct rpc_task *task, size_t size)
 490 {
 491         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 492         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 493         struct rpcrdma_regbuf *rb;
 494         struct rpcrdma_req *req;
 495         size_t min_size;
 496         gfp_t flags;
 497
 498         req = rpcrdma_buffer_get(&r_xprt->rx_buf);
 499         if (req == NULL)
 500                 return NULL;
 501
 502         flags = GFP_NOIO | __GFP_NOWARN;
 503         if (RPC_IS_SWAPPER(task))
 504                 flags = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
 505
 506         if (req->rl_rdmabuf == NULL)
 507                 goto out_rdmabuf;
 508         if (req->rl_sendbuf == NULL)
 509                 goto out_sendbuf;
 510         if (size > req->rl_sendbuf->rg_size)
 511                 goto out_sendbuf;
 512
 513 out:
 514         dprintk("RPC:       %s: size %zd, request 0x%p\n", __func__, size, req);
 515         req->rl_connect_cookie = 0;     /* our reserved value */
 516         return req->rl_sendbuf->rg_base;
 517
 518 out_rdmabuf:
 519         min_size = RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
 520         rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, min_size, flags);
 521         if (IS_ERR(rb))
 522                 goto out_fail;
 523         req->rl_rdmabuf = rb;
 524
 525 out_sendbuf:
 526         /* XDR encoding and RPC/RDMA marshaling of this request has not
 527          * yet occurred. Thus a lower bound is needed to prevent buffer
 528          * overrun during marshaling.
 529          *
 530          * RPC/RDMA marshaling may choose to send payload bearing ops
 531          * inline, if the result is smaller than the inline threshold.
 532          * The value of the "size" argument accounts for header
 533          * requirements but not for the payload in these cases.
 534          *
 535          * Likewise, allocate enough space to receive a reply up to the
 536          * size of the inline threshold.
 537          *
 538          * It's unlikely that both the send header and the received
 539          * reply will be large, but slush is provided here to allow
 540          * flexibility when marshaling.
 541          */
 542         min_size = RPCRDMA_INLINE_READ_THRESHOLD(task->tk_rqstp);
 543         min_size += RPCRDMA_INLINE_WRITE_THRESHOLD(task->tk_rqstp);
 544         if (size < min_size)
 545                 size = min_size;
 546
 547         rb = rpcrdma_alloc_regbuf(&r_xprt->rx_ia, size, flags);
 548         if (IS_ERR(rb))
 549                 goto out_fail;
 550         rb->rg_owner = req;
 551
 552         r_xprt->rx_stats.hardway_register_count += size;
 553         rpcrdma_free_regbuf(&r_xprt->rx_ia, req->rl_sendbuf);
 554         req->rl_sendbuf = rb;
 555         goto out;
 556
 557 out_fail:
 558         rpcrdma_buffer_put(req);
 559         r_xprt->rx_stats.failed_marshal_count++;
 560         return NULL;
 561 }
 562
 563 /*
 564  * This function returns all RDMA resources to the pool.
 565  */
 566 static void
 567 xprt_rdma_free(void *buffer)
 568 {
 569         struct rpcrdma_req *req;
 570         struct rpcrdma_xprt *r_xprt;
 571         struct rpcrdma_regbuf *rb;
 572         int i;
 573
 574         if (buffer == NULL)
 575                 return;
 576
 577         rb = container_of(buffer, struct rpcrdma_regbuf, rg_base[0]);
 578         req = rb->rg_owner;
 579         r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf);
 580
 581         dprintk("RPC:       %s: called on 0x%p\n", __func__, req->rl_reply);
 582
 583         for (i = 0; req->rl_nchunks;) {
 584                 --req->rl_nchunks;
 585                 i += r_xprt->rx_ia.ri_ops->ro_unmap(r_xprt,
 586                                                     &req->rl_segments[i]);
 587         }
 588
 589         rpcrdma_buffer_put(req);
 590 }
 591
 592 /*
 593  * send_request invokes the meat of RPC RDMA. It must do the following:
 594  *  1.  Marshal the RPC request into an RPC RDMA request, which means
 595  *      putting a header in front of data, and creating IOVs for RDMA
 596  *      from those in the request.
 597  *  2.  In marshaling, detect opportunities for RDMA, and use them.
 598  *  3.  Post a recv message to set up asynch completion, then send
 599  *      the request (rpcrdma_ep_post).
 600  *  4.  No partial sends are possible in the RPC-RDMA protocol (as in UDP).
 601  */
 602
 603 static int
 604 xprt_rdma_send_request(struct rpc_task *task)
 605 {
 606         struct rpc_rqst *rqst = task->tk_rqstp;
 607         struct rpc_xprt *xprt = rqst->rq_xprt;
 608         struct rpcrdma_req *req = rpcr_to_rdmar(rqst);
 609         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 610         int rc = 0;
 611
 612         rc = rpcrdma_marshal_req(rqst);
 613         if (rc < 0)
 614                 goto failed_marshal;
 615
 616         if (req->rl_reply == NULL)              /* e.g. reconnection */
 617                 rpcrdma_recv_buffer_get(req);
 618
 619         /* Must suppress retransmit to maintain credits */
 620         if (req->rl_connect_cookie == xprt->connect_cookie)
 621                 goto drop_connection;
 622         req->rl_connect_cookie = xprt->connect_cookie;
 623
 624         if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req))
 625                 goto drop_connection;
 626
 627         rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len;
 628         rqst->rq_bytes_sent = 0;
 629         return 0;
 630
 631 failed_marshal:
 632         r_xprt->rx_stats.failed_marshal_count++;
 633         dprintk("RPC:       %s: rpcrdma_marshal_req failed, status %i\n",
 634                 __func__, rc);
 635         if (rc == -EIO)
 636                 return -EIO;
 637 drop_connection:
 638         xprt_disconnect_done(xprt);
 639         return -ENOTCONN;       /* implies disconnect */
 640 }
 641
 642 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
 643 {
 644         struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 645         long idle_time = 0;
 646
 647         if (xprt_connected(xprt))
 648                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
 649
 650         seq_puts(seq, "\txprt:\trdma ");
 651         seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ",
 652                    0,   /* need a local port? */
 653                    xprt->stat.bind_count,
 654                    xprt->stat.connect_count,
 655                    xprt->stat.connect_time,
 656                    idle_time,
 657                    xprt->stat.sends,
 658                    xprt->stat.recvs,
 659                    xprt->stat.bad_xids,
 660                    xprt->stat.req_u,
 661                    xprt->stat.bklog_u);
 662         seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu\n",
 663                    r_xprt->rx_stats.read_chunk_count,
 664                    r_xprt->rx_stats.write_chunk_count,
 665                    r_xprt->rx_stats.reply_chunk_count,
 666                    r_xprt->rx_stats.total_rdma_request,
 667                    r_xprt->rx_stats.total_rdma_reply,
 668                    r_xprt->rx_stats.pullup_copy_count,
 669                    r_xprt->rx_stats.fixup_copy_count,
 670                    r_xprt->rx_stats.hardway_register_count,
 671                    r_xprt->rx_stats.failed_marshal_count,
 672                    r_xprt->rx_stats.bad_reply_count,
 673                    r_xprt->rx_stats.nomsg_call_count);
 674 }
 675
 676 static int
 677 xprt_rdma_enable_swap(struct rpc_xprt *xprt)
 678 {
 679         return 0;
 680 }
 681
 682 static void
 683 xprt_rdma_disable_swap(struct rpc_xprt *xprt)
 684 {
 685 }
 686
 687 /*
 688  * Plumbing for rpc transport switch and kernel module
 689  */
 690
 691 static struct rpc_xprt_ops xprt_rdma_procs = {
 692         .reserve_xprt           = xprt_reserve_xprt_cong,
 693         .release_xprt           = xprt_release_xprt_cong, /* sunrpc/xprt.c */
 694         .alloc_slot             = xprt_alloc_slot,
 695         .release_request        = xprt_release_rqst_cong,       /* ditto */
 696         .set_retrans_timeout    = xprt_set_retrans_timeout_def, /* ditto */
 697         .rpcbind                = rpcb_getport_async,   /* sunrpc/rpcb_clnt.c */
 698         .set_port               = xprt_rdma_set_port,
 699         .connect                = xprt_rdma_connect,
 700         .buf_alloc              = xprt_rdma_allocate,
 701         .buf_free               = xprt_rdma_free,
 702         .send_request           = xprt_rdma_send_request,
 703         .close                  = xprt_rdma_close,
 704         .destroy                = xprt_rdma_destroy,
 705         .print_stats            = xprt_rdma_print_stats,
 706         .enable_swap            = xprt_rdma_enable_swap,
 707         .disable_swap           = xprt_rdma_disable_swap,
 708         .inject_disconnect      = xprt_rdma_inject_disconnect,
 709 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
 710         .bc_setup               = xprt_rdma_bc_setup,
 711         .bc_up                  = xprt_rdma_bc_up,
 712         .bc_free_rqst           = xprt_rdma_bc_free_rqst,
 713         .bc_destroy             = xprt_rdma_bc_destroy,
 714 #endif
 715 };
 716
 717 static struct xprt_class xprt_rdma = {
 718         .list                   = LIST_HEAD_INIT(xprt_rdma.list),
 719         .name                   = "rdma",
 720         .owner                  = THIS_MODULE,
 721         .ident                  = XPRT_TRANSPORT_RDMA,
 722         .setup                  = xprt_setup_rdma,
 723 };
 724
 725 void xprt_rdma_cleanup(void)
 726 {
 727         int rc;
 728
 729         dprintk("RPCRDMA Module Removed, deregister RPC RDMA transport\n");
 730 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 731         if (sunrpc_table_header) {
 732                 unregister_sysctl_table(sunrpc_table_header);
 733                 sunrpc_table_header = NULL;
 734         }
 735 #endif
 736         rc = xprt_unregister_transport(&xprt_rdma);
 737         if (rc)
 738                 dprintk("RPC:       %s: xprt_unregister returned %i\n",
 739                         __func__, rc);
 740
 741         rpcrdma_destroy_wq();
 742         frwr_destroy_recovery_wq();
 743 }
 744
 745 int xprt_rdma_init(void)
 746 {
 747         int rc;
 748
 749         rc = frwr_alloc_recovery_wq();
 750         if (rc)
 751                 return rc;
 752
 753         rc = rpcrdma_alloc_wq();
 754         if (rc) {
 755                 frwr_destroy_recovery_wq();
 756                 return rc;
 757         }
 758
 759         rc = xprt_register_transport(&xprt_rdma);
 760         if (rc) {
 761                 rpcrdma_destroy_wq();
 762                 frwr_destroy_recovery_wq();
 763                 return rc;
 764         }
 765
 766         dprintk("RPCRDMA Module Init, register RPC RDMA transport\n");
 767
 768         dprintk("Defaults:\n");
 769         dprintk("\tSlots %d\n"
 770                 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n",
 771                 xprt_rdma_slot_table_entries,
 772                 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write);
 773         dprintk("\tPadding %d\n\tMemreg %d\n",
 774                 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy);
 775
 776 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
 777         if (!sunrpc_table_header)
 778                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
 779 #endif
 780         return 0;
 781 }