2 * linux/net/sunrpc/xprt.c
4 * This is a generic RPC call interface supporting congestion avoidance,
5 * and asynchronous calls.
7 * The interface works like this:
9 * - When a process places a call, it allocates a request slot if
10 * one is available. Otherwise, it sleeps on the backlog queue
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
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
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
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.
35 * Copyright (C) 1995-1997, Olaf Kirch <okir@monad.swb.de>
37 * Transport switch API copyright (C) 2005, Chuck Lever <cel@netapp.com>
40 #include <linux/module.h>
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>
48 #include <linux/sunrpc/clnt.h>
49 #include <linux/sunrpc/metrics.h>
50 #include <linux/sunrpc/bc_xprt.h>
51 #include <linux/rcupdate.h>
53 #include <trace/events/sunrpc.h>
61 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
62 # define RPCDBG_FACILITY RPCDBG_XPRT
68 static void xprt_init(struct rpc_xprt *xprt, struct net *net);
69 static void xprt_request_init(struct rpc_task *, struct rpc_xprt *);
70 static void xprt_connect_status(struct rpc_task *task);
71 static int __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
72 static void __xprt_put_cong(struct rpc_xprt *, struct rpc_rqst *);
73 static void xprt_destroy(struct rpc_xprt *xprt);
75 static DEFINE_SPINLOCK(xprt_list_lock);
76 static LIST_HEAD(xprt_list);
79 * xprt_register_transport - register a transport implementation
80 * @transport: transport to register
82 * If a transport implementation is loaded as a kernel module, it can
83 * call this interface to make itself known to the RPC client.
86 * 0: transport successfully registered
87 * -EEXIST: transport already registered
88 * -EINVAL: transport module being unloaded
90 int xprt_register_transport(struct xprt_class *transport)
96 spin_lock(&xprt_list_lock);
97 list_for_each_entry(t, &xprt_list, list) {
98 /* don't register the same transport class twice */
99 if (t->ident == transport->ident)
103 list_add_tail(&transport->list, &xprt_list);
104 printk(KERN_INFO "RPC: Registered %s transport module.\n",
109 spin_unlock(&xprt_list_lock);
112 EXPORT_SYMBOL_GPL(xprt_register_transport);
115 * xprt_unregister_transport - unregister a transport implementation
116 * @transport: transport to unregister
119 * 0: transport successfully unregistered
120 * -ENOENT: transport never registered
122 int xprt_unregister_transport(struct xprt_class *transport)
124 struct xprt_class *t;
128 spin_lock(&xprt_list_lock);
129 list_for_each_entry(t, &xprt_list, list) {
130 if (t == transport) {
132 "RPC: Unregistered %s transport module.\n",
134 list_del_init(&transport->list);
141 spin_unlock(&xprt_list_lock);
144 EXPORT_SYMBOL_GPL(xprt_unregister_transport);
147 * xprt_load_transport - load a transport implementation
148 * @transport_name: transport to load
151 * 0: transport successfully loaded
152 * -ENOENT: transport module not available
154 int xprt_load_transport(const char *transport_name)
156 struct xprt_class *t;
160 spin_lock(&xprt_list_lock);
161 list_for_each_entry(t, &xprt_list, list) {
162 if (strcmp(t->name, transport_name) == 0) {
163 spin_unlock(&xprt_list_lock);
167 spin_unlock(&xprt_list_lock);
168 result = request_module("xprt%s", transport_name);
172 EXPORT_SYMBOL_GPL(xprt_load_transport);
175 * xprt_reserve_xprt - serialize write access to transports
176 * @task: task that is requesting access to the transport
177 * @xprt: pointer to the target transport
179 * This prevents mixing the payload of separate requests, and prevents
180 * transport connects from colliding with writes. No congestion control
183 int xprt_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
185 struct rpc_rqst *req = task->tk_rqstp;
188 if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
189 if (task == xprt->snd_task)
193 xprt->snd_task = task;
200 dprintk("RPC: %5u failed to lock transport %p\n",
202 task->tk_timeout = 0;
203 task->tk_status = -EAGAIN;
205 priority = RPC_PRIORITY_LOW;
206 else if (!req->rq_ntrans)
207 priority = RPC_PRIORITY_NORMAL;
209 priority = RPC_PRIORITY_HIGH;
210 rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
213 EXPORT_SYMBOL_GPL(xprt_reserve_xprt);
215 static void xprt_clear_locked(struct rpc_xprt *xprt)
217 xprt->snd_task = NULL;
218 if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state)) {
219 smp_mb__before_atomic();
220 clear_bit(XPRT_LOCKED, &xprt->state);
221 smp_mb__after_atomic();
223 queue_work(xprtiod_workqueue, &xprt->task_cleanup);
227 * xprt_reserve_xprt_cong - serialize write access to transports
228 * @task: task that is requesting access to the transport
230 * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
231 * integrated into the decision of whether a request is allowed to be
232 * woken up and given access to the transport.
234 int xprt_reserve_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
236 struct rpc_rqst *req = task->tk_rqstp;
239 if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
240 if (task == xprt->snd_task)
245 xprt->snd_task = task;
248 if (__xprt_get_cong(xprt, task)) {
249 xprt->snd_task = task;
253 xprt_clear_locked(xprt);
256 __xprt_put_cong(xprt, req);
257 dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
258 task->tk_timeout = 0;
259 task->tk_status = -EAGAIN;
261 priority = RPC_PRIORITY_LOW;
262 else if (!req->rq_ntrans)
263 priority = RPC_PRIORITY_NORMAL;
265 priority = RPC_PRIORITY_HIGH;
266 rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
269 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
271 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
275 spin_lock_bh(&xprt->transport_lock);
276 retval = xprt->ops->reserve_xprt(xprt, task);
277 spin_unlock_bh(&xprt->transport_lock);
281 static bool __xprt_lock_write_func(struct rpc_task *task, void *data)
283 struct rpc_xprt *xprt = data;
284 struct rpc_rqst *req;
286 req = task->tk_rqstp;
287 xprt->snd_task = task;
293 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
295 if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
298 if (rpc_wake_up_first_on_wq(xprtiod_workqueue, &xprt->sending,
299 __xprt_lock_write_func, xprt))
301 xprt_clear_locked(xprt);
304 static bool __xprt_lock_write_cong_func(struct rpc_task *task, void *data)
306 struct rpc_xprt *xprt = data;
307 struct rpc_rqst *req;
309 req = task->tk_rqstp;
311 xprt->snd_task = task;
314 if (__xprt_get_cong(xprt, task)) {
315 xprt->snd_task = task;
322 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
324 if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
326 if (RPCXPRT_CONGESTED(xprt))
328 if (rpc_wake_up_first_on_wq(xprtiod_workqueue, &xprt->sending,
329 __xprt_lock_write_cong_func, xprt))
332 xprt_clear_locked(xprt);
335 static void xprt_task_clear_bytes_sent(struct rpc_task *task)
338 struct rpc_rqst *req = task->tk_rqstp;
340 req->rq_bytes_sent = 0;
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
349 * Note that "task" can be NULL. No congestion control is provided.
351 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
353 if (xprt->snd_task == task) {
354 xprt_task_clear_bytes_sent(task);
355 xprt_clear_locked(xprt);
356 __xprt_lock_write_next(xprt);
359 EXPORT_SYMBOL_GPL(xprt_release_xprt);
362 * xprt_release_xprt_cong - allow other requests to use a transport
363 * @xprt: transport with other tasks potentially waiting
364 * @task: task that is releasing access to the transport
366 * Note that "task" can be NULL. Another task is awoken to use the
367 * transport if the transport's congestion window allows it.
369 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
371 if (xprt->snd_task == task) {
372 xprt_task_clear_bytes_sent(task);
373 xprt_clear_locked(xprt);
374 __xprt_lock_write_next_cong(xprt);
377 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
379 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
381 spin_lock_bh(&xprt->transport_lock);
382 xprt->ops->release_xprt(xprt, task);
383 spin_unlock_bh(&xprt->transport_lock);
387 * Van Jacobson congestion avoidance. Check if the congestion window
388 * overflowed. Put the task to sleep if this is the case.
391 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
393 struct rpc_rqst *req = task->tk_rqstp;
397 dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
398 task->tk_pid, xprt->cong, xprt->cwnd);
399 if (RPCXPRT_CONGESTED(xprt))
402 xprt->cong += RPC_CWNDSCALE;
407 * Adjust the congestion window, and wake up the next task
408 * that has been sleeping due to congestion
411 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
416 xprt->cong -= RPC_CWNDSCALE;
417 __xprt_lock_write_next_cong(xprt);
421 * xprt_release_rqst_cong - housekeeping when request is complete
422 * @task: RPC request that recently completed
424 * Useful for transports that require congestion control.
426 void xprt_release_rqst_cong(struct rpc_task *task)
428 struct rpc_rqst *req = task->tk_rqstp;
430 __xprt_put_cong(req->rq_xprt, req);
432 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
435 * xprt_adjust_cwnd - adjust transport congestion window
436 * @xprt: pointer to xprt
437 * @task: recently completed RPC request used to adjust window
438 * @result: result code of completed RPC request
440 * The transport code maintains an estimate on the maximum number of out-
441 * standing RPC requests, using a smoothed version of the congestion
442 * avoidance implemented in 44BSD. This is basically the Van Jacobson
443 * congestion algorithm: If a retransmit occurs, the congestion window is
444 * halved; otherwise, it is incremented by 1/cwnd when
446 * - a reply is received and
447 * - a full number of requests are outstanding and
448 * - the congestion window hasn't been updated recently.
450 void xprt_adjust_cwnd(struct rpc_xprt *xprt, struct rpc_task *task, int result)
452 struct rpc_rqst *req = task->tk_rqstp;
453 unsigned long cwnd = xprt->cwnd;
455 if (result >= 0 && cwnd <= xprt->cong) {
456 /* The (cwnd >> 1) term makes sure
457 * the result gets rounded properly. */
458 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
459 if (cwnd > RPC_MAXCWND(xprt))
460 cwnd = RPC_MAXCWND(xprt);
461 __xprt_lock_write_next_cong(xprt);
462 } else if (result == -ETIMEDOUT) {
464 if (cwnd < RPC_CWNDSCALE)
465 cwnd = RPC_CWNDSCALE;
467 dprintk("RPC: cong %ld, cwnd was %ld, now %ld\n",
468 xprt->cong, xprt->cwnd, cwnd);
470 __xprt_put_cong(xprt, req);
472 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
475 * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
476 * @xprt: transport with waiting tasks
477 * @status: result code to plant in each task before waking it
480 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
483 rpc_wake_up_status(&xprt->pending, status);
485 rpc_wake_up(&xprt->pending);
487 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
490 * xprt_wait_for_buffer_space - wait for transport output buffer to clear
491 * @task: task to be put to sleep
492 * @action: function pointer to be executed after wait
494 * Note that we only set the timer for the case of RPC_IS_SOFT(), since
495 * we don't in general want to force a socket disconnection due to
496 * an incomplete RPC call transmission.
498 void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action)
500 struct rpc_rqst *req = task->tk_rqstp;
501 struct rpc_xprt *xprt = req->rq_xprt;
503 task->tk_timeout = RPC_IS_SOFT(task) ? req->rq_timeout : 0;
504 rpc_sleep_on(&xprt->pending, task, action);
506 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
509 * xprt_write_space - wake the task waiting for transport output buffer space
510 * @xprt: transport with waiting tasks
512 * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
514 void xprt_write_space(struct rpc_xprt *xprt)
516 spin_lock_bh(&xprt->transport_lock);
517 if (xprt->snd_task) {
518 dprintk("RPC: write space: waking waiting task on "
520 rpc_wake_up_queued_task(&xprt->pending, xprt->snd_task);
522 spin_unlock_bh(&xprt->transport_lock);
524 EXPORT_SYMBOL_GPL(xprt_write_space);
527 * xprt_set_retrans_timeout_def - set a request's retransmit timeout
528 * @task: task whose timeout is to be set
530 * Set a request's retransmit timeout based on the transport's
531 * default timeout parameters. Used by transports that don't adjust
532 * the retransmit timeout based on round-trip time estimation.
534 void xprt_set_retrans_timeout_def(struct rpc_task *task)
536 task->tk_timeout = task->tk_rqstp->rq_timeout;
538 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
541 * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
542 * @task: task whose timeout is to be set
544 * Set a request's retransmit timeout using the RTT estimator.
546 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
548 int timer = task->tk_msg.rpc_proc->p_timer;
549 struct rpc_clnt *clnt = task->tk_client;
550 struct rpc_rtt *rtt = clnt->cl_rtt;
551 struct rpc_rqst *req = task->tk_rqstp;
552 unsigned long max_timeout = clnt->cl_timeout->to_maxval;
554 task->tk_timeout = rpc_calc_rto(rtt, timer);
555 task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
556 if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
557 task->tk_timeout = max_timeout;
559 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
561 static void xprt_reset_majortimeo(struct rpc_rqst *req)
563 const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
565 req->rq_majortimeo = req->rq_timeout;
566 if (to->to_exponential)
567 req->rq_majortimeo <<= to->to_retries;
569 req->rq_majortimeo += to->to_increment * to->to_retries;
570 if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
571 req->rq_majortimeo = to->to_maxval;
572 req->rq_majortimeo += jiffies;
576 * xprt_adjust_timeout - adjust timeout values for next retransmit
577 * @req: RPC request containing parameters to use for the adjustment
580 int xprt_adjust_timeout(struct rpc_rqst *req)
582 struct rpc_xprt *xprt = req->rq_xprt;
583 const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
586 if (time_before(jiffies, req->rq_majortimeo)) {
587 if (to->to_exponential)
588 req->rq_timeout <<= 1;
590 req->rq_timeout += to->to_increment;
591 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
592 req->rq_timeout = to->to_maxval;
595 req->rq_timeout = to->to_initval;
597 xprt_reset_majortimeo(req);
598 /* Reset the RTT counters == "slow start" */
599 spin_lock_bh(&xprt->transport_lock);
600 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
601 spin_unlock_bh(&xprt->transport_lock);
605 if (req->rq_timeout == 0) {
606 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
607 req->rq_timeout = 5 * HZ;
612 static void xprt_autoclose(struct work_struct *work)
614 struct rpc_xprt *xprt =
615 container_of(work, struct rpc_xprt, task_cleanup);
617 clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
618 xprt->ops->close(xprt);
619 xprt_release_write(xprt, NULL);
620 wake_up_bit(&xprt->state, XPRT_LOCKED);
624 * xprt_disconnect_done - mark a transport as disconnected
625 * @xprt: transport to flag for disconnect
628 void xprt_disconnect_done(struct rpc_xprt *xprt)
630 dprintk("RPC: disconnected transport %p\n", xprt);
631 spin_lock_bh(&xprt->transport_lock);
632 xprt_clear_connected(xprt);
633 xprt_wake_pending_tasks(xprt, -EAGAIN);
634 spin_unlock_bh(&xprt->transport_lock);
636 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
639 * xprt_force_disconnect - force a transport to disconnect
640 * @xprt: transport to disconnect
643 void xprt_force_disconnect(struct rpc_xprt *xprt)
645 /* Don't race with the test_bit() in xprt_clear_locked() */
646 spin_lock_bh(&xprt->transport_lock);
647 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
648 /* Try to schedule an autoclose RPC call */
649 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
650 queue_work(xprtiod_workqueue, &xprt->task_cleanup);
651 xprt_wake_pending_tasks(xprt, -EAGAIN);
652 spin_unlock_bh(&xprt->transport_lock);
654 EXPORT_SYMBOL_GPL(xprt_force_disconnect);
657 * xprt_conditional_disconnect - force a transport to disconnect
658 * @xprt: transport to disconnect
659 * @cookie: 'connection cookie'
661 * This attempts to break the connection if and only if 'cookie' matches
662 * the current transport 'connection cookie'. It ensures that we don't
663 * try to break the connection more than once when we need to retransmit
664 * a batch of RPC requests.
667 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
669 /* Don't race with the test_bit() in xprt_clear_locked() */
670 spin_lock_bh(&xprt->transport_lock);
671 if (cookie != xprt->connect_cookie)
673 if (test_bit(XPRT_CLOSING, &xprt->state))
675 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
676 /* Try to schedule an autoclose RPC call */
677 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
678 queue_work(xprtiod_workqueue, &xprt->task_cleanup);
679 xprt_wake_pending_tasks(xprt, -EAGAIN);
681 spin_unlock_bh(&xprt->transport_lock);
685 xprt_has_timer(const struct rpc_xprt *xprt)
687 return xprt->idle_timeout != 0;
691 xprt_schedule_autodisconnect(struct rpc_xprt *xprt)
692 __must_hold(&xprt->transport_lock)
694 if (list_empty(&xprt->recv) && xprt_has_timer(xprt))
695 mod_timer(&xprt->timer, xprt->last_used + xprt->idle_timeout);
699 xprt_init_autodisconnect(unsigned long data)
701 struct rpc_xprt *xprt = (struct rpc_xprt *)data;
703 spin_lock(&xprt->transport_lock);
704 if (!list_empty(&xprt->recv))
706 /* Reset xprt->last_used to avoid connect/autodisconnect cycling */
707 xprt->last_used = jiffies;
708 if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
710 spin_unlock(&xprt->transport_lock);
711 queue_work(xprtiod_workqueue, &xprt->task_cleanup);
714 spin_unlock(&xprt->transport_lock);
717 bool xprt_lock_connect(struct rpc_xprt *xprt,
718 struct rpc_task *task,
723 spin_lock_bh(&xprt->transport_lock);
724 if (!test_bit(XPRT_LOCKED, &xprt->state))
726 if (xprt->snd_task != task)
728 xprt_task_clear_bytes_sent(task);
729 xprt->snd_task = cookie;
732 spin_unlock_bh(&xprt->transport_lock);
736 void xprt_unlock_connect(struct rpc_xprt *xprt, void *cookie)
738 spin_lock_bh(&xprt->transport_lock);
739 if (xprt->snd_task != cookie)
741 if (!test_bit(XPRT_LOCKED, &xprt->state))
743 xprt->snd_task =NULL;
744 xprt->ops->release_xprt(xprt, NULL);
745 xprt_schedule_autodisconnect(xprt);
747 spin_unlock_bh(&xprt->transport_lock);
748 wake_up_bit(&xprt->state, XPRT_LOCKED);
752 * xprt_connect - schedule a transport connect operation
753 * @task: RPC task that is requesting the connect
756 void xprt_connect(struct rpc_task *task)
758 struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
760 dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
761 xprt, (xprt_connected(xprt) ? "is" : "is not"));
763 if (!xprt_bound(xprt)) {
764 task->tk_status = -EAGAIN;
767 if (!xprt_lock_write(xprt, task))
770 if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state))
771 xprt->ops->close(xprt);
773 if (!xprt_connected(xprt)) {
774 task->tk_rqstp->rq_bytes_sent = 0;
775 task->tk_timeout = task->tk_rqstp->rq_timeout;
776 task->tk_rqstp->rq_connect_cookie = xprt->connect_cookie;
777 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
779 if (test_bit(XPRT_CLOSING, &xprt->state))
781 if (xprt_test_and_set_connecting(xprt))
783 xprt->stat.connect_start = jiffies;
784 xprt->ops->connect(xprt, task);
786 xprt_release_write(xprt, task);
789 static void xprt_connect_status(struct rpc_task *task)
791 struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
793 if (task->tk_status == 0) {
794 xprt->stat.connect_count++;
795 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
796 dprintk("RPC: %5u xprt_connect_status: connection established\n",
801 switch (task->tk_status) {
809 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
812 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
813 "out\n", task->tk_pid);
816 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
817 "server %s\n", task->tk_pid, -task->tk_status,
819 task->tk_status = -EIO;
824 * xprt_lookup_rqst - find an RPC request corresponding to an XID
825 * @xprt: transport on which the original request was transmitted
826 * @xid: RPC XID of incoming reply
829 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
831 struct rpc_rqst *entry;
833 list_for_each_entry(entry, &xprt->recv, rq_list)
834 if (entry->rq_xid == xid) {
835 trace_xprt_lookup_rqst(xprt, xid, 0);
839 dprintk("RPC: xprt_lookup_rqst did not find xid %08x\n",
841 trace_xprt_lookup_rqst(xprt, xid, -ENOENT);
842 xprt->stat.bad_xids++;
845 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
848 * xprt_pin_rqst - Pin a request on the transport receive list
849 * @req: Request to pin
851 * Caller must ensure this is atomic with the call to xprt_lookup_rqst()
852 * so should be holding the xprt transport lock.
854 void xprt_pin_rqst(struct rpc_rqst *req)
856 set_bit(RPC_TASK_MSG_RECV, &req->rq_task->tk_runstate);
858 EXPORT_SYMBOL_GPL(xprt_pin_rqst);
861 * xprt_unpin_rqst - Unpin a request on the transport receive list
862 * @req: Request to pin
864 * Caller should be holding the xprt transport lock.
866 void xprt_unpin_rqst(struct rpc_rqst *req)
868 struct rpc_task *task = req->rq_task;
870 clear_bit(RPC_TASK_MSG_RECV, &task->tk_runstate);
871 if (test_bit(RPC_TASK_MSG_RECV_WAIT, &task->tk_runstate))
872 wake_up_bit(&task->tk_runstate, RPC_TASK_MSG_RECV);
874 EXPORT_SYMBOL_GPL(xprt_unpin_rqst);
876 static void xprt_wait_on_pinned_rqst(struct rpc_rqst *req)
877 __must_hold(&req->rq_xprt->recv_lock)
879 struct rpc_task *task = req->rq_task;
881 if (task && test_bit(RPC_TASK_MSG_RECV, &task->tk_runstate)) {
882 spin_unlock(&req->rq_xprt->recv_lock);
883 set_bit(RPC_TASK_MSG_RECV_WAIT, &task->tk_runstate);
884 wait_on_bit(&task->tk_runstate, RPC_TASK_MSG_RECV,
885 TASK_UNINTERRUPTIBLE);
886 clear_bit(RPC_TASK_MSG_RECV_WAIT, &task->tk_runstate);
887 spin_lock(&req->rq_xprt->recv_lock);
891 static void xprt_update_rtt(struct rpc_task *task)
893 struct rpc_rqst *req = task->tk_rqstp;
894 struct rpc_rtt *rtt = task->tk_client->cl_rtt;
895 unsigned int timer = task->tk_msg.rpc_proc->p_timer;
896 long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt));
899 if (req->rq_ntrans == 1)
900 rpc_update_rtt(rtt, timer, m);
901 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
906 * xprt_complete_rqst - called when reply processing is complete
907 * @task: RPC request that recently completed
908 * @copied: actual number of bytes received from the transport
910 * Caller holds transport lock.
912 void xprt_complete_rqst(struct rpc_task *task, int copied)
914 struct rpc_rqst *req = task->tk_rqstp;
915 struct rpc_xprt *xprt = req->rq_xprt;
917 dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
918 task->tk_pid, ntohl(req->rq_xid), copied);
919 trace_xprt_complete_rqst(xprt, req->rq_xid, copied);
922 req->rq_rtt = ktime_sub(ktime_get(), req->rq_xtime);
923 if (xprt->ops->timer != NULL)
924 xprt_update_rtt(task);
926 list_del_init(&req->rq_list);
927 req->rq_private_buf.len = copied;
928 /* Ensure all writes are done before we update */
929 /* req->rq_reply_bytes_recvd */
931 req->rq_reply_bytes_recvd = copied;
932 rpc_wake_up_queued_task(&xprt->pending, task);
934 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
936 static void xprt_timer(struct rpc_task *task)
938 struct rpc_rqst *req = task->tk_rqstp;
939 struct rpc_xprt *xprt = req->rq_xprt;
941 if (task->tk_status != -ETIMEDOUT)
943 dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
945 if (!req->rq_reply_bytes_recvd) {
946 if (xprt->ops->timer)
947 xprt->ops->timer(xprt, task);
953 * xprt_prepare_transmit - reserve the transport before sending a request
954 * @task: RPC task about to send a request
957 bool xprt_prepare_transmit(struct rpc_task *task)
959 struct rpc_rqst *req = task->tk_rqstp;
960 struct rpc_xprt *xprt = req->rq_xprt;
963 dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
965 spin_lock_bh(&xprt->transport_lock);
966 if (!req->rq_bytes_sent) {
967 if (req->rq_reply_bytes_recvd) {
968 task->tk_status = req->rq_reply_bytes_recvd;
971 if ((task->tk_flags & RPC_TASK_NO_RETRANS_TIMEOUT)
972 && xprt_connected(xprt)
973 && req->rq_connect_cookie == xprt->connect_cookie) {
974 xprt->ops->set_retrans_timeout(task);
975 rpc_sleep_on(&xprt->pending, task, xprt_timer);
979 if (!xprt->ops->reserve_xprt(xprt, task)) {
980 task->tk_status = -EAGAIN;
985 spin_unlock_bh(&xprt->transport_lock);
989 void xprt_end_transmit(struct rpc_task *task)
991 xprt_release_write(task->tk_rqstp->rq_xprt, task);
995 * xprt_transmit - send an RPC request on a transport
996 * @task: controlling RPC task
998 * We have to copy the iovec because sendmsg fiddles with its contents.
1000 void xprt_transmit(struct rpc_task *task)
1002 struct rpc_rqst *req = task->tk_rqstp;
1003 struct rpc_xprt *xprt = req->rq_xprt;
1004 unsigned int connect_cookie;
1005 int status, numreqs;
1007 dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
1009 if (!req->rq_reply_bytes_recvd) {
1010 if (list_empty(&req->rq_list) && rpc_reply_expected(task)) {
1012 * Add to the list only if we're expecting a reply
1014 /* Update the softirq receive buffer */
1015 memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
1016 sizeof(req->rq_private_buf));
1017 /* Add request to the receive list */
1018 spin_lock(&xprt->recv_lock);
1019 list_add_tail(&req->rq_list, &xprt->recv);
1020 spin_unlock(&xprt->recv_lock);
1021 xprt_reset_majortimeo(req);
1022 /* Turn off autodisconnect */
1023 del_singleshot_timer_sync(&xprt->timer);
1025 } else if (!req->rq_bytes_sent)
1028 connect_cookie = xprt->connect_cookie;
1029 req->rq_xtime = ktime_get();
1030 status = xprt->ops->send_request(task);
1031 trace_xprt_transmit(xprt, req->rq_xid, status);
1033 task->tk_status = status;
1036 xprt_inject_disconnect(xprt);
1038 dprintk("RPC: %5u xmit complete\n", task->tk_pid);
1039 task->tk_flags |= RPC_TASK_SENT;
1040 spin_lock_bh(&xprt->transport_lock);
1042 xprt->ops->set_retrans_timeout(task);
1044 numreqs = atomic_read(&xprt->num_reqs);
1045 if (numreqs > xprt->stat.max_slots)
1046 xprt->stat.max_slots = numreqs;
1048 xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
1049 xprt->stat.bklog_u += xprt->backlog.qlen;
1050 xprt->stat.sending_u += xprt->sending.qlen;
1051 xprt->stat.pending_u += xprt->pending.qlen;
1052 spin_unlock_bh(&xprt->transport_lock);
1054 req->rq_connect_cookie = connect_cookie;
1055 if (rpc_reply_expected(task) && !READ_ONCE(req->rq_reply_bytes_recvd)) {
1057 * Sleep on the pending queue if we're expecting a reply.
1058 * The spinlock ensures atomicity between the test of
1059 * req->rq_reply_bytes_recvd, and the call to rpc_sleep_on().
1061 spin_lock(&xprt->recv_lock);
1062 if (!req->rq_reply_bytes_recvd) {
1063 rpc_sleep_on(&xprt->pending, task, xprt_timer);
1065 * Send an extra queue wakeup call if the
1066 * connection was dropped in case the call to
1067 * rpc_sleep_on() raced.
1069 if (!xprt_connected(xprt))
1070 xprt_wake_pending_tasks(xprt, -ENOTCONN);
1072 spin_unlock(&xprt->recv_lock);
1076 static void xprt_add_backlog(struct rpc_xprt *xprt, struct rpc_task *task)
1078 set_bit(XPRT_CONGESTED, &xprt->state);
1079 rpc_sleep_on(&xprt->backlog, task, NULL);
1082 static void xprt_wake_up_backlog(struct rpc_xprt *xprt)
1084 if (rpc_wake_up_next(&xprt->backlog) == NULL)
1085 clear_bit(XPRT_CONGESTED, &xprt->state);
1088 static bool xprt_throttle_congested(struct rpc_xprt *xprt, struct rpc_task *task)
1092 if (!test_bit(XPRT_CONGESTED, &xprt->state))
1094 spin_lock(&xprt->reserve_lock);
1095 if (test_bit(XPRT_CONGESTED, &xprt->state)) {
1096 rpc_sleep_on(&xprt->backlog, task, NULL);
1099 spin_unlock(&xprt->reserve_lock);
1104 static struct rpc_rqst *xprt_dynamic_alloc_slot(struct rpc_xprt *xprt)
1106 struct rpc_rqst *req = ERR_PTR(-EAGAIN);
1108 if (!atomic_add_unless(&xprt->num_reqs, 1, xprt->max_reqs))
1110 spin_unlock(&xprt->reserve_lock);
1111 req = kzalloc(sizeof(struct rpc_rqst), GFP_NOFS);
1112 spin_lock(&xprt->reserve_lock);
1115 atomic_dec(&xprt->num_reqs);
1116 req = ERR_PTR(-ENOMEM);
1121 static bool xprt_dynamic_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1123 if (atomic_add_unless(&xprt->num_reqs, -1, xprt->min_reqs)) {
1130 void xprt_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1132 struct rpc_rqst *req;
1134 spin_lock(&xprt->reserve_lock);
1135 if (!list_empty(&xprt->free)) {
1136 req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
1137 list_del(&req->rq_list);
1140 req = xprt_dynamic_alloc_slot(xprt);
1143 switch (PTR_ERR(req)) {
1145 dprintk("RPC: dynamic allocation of request slot "
1146 "failed! Retrying\n");
1147 task->tk_status = -ENOMEM;
1150 xprt_add_backlog(xprt, task);
1151 dprintk("RPC: waiting for request slot\n");
1154 task->tk_status = -EAGAIN;
1156 spin_unlock(&xprt->reserve_lock);
1159 task->tk_status = 0;
1160 task->tk_rqstp = req;
1161 xprt_request_init(task, xprt);
1162 spin_unlock(&xprt->reserve_lock);
1164 EXPORT_SYMBOL_GPL(xprt_alloc_slot);
1166 void xprt_lock_and_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1168 /* Note: grabbing the xprt_lock_write() ensures that we throttle
1169 * new slot allocation if the transport is congested (i.e. when
1170 * reconnecting a stream transport or when out of socket write
1173 if (xprt_lock_write(xprt, task)) {
1174 xprt_alloc_slot(xprt, task);
1175 xprt_release_write(xprt, task);
1178 EXPORT_SYMBOL_GPL(xprt_lock_and_alloc_slot);
1180 static void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1182 spin_lock(&xprt->reserve_lock);
1183 if (!xprt_dynamic_free_slot(xprt, req)) {
1184 memset(req, 0, sizeof(*req)); /* mark unused */
1185 list_add(&req->rq_list, &xprt->free);
1187 xprt_wake_up_backlog(xprt);
1188 spin_unlock(&xprt->reserve_lock);
1191 static void xprt_free_all_slots(struct rpc_xprt *xprt)
1193 struct rpc_rqst *req;
1194 while (!list_empty(&xprt->free)) {
1195 req = list_first_entry(&xprt->free, struct rpc_rqst, rq_list);
1196 list_del(&req->rq_list);
1201 struct rpc_xprt *xprt_alloc(struct net *net, size_t size,
1202 unsigned int num_prealloc,
1203 unsigned int max_alloc)
1205 struct rpc_xprt *xprt;
1206 struct rpc_rqst *req;
1209 xprt = kzalloc(size, GFP_KERNEL);
1213 xprt_init(xprt, net);
1215 for (i = 0; i < num_prealloc; i++) {
1216 req = kzalloc(sizeof(struct rpc_rqst), GFP_KERNEL);
1219 list_add(&req->rq_list, &xprt->free);
1221 if (max_alloc > num_prealloc)
1222 xprt->max_reqs = max_alloc;
1224 xprt->max_reqs = num_prealloc;
1225 xprt->min_reqs = num_prealloc;
1226 atomic_set(&xprt->num_reqs, num_prealloc);
1235 EXPORT_SYMBOL_GPL(xprt_alloc);
1237 void xprt_free(struct rpc_xprt *xprt)
1239 put_net(xprt->xprt_net);
1240 xprt_free_all_slots(xprt);
1241 kfree_rcu(xprt, rcu);
1243 EXPORT_SYMBOL_GPL(xprt_free);
1246 * xprt_reserve - allocate an RPC request slot
1247 * @task: RPC task requesting a slot allocation
1249 * If the transport is marked as being congested, or if no more
1250 * slots are available, place the task on the transport's
1253 void xprt_reserve(struct rpc_task *task)
1255 struct rpc_xprt *xprt = task->tk_xprt;
1257 task->tk_status = 0;
1258 if (task->tk_rqstp != NULL)
1261 task->tk_timeout = 0;
1262 task->tk_status = -EAGAIN;
1263 if (!xprt_throttle_congested(xprt, task))
1264 xprt->ops->alloc_slot(xprt, task);
1268 * xprt_retry_reserve - allocate an RPC request slot
1269 * @task: RPC task requesting a slot allocation
1271 * If no more slots are available, place the task on the transport's
1273 * Note that the only difference with xprt_reserve is that we now
1274 * ignore the value of the XPRT_CONGESTED flag.
1276 void xprt_retry_reserve(struct rpc_task *task)
1278 struct rpc_xprt *xprt = task->tk_xprt;
1280 task->tk_status = 0;
1281 if (task->tk_rqstp != NULL)
1284 task->tk_timeout = 0;
1285 task->tk_status = -EAGAIN;
1286 xprt->ops->alloc_slot(xprt, task);
1289 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
1291 return (__force __be32)xprt->xid++;
1294 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1296 xprt->xid = prandom_u32();
1299 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1301 struct rpc_rqst *req = task->tk_rqstp;
1303 INIT_LIST_HEAD(&req->rq_list);
1304 req->rq_timeout = task->tk_client->cl_timeout->to_initval;
1305 req->rq_task = task;
1306 req->rq_xprt = xprt;
1307 req->rq_buffer = NULL;
1308 req->rq_xid = xprt_alloc_xid(xprt);
1309 req->rq_connect_cookie = xprt->connect_cookie - 1;
1310 req->rq_bytes_sent = 0;
1311 req->rq_snd_buf.len = 0;
1312 req->rq_snd_buf.buflen = 0;
1313 req->rq_rcv_buf.len = 0;
1314 req->rq_rcv_buf.buflen = 0;
1315 req->rq_release_snd_buf = NULL;
1316 xprt_reset_majortimeo(req);
1317 dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
1318 req, ntohl(req->rq_xid));
1322 * xprt_release - release an RPC request slot
1323 * @task: task which is finished with the slot
1326 void xprt_release(struct rpc_task *task)
1328 struct rpc_xprt *xprt;
1329 struct rpc_rqst *req = task->tk_rqstp;
1332 if (task->tk_client) {
1333 xprt = task->tk_xprt;
1334 if (xprt->snd_task == task)
1335 xprt_release_write(xprt, task);
1340 xprt = req->rq_xprt;
1341 if (task->tk_ops->rpc_count_stats != NULL)
1342 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
1343 else if (task->tk_client)
1344 rpc_count_iostats(task, task->tk_client->cl_metrics);
1345 spin_lock(&xprt->recv_lock);
1346 if (!list_empty(&req->rq_list)) {
1347 list_del(&req->rq_list);
1348 xprt_wait_on_pinned_rqst(req);
1350 spin_unlock(&xprt->recv_lock);
1351 spin_lock_bh(&xprt->transport_lock);
1352 xprt->ops->release_xprt(xprt, task);
1353 if (xprt->ops->release_request)
1354 xprt->ops->release_request(task);
1355 xprt->last_used = jiffies;
1356 xprt_schedule_autodisconnect(xprt);
1357 spin_unlock_bh(&xprt->transport_lock);
1359 xprt->ops->buf_free(task);
1360 xprt_inject_disconnect(xprt);
1361 if (req->rq_cred != NULL)
1362 put_rpccred(req->rq_cred);
1363 task->tk_rqstp = NULL;
1364 if (req->rq_release_snd_buf)
1365 req->rq_release_snd_buf(req);
1367 dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1368 if (likely(!bc_prealloc(req)))
1369 xprt_free_slot(xprt, req);
1371 xprt_free_bc_request(req);
1374 static void xprt_init(struct rpc_xprt *xprt, struct net *net)
1376 kref_init(&xprt->kref);
1378 spin_lock_init(&xprt->transport_lock);
1379 spin_lock_init(&xprt->reserve_lock);
1380 spin_lock_init(&xprt->recv_lock);
1382 INIT_LIST_HEAD(&xprt->free);
1383 INIT_LIST_HEAD(&xprt->recv);
1384 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1385 spin_lock_init(&xprt->bc_pa_lock);
1386 INIT_LIST_HEAD(&xprt->bc_pa_list);
1387 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1388 INIT_LIST_HEAD(&xprt->xprt_switch);
1390 xprt->last_used = jiffies;
1391 xprt->cwnd = RPC_INITCWND;
1392 xprt->bind_index = 0;
1394 rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1395 rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1396 rpc_init_priority_wait_queue(&xprt->sending, "xprt_sending");
1397 rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1399 xprt_init_xid(xprt);
1401 xprt->xprt_net = get_net(net);
1405 * xprt_create_transport - create an RPC transport
1406 * @args: rpc transport creation arguments
1409 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1411 struct rpc_xprt *xprt;
1412 struct xprt_class *t;
1414 spin_lock(&xprt_list_lock);
1415 list_for_each_entry(t, &xprt_list, list) {
1416 if (t->ident == args->ident) {
1417 spin_unlock(&xprt_list_lock);
1421 spin_unlock(&xprt_list_lock);
1422 dprintk("RPC: transport (%d) not supported\n", args->ident);
1423 return ERR_PTR(-EIO);
1426 xprt = t->setup(args);
1428 dprintk("RPC: xprt_create_transport: failed, %ld\n",
1432 if (args->flags & XPRT_CREATE_NO_IDLE_TIMEOUT)
1433 xprt->idle_timeout = 0;
1434 INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1435 if (xprt_has_timer(xprt))
1436 setup_timer(&xprt->timer, xprt_init_autodisconnect,
1437 (unsigned long)xprt);
1439 init_timer(&xprt->timer);
1441 if (strlen(args->servername) > RPC_MAXNETNAMELEN) {
1443 return ERR_PTR(-EINVAL);
1445 xprt->servername = kstrdup(args->servername, GFP_KERNEL);
1446 if (xprt->servername == NULL) {
1448 return ERR_PTR(-ENOMEM);
1451 rpc_xprt_debugfs_register(xprt);
1453 dprintk("RPC: created transport %p with %u slots\n", xprt,
1460 * xprt_destroy - destroy an RPC transport, killing off all requests.
1461 * @xprt: transport to destroy
1464 static void xprt_destroy(struct rpc_xprt *xprt)
1466 dprintk("RPC: destroying transport %p\n", xprt);
1468 /* Exclude transport connect/disconnect handlers */
1469 wait_on_bit_lock(&xprt->state, XPRT_LOCKED, TASK_UNINTERRUPTIBLE);
1471 del_timer_sync(&xprt->timer);
1473 rpc_xprt_debugfs_unregister(xprt);
1474 rpc_destroy_wait_queue(&xprt->binding);
1475 rpc_destroy_wait_queue(&xprt->pending);
1476 rpc_destroy_wait_queue(&xprt->sending);
1477 rpc_destroy_wait_queue(&xprt->backlog);
1478 cancel_work_sync(&xprt->task_cleanup);
1479 kfree(xprt->servername);
1481 * Tear down transport state and free the rpc_xprt
1483 xprt->ops->destroy(xprt);
1486 static void xprt_destroy_kref(struct kref *kref)
1488 xprt_destroy(container_of(kref, struct rpc_xprt, kref));
1492 * xprt_get - return a reference to an RPC transport.
1493 * @xprt: pointer to the transport
1496 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1498 if (xprt != NULL && kref_get_unless_zero(&xprt->kref))
1502 EXPORT_SYMBOL_GPL(xprt_get);
1505 * xprt_put - release a reference to an RPC transport.
1506 * @xprt: pointer to the transport
1509 void xprt_put(struct rpc_xprt *xprt)
1512 kref_put(&xprt->kref, xprt_destroy_kref);
1514 EXPORT_SYMBOL_GPL(xprt_put);