1 /* 2 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the BSD-type 8 * license below: 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 14 * Redistributions of source code must retain the above copyright 15 * notice, this list of conditions and the following disclaimer. 16 * 17 * Redistributions in binary form must reproduce the above 18 * copyright notice, this list of conditions and the following 19 * disclaimer in the documentation and/or other materials provided 20 * with the distribution. 21 * 22 * Neither the name of the Network Appliance, Inc. nor the names of 23 * its contributors may be used to endorse or promote products 24 * derived from this software without specific prior written 25 * permission. 26 * 27 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 28 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 29 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 30 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 31 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 32 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 33 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 34 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 35 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 36 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 37 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 38 */ 39 40 /* 41 * transport.c 42 * 43 * This file contains the top-level implementation of an RPC RDMA 44 * transport. 45 * 46 * Naming convention: functions beginning with xprt_ are part of the 47 * transport switch. All others are RPC RDMA internal. 48 */ 49 50 #include <linux/module.h> 51 #include <linux/init.h> 52 #include <linux/slab.h> 53 #include <linux/seq_file.h> 54 55 #include "xprt_rdma.h" 56 57 #ifdef RPC_DEBUG 58 # define RPCDBG_FACILITY RPCDBG_TRANS 59 #endif 60 61 MODULE_LICENSE("Dual BSD/GPL"); 62 63 MODULE_DESCRIPTION("RPC/RDMA Transport for Linux kernel NFS"); 64 MODULE_AUTHOR("Network Appliance, Inc."); 65 66 /* 67 * tunables 68 */ 69 70 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE; 71 static unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE; 72 static unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE; 73 static unsigned int xprt_rdma_inline_write_padding; 74 static unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRMR; 75 int xprt_rdma_pad_optimize = 0; 76 77 #ifdef RPC_DEBUG 78 79 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE; 80 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE; 81 static unsigned int zero; 82 static unsigned int max_padding = PAGE_SIZE; 83 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS; 84 static unsigned int max_memreg = RPCRDMA_LAST - 1; 85 86 static struct ctl_table_header *sunrpc_table_header; 87 88 static ctl_table xr_tunables_table[] = { 89 { 90 .procname = "rdma_slot_table_entries", 91 .data = &xprt_rdma_slot_table_entries, 92 .maxlen = sizeof(unsigned int), 93 .mode = 0644, 94 .proc_handler = proc_dointvec_minmax, 95 .extra1 = &min_slot_table_size, 96 .extra2 = &max_slot_table_size 97 }, 98 { 99 .procname = "rdma_max_inline_read", 100 .data = &xprt_rdma_max_inline_read, 101 .maxlen = sizeof(unsigned int), 102 .mode = 0644, 103 .proc_handler = proc_dointvec, 104 }, 105 { 106 .procname = "rdma_max_inline_write", 107 .data = &xprt_rdma_max_inline_write, 108 .maxlen = sizeof(unsigned int), 109 .mode = 0644, 110 .proc_handler = proc_dointvec, 111 }, 112 { 113 .procname = "rdma_inline_write_padding", 114 .data = &xprt_rdma_inline_write_padding, 115 .maxlen = sizeof(unsigned int), 116 .mode = 0644, 117 .proc_handler = proc_dointvec_minmax, 118 .extra1 = &zero, 119 .extra2 = &max_padding, 120 }, 121 { 122 .procname = "rdma_memreg_strategy", 123 .data = &xprt_rdma_memreg_strategy, 124 .maxlen = sizeof(unsigned int), 125 .mode = 0644, 126 .proc_handler = proc_dointvec_minmax, 127 .extra1 = &min_memreg, 128 .extra2 = &max_memreg, 129 }, 130 { 131 .procname = "rdma_pad_optimize", 132 .data = &xprt_rdma_pad_optimize, 133 .maxlen = sizeof(unsigned int), 134 .mode = 0644, 135 .proc_handler = proc_dointvec, 136 }, 137 { }, 138 }; 139 140 static ctl_table sunrpc_table[] = { 141 { 142 .procname = "sunrpc", 143 .mode = 0555, 144 .child = xr_tunables_table 145 }, 146 { }, 147 }; 148 149 #endif 150 151 static struct rpc_xprt_ops xprt_rdma_procs; /* forward reference */ 152 153 static void 154 xprt_rdma_format_addresses(struct rpc_xprt *xprt) 155 { 156 struct sockaddr *sap = (struct sockaddr *) 157 &rpcx_to_rdmad(xprt).addr; 158 struct sockaddr_in *sin = (struct sockaddr_in *)sap; 159 char buf[64]; 160 161 (void)rpc_ntop(sap, buf, sizeof(buf)); 162 xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL); 163 164 snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap)); 165 xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL); 166 167 xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma"; 168 169 snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr)); 170 xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL); 171 172 snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap)); 173 xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL); 174 175 /* netid */ 176 xprt->address_strings[RPC_DISPLAY_NETID] = "rdma"; 177 } 178 179 static void 180 xprt_rdma_free_addresses(struct rpc_xprt *xprt) 181 { 182 unsigned int i; 183 184 for (i = 0; i < RPC_DISPLAY_MAX; i++) 185 switch (i) { 186 case RPC_DISPLAY_PROTO: 187 case RPC_DISPLAY_NETID: 188 continue; 189 default: 190 kfree(xprt->address_strings[i]); 191 } 192 } 193 194 static void 195 xprt_rdma_connect_worker(struct work_struct *work) 196 { 197 struct rpcrdma_xprt *r_xprt = 198 container_of(work, struct rpcrdma_xprt, rdma_connect.work); 199 struct rpc_xprt *xprt = &r_xprt->xprt; 200 int rc = 0; 201 202 if (!xprt->shutdown) { 203 xprt_clear_connected(xprt); 204 205 dprintk("RPC: %s: %sconnect\n", __func__, 206 r_xprt->rx_ep.rep_connected != 0 ? "re" : ""); 207 rc = rpcrdma_ep_connect(&r_xprt->rx_ep, &r_xprt->rx_ia); 208 if (rc) 209 goto out; 210 } 211 goto out_clear; 212 213 out: 214 xprt_wake_pending_tasks(xprt, rc); 215 216 out_clear: 217 dprintk("RPC: %s: exit\n", __func__); 218 xprt_clear_connecting(xprt); 219 } 220 221 /* 222 * xprt_rdma_destroy 223 * 224 * Destroy the xprt. 225 * Free all memory associated with the object, including its own. 226 * NOTE: none of the *destroy methods free memory for their top-level 227 * objects, even though they may have allocated it (they do free 228 * private memory). It's up to the caller to handle it. In this 229 * case (RDMA transport), all structure memory is inlined with the 230 * struct rpcrdma_xprt. 231 */ 232 static void 233 xprt_rdma_destroy(struct rpc_xprt *xprt) 234 { 235 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 236 int rc; 237 238 dprintk("RPC: %s: called\n", __func__); 239 240 cancel_delayed_work_sync(&r_xprt->rdma_connect); 241 242 xprt_clear_connected(xprt); 243 244 rpcrdma_buffer_destroy(&r_xprt->rx_buf); 245 rc = rpcrdma_ep_destroy(&r_xprt->rx_ep, &r_xprt->rx_ia); 246 if (rc) 247 dprintk("RPC: %s: rpcrdma_ep_destroy returned %i\n", 248 __func__, rc); 249 rpcrdma_ia_close(&r_xprt->rx_ia); 250 251 xprt_rdma_free_addresses(xprt); 252 253 xprt_free(xprt); 254 255 dprintk("RPC: %s: returning\n", __func__); 256 257 module_put(THIS_MODULE); 258 } 259 260 static const struct rpc_timeout xprt_rdma_default_timeout = { 261 .to_initval = 60 * HZ, 262 .to_maxval = 60 * HZ, 263 }; 264 265 /** 266 * xprt_setup_rdma - Set up transport to use RDMA 267 * 268 * @args: rpc transport arguments 269 */ 270 static struct rpc_xprt * 271 xprt_setup_rdma(struct xprt_create *args) 272 { 273 struct rpcrdma_create_data_internal cdata; 274 struct rpc_xprt *xprt; 275 struct rpcrdma_xprt *new_xprt; 276 struct rpcrdma_ep *new_ep; 277 struct sockaddr_in *sin; 278 int rc; 279 280 if (args->addrlen > sizeof(xprt->addr)) { 281 dprintk("RPC: %s: address too large\n", __func__); 282 return ERR_PTR(-EBADF); 283 } 284 285 xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 286 xprt_rdma_slot_table_entries, 287 xprt_rdma_slot_table_entries); 288 if (xprt == NULL) { 289 dprintk("RPC: %s: couldn't allocate rpcrdma_xprt\n", 290 __func__); 291 return ERR_PTR(-ENOMEM); 292 } 293 294 /* 60 second timeout, no retries */ 295 xprt->timeout = &xprt_rdma_default_timeout; 296 xprt->bind_timeout = (60U * HZ); 297 xprt->reestablish_timeout = (5U * HZ); 298 xprt->idle_timeout = (5U * 60 * HZ); 299 300 xprt->resvport = 0; /* privileged port not needed */ 301 xprt->tsh_size = 0; /* RPC-RDMA handles framing */ 302 xprt->max_payload = RPCRDMA_MAX_DATA_SEGS * PAGE_SIZE; 303 xprt->ops = &xprt_rdma_procs; 304 305 /* 306 * Set up RDMA-specific connect data. 307 */ 308 309 /* Put server RDMA address in local cdata */ 310 memcpy(&cdata.addr, args->dstaddr, args->addrlen); 311 312 /* Ensure xprt->addr holds valid server TCP (not RDMA) 313 * address, for any side protocols which peek at it */ 314 xprt->prot = IPPROTO_TCP; 315 xprt->addrlen = args->addrlen; 316 memcpy(&xprt->addr, &cdata.addr, xprt->addrlen); 317 318 sin = (struct sockaddr_in *)&cdata.addr; 319 if (ntohs(sin->sin_port) != 0) 320 xprt_set_bound(xprt); 321 322 dprintk("RPC: %s: %pI4:%u\n", 323 __func__, &sin->sin_addr.s_addr, ntohs(sin->sin_port)); 324 325 /* Set max requests */ 326 cdata.max_requests = xprt->max_reqs; 327 328 /* Set some length limits */ 329 cdata.rsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA write max */ 330 cdata.wsize = RPCRDMA_MAX_SEGS * PAGE_SIZE; /* RDMA read max */ 331 332 cdata.inline_wsize = xprt_rdma_max_inline_write; 333 if (cdata.inline_wsize > cdata.wsize) 334 cdata.inline_wsize = cdata.wsize; 335 336 cdata.inline_rsize = xprt_rdma_max_inline_read; 337 if (cdata.inline_rsize > cdata.rsize) 338 cdata.inline_rsize = cdata.rsize; 339 340 cdata.padding = xprt_rdma_inline_write_padding; 341 342 /* 343 * Create new transport instance, which includes initialized 344 * o ia 345 * o endpoint 346 * o buffers 347 */ 348 349 new_xprt = rpcx_to_rdmax(xprt); 350 351 rc = rpcrdma_ia_open(new_xprt, (struct sockaddr *) &cdata.addr, 352 xprt_rdma_memreg_strategy); 353 if (rc) 354 goto out1; 355 356 /* 357 * initialize and create ep 358 */ 359 new_xprt->rx_data = cdata; 360 new_ep = &new_xprt->rx_ep; 361 new_ep->rep_remote_addr = cdata.addr; 362 363 rc = rpcrdma_ep_create(&new_xprt->rx_ep, 364 &new_xprt->rx_ia, &new_xprt->rx_data); 365 if (rc) 366 goto out2; 367 368 /* 369 * Allocate pre-registered send and receive buffers for headers and 370 * any inline data. Also specify any padding which will be provided 371 * from a preregistered zero buffer. 372 */ 373 rc = rpcrdma_buffer_create(&new_xprt->rx_buf, new_ep, &new_xprt->rx_ia, 374 &new_xprt->rx_data); 375 if (rc) 376 goto out3; 377 378 /* 379 * Register a callback for connection events. This is necessary because 380 * connection loss notification is async. We also catch connection loss 381 * when reaping receives. 382 */ 383 INIT_DELAYED_WORK(&new_xprt->rdma_connect, xprt_rdma_connect_worker); 384 new_ep->rep_func = rpcrdma_conn_func; 385 new_ep->rep_xprt = xprt; 386 387 xprt_rdma_format_addresses(xprt); 388 389 if (!try_module_get(THIS_MODULE)) 390 goto out4; 391 392 return xprt; 393 394 out4: 395 xprt_rdma_free_addresses(xprt); 396 rc = -EINVAL; 397 out3: 398 (void) rpcrdma_ep_destroy(new_ep, &new_xprt->rx_ia); 399 out2: 400 rpcrdma_ia_close(&new_xprt->rx_ia); 401 out1: 402 xprt_free(xprt); 403 return ERR_PTR(rc); 404 } 405 406 /* 407 * Close a connection, during shutdown or timeout/reconnect 408 */ 409 static void 410 xprt_rdma_close(struct rpc_xprt *xprt) 411 { 412 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 413 414 dprintk("RPC: %s: closing\n", __func__); 415 if (r_xprt->rx_ep.rep_connected > 0) 416 xprt->reestablish_timeout = 0; 417 xprt_disconnect_done(xprt); 418 (void) rpcrdma_ep_disconnect(&r_xprt->rx_ep, &r_xprt->rx_ia); 419 } 420 421 static void 422 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port) 423 { 424 struct sockaddr_in *sap; 425 426 sap = (struct sockaddr_in *)&xprt->addr; 427 sap->sin_port = htons(port); 428 sap = (struct sockaddr_in *)&rpcx_to_rdmad(xprt).addr; 429 sap->sin_port = htons(port); 430 dprintk("RPC: %s: %u\n", __func__, port); 431 } 432 433 static void 434 xprt_rdma_connect(struct rpc_task *task) 435 { 436 struct rpc_xprt *xprt = (struct rpc_xprt *)task->tk_xprt; 437 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 438 439 if (r_xprt->rx_ep.rep_connected != 0) { 440 /* Reconnect */ 441 schedule_delayed_work(&r_xprt->rdma_connect, 442 xprt->reestablish_timeout); 443 xprt->reestablish_timeout <<= 1; 444 if (xprt->reestablish_timeout > (30 * HZ)) 445 xprt->reestablish_timeout = (30 * HZ); 446 else if (xprt->reestablish_timeout < (5 * HZ)) 447 xprt->reestablish_timeout = (5 * HZ); 448 } else { 449 schedule_delayed_work(&r_xprt->rdma_connect, 0); 450 if (!RPC_IS_ASYNC(task)) 451 flush_delayed_work(&r_xprt->rdma_connect); 452 } 453 } 454 455 static int 456 xprt_rdma_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task) 457 { 458 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 459 int credits = atomic_read(&r_xprt->rx_buf.rb_credits); 460 461 /* == RPC_CWNDSCALE @ init, but *after* setup */ 462 if (r_xprt->rx_buf.rb_cwndscale == 0UL) { 463 r_xprt->rx_buf.rb_cwndscale = xprt->cwnd; 464 dprintk("RPC: %s: cwndscale %lu\n", __func__, 465 r_xprt->rx_buf.rb_cwndscale); 466 BUG_ON(r_xprt->rx_buf.rb_cwndscale <= 0); 467 } 468 xprt->cwnd = credits * r_xprt->rx_buf.rb_cwndscale; 469 return xprt_reserve_xprt_cong(xprt, task); 470 } 471 472 /* 473 * The RDMA allocate/free functions need the task structure as a place 474 * to hide the struct rpcrdma_req, which is necessary for the actual send/recv 475 * sequence. For this reason, the recv buffers are attached to send 476 * buffers for portions of the RPC. Note that the RPC layer allocates 477 * both send and receive buffers in the same call. We may register 478 * the receive buffer portion when using reply chunks. 479 */ 480 static void * 481 xprt_rdma_allocate(struct rpc_task *task, size_t size) 482 { 483 struct rpc_xprt *xprt = task->tk_xprt; 484 struct rpcrdma_req *req, *nreq; 485 486 req = rpcrdma_buffer_get(&rpcx_to_rdmax(xprt)->rx_buf); 487 BUG_ON(NULL == req); 488 489 if (size > req->rl_size) { 490 dprintk("RPC: %s: size %zd too large for buffer[%zd]: " 491 "prog %d vers %d proc %d\n", 492 __func__, size, req->rl_size, 493 task->tk_client->cl_prog, task->tk_client->cl_vers, 494 task->tk_msg.rpc_proc->p_proc); 495 /* 496 * Outgoing length shortage. Our inline write max must have 497 * been configured to perform direct i/o. 498 * 499 * This is therefore a large metadata operation, and the 500 * allocate call was made on the maximum possible message, 501 * e.g. containing long filename(s) or symlink data. In 502 * fact, while these metadata operations *might* carry 503 * large outgoing payloads, they rarely *do*. However, we 504 * have to commit to the request here, so reallocate and 505 * register it now. The data path will never require this 506 * reallocation. 507 * 508 * If the allocation or registration fails, the RPC framework 509 * will (doggedly) retry. 510 */ 511 if (rpcx_to_rdmax(xprt)->rx_ia.ri_memreg_strategy == 512 RPCRDMA_BOUNCEBUFFERS) { 513 /* forced to "pure inline" */ 514 dprintk("RPC: %s: too much data (%zd) for inline " 515 "(r/w max %d/%d)\n", __func__, size, 516 rpcx_to_rdmad(xprt).inline_rsize, 517 rpcx_to_rdmad(xprt).inline_wsize); 518 size = req->rl_size; 519 rpc_exit(task, -EIO); /* fail the operation */ 520 rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++; 521 goto out; 522 } 523 if (task->tk_flags & RPC_TASK_SWAPPER) 524 nreq = kmalloc(sizeof *req + size, GFP_ATOMIC); 525 else 526 nreq = kmalloc(sizeof *req + size, GFP_NOFS); 527 if (nreq == NULL) 528 goto outfail; 529 530 if (rpcrdma_register_internal(&rpcx_to_rdmax(xprt)->rx_ia, 531 nreq->rl_base, size + sizeof(struct rpcrdma_req) 532 - offsetof(struct rpcrdma_req, rl_base), 533 &nreq->rl_handle, &nreq->rl_iov)) { 534 kfree(nreq); 535 goto outfail; 536 } 537 rpcx_to_rdmax(xprt)->rx_stats.hardway_register_count += size; 538 nreq->rl_size = size; 539 nreq->rl_niovs = 0; 540 nreq->rl_nchunks = 0; 541 nreq->rl_buffer = (struct rpcrdma_buffer *)req; 542 nreq->rl_reply = req->rl_reply; 543 memcpy(nreq->rl_segments, 544 req->rl_segments, sizeof nreq->rl_segments); 545 /* flag the swap with an unused field */ 546 nreq->rl_iov.length = 0; 547 req->rl_reply = NULL; 548 req = nreq; 549 } 550 dprintk("RPC: %s: size %zd, request 0x%p\n", __func__, size, req); 551 out: 552 req->rl_connect_cookie = 0; /* our reserved value */ 553 return req->rl_xdr_buf; 554 555 outfail: 556 rpcrdma_buffer_put(req); 557 rpcx_to_rdmax(xprt)->rx_stats.failed_marshal_count++; 558 return NULL; 559 } 560 561 /* 562 * This function returns all RDMA resources to the pool. 563 */ 564 static void 565 xprt_rdma_free(void *buffer) 566 { 567 struct rpcrdma_req *req; 568 struct rpcrdma_xprt *r_xprt; 569 struct rpcrdma_rep *rep; 570 int i; 571 572 if (buffer == NULL) 573 return; 574 575 req = container_of(buffer, struct rpcrdma_req, rl_xdr_buf[0]); 576 if (req->rl_iov.length == 0) { /* see allocate above */ 577 r_xprt = container_of(((struct rpcrdma_req *) req->rl_buffer)->rl_buffer, 578 struct rpcrdma_xprt, rx_buf); 579 } else 580 r_xprt = container_of(req->rl_buffer, struct rpcrdma_xprt, rx_buf); 581 rep = req->rl_reply; 582 583 dprintk("RPC: %s: called on 0x%p%s\n", 584 __func__, rep, (rep && rep->rr_func) ? " (with waiter)" : ""); 585 586 /* 587 * Finish the deregistration. When using mw bind, this was 588 * begun in rpcrdma_reply_handler(). In all other modes, we 589 * do it here, in thread context. The process is considered 590 * complete when the rr_func vector becomes NULL - this 591 * was put in place during rpcrdma_reply_handler() - the wait 592 * call below will not block if the dereg is "done". If 593 * interrupted, our framework will clean up. 594 */ 595 for (i = 0; req->rl_nchunks;) { 596 --req->rl_nchunks; 597 i += rpcrdma_deregister_external( 598 &req->rl_segments[i], r_xprt, NULL); 599 } 600 601 if (rep && wait_event_interruptible(rep->rr_unbind, !rep->rr_func)) { 602 rep->rr_func = NULL; /* abandon the callback */ 603 req->rl_reply = NULL; 604 } 605 606 if (req->rl_iov.length == 0) { /* see allocate above */ 607 struct rpcrdma_req *oreq = (struct rpcrdma_req *)req->rl_buffer; 608 oreq->rl_reply = req->rl_reply; 609 (void) rpcrdma_deregister_internal(&r_xprt->rx_ia, 610 req->rl_handle, 611 &req->rl_iov); 612 kfree(req); 613 req = oreq; 614 } 615 616 /* Put back request+reply buffers */ 617 rpcrdma_buffer_put(req); 618 } 619 620 /* 621 * send_request invokes the meat of RPC RDMA. It must do the following: 622 * 1. Marshal the RPC request into an RPC RDMA request, which means 623 * putting a header in front of data, and creating IOVs for RDMA 624 * from those in the request. 625 * 2. In marshaling, detect opportunities for RDMA, and use them. 626 * 3. Post a recv message to set up asynch completion, then send 627 * the request (rpcrdma_ep_post). 628 * 4. No partial sends are possible in the RPC-RDMA protocol (as in UDP). 629 */ 630 631 static int 632 xprt_rdma_send_request(struct rpc_task *task) 633 { 634 struct rpc_rqst *rqst = task->tk_rqstp; 635 struct rpc_xprt *xprt = task->tk_xprt; 636 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 637 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 638 639 /* marshal the send itself */ 640 if (req->rl_niovs == 0 && rpcrdma_marshal_req(rqst) != 0) { 641 r_xprt->rx_stats.failed_marshal_count++; 642 dprintk("RPC: %s: rpcrdma_marshal_req failed\n", 643 __func__); 644 return -EIO; 645 } 646 647 if (req->rl_reply == NULL) /* e.g. reconnection */ 648 rpcrdma_recv_buffer_get(req); 649 650 if (req->rl_reply) { 651 req->rl_reply->rr_func = rpcrdma_reply_handler; 652 /* this need only be done once, but... */ 653 req->rl_reply->rr_xprt = xprt; 654 } 655 656 /* Must suppress retransmit to maintain credits */ 657 if (req->rl_connect_cookie == xprt->connect_cookie) 658 goto drop_connection; 659 req->rl_connect_cookie = xprt->connect_cookie; 660 661 if (rpcrdma_ep_post(&r_xprt->rx_ia, &r_xprt->rx_ep, req)) 662 goto drop_connection; 663 664 rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len; 665 rqst->rq_bytes_sent = 0; 666 return 0; 667 668 drop_connection: 669 xprt_disconnect_done(xprt); 670 return -ENOTCONN; /* implies disconnect */ 671 } 672 673 static void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq) 674 { 675 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 676 long idle_time = 0; 677 678 if (xprt_connected(xprt)) 679 idle_time = (long)(jiffies - xprt->last_used) / HZ; 680 681 seq_printf(seq, 682 "\txprt:\trdma %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu " 683 "%lu %lu %lu %Lu %Lu %Lu %Lu %lu %lu %lu\n", 684 685 0, /* need a local port? */ 686 xprt->stat.bind_count, 687 xprt->stat.connect_count, 688 xprt->stat.connect_time, 689 idle_time, 690 xprt->stat.sends, 691 xprt->stat.recvs, 692 xprt->stat.bad_xids, 693 xprt->stat.req_u, 694 xprt->stat.bklog_u, 695 696 r_xprt->rx_stats.read_chunk_count, 697 r_xprt->rx_stats.write_chunk_count, 698 r_xprt->rx_stats.reply_chunk_count, 699 r_xprt->rx_stats.total_rdma_request, 700 r_xprt->rx_stats.total_rdma_reply, 701 r_xprt->rx_stats.pullup_copy_count, 702 r_xprt->rx_stats.fixup_copy_count, 703 r_xprt->rx_stats.hardway_register_count, 704 r_xprt->rx_stats.failed_marshal_count, 705 r_xprt->rx_stats.bad_reply_count); 706 } 707 708 /* 709 * Plumbing for rpc transport switch and kernel module 710 */ 711 712 static struct rpc_xprt_ops xprt_rdma_procs = { 713 .reserve_xprt = xprt_rdma_reserve_xprt, 714 .release_xprt = xprt_release_xprt_cong, /* sunrpc/xprt.c */ 715 .release_request = xprt_release_rqst_cong, /* ditto */ 716 .set_retrans_timeout = xprt_set_retrans_timeout_def, /* ditto */ 717 .rpcbind = rpcb_getport_async, /* sunrpc/rpcb_clnt.c */ 718 .set_port = xprt_rdma_set_port, 719 .connect = xprt_rdma_connect, 720 .buf_alloc = xprt_rdma_allocate, 721 .buf_free = xprt_rdma_free, 722 .send_request = xprt_rdma_send_request, 723 .close = xprt_rdma_close, 724 .destroy = xprt_rdma_destroy, 725 .print_stats = xprt_rdma_print_stats 726 }; 727 728 static struct xprt_class xprt_rdma = { 729 .list = LIST_HEAD_INIT(xprt_rdma.list), 730 .name = "rdma", 731 .owner = THIS_MODULE, 732 .ident = XPRT_TRANSPORT_RDMA, 733 .setup = xprt_setup_rdma, 734 }; 735 736 static void __exit xprt_rdma_cleanup(void) 737 { 738 int rc; 739 740 dprintk(KERN_INFO "RPCRDMA Module Removed, deregister RPC RDMA transport\n"); 741 #ifdef RPC_DEBUG 742 if (sunrpc_table_header) { 743 unregister_sysctl_table(sunrpc_table_header); 744 sunrpc_table_header = NULL; 745 } 746 #endif 747 rc = xprt_unregister_transport(&xprt_rdma); 748 if (rc) 749 dprintk("RPC: %s: xprt_unregister returned %i\n", 750 __func__, rc); 751 } 752 753 static int __init xprt_rdma_init(void) 754 { 755 int rc; 756 757 rc = xprt_register_transport(&xprt_rdma); 758 759 if (rc) 760 return rc; 761 762 dprintk(KERN_INFO "RPCRDMA Module Init, register RPC RDMA transport\n"); 763 764 dprintk(KERN_INFO "Defaults:\n"); 765 dprintk(KERN_INFO "\tSlots %d\n" 766 "\tMaxInlineRead %d\n\tMaxInlineWrite %d\n", 767 xprt_rdma_slot_table_entries, 768 xprt_rdma_max_inline_read, xprt_rdma_max_inline_write); 769 dprintk(KERN_INFO "\tPadding %d\n\tMemreg %d\n", 770 xprt_rdma_inline_write_padding, xprt_rdma_memreg_strategy); 771 772 #ifdef RPC_DEBUG 773 if (!sunrpc_table_header) 774 sunrpc_table_header = register_sysctl_table(sunrpc_table); 775 #endif 776 return 0; 777 } 778 779 module_init(xprt_rdma_init); 780 module_exit(xprt_rdma_cleanup); 781