xprt.c revision 21de0a955f3af29fa1100d96f66e6adade89e77a
1/* 2 * linux/net/sunrpc/xprt.c 3 * 4 * This is a generic RPC call interface supporting congestion avoidance, 5 * and asynchronous calls. 6 * 7 * The interface works like this: 8 * 9 * - When a process places a call, it allocates a request slot if 10 * one is available. Otherwise, it sleeps on the backlog queue 11 * (xprt_reserve). 12 * - Next, the caller puts together the RPC message, stuffs it into 13 * the request struct, and calls xprt_transmit(). 14 * - xprt_transmit sends the message and installs the caller on the 15 * transport's wait list. At the same time, if a reply is expected, 16 * it installs a timer that is run after the packet's timeout has 17 * expired. 18 * - When a packet arrives, the data_ready handler walks the list of 19 * pending requests for that transport. If a matching XID is found, the 20 * caller is woken up, and the timer removed. 21 * - When no reply arrives within the timeout interval, the timer is 22 * fired by the kernel and runs xprt_timer(). It either adjusts the 23 * timeout values (minor timeout) or wakes up the caller with a status 24 * of -ETIMEDOUT. 25 * - When the caller receives a notification from RPC that a reply arrived, 26 * it should release the RPC slot, and process the reply. 27 * If the call timed out, it may choose to retry the operation by 28 * adjusting the initial timeout value, and simply calling rpc_call 29 * again. 30 * 31 * Support for async RPC is done through a set of RPC-specific scheduling 32 * primitives that `transparently' work for processes as well as async 33 * tasks that rely on callbacks. 34 * 35 * Copyright (C) 1995-1997, Olaf Kirch <okir@monad.swb.de> 36 * 37 * Transport switch API copyright (C) 2005, Chuck Lever <cel@netapp.com> 38 */ 39 40#include <linux/module.h> 41 42#include <linux/types.h> 43#include <linux/interrupt.h> 44#include <linux/workqueue.h> 45#include <linux/net.h> 46#include <linux/ktime.h> 47 48#include <linux/sunrpc/clnt.h> 49#include <linux/sunrpc/metrics.h> 50#include <linux/sunrpc/bc_xprt.h> 51 52#include "sunrpc.h" 53 54/* 55 * Local variables 56 */ 57 58#ifdef RPC_DEBUG 59# define RPCDBG_FACILITY RPCDBG_XPRT 60#endif 61 62/* 63 * Local functions 64 */ 65static void xprt_init(struct rpc_xprt *xprt, struct net *net); 66static void xprt_request_init(struct rpc_task *, struct rpc_xprt *); 67static void xprt_connect_status(struct rpc_task *task); 68static int __xprt_get_cong(struct rpc_xprt *, struct rpc_task *); 69 70static DEFINE_SPINLOCK(xprt_list_lock); 71static LIST_HEAD(xprt_list); 72 73/* 74 * The transport code maintains an estimate on the maximum number of out- 75 * standing RPC requests, using a smoothed version of the congestion 76 * avoidance implemented in 44BSD. This is basically the Van Jacobson 77 * congestion algorithm: If a retransmit occurs, the congestion window is 78 * halved; otherwise, it is incremented by 1/cwnd when 79 * 80 * - a reply is received and 81 * - a full number of requests are outstanding and 82 * - the congestion window hasn't been updated recently. 83 */ 84#define RPC_CWNDSHIFT (8U) 85#define RPC_CWNDSCALE (1U << RPC_CWNDSHIFT) 86#define RPC_INITCWND RPC_CWNDSCALE 87#define RPC_MAXCWND(xprt) ((xprt)->max_reqs << RPC_CWNDSHIFT) 88 89#define RPCXPRT_CONGESTED(xprt) ((xprt)->cong >= (xprt)->cwnd) 90 91/** 92 * xprt_register_transport - register a transport implementation 93 * @transport: transport to register 94 * 95 * If a transport implementation is loaded as a kernel module, it can 96 * call this interface to make itself known to the RPC client. 97 * 98 * Returns: 99 * 0: transport successfully registered 100 * -EEXIST: transport already registered 101 * -EINVAL: transport module being unloaded 102 */ 103int xprt_register_transport(struct xprt_class *transport) 104{ 105 struct xprt_class *t; 106 int result; 107 108 result = -EEXIST; 109 spin_lock(&xprt_list_lock); 110 list_for_each_entry(t, &xprt_list, list) { 111 /* don't register the same transport class twice */ 112 if (t->ident == transport->ident) 113 goto out; 114 } 115 116 list_add_tail(&transport->list, &xprt_list); 117 printk(KERN_INFO "RPC: Registered %s transport module.\n", 118 transport->name); 119 result = 0; 120 121out: 122 spin_unlock(&xprt_list_lock); 123 return result; 124} 125EXPORT_SYMBOL_GPL(xprt_register_transport); 126 127/** 128 * xprt_unregister_transport - unregister a transport implementation 129 * @transport: transport to unregister 130 * 131 * Returns: 132 * 0: transport successfully unregistered 133 * -ENOENT: transport never registered 134 */ 135int xprt_unregister_transport(struct xprt_class *transport) 136{ 137 struct xprt_class *t; 138 int result; 139 140 result = 0; 141 spin_lock(&xprt_list_lock); 142 list_for_each_entry(t, &xprt_list, list) { 143 if (t == transport) { 144 printk(KERN_INFO 145 "RPC: Unregistered %s transport module.\n", 146 transport->name); 147 list_del_init(&transport->list); 148 goto out; 149 } 150 } 151 result = -ENOENT; 152 153out: 154 spin_unlock(&xprt_list_lock); 155 return result; 156} 157EXPORT_SYMBOL_GPL(xprt_unregister_transport); 158 159/** 160 * xprt_load_transport - load a transport implementation 161 * @transport_name: transport to load 162 * 163 * Returns: 164 * 0: transport successfully loaded 165 * -ENOENT: transport module not available 166 */ 167int xprt_load_transport(const char *transport_name) 168{ 169 struct xprt_class *t; 170 int result; 171 172 result = 0; 173 spin_lock(&xprt_list_lock); 174 list_for_each_entry(t, &xprt_list, list) { 175 if (strcmp(t->name, transport_name) == 0) { 176 spin_unlock(&xprt_list_lock); 177 goto out; 178 } 179 } 180 spin_unlock(&xprt_list_lock); 181 result = request_module("xprt%s", transport_name); 182out: 183 return result; 184} 185EXPORT_SYMBOL_GPL(xprt_load_transport); 186 187/** 188 * xprt_reserve_xprt - serialize write access to transports 189 * @task: task that is requesting access to the transport 190 * 191 * This prevents mixing the payload of separate requests, and prevents 192 * transport connects from colliding with writes. No congestion control 193 * is provided. 194 */ 195int xprt_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task) 196{ 197 struct rpc_rqst *req = task->tk_rqstp; 198 199 if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) { 200 if (task == xprt->snd_task) 201 return 1; 202 goto out_sleep; 203 } 204 xprt->snd_task = task; 205 if (req != NULL) { 206 req->rq_bytes_sent = 0; 207 req->rq_ntrans++; 208 } 209 210 return 1; 211 212out_sleep: 213 dprintk("RPC: %5u failed to lock transport %p\n", 214 task->tk_pid, xprt); 215 task->tk_timeout = 0; 216 task->tk_status = -EAGAIN; 217 if (req != NULL && req->rq_ntrans) 218 rpc_sleep_on(&xprt->resend, task, NULL); 219 else 220 rpc_sleep_on(&xprt->sending, task, NULL); 221 return 0; 222} 223EXPORT_SYMBOL_GPL(xprt_reserve_xprt); 224 225static void xprt_clear_locked(struct rpc_xprt *xprt) 226{ 227 xprt->snd_task = NULL; 228 if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state) || xprt->shutdown) { 229 smp_mb__before_clear_bit(); 230 clear_bit(XPRT_LOCKED, &xprt->state); 231 smp_mb__after_clear_bit(); 232 } else 233 queue_work(rpciod_workqueue, &xprt->task_cleanup); 234} 235 236/* 237 * xprt_reserve_xprt_cong - serialize write access to transports 238 * @task: task that is requesting access to the transport 239 * 240 * Same as xprt_reserve_xprt, but Van Jacobson congestion control is 241 * integrated into the decision of whether a request is allowed to be 242 * woken up and given access to the transport. 243 */ 244int xprt_reserve_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task) 245{ 246 struct rpc_rqst *req = task->tk_rqstp; 247 248 if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) { 249 if (task == xprt->snd_task) 250 return 1; 251 goto out_sleep; 252 } 253 if (req == NULL) { 254 xprt->snd_task = task; 255 return 1; 256 } 257 if (__xprt_get_cong(xprt, task)) { 258 xprt->snd_task = task; 259 req->rq_bytes_sent = 0; 260 req->rq_ntrans++; 261 return 1; 262 } 263 xprt_clear_locked(xprt); 264out_sleep: 265 dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt); 266 task->tk_timeout = 0; 267 task->tk_status = -EAGAIN; 268 if (req != NULL && req->rq_ntrans) 269 rpc_sleep_on(&xprt->resend, task, NULL); 270 else 271 rpc_sleep_on(&xprt->sending, task, NULL); 272 return 0; 273} 274EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong); 275 276static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task) 277{ 278 int retval; 279 280 spin_lock_bh(&xprt->transport_lock); 281 retval = xprt->ops->reserve_xprt(xprt, task); 282 spin_unlock_bh(&xprt->transport_lock); 283 return retval; 284} 285 286static void __xprt_lock_write_next(struct rpc_xprt *xprt) 287{ 288 struct rpc_task *task; 289 struct rpc_rqst *req; 290 291 if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) 292 return; 293 294 task = rpc_wake_up_next(&xprt->resend); 295 if (!task) { 296 task = rpc_wake_up_next(&xprt->sending); 297 if (task == NULL) 298 goto out_unlock; 299 } 300 301 req = task->tk_rqstp; 302 xprt->snd_task = task; 303 if (req) { 304 req->rq_bytes_sent = 0; 305 req->rq_ntrans++; 306 } 307 return; 308 309out_unlock: 310 xprt_clear_locked(xprt); 311} 312 313static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt) 314{ 315 struct rpc_task *task; 316 struct rpc_rqst *req; 317 318 if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) 319 return; 320 if (RPCXPRT_CONGESTED(xprt)) 321 goto out_unlock; 322 task = rpc_wake_up_next(&xprt->resend); 323 if (!task) { 324 task = rpc_wake_up_next(&xprt->sending); 325 if (task == NULL) 326 goto out_unlock; 327 } 328 329 req = task->tk_rqstp; 330 if (req == NULL) { 331 xprt->snd_task = task; 332 return; 333 } 334 if (__xprt_get_cong(xprt, task)) { 335 xprt->snd_task = task; 336 req->rq_bytes_sent = 0; 337 req->rq_ntrans++; 338 return; 339 } 340out_unlock: 341 xprt_clear_locked(xprt); 342} 343 344/** 345 * xprt_release_xprt - allow other requests to use a transport 346 * @xprt: transport with other tasks potentially waiting 347 * @task: task that is releasing access to the transport 348 * 349 * Note that "task" can be NULL. No congestion control is provided. 350 */ 351void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task) 352{ 353 if (xprt->snd_task == task) { 354 xprt_clear_locked(xprt); 355 __xprt_lock_write_next(xprt); 356 } 357} 358EXPORT_SYMBOL_GPL(xprt_release_xprt); 359 360/** 361 * xprt_release_xprt_cong - allow other requests to use a transport 362 * @xprt: transport with other tasks potentially waiting 363 * @task: task that is releasing access to the transport 364 * 365 * Note that "task" can be NULL. Another task is awoken to use the 366 * transport if the transport's congestion window allows it. 367 */ 368void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task) 369{ 370 if (xprt->snd_task == task) { 371 xprt_clear_locked(xprt); 372 __xprt_lock_write_next_cong(xprt); 373 } 374} 375EXPORT_SYMBOL_GPL(xprt_release_xprt_cong); 376 377static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task) 378{ 379 spin_lock_bh(&xprt->transport_lock); 380 xprt->ops->release_xprt(xprt, task); 381 spin_unlock_bh(&xprt->transport_lock); 382} 383 384/* 385 * Van Jacobson congestion avoidance. Check if the congestion window 386 * overflowed. Put the task to sleep if this is the case. 387 */ 388static int 389__xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task) 390{ 391 struct rpc_rqst *req = task->tk_rqstp; 392 393 if (req->rq_cong) 394 return 1; 395 dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n", 396 task->tk_pid, xprt->cong, xprt->cwnd); 397 if (RPCXPRT_CONGESTED(xprt)) 398 return 0; 399 req->rq_cong = 1; 400 xprt->cong += RPC_CWNDSCALE; 401 return 1; 402} 403 404/* 405 * Adjust the congestion window, and wake up the next task 406 * that has been sleeping due to congestion 407 */ 408static void 409__xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req) 410{ 411 if (!req->rq_cong) 412 return; 413 req->rq_cong = 0; 414 xprt->cong -= RPC_CWNDSCALE; 415 __xprt_lock_write_next_cong(xprt); 416} 417 418/** 419 * xprt_release_rqst_cong - housekeeping when request is complete 420 * @task: RPC request that recently completed 421 * 422 * Useful for transports that require congestion control. 423 */ 424void xprt_release_rqst_cong(struct rpc_task *task) 425{ 426 __xprt_put_cong(task->tk_xprt, task->tk_rqstp); 427} 428EXPORT_SYMBOL_GPL(xprt_release_rqst_cong); 429 430/** 431 * xprt_adjust_cwnd - adjust transport congestion window 432 * @task: recently completed RPC request used to adjust window 433 * @result: result code of completed RPC request 434 * 435 * We use a time-smoothed congestion estimator to avoid heavy oscillation. 436 */ 437void xprt_adjust_cwnd(struct rpc_task *task, int result) 438{ 439 struct rpc_rqst *req = task->tk_rqstp; 440 struct rpc_xprt *xprt = task->tk_xprt; 441 unsigned long cwnd = xprt->cwnd; 442 443 if (result >= 0 && cwnd <= xprt->cong) { 444 /* The (cwnd >> 1) term makes sure 445 * the result gets rounded properly. */ 446 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd; 447 if (cwnd > RPC_MAXCWND(xprt)) 448 cwnd = RPC_MAXCWND(xprt); 449 __xprt_lock_write_next_cong(xprt); 450 } else if (result == -ETIMEDOUT) { 451 cwnd >>= 1; 452 if (cwnd < RPC_CWNDSCALE) 453 cwnd = RPC_CWNDSCALE; 454 } 455 dprintk("RPC: cong %ld, cwnd was %ld, now %ld\n", 456 xprt->cong, xprt->cwnd, cwnd); 457 xprt->cwnd = cwnd; 458 __xprt_put_cong(xprt, req); 459} 460EXPORT_SYMBOL_GPL(xprt_adjust_cwnd); 461 462/** 463 * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue 464 * @xprt: transport with waiting tasks 465 * @status: result code to plant in each task before waking it 466 * 467 */ 468void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status) 469{ 470 if (status < 0) 471 rpc_wake_up_status(&xprt->pending, status); 472 else 473 rpc_wake_up(&xprt->pending); 474} 475EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks); 476 477/** 478 * xprt_wait_for_buffer_space - wait for transport output buffer to clear 479 * @task: task to be put to sleep 480 * @action: function pointer to be executed after wait 481 */ 482void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action) 483{ 484 struct rpc_rqst *req = task->tk_rqstp; 485 struct rpc_xprt *xprt = req->rq_xprt; 486 487 task->tk_timeout = req->rq_timeout; 488 rpc_sleep_on(&xprt->pending, task, action); 489} 490EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space); 491 492/** 493 * xprt_write_space - wake the task waiting for transport output buffer space 494 * @xprt: transport with waiting tasks 495 * 496 * Can be called in a soft IRQ context, so xprt_write_space never sleeps. 497 */ 498void xprt_write_space(struct rpc_xprt *xprt) 499{ 500 if (unlikely(xprt->shutdown)) 501 return; 502 503 spin_lock_bh(&xprt->transport_lock); 504 if (xprt->snd_task) { 505 dprintk("RPC: write space: waking waiting task on " 506 "xprt %p\n", xprt); 507 rpc_wake_up_queued_task(&xprt->pending, xprt->snd_task); 508 } 509 spin_unlock_bh(&xprt->transport_lock); 510} 511EXPORT_SYMBOL_GPL(xprt_write_space); 512 513/** 514 * xprt_set_retrans_timeout_def - set a request's retransmit timeout 515 * @task: task whose timeout is to be set 516 * 517 * Set a request's retransmit timeout based on the transport's 518 * default timeout parameters. Used by transports that don't adjust 519 * the retransmit timeout based on round-trip time estimation. 520 */ 521void xprt_set_retrans_timeout_def(struct rpc_task *task) 522{ 523 task->tk_timeout = task->tk_rqstp->rq_timeout; 524} 525EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def); 526 527/* 528 * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout 529 * @task: task whose timeout is to be set 530 * 531 * Set a request's retransmit timeout using the RTT estimator. 532 */ 533void xprt_set_retrans_timeout_rtt(struct rpc_task *task) 534{ 535 int timer = task->tk_msg.rpc_proc->p_timer; 536 struct rpc_clnt *clnt = task->tk_client; 537 struct rpc_rtt *rtt = clnt->cl_rtt; 538 struct rpc_rqst *req = task->tk_rqstp; 539 unsigned long max_timeout = clnt->cl_timeout->to_maxval; 540 541 task->tk_timeout = rpc_calc_rto(rtt, timer); 542 task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries; 543 if (task->tk_timeout > max_timeout || task->tk_timeout == 0) 544 task->tk_timeout = max_timeout; 545} 546EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt); 547 548static void xprt_reset_majortimeo(struct rpc_rqst *req) 549{ 550 const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout; 551 552 req->rq_majortimeo = req->rq_timeout; 553 if (to->to_exponential) 554 req->rq_majortimeo <<= to->to_retries; 555 else 556 req->rq_majortimeo += to->to_increment * to->to_retries; 557 if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0) 558 req->rq_majortimeo = to->to_maxval; 559 req->rq_majortimeo += jiffies; 560} 561 562/** 563 * xprt_adjust_timeout - adjust timeout values for next retransmit 564 * @req: RPC request containing parameters to use for the adjustment 565 * 566 */ 567int xprt_adjust_timeout(struct rpc_rqst *req) 568{ 569 struct rpc_xprt *xprt = req->rq_xprt; 570 const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout; 571 int status = 0; 572 573 if (time_before(jiffies, req->rq_majortimeo)) { 574 if (to->to_exponential) 575 req->rq_timeout <<= 1; 576 else 577 req->rq_timeout += to->to_increment; 578 if (to->to_maxval && req->rq_timeout >= to->to_maxval) 579 req->rq_timeout = to->to_maxval; 580 req->rq_retries++; 581 } else { 582 req->rq_timeout = to->to_initval; 583 req->rq_retries = 0; 584 xprt_reset_majortimeo(req); 585 /* Reset the RTT counters == "slow start" */ 586 spin_lock_bh(&xprt->transport_lock); 587 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval); 588 spin_unlock_bh(&xprt->transport_lock); 589 status = -ETIMEDOUT; 590 } 591 592 if (req->rq_timeout == 0) { 593 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n"); 594 req->rq_timeout = 5 * HZ; 595 } 596 return status; 597} 598 599static void xprt_autoclose(struct work_struct *work) 600{ 601 struct rpc_xprt *xprt = 602 container_of(work, struct rpc_xprt, task_cleanup); 603 604 xprt->ops->close(xprt); 605 clear_bit(XPRT_CLOSE_WAIT, &xprt->state); 606 xprt_release_write(xprt, NULL); 607} 608 609/** 610 * xprt_disconnect_done - mark a transport as disconnected 611 * @xprt: transport to flag for disconnect 612 * 613 */ 614void xprt_disconnect_done(struct rpc_xprt *xprt) 615{ 616 dprintk("RPC: disconnected transport %p\n", xprt); 617 spin_lock_bh(&xprt->transport_lock); 618 xprt_clear_connected(xprt); 619 xprt_wake_pending_tasks(xprt, -EAGAIN); 620 spin_unlock_bh(&xprt->transport_lock); 621} 622EXPORT_SYMBOL_GPL(xprt_disconnect_done); 623 624/** 625 * xprt_force_disconnect - force a transport to disconnect 626 * @xprt: transport to disconnect 627 * 628 */ 629void xprt_force_disconnect(struct rpc_xprt *xprt) 630{ 631 /* Don't race with the test_bit() in xprt_clear_locked() */ 632 spin_lock_bh(&xprt->transport_lock); 633 set_bit(XPRT_CLOSE_WAIT, &xprt->state); 634 /* Try to schedule an autoclose RPC call */ 635 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0) 636 queue_work(rpciod_workqueue, &xprt->task_cleanup); 637 xprt_wake_pending_tasks(xprt, -EAGAIN); 638 spin_unlock_bh(&xprt->transport_lock); 639} 640 641/** 642 * xprt_conditional_disconnect - force a transport to disconnect 643 * @xprt: transport to disconnect 644 * @cookie: 'connection cookie' 645 * 646 * This attempts to break the connection if and only if 'cookie' matches 647 * the current transport 'connection cookie'. It ensures that we don't 648 * try to break the connection more than once when we need to retransmit 649 * a batch of RPC requests. 650 * 651 */ 652void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie) 653{ 654 /* Don't race with the test_bit() in xprt_clear_locked() */ 655 spin_lock_bh(&xprt->transport_lock); 656 if (cookie != xprt->connect_cookie) 657 goto out; 658 if (test_bit(XPRT_CLOSING, &xprt->state) || !xprt_connected(xprt)) 659 goto out; 660 set_bit(XPRT_CLOSE_WAIT, &xprt->state); 661 /* Try to schedule an autoclose RPC call */ 662 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0) 663 queue_work(rpciod_workqueue, &xprt->task_cleanup); 664 xprt_wake_pending_tasks(xprt, -EAGAIN); 665out: 666 spin_unlock_bh(&xprt->transport_lock); 667} 668 669static void 670xprt_init_autodisconnect(unsigned long data) 671{ 672 struct rpc_xprt *xprt = (struct rpc_xprt *)data; 673 674 spin_lock(&xprt->transport_lock); 675 if (!list_empty(&xprt->recv) || xprt->shutdown) 676 goto out_abort; 677 if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) 678 goto out_abort; 679 spin_unlock(&xprt->transport_lock); 680 set_bit(XPRT_CONNECTION_CLOSE, &xprt->state); 681 queue_work(rpciod_workqueue, &xprt->task_cleanup); 682 return; 683out_abort: 684 spin_unlock(&xprt->transport_lock); 685} 686 687/** 688 * xprt_connect - schedule a transport connect operation 689 * @task: RPC task that is requesting the connect 690 * 691 */ 692void xprt_connect(struct rpc_task *task) 693{ 694 struct rpc_xprt *xprt = task->tk_xprt; 695 696 dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid, 697 xprt, (xprt_connected(xprt) ? "is" : "is not")); 698 699 if (!xprt_bound(xprt)) { 700 task->tk_status = -EAGAIN; 701 return; 702 } 703 if (!xprt_lock_write(xprt, task)) 704 return; 705 706 if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state)) 707 xprt->ops->close(xprt); 708 709 if (xprt_connected(xprt)) 710 xprt_release_write(xprt, task); 711 else { 712 if (task->tk_rqstp) 713 task->tk_rqstp->rq_bytes_sent = 0; 714 715 task->tk_timeout = task->tk_rqstp->rq_timeout; 716 rpc_sleep_on(&xprt->pending, task, xprt_connect_status); 717 718 if (test_bit(XPRT_CLOSING, &xprt->state)) 719 return; 720 if (xprt_test_and_set_connecting(xprt)) 721 return; 722 xprt->stat.connect_start = jiffies; 723 xprt->ops->connect(task); 724 } 725} 726 727static void xprt_connect_status(struct rpc_task *task) 728{ 729 struct rpc_xprt *xprt = task->tk_xprt; 730 731 if (task->tk_status == 0) { 732 xprt->stat.connect_count++; 733 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start; 734 dprintk("RPC: %5u xprt_connect_status: connection established\n", 735 task->tk_pid); 736 return; 737 } 738 739 switch (task->tk_status) { 740 case -EAGAIN: 741 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid); 742 break; 743 case -ETIMEDOUT: 744 dprintk("RPC: %5u xprt_connect_status: connect attempt timed " 745 "out\n", task->tk_pid); 746 break; 747 default: 748 dprintk("RPC: %5u xprt_connect_status: error %d connecting to " 749 "server %s\n", task->tk_pid, -task->tk_status, 750 task->tk_client->cl_server); 751 xprt_release_write(xprt, task); 752 task->tk_status = -EIO; 753 } 754} 755 756/** 757 * xprt_lookup_rqst - find an RPC request corresponding to an XID 758 * @xprt: transport on which the original request was transmitted 759 * @xid: RPC XID of incoming reply 760 * 761 */ 762struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid) 763{ 764 struct rpc_rqst *entry; 765 766 list_for_each_entry(entry, &xprt->recv, rq_list) 767 if (entry->rq_xid == xid) 768 return entry; 769 770 dprintk("RPC: xprt_lookup_rqst did not find xid %08x\n", 771 ntohl(xid)); 772 xprt->stat.bad_xids++; 773 return NULL; 774} 775EXPORT_SYMBOL_GPL(xprt_lookup_rqst); 776 777static void xprt_update_rtt(struct rpc_task *task) 778{ 779 struct rpc_rqst *req = task->tk_rqstp; 780 struct rpc_rtt *rtt = task->tk_client->cl_rtt; 781 unsigned timer = task->tk_msg.rpc_proc->p_timer; 782 long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt)); 783 784 if (timer) { 785 if (req->rq_ntrans == 1) 786 rpc_update_rtt(rtt, timer, m); 787 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1); 788 } 789} 790 791/** 792 * xprt_complete_rqst - called when reply processing is complete 793 * @task: RPC request that recently completed 794 * @copied: actual number of bytes received from the transport 795 * 796 * Caller holds transport lock. 797 */ 798void xprt_complete_rqst(struct rpc_task *task, int copied) 799{ 800 struct rpc_rqst *req = task->tk_rqstp; 801 struct rpc_xprt *xprt = req->rq_xprt; 802 803 dprintk("RPC: %5u xid %08x complete (%d bytes received)\n", 804 task->tk_pid, ntohl(req->rq_xid), copied); 805 806 xprt->stat.recvs++; 807 req->rq_rtt = ktime_sub(ktime_get(), req->rq_xtime); 808 if (xprt->ops->timer != NULL) 809 xprt_update_rtt(task); 810 811 list_del_init(&req->rq_list); 812 req->rq_private_buf.len = copied; 813 /* Ensure all writes are done before we update */ 814 /* req->rq_reply_bytes_recvd */ 815 smp_wmb(); 816 req->rq_reply_bytes_recvd = copied; 817 rpc_wake_up_queued_task(&xprt->pending, task); 818} 819EXPORT_SYMBOL_GPL(xprt_complete_rqst); 820 821static void xprt_timer(struct rpc_task *task) 822{ 823 struct rpc_rqst *req = task->tk_rqstp; 824 struct rpc_xprt *xprt = req->rq_xprt; 825 826 if (task->tk_status != -ETIMEDOUT) 827 return; 828 dprintk("RPC: %5u xprt_timer\n", task->tk_pid); 829 830 spin_lock_bh(&xprt->transport_lock); 831 if (!req->rq_reply_bytes_recvd) { 832 if (xprt->ops->timer) 833 xprt->ops->timer(task); 834 } else 835 task->tk_status = 0; 836 spin_unlock_bh(&xprt->transport_lock); 837} 838 839static inline int xprt_has_timer(struct rpc_xprt *xprt) 840{ 841 return xprt->idle_timeout != 0; 842} 843 844/** 845 * xprt_prepare_transmit - reserve the transport before sending a request 846 * @task: RPC task about to send a request 847 * 848 */ 849int xprt_prepare_transmit(struct rpc_task *task) 850{ 851 struct rpc_rqst *req = task->tk_rqstp; 852 struct rpc_xprt *xprt = req->rq_xprt; 853 int err = 0; 854 855 dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid); 856 857 spin_lock_bh(&xprt->transport_lock); 858 if (req->rq_reply_bytes_recvd && !req->rq_bytes_sent) { 859 err = req->rq_reply_bytes_recvd; 860 goto out_unlock; 861 } 862 if (!xprt->ops->reserve_xprt(xprt, task)) 863 err = -EAGAIN; 864out_unlock: 865 spin_unlock_bh(&xprt->transport_lock); 866 return err; 867} 868 869void xprt_end_transmit(struct rpc_task *task) 870{ 871 xprt_release_write(task->tk_rqstp->rq_xprt, task); 872} 873 874/** 875 * xprt_transmit - send an RPC request on a transport 876 * @task: controlling RPC task 877 * 878 * We have to copy the iovec because sendmsg fiddles with its contents. 879 */ 880void xprt_transmit(struct rpc_task *task) 881{ 882 struct rpc_rqst *req = task->tk_rqstp; 883 struct rpc_xprt *xprt = req->rq_xprt; 884 int status; 885 886 dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen); 887 888 if (!req->rq_reply_bytes_recvd) { 889 if (list_empty(&req->rq_list) && rpc_reply_expected(task)) { 890 /* 891 * Add to the list only if we're expecting a reply 892 */ 893 spin_lock_bh(&xprt->transport_lock); 894 /* Update the softirq receive buffer */ 895 memcpy(&req->rq_private_buf, &req->rq_rcv_buf, 896 sizeof(req->rq_private_buf)); 897 /* Add request to the receive list */ 898 list_add_tail(&req->rq_list, &xprt->recv); 899 spin_unlock_bh(&xprt->transport_lock); 900 xprt_reset_majortimeo(req); 901 /* Turn off autodisconnect */ 902 del_singleshot_timer_sync(&xprt->timer); 903 } 904 } else if (!req->rq_bytes_sent) 905 return; 906 907 req->rq_connect_cookie = xprt->connect_cookie; 908 req->rq_xtime = ktime_get(); 909 status = xprt->ops->send_request(task); 910 if (status != 0) { 911 task->tk_status = status; 912 return; 913 } 914 915 dprintk("RPC: %5u xmit complete\n", task->tk_pid); 916 task->tk_flags |= RPC_TASK_SENT; 917 spin_lock_bh(&xprt->transport_lock); 918 919 xprt->ops->set_retrans_timeout(task); 920 921 xprt->stat.sends++; 922 xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs; 923 xprt->stat.bklog_u += xprt->backlog.qlen; 924 925 /* Don't race with disconnect */ 926 if (!xprt_connected(xprt)) 927 task->tk_status = -ENOTCONN; 928 else if (!req->rq_reply_bytes_recvd && rpc_reply_expected(task)) { 929 /* 930 * Sleep on the pending queue since 931 * we're expecting a reply. 932 */ 933 rpc_sleep_on(&xprt->pending, task, xprt_timer); 934 } 935 spin_unlock_bh(&xprt->transport_lock); 936} 937 938static void xprt_alloc_slot(struct rpc_task *task) 939{ 940 struct rpc_xprt *xprt = task->tk_xprt; 941 942 task->tk_status = 0; 943 if (!list_empty(&xprt->free)) { 944 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list); 945 list_del_init(&req->rq_list); 946 task->tk_rqstp = req; 947 xprt_request_init(task, xprt); 948 return; 949 } 950 dprintk("RPC: waiting for request slot\n"); 951 task->tk_status = -EAGAIN; 952 rpc_sleep_on(&xprt->backlog, task, NULL); 953} 954 955static void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req) 956{ 957 memset(req, 0, sizeof(*req)); /* mark unused */ 958 959 spin_lock(&xprt->reserve_lock); 960 list_add(&req->rq_list, &xprt->free); 961 rpc_wake_up_next(&xprt->backlog); 962 spin_unlock(&xprt->reserve_lock); 963} 964 965static void xprt_free_all_slots(struct rpc_xprt *xprt) 966{ 967 struct rpc_rqst *req; 968 while (!list_empty(&xprt->free)) { 969 req = list_first_entry(&xprt->free, struct rpc_rqst, rq_list); 970 list_del(&req->rq_list); 971 kfree(req); 972 } 973} 974 975struct rpc_xprt *xprt_alloc(struct net *net, int size, int num_prealloc) 976{ 977 struct rpc_xprt *xprt; 978 struct rpc_rqst *req; 979 int i; 980 981 xprt = kzalloc(size, GFP_KERNEL); 982 if (xprt == NULL) 983 goto out; 984 985 xprt_init(xprt, net); 986 987 for (i = 0; i < num_prealloc; i++) { 988 req = kzalloc(sizeof(struct rpc_rqst), GFP_KERNEL); 989 if (!req) 990 break; 991 list_add(&req->rq_list, &xprt->free); 992 } 993 if (i < num_prealloc) 994 goto out_free; 995 xprt->max_reqs = num_prealloc; 996 997 return xprt; 998 999out_free: 1000 xprt_free(xprt); 1001out: 1002 return NULL; 1003} 1004EXPORT_SYMBOL_GPL(xprt_alloc); 1005 1006void xprt_free(struct rpc_xprt *xprt) 1007{ 1008 put_net(xprt->xprt_net); 1009 xprt_free_all_slots(xprt); 1010 kfree(xprt); 1011} 1012EXPORT_SYMBOL_GPL(xprt_free); 1013 1014/** 1015 * xprt_reserve - allocate an RPC request slot 1016 * @task: RPC task requesting a slot allocation 1017 * 1018 * If no more slots are available, place the task on the transport's 1019 * backlog queue. 1020 */ 1021void xprt_reserve(struct rpc_task *task) 1022{ 1023 struct rpc_xprt *xprt = task->tk_xprt; 1024 1025 task->tk_status = 0; 1026 if (task->tk_rqstp != NULL) 1027 return; 1028 1029 /* Note: grabbing the xprt_lock_write() here is not strictly needed, 1030 * but ensures that we throttle new slot allocation if the transport 1031 * is congested (e.g. if reconnecting or if we're out of socket 1032 * write buffer space). 1033 */ 1034 task->tk_timeout = 0; 1035 task->tk_status = -EAGAIN; 1036 if (!xprt_lock_write(xprt, task)) 1037 return; 1038 1039 task->tk_status = -EIO; 1040 spin_lock(&xprt->reserve_lock); 1041 xprt_alloc_slot(task); 1042 spin_unlock(&xprt->reserve_lock); 1043 xprt_release_write(xprt, task); 1044} 1045 1046static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt) 1047{ 1048 return (__force __be32)xprt->xid++; 1049} 1050 1051static inline void xprt_init_xid(struct rpc_xprt *xprt) 1052{ 1053 xprt->xid = net_random(); 1054} 1055 1056static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt) 1057{ 1058 struct rpc_rqst *req = task->tk_rqstp; 1059 1060 req->rq_timeout = task->tk_client->cl_timeout->to_initval; 1061 req->rq_task = task; 1062 req->rq_xprt = xprt; 1063 req->rq_buffer = NULL; 1064 req->rq_xid = xprt_alloc_xid(xprt); 1065 req->rq_release_snd_buf = NULL; 1066 xprt_reset_majortimeo(req); 1067 dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid, 1068 req, ntohl(req->rq_xid)); 1069} 1070 1071/** 1072 * xprt_release - release an RPC request slot 1073 * @task: task which is finished with the slot 1074 * 1075 */ 1076void xprt_release(struct rpc_task *task) 1077{ 1078 struct rpc_xprt *xprt; 1079 struct rpc_rqst *req; 1080 1081 if (!(req = task->tk_rqstp)) 1082 return; 1083 1084 xprt = req->rq_xprt; 1085 rpc_count_iostats(task); 1086 spin_lock_bh(&xprt->transport_lock); 1087 xprt->ops->release_xprt(xprt, task); 1088 if (xprt->ops->release_request) 1089 xprt->ops->release_request(task); 1090 if (!list_empty(&req->rq_list)) 1091 list_del(&req->rq_list); 1092 xprt->last_used = jiffies; 1093 if (list_empty(&xprt->recv) && xprt_has_timer(xprt)) 1094 mod_timer(&xprt->timer, 1095 xprt->last_used + xprt->idle_timeout); 1096 spin_unlock_bh(&xprt->transport_lock); 1097 if (req->rq_buffer) 1098 xprt->ops->buf_free(req->rq_buffer); 1099 if (req->rq_cred != NULL) 1100 put_rpccred(req->rq_cred); 1101 task->tk_rqstp = NULL; 1102 if (req->rq_release_snd_buf) 1103 req->rq_release_snd_buf(req); 1104 1105 dprintk("RPC: %5u release request %p\n", task->tk_pid, req); 1106 if (likely(!bc_prealloc(req))) 1107 xprt_free_slot(xprt, req); 1108 else 1109 xprt_free_bc_request(req); 1110} 1111 1112static void xprt_init(struct rpc_xprt *xprt, struct net *net) 1113{ 1114 atomic_set(&xprt->count, 1); 1115 1116 spin_lock_init(&xprt->transport_lock); 1117 spin_lock_init(&xprt->reserve_lock); 1118 1119 INIT_LIST_HEAD(&xprt->free); 1120 INIT_LIST_HEAD(&xprt->recv); 1121#if defined(CONFIG_SUNRPC_BACKCHANNEL) 1122 spin_lock_init(&xprt->bc_pa_lock); 1123 INIT_LIST_HEAD(&xprt->bc_pa_list); 1124#endif /* CONFIG_SUNRPC_BACKCHANNEL */ 1125 1126 xprt->last_used = jiffies; 1127 xprt->cwnd = RPC_INITCWND; 1128 xprt->bind_index = 0; 1129 1130 rpc_init_wait_queue(&xprt->binding, "xprt_binding"); 1131 rpc_init_wait_queue(&xprt->pending, "xprt_pending"); 1132 rpc_init_wait_queue(&xprt->sending, "xprt_sending"); 1133 rpc_init_wait_queue(&xprt->resend, "xprt_resend"); 1134 rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog"); 1135 1136 xprt_init_xid(xprt); 1137 1138 xprt->xprt_net = get_net(net); 1139} 1140 1141/** 1142 * xprt_create_transport - create an RPC transport 1143 * @args: rpc transport creation arguments 1144 * 1145 */ 1146struct rpc_xprt *xprt_create_transport(struct xprt_create *args) 1147{ 1148 struct rpc_xprt *xprt; 1149 struct xprt_class *t; 1150 1151 spin_lock(&xprt_list_lock); 1152 list_for_each_entry(t, &xprt_list, list) { 1153 if (t->ident == args->ident) { 1154 spin_unlock(&xprt_list_lock); 1155 goto found; 1156 } 1157 } 1158 spin_unlock(&xprt_list_lock); 1159 printk(KERN_ERR "RPC: transport (%d) not supported\n", args->ident); 1160 return ERR_PTR(-EIO); 1161 1162found: 1163 xprt = t->setup(args); 1164 if (IS_ERR(xprt)) { 1165 dprintk("RPC: xprt_create_transport: failed, %ld\n", 1166 -PTR_ERR(xprt)); 1167 goto out; 1168 } 1169 INIT_WORK(&xprt->task_cleanup, xprt_autoclose); 1170 if (xprt_has_timer(xprt)) 1171 setup_timer(&xprt->timer, xprt_init_autodisconnect, 1172 (unsigned long)xprt); 1173 else 1174 init_timer(&xprt->timer); 1175 dprintk("RPC: created transport %p with %u slots\n", xprt, 1176 xprt->max_reqs); 1177out: 1178 return xprt; 1179} 1180 1181/** 1182 * xprt_destroy - destroy an RPC transport, killing off all requests. 1183 * @xprt: transport to destroy 1184 * 1185 */ 1186static void xprt_destroy(struct rpc_xprt *xprt) 1187{ 1188 dprintk("RPC: destroying transport %p\n", xprt); 1189 xprt->shutdown = 1; 1190 del_timer_sync(&xprt->timer); 1191 1192 rpc_destroy_wait_queue(&xprt->binding); 1193 rpc_destroy_wait_queue(&xprt->pending); 1194 rpc_destroy_wait_queue(&xprt->sending); 1195 rpc_destroy_wait_queue(&xprt->resend); 1196 rpc_destroy_wait_queue(&xprt->backlog); 1197 cancel_work_sync(&xprt->task_cleanup); 1198 /* 1199 * Tear down transport state and free the rpc_xprt 1200 */ 1201 xprt->ops->destroy(xprt); 1202} 1203 1204/** 1205 * xprt_put - release a reference to an RPC transport. 1206 * @xprt: pointer to the transport 1207 * 1208 */ 1209void xprt_put(struct rpc_xprt *xprt) 1210{ 1211 if (atomic_dec_and_test(&xprt->count)) 1212 xprt_destroy(xprt); 1213} 1214 1215/** 1216 * xprt_get - return a reference to an RPC transport. 1217 * @xprt: pointer to the transport 1218 * 1219 */ 1220struct rpc_xprt *xprt_get(struct rpc_xprt *xprt) 1221{ 1222 if (atomic_inc_not_zero(&xprt->count)) 1223 return xprt; 1224 return NULL; 1225} 1226