1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause 2 /* 3 * Copyright (c) 2014-2017 Oracle. All rights reserved. 4 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved. 5 * 6 * This software is available to you under a choice of one of two 7 * licenses. You may choose to be licensed under the terms of the GNU 8 * General Public License (GPL) Version 2, available from the file 9 * COPYING in the main directory of this source tree, or the BSD-type 10 * license below: 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 16 * Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 19 * Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials provided 22 * with the distribution. 23 * 24 * Neither the name of the Network Appliance, Inc. nor the names of 25 * its contributors may be used to endorse or promote products 26 * derived from this software without specific prior written 27 * permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 40 */ 41 42 /* 43 * transport.c 44 * 45 * This file contains the top-level implementation of an RPC RDMA 46 * transport. 47 * 48 * Naming convention: functions beginning with xprt_ are part of the 49 * transport switch. All others are RPC RDMA internal. 50 */ 51 52 #include <linux/module.h> 53 #include <linux/slab.h> 54 #include <linux/seq_file.h> 55 #include <linux/smp.h> 56 57 #include <linux/sunrpc/addr.h> 58 #include <linux/sunrpc/svc_rdma.h> 59 60 #include "xprt_rdma.h" 61 #include <trace/events/rpcrdma.h> 62 63 /* 64 * tunables 65 */ 66 67 static unsigned int xprt_rdma_slot_table_entries = RPCRDMA_DEF_SLOT_TABLE; 68 unsigned int xprt_rdma_max_inline_read = RPCRDMA_DEF_INLINE; 69 unsigned int xprt_rdma_max_inline_write = RPCRDMA_DEF_INLINE; 70 unsigned int xprt_rdma_memreg_strategy = RPCRDMA_FRWR; 71 int xprt_rdma_pad_optimize; 72 static struct xprt_class xprt_rdma; 73 74 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) 75 76 static unsigned int min_slot_table_size = RPCRDMA_MIN_SLOT_TABLE; 77 static unsigned int max_slot_table_size = RPCRDMA_MAX_SLOT_TABLE; 78 static unsigned int min_inline_size = RPCRDMA_MIN_INLINE; 79 static unsigned int max_inline_size = RPCRDMA_MAX_INLINE; 80 static unsigned int max_padding = PAGE_SIZE; 81 static unsigned int min_memreg = RPCRDMA_BOUNCEBUFFERS; 82 static unsigned int max_memreg = RPCRDMA_LAST - 1; 83 static unsigned int dummy; 84 85 static struct ctl_table_header *sunrpc_table_header; 86 87 static struct ctl_table xr_tunables_table[] = { 88 { 89 .procname = "rdma_slot_table_entries", 90 .data = &xprt_rdma_slot_table_entries, 91 .maxlen = sizeof(unsigned int), 92 .mode = 0644, 93 .proc_handler = proc_dointvec_minmax, 94 .extra1 = &min_slot_table_size, 95 .extra2 = &max_slot_table_size 96 }, 97 { 98 .procname = "rdma_max_inline_read", 99 .data = &xprt_rdma_max_inline_read, 100 .maxlen = sizeof(unsigned int), 101 .mode = 0644, 102 .proc_handler = proc_dointvec_minmax, 103 .extra1 = &min_inline_size, 104 .extra2 = &max_inline_size, 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_minmax, 112 .extra1 = &min_inline_size, 113 .extra2 = &max_inline_size, 114 }, 115 { 116 .procname = "rdma_inline_write_padding", 117 .data = &dummy, 118 .maxlen = sizeof(unsigned int), 119 .mode = 0644, 120 .proc_handler = proc_dointvec_minmax, 121 .extra1 = SYSCTL_ZERO, 122 .extra2 = &max_padding, 123 }, 124 { 125 .procname = "rdma_memreg_strategy", 126 .data = &xprt_rdma_memreg_strategy, 127 .maxlen = sizeof(unsigned int), 128 .mode = 0644, 129 .proc_handler = proc_dointvec_minmax, 130 .extra1 = &min_memreg, 131 .extra2 = &max_memreg, 132 }, 133 { 134 .procname = "rdma_pad_optimize", 135 .data = &xprt_rdma_pad_optimize, 136 .maxlen = sizeof(unsigned int), 137 .mode = 0644, 138 .proc_handler = proc_dointvec, 139 }, 140 }; 141 142 #endif 143 144 static const struct rpc_xprt_ops xprt_rdma_procs; 145 146 static void 147 xprt_rdma_format_addresses4(struct rpc_xprt *xprt, struct sockaddr *sap) 148 { 149 struct sockaddr_in *sin = (struct sockaddr_in *)sap; 150 char buf[20]; 151 152 snprintf(buf, sizeof(buf), "%08x", ntohl(sin->sin_addr.s_addr)); 153 xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL); 154 155 xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA; 156 } 157 158 static void 159 xprt_rdma_format_addresses6(struct rpc_xprt *xprt, struct sockaddr *sap) 160 { 161 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap; 162 char buf[40]; 163 164 snprintf(buf, sizeof(buf), "%pi6", &sin6->sin6_addr); 165 xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = kstrdup(buf, GFP_KERNEL); 166 167 xprt->address_strings[RPC_DISPLAY_NETID] = RPCBIND_NETID_RDMA6; 168 } 169 170 void 171 xprt_rdma_format_addresses(struct rpc_xprt *xprt, struct sockaddr *sap) 172 { 173 char buf[128]; 174 175 switch (sap->sa_family) { 176 case AF_INET: 177 xprt_rdma_format_addresses4(xprt, sap); 178 break; 179 case AF_INET6: 180 xprt_rdma_format_addresses6(xprt, sap); 181 break; 182 default: 183 pr_err("rpcrdma: Unrecognized address family\n"); 184 return; 185 } 186 187 (void)rpc_ntop(sap, buf, sizeof(buf)); 188 xprt->address_strings[RPC_DISPLAY_ADDR] = kstrdup(buf, GFP_KERNEL); 189 190 snprintf(buf, sizeof(buf), "%u", rpc_get_port(sap)); 191 xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL); 192 193 snprintf(buf, sizeof(buf), "%4hx", rpc_get_port(sap)); 194 xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL); 195 196 xprt->address_strings[RPC_DISPLAY_PROTO] = "rdma"; 197 } 198 199 void 200 xprt_rdma_free_addresses(struct rpc_xprt *xprt) 201 { 202 unsigned int i; 203 204 for (i = 0; i < RPC_DISPLAY_MAX; i++) 205 switch (i) { 206 case RPC_DISPLAY_PROTO: 207 case RPC_DISPLAY_NETID: 208 continue; 209 default: 210 kfree(xprt->address_strings[i]); 211 } 212 } 213 214 /** 215 * xprt_rdma_connect_worker - establish connection in the background 216 * @work: worker thread context 217 * 218 * Requester holds the xprt's send lock to prevent activity on this 219 * transport while a fresh connection is being established. RPC tasks 220 * sleep on the xprt's pending queue waiting for connect to complete. 221 */ 222 static void 223 xprt_rdma_connect_worker(struct work_struct *work) 224 { 225 struct rpcrdma_xprt *r_xprt = container_of(work, struct rpcrdma_xprt, 226 rx_connect_worker.work); 227 struct rpc_xprt *xprt = &r_xprt->rx_xprt; 228 unsigned int pflags = current->flags; 229 int rc; 230 231 if (atomic_read(&xprt->swapper)) 232 current->flags |= PF_MEMALLOC; 233 rc = rpcrdma_xprt_connect(r_xprt); 234 xprt_clear_connecting(xprt); 235 if (!rc) { 236 xprt->connect_cookie++; 237 xprt->stat.connect_count++; 238 xprt->stat.connect_time += (long)jiffies - 239 xprt->stat.connect_start; 240 xprt_set_connected(xprt); 241 rc = -EAGAIN; 242 } else 243 rpcrdma_xprt_disconnect(r_xprt); 244 xprt_unlock_connect(xprt, r_xprt); 245 xprt_wake_pending_tasks(xprt, rc); 246 current_restore_flags(pflags, PF_MEMALLOC); 247 } 248 249 /** 250 * xprt_rdma_inject_disconnect - inject a connection fault 251 * @xprt: transport context 252 * 253 * If @xprt is connected, disconnect it to simulate spurious 254 * connection loss. Caller must hold @xprt's send lock to 255 * ensure that data structures and hardware resources are 256 * stable during the rdma_disconnect() call. 257 */ 258 static void 259 xprt_rdma_inject_disconnect(struct rpc_xprt *xprt) 260 { 261 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 262 263 trace_xprtrdma_op_inject_dsc(r_xprt); 264 rdma_disconnect(r_xprt->rx_ep->re_id); 265 } 266 267 /** 268 * xprt_rdma_destroy - Full tear down of transport 269 * @xprt: doomed transport context 270 * 271 * Caller guarantees there will be no more calls to us with 272 * this @xprt. 273 */ 274 static void 275 xprt_rdma_destroy(struct rpc_xprt *xprt) 276 { 277 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 278 279 cancel_delayed_work_sync(&r_xprt->rx_connect_worker); 280 281 rpcrdma_xprt_disconnect(r_xprt); 282 283 /* The disconnect's sendctx drain can return bc_prealloc reqs 284 * to bc_pa_list after xprt_destroy_backchannel() emptied it. 285 */ 286 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 287 xprt_rdma_bc_destroy(xprt, 0); 288 #endif 289 rpcrdma_buffer_destroy(&r_xprt->rx_buf); 290 291 xprt_rdma_free_addresses(xprt); 292 xprt_free(xprt); 293 294 module_put(THIS_MODULE); 295 } 296 297 /* 60 second timeout, no retries */ 298 static const struct rpc_timeout xprt_rdma_default_timeout = { 299 .to_initval = 60 * HZ, 300 .to_maxval = 60 * HZ, 301 }; 302 303 /** 304 * xprt_setup_rdma - Set up transport to use RDMA 305 * 306 * @args: rpc transport arguments 307 */ 308 static struct rpc_xprt * 309 xprt_setup_rdma(struct xprt_create *args) 310 { 311 struct rpc_xprt *xprt; 312 struct rpcrdma_xprt *new_xprt; 313 struct sockaddr *sap; 314 int rc; 315 316 if (args->addrlen > sizeof(xprt->addr)) 317 return ERR_PTR(-EBADF); 318 319 if (!try_module_get(THIS_MODULE)) 320 return ERR_PTR(-EIO); 321 322 xprt = xprt_alloc(args->net, sizeof(struct rpcrdma_xprt), 0, 323 xprt_rdma_slot_table_entries); 324 if (!xprt) { 325 module_put(THIS_MODULE); 326 return ERR_PTR(-ENOMEM); 327 } 328 329 xprt->timeout = &xprt_rdma_default_timeout; 330 xprt->connect_timeout = xprt->timeout->to_initval; 331 xprt->max_reconnect_timeout = xprt->timeout->to_maxval; 332 xprt->bind_timeout = RPCRDMA_BIND_TO; 333 xprt->reestablish_timeout = RPCRDMA_INIT_REEST_TO; 334 xprt->idle_timeout = RPCRDMA_IDLE_DISC_TO; 335 336 xprt->resvport = 0; /* privileged port not needed */ 337 xprt->ops = &xprt_rdma_procs; 338 339 /* 340 * Set up RDMA-specific connect data. 341 */ 342 sap = args->dstaddr; 343 344 /* Ensure xprt->addr holds valid server TCP (not RDMA) 345 * address, for any side protocols which peek at it */ 346 xprt->prot = IPPROTO_TCP; 347 xprt->xprt_class = &xprt_rdma; 348 xprt->addrlen = args->addrlen; 349 memcpy(&xprt->addr, sap, xprt->addrlen); 350 351 if (rpc_get_port(sap)) 352 xprt_set_bound(xprt); 353 xprt_rdma_format_addresses(xprt, sap); 354 355 new_xprt = rpcx_to_rdmax(xprt); 356 rc = rpcrdma_buffer_create(new_xprt); 357 if (rc) { 358 xprt_rdma_free_addresses(xprt); 359 xprt_free(xprt); 360 module_put(THIS_MODULE); 361 return ERR_PTR(rc); 362 } 363 364 INIT_DELAYED_WORK(&new_xprt->rx_connect_worker, 365 xprt_rdma_connect_worker); 366 367 xprt->max_payload = RPCRDMA_MAX_DATA_SEGS << PAGE_SHIFT; 368 369 return xprt; 370 } 371 372 /** 373 * xprt_rdma_close - close a transport connection 374 * @xprt: transport context 375 * 376 * Called during autoclose or device removal. 377 * 378 * Caller holds @xprt's send lock to prevent activity on this 379 * transport while the connection is torn down. 380 */ 381 void xprt_rdma_close(struct rpc_xprt *xprt) 382 { 383 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 384 385 rpcrdma_xprt_disconnect(r_xprt); 386 387 xprt->reestablish_timeout = 0; 388 ++xprt->connect_cookie; 389 xprt_disconnect_done(xprt); 390 } 391 392 /** 393 * xprt_rdma_set_port - update server port with rpcbind result 394 * @xprt: controlling RPC transport 395 * @port: new port value 396 * 397 * Transport connect status is unchanged. 398 */ 399 static void 400 xprt_rdma_set_port(struct rpc_xprt *xprt, u16 port) 401 { 402 struct sockaddr *sap = (struct sockaddr *)&xprt->addr; 403 char buf[8]; 404 405 rpc_set_port(sap, port); 406 407 kfree(xprt->address_strings[RPC_DISPLAY_PORT]); 408 snprintf(buf, sizeof(buf), "%u", port); 409 xprt->address_strings[RPC_DISPLAY_PORT] = kstrdup(buf, GFP_KERNEL); 410 411 kfree(xprt->address_strings[RPC_DISPLAY_HEX_PORT]); 412 snprintf(buf, sizeof(buf), "%4hx", port); 413 xprt->address_strings[RPC_DISPLAY_HEX_PORT] = kstrdup(buf, GFP_KERNEL); 414 } 415 416 /** 417 * xprt_rdma_timer - invoked when an RPC times out 418 * @xprt: controlling RPC transport 419 * @task: RPC task that timed out 420 * 421 * Invoked when the transport is still connected, but an RPC 422 * retransmit timeout occurs. 423 * 424 * Since RDMA connections don't have a keep-alive, forcibly 425 * disconnect and retry to connect. This drives full 426 * detection of the network path, and retransmissions of 427 * all pending RPCs. 428 */ 429 static void 430 xprt_rdma_timer(struct rpc_xprt *xprt, struct rpc_task *task) 431 { 432 xprt_force_disconnect(xprt); 433 } 434 435 /** 436 * xprt_rdma_set_connect_timeout - set timeouts for establishing a connection 437 * @xprt: controlling transport instance 438 * @connect_timeout: reconnect timeout after client disconnects 439 * @reconnect_timeout: reconnect timeout after server disconnects 440 * 441 */ 442 static void xprt_rdma_set_connect_timeout(struct rpc_xprt *xprt, 443 unsigned long connect_timeout, 444 unsigned long reconnect_timeout) 445 { 446 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 447 448 trace_xprtrdma_op_set_cto(r_xprt, connect_timeout, reconnect_timeout); 449 450 spin_lock(&xprt->transport_lock); 451 452 if (connect_timeout < xprt->connect_timeout) { 453 struct rpc_timeout to; 454 unsigned long initval; 455 456 to = *xprt->timeout; 457 initval = connect_timeout; 458 if (initval < RPCRDMA_INIT_REEST_TO << 1) 459 initval = RPCRDMA_INIT_REEST_TO << 1; 460 to.to_initval = initval; 461 to.to_maxval = initval; 462 r_xprt->rx_timeout = to; 463 xprt->timeout = &r_xprt->rx_timeout; 464 xprt->connect_timeout = connect_timeout; 465 } 466 467 if (reconnect_timeout < xprt->max_reconnect_timeout) 468 xprt->max_reconnect_timeout = reconnect_timeout; 469 470 spin_unlock(&xprt->transport_lock); 471 } 472 473 /** 474 * xprt_rdma_connect - schedule an attempt to reconnect 475 * @xprt: transport state 476 * @task: RPC scheduler context (unused) 477 * 478 */ 479 static void 480 xprt_rdma_connect(struct rpc_xprt *xprt, struct rpc_task *task) 481 { 482 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 483 struct rpcrdma_ep *ep = r_xprt->rx_ep; 484 unsigned long delay; 485 486 WARN_ON_ONCE(!xprt_lock_connect(xprt, task, r_xprt)); 487 488 delay = 0; 489 if (ep && ep->re_connect_status != 0) { 490 delay = xprt_reconnect_delay(xprt); 491 xprt_reconnect_backoff(xprt, RPCRDMA_INIT_REEST_TO); 492 } 493 trace_xprtrdma_op_connect(r_xprt, delay); 494 queue_delayed_work(system_dfl_long_wq, &r_xprt->rx_connect_worker, 495 delay); 496 } 497 498 /* rl_kref has two owners while a Send is outstanding: the rpc_rqst 499 * owner and the sendctx. Replies complete the RPC but do not drop 500 * either reference. The req returns to its free pool only after 501 * xprt_rdma_free_slot() or xprt_rdma_bc_free_rqst() has dropped the 502 * RPC-layer reference and rpcrdma_sendctx_unmap() has dropped the 503 * Send-side reference. 504 */ 505 static void rpcrdma_req_release(struct kref *kref) 506 { 507 struct rpcrdma_req *req = 508 container_of(kref, struct rpcrdma_req, rl_kref); 509 struct rpc_rqst *rqst = &req->rl_slot; 510 struct rpc_xprt *xprt = rqst->rq_xprt; 511 struct rpcrdma_xprt *r_xprt; 512 513 /* I4: both the RPC-layer and Send-side owners have dropped, 514 * so rpcrdma_sendctx_unmap() has cleared rl_sendctx. 515 */ 516 WARN_ON_ONCE(req->rl_sendctx); 517 518 kref_init(&req->rl_kref); 519 520 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 521 if (bc_prealloc(rqst)) { 522 spin_lock(&xprt->bc_pa_lock); 523 list_add_tail(&rqst->rq_bc_pa_list, &xprt->bc_pa_list); 524 spin_unlock(&xprt->bc_pa_lock); 525 return; 526 } 527 #endif 528 529 if (xprt_wake_up_backlog(xprt, rqst)) 530 return; 531 532 r_xprt = rpcx_to_rdmax(xprt); 533 memset(rqst, 0, sizeof(*rqst)); 534 rpcrdma_buffer_put(&r_xprt->rx_buf, req); 535 } 536 537 void rpcrdma_req_put(struct rpcrdma_req *req) 538 { 539 kref_put(&req->rl_kref, rpcrdma_req_release); 540 } 541 542 /** 543 * xprt_rdma_alloc_slot - allocate an rpc_rqst 544 * @xprt: controlling RPC transport 545 * @task: RPC task requesting a fresh rpc_rqst 546 * 547 * tk_status values: 548 * %0 if task->tk_rqstp points to a fresh rpc_rqst 549 * %-EAGAIN if no rpc_rqst is available; queued on backlog 550 */ 551 static void 552 xprt_rdma_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task) 553 { 554 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 555 struct rpcrdma_req *req; 556 557 req = rpcrdma_buffer_get(&r_xprt->rx_buf); 558 if (!req) 559 goto out_sleep; 560 kref_init(&req->rl_kref); 561 task->tk_rqstp = &req->rl_slot; 562 task->tk_status = 0; 563 return; 564 565 out_sleep: 566 task->tk_status = -EAGAIN; 567 xprt_add_backlog_noncongested(xprt, task); 568 /* A buffer freed between buffer_get and rpc_sleep_on 569 * goes back to the pool with no waiter to wake. 570 * Re-check after joining the backlog to close that gap. 571 */ 572 req = rpcrdma_buffer_get(&r_xprt->rx_buf); 573 if (req) { 574 struct rpc_rqst *rqst = &req->rl_slot; 575 576 kref_init(&req->rl_kref); 577 if (!xprt_wake_up_backlog(xprt, rqst)) { 578 memset(rqst, 0, sizeof(*rqst)); 579 rpcrdma_buffer_put(&r_xprt->rx_buf, req); 580 } 581 } 582 } 583 584 /** 585 * xprt_rdma_free_slot - release an rpc_rqst 586 * @xprt: controlling RPC transport 587 * @rqst: rpc_rqst to release 588 * 589 */ 590 static void 591 xprt_rdma_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *rqst) 592 { 593 struct rpcrdma_xprt *r_xprt = 594 container_of(xprt, struct rpcrdma_xprt, rx_xprt); 595 596 rpcrdma_reply_put(&r_xprt->rx_buf, rpcr_to_rdmar(rqst)); 597 rpcrdma_req_put(rpcr_to_rdmar(rqst)); 598 } 599 600 static bool rpcrdma_check_regbuf(struct rpcrdma_xprt *r_xprt, 601 struct rpcrdma_regbuf *rb, size_t size, 602 gfp_t flags) 603 { 604 if (unlikely(rdmab_length(rb) < size)) { 605 if (!rpcrdma_regbuf_realloc(rb, size, flags)) 606 return false; 607 r_xprt->rx_stats.hardway_register_count += size; 608 } 609 return true; 610 } 611 612 /** 613 * xprt_rdma_allocate - allocate transport resources for an RPC 614 * @task: RPC task 615 * 616 * Return values: 617 * 0: Success; rq_buffer points to RPC buffer to use 618 * ENOMEM: Out of memory, call again later 619 * EIO: A permanent error occurred, do not retry 620 */ 621 static int 622 xprt_rdma_allocate(struct rpc_task *task) 623 { 624 struct rpc_rqst *rqst = task->tk_rqstp; 625 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(rqst->rq_xprt); 626 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 627 gfp_t flags = rpc_task_gfp_mask(); 628 629 if (!rpcrdma_check_regbuf(r_xprt, req->rl_sendbuf, rqst->rq_callsize, 630 flags)) 631 goto out_fail; 632 if (!rpcrdma_check_regbuf(r_xprt, req->rl_recvbuf, rqst->rq_rcvsize, 633 flags)) 634 goto out_fail; 635 636 rqst->rq_buffer = rdmab_data(req->rl_sendbuf); 637 rqst->rq_rbuffer = rdmab_data(req->rl_recvbuf); 638 return 0; 639 640 out_fail: 641 return -ENOMEM; 642 } 643 644 /** 645 * xprt_rdma_free - release resources allocated by xprt_rdma_allocate 646 * @task: RPC task 647 * 648 * Caller guarantees rqst->rq_buffer is non-NULL. 649 */ 650 static void 651 xprt_rdma_free(struct rpc_task *task) 652 { 653 struct rpc_rqst *rqst = task->tk_rqstp; 654 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 655 656 if (unlikely(!list_empty(&req->rl_registered))) { 657 trace_xprtrdma_mrs_zap(task); 658 frwr_unmap_sync(rpcx_to_rdmax(rqst->rq_xprt), req); 659 } 660 661 /* The Send-side rl_kref owner keeps req out of its free pool 662 * until rpcrdma_sendctx_unmap() has fired -- see I4 above 663 * rpcrdma_reply_handler() -- so signal-driven release here 664 * does not let the HCA touch a recycled send buffer. 665 */ 666 } 667 668 /** 669 * xprt_rdma_send_request - marshal and send an RPC request 670 * @rqst: RPC message in rq_snd_buf 671 * 672 * Caller holds the transport's write lock. 673 * 674 * Returns: 675 * %0 if the RPC message has been sent 676 * %-ENOTCONN if the caller should reconnect and call again 677 * %-EAGAIN if the caller should call again 678 * %-ENOBUFS if the caller should call again after a delay 679 * %-EMSGSIZE if encoding ran out of buffer space. The request 680 * was not sent. Do not try to send this message again. 681 * %-EIO if an I/O error occurred. The request was not sent. 682 * Do not try to send this message again. 683 */ 684 static int 685 xprt_rdma_send_request(struct rpc_rqst *rqst) 686 { 687 struct rpc_xprt *xprt = rqst->rq_xprt; 688 struct rpcrdma_req *req = rpcr_to_rdmar(rqst); 689 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 690 int rc = 0; 691 692 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 693 if (unlikely(!rqst->rq_buffer)) 694 return xprt_rdma_bc_send_reply(rqst); 695 #endif /* CONFIG_SUNRPC_BACKCHANNEL */ 696 697 if (!xprt_connected(xprt)) 698 return -ENOTCONN; 699 700 if (!xprt_request_get_cong(xprt, rqst)) 701 return -EBADSLT; 702 703 rc = rpcrdma_marshal_req(r_xprt, rqst); 704 if (rc < 0) 705 goto failed_marshal; 706 707 /* Must suppress retransmit to maintain credits */ 708 if (rqst->rq_connect_cookie == xprt->connect_cookie) 709 goto drop_connection; 710 rqst->rq_xtime = ktime_get(); 711 712 if (frwr_send(r_xprt, req)) 713 goto drop_connection; 714 715 rqst->rq_xmit_bytes_sent += rqst->rq_snd_buf.len; 716 717 /* An RPC with no reply will throw off credit accounting, 718 * so drop the connection to reset the credit grant. 719 */ 720 if (!rpc_reply_expected(rqst->rq_task)) 721 goto drop_connection; 722 return 0; 723 724 failed_marshal: 725 if (rc != -ENOTCONN) 726 return rc; 727 drop_connection: 728 xprt_rdma_close(xprt); 729 return -ENOTCONN; 730 } 731 732 void xprt_rdma_print_stats(struct rpc_xprt *xprt, struct seq_file *seq) 733 { 734 struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt); 735 long idle_time = 0; 736 737 if (xprt_connected(xprt)) 738 idle_time = (long)(jiffies - xprt->last_used) / HZ; 739 740 seq_puts(seq, "\txprt:\trdma "); 741 seq_printf(seq, "%u %lu %lu %lu %ld %lu %lu %lu %llu %llu ", 742 0, /* need a local port? */ 743 xprt->stat.bind_count, 744 xprt->stat.connect_count, 745 xprt->stat.connect_time / HZ, 746 idle_time, 747 xprt->stat.sends, 748 xprt->stat.recvs, 749 xprt->stat.bad_xids, 750 xprt->stat.req_u, 751 xprt->stat.bklog_u); 752 seq_printf(seq, "%lu %lu %lu %llu %llu %llu %llu %lu %lu %lu %lu ", 753 r_xprt->rx_stats.read_chunk_count, 754 r_xprt->rx_stats.write_chunk_count, 755 r_xprt->rx_stats.reply_chunk_count, 756 r_xprt->rx_stats.total_rdma_request, 757 r_xprt->rx_stats.total_rdma_reply, 758 r_xprt->rx_stats.pullup_copy_count, 759 r_xprt->rx_stats.fixup_copy_count, 760 r_xprt->rx_stats.hardway_register_count, 761 r_xprt->rx_stats.failed_marshal_count, 762 r_xprt->rx_stats.bad_reply_count, 763 r_xprt->rx_stats.nomsg_call_count); 764 seq_printf(seq, "%lu %lu %lu %lu %lu %lu\n", 765 r_xprt->rx_stats.mrs_recycled, 766 r_xprt->rx_stats.mrs_orphaned, 767 r_xprt->rx_stats.mrs_allocated, 768 r_xprt->rx_stats.local_inv_needed, 769 r_xprt->rx_stats.empty_sendctx_q, 770 0LU); /* was reply_waits_for_send; column preserved */ 771 } 772 773 static int 774 xprt_rdma_enable_swap(struct rpc_xprt *xprt) 775 { 776 return 0; 777 } 778 779 static void 780 xprt_rdma_disable_swap(struct rpc_xprt *xprt) 781 { 782 } 783 784 /* 785 * Plumbing for rpc transport switch and kernel module 786 */ 787 788 static const struct rpc_xprt_ops xprt_rdma_procs = { 789 .reserve_xprt = xprt_reserve_xprt_cong, 790 .release_xprt = xprt_release_xprt_cong, /* sunrpc/xprt.c */ 791 .alloc_slot = xprt_rdma_alloc_slot, 792 .free_slot = xprt_rdma_free_slot, 793 .release_request = xprt_release_rqst_cong, /* ditto */ 794 .wait_for_reply_request = xprt_wait_for_reply_request_def, /* ditto */ 795 .timer = xprt_rdma_timer, 796 .rpcbind = rpcb_getport_async, /* sunrpc/rpcb_clnt.c */ 797 .set_port = xprt_rdma_set_port, 798 .connect = xprt_rdma_connect, 799 .buf_alloc = xprt_rdma_allocate, 800 .buf_free = xprt_rdma_free, 801 .send_request = xprt_rdma_send_request, 802 .close = xprt_rdma_close, 803 .destroy = xprt_rdma_destroy, 804 .set_connect_timeout = xprt_rdma_set_connect_timeout, 805 .print_stats = xprt_rdma_print_stats, 806 .enable_swap = xprt_rdma_enable_swap, 807 .disable_swap = xprt_rdma_disable_swap, 808 .inject_disconnect = xprt_rdma_inject_disconnect, 809 #if defined(CONFIG_SUNRPC_BACKCHANNEL) 810 .bc_setup = xprt_rdma_bc_setup, 811 .bc_maxpayload = xprt_rdma_bc_maxpayload, 812 .bc_num_slots = xprt_rdma_bc_max_slots, 813 .bc_free_rqst = xprt_rdma_bc_free_rqst, 814 .bc_destroy = xprt_rdma_bc_destroy, 815 #endif 816 }; 817 818 static struct xprt_class xprt_rdma = { 819 .list = LIST_HEAD_INIT(xprt_rdma.list), 820 .name = "rdma", 821 .owner = THIS_MODULE, 822 .ident = XPRT_TRANSPORT_RDMA, 823 .setup = xprt_setup_rdma, 824 .netid = { "rdma", "rdma6", "" }, 825 }; 826 827 void xprt_rdma_cleanup(void) 828 { 829 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) 830 if (sunrpc_table_header) { 831 unregister_sysctl_table(sunrpc_table_header); 832 sunrpc_table_header = NULL; 833 } 834 #endif 835 836 xprt_unregister_transport(&xprt_rdma); 837 xprt_unregister_transport(&xprt_rdma_bc); 838 } 839 840 int xprt_rdma_init(void) 841 { 842 int rc; 843 844 rc = xprt_register_transport(&xprt_rdma); 845 if (rc) 846 return rc; 847 848 rc = xprt_register_transport(&xprt_rdma_bc); 849 if (rc) { 850 xprt_unregister_transport(&xprt_rdma); 851 return rc; 852 } 853 854 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) 855 if (!sunrpc_table_header) 856 sunrpc_table_header = register_sysctl("sunrpc", xr_tunables_table); 857 #endif 858 return 0; 859 } 860