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/seq_file.h> 53 54 #include "xprt_rdma.h" 55 56 #ifdef RPC_DEBUG 57 # define RPCDBG_FACILITY RPCDBG_TRANS 58 #endif 59 60 MODULE_LICENSE("Dual BSD/GPL"); 61 62 MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS"); 63 MODULE_AUTHOR("Network Appliance, Inc."); 64 65 /* 66 * tunables 67 */ 68 69 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE; 70 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE; 71 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE; 72 static unsigned int xprt_rdma_inline_write_padding; 73 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR; 74 int xprt_rdma_pad_optimize = 0; 75 76 #ifdef RPC_DEBUG 77 78 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE; 79 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE; 80 static unsigned int zero; 81 static unsigned int max_padding = PAGE_SIZE; 82 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS; 83 static unsigned int max_memreg = RPCRDMA_LAST - 1; 84 85 static struct ctl_table_header *sunrpc_table_header; 86 87 static ctl_table xr_tunables_table[] = { 88 { 89 .ctl_name = CTL_UNNUMBERED, 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 .strategy = &sysctl_intvec, 96 .extra1 = &min_slot_table_size, 97 .extra2 = &max_slot_table_size 98 }, 99 { 100 .ctl_name = CTL_UNNUMBERED, 101 .procname = "rdma_max_inline_read", 102 .data = &xprt_rdma_max_inline_read, 103 .maxlen = sizeof(unsigned int), 104 .mode = 0644, 105 .proc_handler = &proc_dointvec, 106 .strategy = &sysctl_intvec, 107 }, 108 { 109 .ctl_name = CTL_UNNUMBERED, 110 .procname = "rdma_max_inline_write", 111 .data = &xprt_rdma_max_inline_write, 112 .maxlen = sizeof(unsigned int), 113 .mode = 0644, 114 .proc_handler = &proc_dointvec, 115 .strategy = &sysctl_intvec, 116 }, 117 { 118 .ctl_name = CTL_UNNUMBERED, 119 .procname = "rdma_inline_write_padding", 120 .data = &xprt_rdma_inline_write_padding, 121 .maxlen = sizeof(unsigned int), 122 .mode = 0644, 123 .proc_handler = &proc_dointvec_minmax, 124 .strategy = &sysctl_intvec, 125 .extra1 = &zero, 126 .extra2 = &max_padding, 127 }, 128 { 129 .ctl_name = CTL_UNNUMBERED, 130 .procname = "rdma_memreg_strategy", 131 .data = &xprt_rdma_memreg_strategy, 132 .maxlen = sizeof(unsigned int), 133 .mode = 0644, 134 .proc_handler = &proc_dointvec_minmax, 135 .strategy = &sysctl_intvec, 136 .extra1 = &min_memreg, 137 .extra2 = &max_memreg, 138 }, 139 { 140 .ctl_name = CTL_UNNUMBERED, 141 .procname = "rdma_pad_optimize", 142 .data = &xprt_rdma_pad_optimize, 143 .maxlen = sizeof(unsigned int), 144 .mode = 0644, 145 .proc_handler = &proc_dointvec, 146 }, 147 { 148 .ctl_name = 0, 149 }, 150 }; 151 152 static ctl_table sunrpc_table[] = { 153 { 154 .ctl_name = CTL_SUNRPC, 155 .procname = "sunrpc", 156 .mode = 0555, 157 .child = xr_tunables_table 158 }, 159 { 160 .ctl_name = 0, 161 }, 162 }; 163 164 #endif 165 166 static struct rpc_xprt_ops xprt_rdma_procs; /* forward reference */ 167 168 static void 169 xprt_rdma_format_addresses(struct rpc_xprt *xprt) 170 { 171 struct sockaddr_in *addr = (struct sockaddr_in *) 172 &rpcx_to_rdmad(xprt).addr; 173 char *buf; 174 175 buf = kzalloc(20, GFP_KERNEL); 176 if (buf) 177 snprintf(buf, 20, "%pI4", &addr->sin_addr.s_addr); 178 xprt->address_strings[RPC_DISPLAY_ADDR] = buf; 179 180 buf = kzalloc(8, GFP_KERNEL); 181 if (buf) 182 snprintf(buf, 8, "%u", ntohs(addr->sin_port)); 183 xprt->address_strings[RPC_DISPLAY_PORT] = buf; 184 185 xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma"; 186 187 buf = kzalloc(48, GFP_KERNEL); 188 if (buf) 189 snprintf(buf, 48, "addr=%pI4 port=%u proto=%s", 190 &addr->sin_addr.s_addr, 191 ntohs(addr->sin_port), "rdma"); 192 xprt->address_strings[RPC_DISPLAY_ALL] = buf; 193 194 buf = kzalloc(10, GFP_KERNEL); 195 if (buf) 196 snprintf(buf, 10, "%02x%02x%02x%02x", 197 NIPQUAD(addr->sin_addr.s_addr)); 198 xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = buf; 199 200 buf = kzalloc(8, GFP_KERNEL); 201 if (buf) 202 snprintf(buf, 8, "%4hx", ntohs(addr->sin_port)); 203 xprt->address_strings[RPC_DISPLAY_HEX_PORT] = buf; 204 205 buf = kzalloc(30, GFP_KERNEL); 206 if (buf) 207 snprintf(buf, 30, "%pI4.%u.%u", 208 &addr->sin_addr.s_addr, 209 ntohs(addr->sin_port) >> 8, 210 ntohs(addr->sin_port) & 0xff); 211 xprt->address_strings[RPC_DISPLAY_UNIVERSAL_ADDR] = buf; 212 213 /* netid */ 214 xprt->address_strings[RPC_DISPLAY_NETID] = "rdma"; 215 } 216 217 static void 218 xprt_rdma_free_addresses(struct rpc_xprt *xprt) 219 { 220 unsigned int i; 221 222 for (i = 0; i < RPC_DISPLAY_MAX; i++) 223 switch (i) { 224 case RPC_DISPLAY_PROTO: 225 case RPC_DISPLAY_NETID: 226 continue; 227 default: 228 kfree(xprt->address_strings[i]); 229 } 230 } 231 232 static void 233 xprt_rdma_connect_worker(struct work_struct *work) 234 { 235 struct rpcrdma_xprt *r_xprt = 236 container_of(work, struct rpcrdma_xprt, rdma_connect.work); 237 struct rpc_xprt *xprt = &r_xprt->xprt; 238 int rc = 0; 239 240 if (!xprt->shutdown) { 241 xprt_clear_connected(xprt); 242 243 dprintk("RPC: %s: %sconnect\n", __func__, 244 r_xprt->rx_ep.rep_connected != 0 ? "re" : ""); 245 rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia); 246 if (rc) 247 goto out; 248 } 249 goto out_clear; 250 251 out: 252 xprt_wake_pending_tasks(xprt, rc); 253 254 out_clear: 255 dprintk("RPC: %s: exit\n", __func__); 256 xprt_clear_connecting(xprt); 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 int rc; 275 276 dprintk("RPC: %s: called\n", __func__); 277 278 cancel_delayed_work(&r_xprt->rdma_connect); 279 flush_scheduled_work(); 280 281 xprt_clear_connected(xprt); 282 283 rpcrdma_buffer_destroy(&r_xprt->rx_buf); 284 rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia); 285 if (rc) 286 dprintk("RPC: %s: rpcrdma_ep_destroy returned %i\n", 287 __func__, rc); 288 rpcrdma_ia_close(&r_xprt->rx_ia); 289 290 xprt_rdma_free_addresses(xprt); 291 292 kfree(xprt->slot); 293 xprt->slot = NULL; 294 kfree(xprt); 295 296 dprintk("RPC: %s: returning\n", __func__); 297 298 module_put(THIS_MODULE); 299 } 300 301 static const struct rpc_timeout xprt_rdma_default_timeout = { 302 .to_initval = 60 * HZ, 303 .to_maxval = 60 * HZ, 304 }; 305 306 /** 307 * xprt_setup_rdma - Set up transport to use RDMA 308 * 309 * @args: rpc transport arguments 310 */ 311 static struct rpc_xprt * 312 xprt_setup_rdma(struct xprt_create *args) 313 { 314 struct rpcrdma_create_data_internal cdata; 315 struct rpc_xprt *xprt; 316 struct rpcrdma_xprt *new_xprt; 317 struct rpcrdma_ep *new_ep; 318 struct sockaddr_in *sin; 319 int rc; 320 321 if (args->addrlen > sizeof(xprt->addr)) { 322 dprintk("RPC: %s: address too large\n", __func__); 323 return ERR_PTR(-EBADF); 324 } 325 326 xprt = kzalloc(sizeof(struct rpcrdma_xprt), GFP_KERNEL); 327 if (xprt == NULL) { 328 dprintk("RPC: %s: couldn't allocate rpcrdma_xprt\n", 329 __func__); 330 return ERR_PTR(-ENOMEM); 331 } 332 333 xprt->max_reqs = xprt_rdma_slot_table_entries; 334 xprt->slot = kcalloc(xprt->max_reqs, 335 sizeof(struct rpc_rqst), GFP_KERNEL); 336 if (xprt->slot == NULL) { 337 dprintk("RPC: %s: couldn't allocate %d slots\n", 338 __func__, xprt->max_reqs); 339 kfree(xprt); 340 return ERR_PTR(-ENOMEM); 341 } 342 343 /* 60 second timeout, no retries */ 344 xprt->timeout = &xprt_rdma_default_timeout; 345 xprt->bind_timeout = (60U * HZ); 346 xprt->connect_timeout = (60U * HZ); 347 xprt->reestablish_timeout = (5U * HZ); 348 xprt->idle_timeout = (5U * 60 * HZ); 349 350 xprt->resvport = 0; /* privileged port not needed */ 351 xprt->tsh_size = 0; /* RPC-RDMA handles framing */ 352 xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE; 353 xprt->ops = &xprt_rdma_procs; 354 355 /* 356 * Set up RDMA-specific connect data. 357 */ 358 359 /* Put server RDMA address in local cdata */ 360 memcpy(&cdata.addr, args->dstaddr, args->addrlen); 361 362 /* Ensure xprt->addr holds valid server TCP (not RDMA) 363 * address, for any side protocols which peek at it */ 364 xprt->prot = IPPROTO_TCP; 365 xprt->addrlen = args->addrlen; 366 memcpy(&xprt->addr, &cdata.addr, xprt->addrlen); 367 368 sin = (struct sockaddr_in *)&cdata.addr; 369 if (ntohs(sin->sin_port) != 0) 370 xprt_set_bound(xprt); 371 372 dprintk("RPC: %s: %pI4:%u\n", 373 __func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port)); 374 375 /* Set max requests */ 376 cdata.max_requests = xprt->max_reqs; 377 378 /* Set some length limits */ 379 cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */ 380 cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */ 381 382 cdata.inline_wsize = xprt_rdma_max_inline_write; 383 if (cdata.inline_wsize > cdata.wsize) 384 cdata.inline_wsize = cdata.wsize; 385 386 cdata.inline_rsize = xprt_rdma_max_inline_read; 387 if (cdata.inline_rsize > cdata.rsize) 388 cdata.inline_rsize = cdata.rsize; 389 390 cdata.padding = xprt_rdma_inline_write_padding; 391 392 /* 393 * Create new transport instance, which includes initialized 394 * o ia 395 * o endpoint 396 * o buffers 397 */ 398 399 new_xprt = rpcx_to_rdmax(xprt); 400 401 rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr, 402 xprt_rdma_memreg_strategy); 403 if (rc) 404 goto out1; 405 406 /* 407 * initialize and create ep 408 */ 409 new_xprt->rx_data = cdata; 410 new_ep = &new_xprt->rx_ep; 411 new_ep->rep_remote_addr = cdata.addr; 412 413 rc = rpcrdma_ep_create(&new_xprt->rx_ep, 414 &new_xprt->rx_ia, &new_xprt->rx_data); 415 if (rc) 416 goto out2; 417 418 /* 419 * Allocate pre-registered send and receive buffers for headers and 420 * any inline data. Also specify any padding which will be provided 421 * from a preregistered zero buffer. 422 */ 423 rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia, 424 &new_xprt->rx_data); 425 if (rc) 426 goto out3; 427 428 /* 429 * Register a callback for connection events. This is necessary because 430 * connection loss notification is async. We also catch connection loss 431 * when reaping receives. 432 */ 433 INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker); 434 new_ep->rep_func = rpcrdma_conn_func; 435 new_ep->rep_xprt = xprt; 436 437 xprt_rdma_format_addresses(xprt); 438 439 if (!try_module_get(THIS_MODULE)) 440 goto out4; 441 442 return xprt; 443 444 out4: 445 xprt_rdma_free_addresses(xprt); 446 rc = -EINVAL; 447 out3: 448 (void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia); 449 out2: 450 rpcrdma_ia_close(&new_xprt->rx_ia); 451 out1: 452 kfree(xprt->slot); 453 kfree(xprt); 454 return ERR_PTR(rc); 455 } 456 457 /* 458 * Close a connection, during shutdown or timeout/reconnect 459 */ 460 static void 461 xprt_rdma_close(struct rpc_xprt *xprt) 462 { 463 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 464 465 dprintk("RPC: %s: closing\n", __func__); 466 if (r_xprt->rx_ep.rep_connected > 0) 467 xprt->reestablish_timeout = 0; 468 xprt_disconnect_done(xprt); 469 (void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia); 470 } 471 472 static void 473 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port) 474 { 475 struct sockaddr_in *sap; 476 477 sap = (struct sockaddr_in *)&xprt->addr; 478 sap->sin_port = htons(port); 479 sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr; 480 sap->sin_port = htons(port); 481 dprintk("RPC: %s: %u\n", __func__, port); 482 } 483 484 static void 485 xprt_rdma_connect(struct rpc_task *task) 486 { 487 struct rpc_xprt *xprt = (struct rpc_xprt *)task->tk_xprt; 488 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 489 490 if (!xprt_test_and_set_connecting(xprt)) { 491 if (r_xprt->rx_ep.rep_connected != 0) { 492 /* Reconnect */ 493 schedule_delayed_work(&r_xprt->rdma_connect, 494 xprt->reestablish_timeout); 495 xprt->reestablish_timeout <<= 1; 496 if (xprt->reestablish_timeout > (30 * HZ)) 497 xprt->reestablish_timeout = (30 * HZ); 498 else if (xprt->reestablish_timeout < (5 * HZ)) 499 xprt->reestablish_timeout = (5 * HZ); 500 } else { 501 schedule_delayed_work(&r_xprt->rdma_connect, 0); 502 if (!RPC_IS_ASYNC(task)) 503 flush_scheduled_work(); 504 } 505 } 506 } 507 508 static int 509 xprt_rdma_reserve_xprt(struct rpc_task *task) 510 { 511 struct rpc_xprt *xprt = task->tk_xprt; 512 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 513 int credits = atomic_read(&r_xprt->rx_buf.rb_credits); 514 515 /* == RPC_CWNDSCALE @ init, but *after* setup */ 516 if (r_xprt->rx_buf.rb_cwndscale == 0UL) { 517 r_xprt->rx_buf.rb_cwndscale = xprt->cwnd; 518 dprintk("RPC: %s: cwndscale %lu\n", __func__, 519 r_xprt->rx_buf.rb_cwndscale); 520 BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0); 521 } 522 xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale; 523 return xprt_reserve_xprt_cong(task); 524 } 525 526 /* 527 * The RDMA allocate/free functions need the task structure as a place 528 * to hide the struct rpcrdma_req, which is necessary for the actual send/recv 529 * sequence. For this reason, the recv buffers are attached to send 530 * buffers for portions of the RPC. Note that the RPC layer allocates 531 * both send and receive buffers in the same call. We may register 532 * the receive buffer portion when using reply chunks. 533 */ 534 static void * 535 xprt_rdma_allocate(struct rpc_task *task, size_t size) 536 { 537 struct rpc_xprt *xprt = task->tk_xprt; 538 struct rpcrdma_req *req, *nreq; 539 540 req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf); 541 BUG_ON(NULL == req); 542 543 if (size > req->rl_size) { 544 dprintk("RPC: %s: size %zd too large for buffer[%zd]: " 545 "prog %d vers %d proc %d\n", 546 __func__, size, req->rl_size, 547 task->tk_client->cl_prog, task->tk_client->cl_vers, 548 task->tk_msg.rpc_proc->p_proc); 549 /* 550 * Outgoing length shortage. Our inline write max must have 551 * been configured to perform direct i/o. 552 * 553 * This is therefore a large metadata operation, and the 554 * allocate call was made on the maximum possible message, 555 * e.g. containing long filename(s) or symlink data. In 556 * fact, while these metadata operations *might* carry 557 * large outgoing payloads, they rarely *do*. However, we 558 * have to commit to the request here, so reallocate and 559 * register it now. The data path will never require this 560 * reallocation. 561 * 562 * If the allocation or registration fails, the RPC framework 563 * will (doggedly) retry. 564 */ 565 if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy == 566 RPCRDMA_BOUNCEBUFFERS) { 567 /* forced to "pure inline" */ 568 dprintk("RPC: %s: too much data (%zd) for inline " 569 "(r/w max %d/%d)\n", __func__, size, 570 rpcx_to_rdmad(xprt).inline_rsize, 571 rpcx_to_rdmad(xprt).inline_wsize); 572 size = req->rl_size; 573 rpc_exit(task, -EIO); /* fail the operation */ 574 rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++; 575 goto out; 576 } 577 if (task->tk_flags & RPC_TASK_SWAPPER) 578 nreq = kmalloc(sizeof *req + size, GFP_ATOMIC); 579 else 580 nreq = kmalloc(sizeof *req + size, GFP_NOFS); 581 if (nreq == NULL) 582 goto outfail; 583 584 if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia, 585 nreq->rl_base, size + sizeof(struct rpcrdma_req) 586 - offsetof(struct rpcrdma_req, rl_base), 587 &nreq->rl_handle, &nreq->rl_iov)) { 588 kfree(nreq); 589 goto outfail; 590 } 591 rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size; 592 nreq->rl_size = size; 593 nreq->rl_niovs = 0; 594 nreq->rl_nchunks = 0; 595 nreq->rl_buffer = (struct rpcrdma_buffer *)req; 596 nreq->rl_reply = req->rl_reply; 597 memcpy(nreq->rl_segments, 598 req->rl_segments, sizeof nreq->rl_segments); 599 /* flag the swap with an unused field */ 600 nreq->rl_iov.length = 0; 601 req->rl_reply = NULL; 602 req = nreq; 603 } 604 dprintk("RPC: %s: size %zd, request 0x%p\n", __func__, size, req); 605 out: 606 req->rl_connect_cookie = 0; /* our reserved value */ 607 return req->rl_xdr_buf; 608 609 outfail: 610 rpcrdma_buffer_put(req); 611 rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++; 612 return NULL; 613 } 614 615 /* 616 * This function returns all RDMA resources to the pool. 617 */ 618 static void 619 xprt_rdma_free(void *buffer) 620 { 621 struct rpcrdma_req *req; 622 struct rpcrdma_xprt *r_xprt; 623 struct rpcrdma_rep *rep; 624 int i; 625 626 if (buffer == NULL) 627 return; 628 629 req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]); 630 if (req->rl_iov.length == 0) { /* see allocate above */ 631 r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer, 632 struct rpcrdma_xprt, rx_buf); 633 } else 634 r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf); 635 rep = req->rl_reply; 636 637 dprintk("RPC: %s: called on 0x%p%s\n", 638 __func__, rep, (rep && rep->rr_func) ? " (with waiter)" : ""); 639 640 /* 641 * Finish the deregistration. When using mw bind, this was 642 * begun in rpcrdma_reply_handler(). In all other modes, we 643 * do it here, in thread context. The process is considered 644 * complete when the rr_func vector becomes NULL - this 645 * was put in place during rpcrdma_reply_handler() - the wait 646 * call below will not block if the dereg is "done". If 647 * interrupted, our framework will clean up. 648 */ 649 for (i = 0; req->rl_nchunks;) { 650 --req->rl_nchunks; 651 i += rpcrdma_deregister_external( 652 &req->rl_segments[i], r_xprt, NULL); 653 } 654 655 if (rep && wait_event_interruptible(rep->rr_unbind, !rep->rr_func)) { 656 rep->rr_func = NULL; /* abandon the callback */ 657 req->rl_reply = NULL; 658 } 659 660 if (req->rl_iov.length == 0) { /* see allocate above */ 661 struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer; 662 oreq->rl_reply = req->rl_reply; 663 (void) rpcrdma_deregister_internal(&r_xprt->rx_ia, 664 req->rl_handle, 665 &req->rl_iov); 666 kfree(req); 667 req = oreq; 668 } 669 670 /* Put back request+reply buffers */ 671 rpcrdma_buffer_put(req); 672 } 673 674 /* 675 * send_request invokes the meat of RPC RDMA. It must do the following: 676 * 1. Marshal the RPC request into an RPC RDMA request, which means 677 * putting a header in front of data, and creating IOVs for RDMA 678 * from those in the request. 679 * 2. In marshaling, detect opportunities for RDMA, and use them. 680 * 3. Post a recv message to set up asynch completion, then send 681 * the request (rpcrdma_ep_post). 682 * 4. No partial sends are possible in the RPC-RDMA protocol (as in UDP). 683 */ 684 685 static int 686 xprt_rdma_send_request(struct rpc_task *task) 687 { 688 struct rpc_rqst *rqst = task->tk_rqstp; 689 struct rpc_xprt *xprt = task->tk_xprt; 690 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 691 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 692 693 /* marshal the send itself */ 694 if (req->rl_niovs == 0 && rpcrdma_marshal_req(rqst) != 0) { 695 r_xprt->rx_stats.failed_marshal_count++; 696 dprintk("RPC: %s: rpcrdma_marshal_req failed\n", 697 __func__); 698 return -EIO; 699 } 700 701 if (req->rl_reply == NULL) /* e.g. reconnection */ 702 rpcrdma_recv_buffer_get(req); 703 704 if (req->rl_reply) { 705 req->rl_reply->rr_func = rpcrdma_reply_handler; 706 /* this need only be done once, but... */ 707 req->rl_reply->rr_xprt = xprt; 708 } 709 710 /* Must suppress retransmit to maintain credits */ 711 if (req->rl_connect_cookie == xprt->connect_cookie) 712 goto drop_connection; 713 req->rl_connect_cookie = xprt->connect_cookie; 714 715 if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req)) 716 goto drop_connection; 717 718 task->tk_bytes_sent += rqst->rq_snd_buf.len; 719 rqst->rq_bytes_sent = 0; 720 return 0; 721 722 drop_connection: 723 xprt_disconnect_done(xprt); 724 return -ENOTCONN; /* implies disconnect */ 725 } 726 727 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq) 728 { 729 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 730 long idle_time = 0; 731 732 if (xprt_connected(xprt)) 733 idle_time = (long)(jiffies - xprt->last_used) / HZ; 734 735 seq_printf(seq, 736 "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu " 737 "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n", 738 739 0, /* need a local port? */ 740 xprt->stat.bind_count, 741 xprt->stat.connect_count, 742 xprt->stat.connect_time, 743 idle_time, 744 xprt->stat.sends, 745 xprt->stat.recvs, 746 xprt->stat.bad_xids, 747 xprt->stat.req_u, 748 xprt->stat.bklog_u, 749 750 r_xprt->rx_stats.read_chunk_count, 751 r_xprt->rx_stats.write_chunk_count, 752 r_xprt->rx_stats.reply_chunk_count, 753 r_xprt->rx_stats.total_rdma_request, 754 r_xprt->rx_stats.total_rdma_reply, 755 r_xprt->rx_stats.pullup_copy_count, 756 r_xprt->rx_stats.fixup_copy_count, 757 r_xprt->rx_stats.hardway_register_count, 758 r_xprt->rx_stats.failed_marshal_count, 759 r_xprt->rx_stats.bad_reply_count); 760 } 761 762 /* 763 * Plumbing for rpc transport switch and kernel module 764 */ 765 766 static struct rpc_xprt_ops xprt_rdma_procs = { 767 .reserve_xprt = xprt_rdma_reserve_xprt, 768 .release_xprt = xprt_release_xprt_cong, /* sunrpc/xprt.c */ 769 .release_request = xprt_release_rqst_cong, /* ditto */ 770 .set_retrans_timeout = xprt_set_retrans_timeout_def, /* ditto */ 771 .rpcbind = rpcb_getport_async, /* sunrpc/rpcb_clnt.c */ 772 .set_port = xprt_rdma_set_port, 773 .connect = xprt_rdma_connect, 774 .buf_alloc = xprt_rdma_allocate, 775 .buf_free = xprt_rdma_free, 776 .send_request = xprt_rdma_send_request, 777 .close = xprt_rdma_close, 778 .destroy = xprt_rdma_destroy, 779 .print_stats = xprt_rdma_print_stats 780 }; 781 782 static struct xprt_class xprt_rdma = { 783 .list = LIST_HEAD_INIT(xprt_rdma.list), 784 .name = "rdma", 785 .owner = THIS_MODULE, 786 .ident = XPRT_TRANSPORT_RDMA, 787 .setup = xprt_setup_rdma, 788 }; 789 790 static void __exit xprt_rdma_cleanup(void) 791 { 792 int rc; 793 794 dprintk(KERN_INFO "RPCRDMA Module Removed, deregister RPC RDMA transport\n"); 795 #ifdef RPC_DEBUG 796 if (sunrpc_table_header) { 797 unregister_sysctl_table(sunrpc_table_header); 798 sunrpc_table_header = NULL; 799 } 800 #endif 801 rc = xprt_unregister_transport(&xprt_rdma); 802 if (rc) 803 dprintk("RPC: %s: xprt_unregister returned %i\n", 804 __func__, rc); 805 } 806 807 static int __init xprt_rdma_init(void) 808 { 809 int rc; 810 811 rc = xprt_register_transport(&xprt_rdma); 812 813 if (rc) 814 return rc; 815 816 dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n"); 817 818 dprintk(KERN_INFO "Defaults:\n"); 819 dprintk(KERN_INFO "\tSlots %d\n" 820 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n", 821 xprt_rdma_slot_table_entries, 822 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write); 823 dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n", 824 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy); 825 826 #ifdef RPC_DEBUG 827 if (!sunrpc_table_header) 828 sunrpc_table_header = register_sysctl_table(sunrpc_table); 829 #endif 830 return 0; 831 } 832 833 module_init(xprt_rdma_init); 834 module_exit(xprt_rdma_cleanup); 835