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>
59 # define RPCDBG_FACILITY RPCDBG_XPRT
65 static void xprt_init(struct rpc_xprt *xprt, struct net *net);
66 static void xprt_request_init(struct rpc_task *, struct rpc_xprt *);
67 static void xprt_connect_status(struct rpc_task *task);
68 static int __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
69 static void xprt_destroy(struct rpc_xprt *xprt);
71 static DEFINE_SPINLOCK(xprt_list_lock);
72 static LIST_HEAD(xprt_list);
75 * The transport code maintains an estimate on the maximum number of out-
76 * standing RPC requests, using a smoothed version of the congestion
77 * avoidance implemented in 44BSD. This is basically the Van Jacobson
78 * congestion algorithm: If a retransmit occurs, the congestion window is
79 * halved; otherwise, it is incremented by 1/cwnd when
81 * - a reply is received and
82 * - a full number of requests are outstanding and
83 * - the congestion window hasn't been updated recently.
85 #define RPC_CWNDSHIFT (8U)
86 #define RPC_CWNDSCALE (1U << RPC_CWNDSHIFT)
87 #define RPC_INITCWND RPC_CWNDSCALE
88 #define RPC_MAXCWND(xprt) ((xprt)->max_reqs << RPC_CWNDSHIFT)
90 #define RPCXPRT_CONGESTED(xprt) ((xprt)->cong >= (xprt)->cwnd)
93 * xprt_register_transport - register a transport implementation
94 * @transport: transport to register
96 * If a transport implementation is loaded as a kernel module, it can
97 * call this interface to make itself known to the RPC client.
100 * 0: transport successfully registered
101 * -EEXIST: transport already registered
102 * -EINVAL: transport module being unloaded
104 int xprt_register_transport(struct xprt_class *transport)
106 struct xprt_class *t;
110 spin_lock(&xprt_list_lock);
111 list_for_each_entry(t, &xprt_list, list) {
112 /* don't register the same transport class twice */
113 if (t->ident == transport->ident)
117 list_add_tail(&transport->list, &xprt_list);
118 printk(KERN_INFO "RPC: Registered %s transport module.\n",
123 spin_unlock(&xprt_list_lock);
126 EXPORT_SYMBOL_GPL(xprt_register_transport);
129 * xprt_unregister_transport - unregister a transport implementation
130 * @transport: transport to unregister
133 * 0: transport successfully unregistered
134 * -ENOENT: transport never registered
136 int xprt_unregister_transport(struct xprt_class *transport)
138 struct xprt_class *t;
142 spin_lock(&xprt_list_lock);
143 list_for_each_entry(t, &xprt_list, list) {
144 if (t == transport) {
146 "RPC: Unregistered %s transport module.\n",
148 list_del_init(&transport->list);
155 spin_unlock(&xprt_list_lock);
158 EXPORT_SYMBOL_GPL(xprt_unregister_transport);
161 * xprt_load_transport - load a transport implementation
162 * @transport_name: transport to load
165 * 0: transport successfully loaded
166 * -ENOENT: transport module not available
168 int xprt_load_transport(const char *transport_name)
170 struct xprt_class *t;
174 spin_lock(&xprt_list_lock);
175 list_for_each_entry(t, &xprt_list, list) {
176 if (strcmp(t->name, transport_name) == 0) {
177 spin_unlock(&xprt_list_lock);
181 spin_unlock(&xprt_list_lock);
182 result = request_module("xprt%s", transport_name);
186 EXPORT_SYMBOL_GPL(xprt_load_transport);
189 * xprt_reserve_xprt - serialize write access to transports
190 * @task: task that is requesting access to the transport
191 * @xprt: pointer to the target transport
193 * This prevents mixing the payload of separate requests, and prevents
194 * transport connects from colliding with writes. No congestion control
197 int xprt_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
199 struct rpc_rqst *req = task->tk_rqstp;
202 if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
203 if (task == xprt->snd_task)
207 xprt->snd_task = task;
209 req->rq_bytes_sent = 0;
216 dprintk("RPC: %5u failed to lock transport %p\n",
218 task->tk_timeout = 0;
219 task->tk_status = -EAGAIN;
221 priority = RPC_PRIORITY_LOW;
222 else if (!req->rq_ntrans)
223 priority = RPC_PRIORITY_NORMAL;
225 priority = RPC_PRIORITY_HIGH;
226 rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
229 EXPORT_SYMBOL_GPL(xprt_reserve_xprt);
231 static void xprt_clear_locked(struct rpc_xprt *xprt)
233 xprt->snd_task = NULL;
234 if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state)) {
235 smp_mb__before_clear_bit();
236 clear_bit(XPRT_LOCKED, &xprt->state);
237 smp_mb__after_clear_bit();
239 queue_work(rpciod_workqueue, &xprt->task_cleanup);
243 * xprt_reserve_xprt_cong - serialize write access to transports
244 * @task: task that is requesting access to the transport
246 * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
247 * integrated into the decision of whether a request is allowed to be
248 * woken up and given access to the transport.
250 int xprt_reserve_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
252 struct rpc_rqst *req = task->tk_rqstp;
255 if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
256 if (task == xprt->snd_task)
261 xprt->snd_task = task;
264 if (__xprt_get_cong(xprt, task)) {
265 xprt->snd_task = task;
266 req->rq_bytes_sent = 0;
270 xprt_clear_locked(xprt);
272 dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
273 task->tk_timeout = 0;
274 task->tk_status = -EAGAIN;
276 priority = RPC_PRIORITY_LOW;
277 else if (!req->rq_ntrans)
278 priority = RPC_PRIORITY_NORMAL;
280 priority = RPC_PRIORITY_HIGH;
281 rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
284 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
286 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
290 spin_lock_bh(&xprt->transport_lock);
291 retval = xprt->ops->reserve_xprt(xprt, task);
292 spin_unlock_bh(&xprt->transport_lock);
296 static bool __xprt_lock_write_func(struct rpc_task *task, void *data)
298 struct rpc_xprt *xprt = data;
299 struct rpc_rqst *req;
301 req = task->tk_rqstp;
302 xprt->snd_task = task;
304 req->rq_bytes_sent = 0;
310 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
312 if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
315 if (rpc_wake_up_first(&xprt->sending, __xprt_lock_write_func, xprt))
317 xprt_clear_locked(xprt);
320 static bool __xprt_lock_write_cong_func(struct rpc_task *task, void *data)
322 struct rpc_xprt *xprt = data;
323 struct rpc_rqst *req;
325 req = task->tk_rqstp;
327 xprt->snd_task = task;
330 if (__xprt_get_cong(xprt, task)) {
331 xprt->snd_task = task;
332 req->rq_bytes_sent = 0;
339 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
341 if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
343 if (RPCXPRT_CONGESTED(xprt))
345 if (rpc_wake_up_first(&xprt->sending, __xprt_lock_write_cong_func, xprt))
348 xprt_clear_locked(xprt);
352 * xprt_release_xprt - allow other requests to use a transport
353 * @xprt: transport with other tasks potentially waiting
354 * @task: task that is releasing access to the transport
356 * Note that "task" can be NULL. No congestion control is provided.
358 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
360 if (xprt->snd_task == task) {
362 struct rpc_rqst *req = task->tk_rqstp;
364 req->rq_bytes_sent = 0;
366 xprt_clear_locked(xprt);
367 __xprt_lock_write_next(xprt);
370 EXPORT_SYMBOL_GPL(xprt_release_xprt);
373 * xprt_release_xprt_cong - allow other requests to use a transport
374 * @xprt: transport with other tasks potentially waiting
375 * @task: task that is releasing access to the transport
377 * Note that "task" can be NULL. Another task is awoken to use the
378 * transport if the transport's congestion window allows it.
380 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
382 if (xprt->snd_task == task) {
384 struct rpc_rqst *req = task->tk_rqstp;
386 req->rq_bytes_sent = 0;
388 xprt_clear_locked(xprt);
389 __xprt_lock_write_next_cong(xprt);
392 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
394 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
396 spin_lock_bh(&xprt->transport_lock);
397 xprt->ops->release_xprt(xprt, task);
398 spin_unlock_bh(&xprt->transport_lock);
402 * Van Jacobson congestion avoidance. Check if the congestion window
403 * overflowed. Put the task to sleep if this is the case.
406 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
408 struct rpc_rqst *req = task->tk_rqstp;
412 dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
413 task->tk_pid, xprt->cong, xprt->cwnd);
414 if (RPCXPRT_CONGESTED(xprt))
417 xprt->cong += RPC_CWNDSCALE;
422 * Adjust the congestion window, and wake up the next task
423 * that has been sleeping due to congestion
426 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
431 xprt->cong -= RPC_CWNDSCALE;
432 __xprt_lock_write_next_cong(xprt);
436 * xprt_release_rqst_cong - housekeeping when request is complete
437 * @task: RPC request that recently completed
439 * Useful for transports that require congestion control.
441 void xprt_release_rqst_cong(struct rpc_task *task)
443 struct rpc_rqst *req = task->tk_rqstp;
445 __xprt_put_cong(req->rq_xprt, req);
447 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
450 * xprt_adjust_cwnd - adjust transport congestion window
451 * @xprt: pointer to xprt
452 * @task: recently completed RPC request used to adjust window
453 * @result: result code of completed RPC request
455 * We use a time-smoothed congestion estimator to avoid heavy oscillation.
457 void xprt_adjust_cwnd(struct rpc_xprt *xprt, struct rpc_task *task, int result)
459 struct rpc_rqst *req = task->tk_rqstp;
460 unsigned long cwnd = xprt->cwnd;
462 if (result >= 0 && cwnd <= xprt->cong) {
463 /* The (cwnd >> 1) term makes sure
464 * the result gets rounded properly. */
465 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
466 if (cwnd > RPC_MAXCWND(xprt))
467 cwnd = RPC_MAXCWND(xprt);
468 __xprt_lock_write_next_cong(xprt);
469 } else if (result == -ETIMEDOUT) {
471 if (cwnd < RPC_CWNDSCALE)
472 cwnd = RPC_CWNDSCALE;
474 dprintk("RPC: cong %ld, cwnd was %ld, now %ld\n",
475 xprt->cong, xprt->cwnd, cwnd);
477 __xprt_put_cong(xprt, req);
479 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
482 * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
483 * @xprt: transport with waiting tasks
484 * @status: result code to plant in each task before waking it
487 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
490 rpc_wake_up_status(&xprt->pending, status);
492 rpc_wake_up(&xprt->pending);
494 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
497 * xprt_wait_for_buffer_space - wait for transport output buffer to clear
498 * @task: task to be put to sleep
499 * @action: function pointer to be executed after wait
501 * Note that we only set the timer for the case of RPC_IS_SOFT(), since
502 * we don't in general want to force a socket disconnection due to
503 * an incomplete RPC call transmission.
505 void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action)
507 struct rpc_rqst *req = task->tk_rqstp;
508 struct rpc_xprt *xprt = req->rq_xprt;
510 task->tk_timeout = RPC_IS_SOFT(task) ? req->rq_timeout : 0;
511 rpc_sleep_on(&xprt->pending, task, action);
513 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
516 * xprt_write_space - wake the task waiting for transport output buffer space
517 * @xprt: transport with waiting tasks
519 * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
521 void xprt_write_space(struct rpc_xprt *xprt)
523 spin_lock_bh(&xprt->transport_lock);
524 if (xprt->snd_task) {
525 dprintk("RPC: write space: waking waiting task on "
527 rpc_wake_up_queued_task(&xprt->pending, xprt->snd_task);
529 spin_unlock_bh(&xprt->transport_lock);
531 EXPORT_SYMBOL_GPL(xprt_write_space);
534 * xprt_set_retrans_timeout_def - set a request's retransmit timeout
535 * @task: task whose timeout is to be set
537 * Set a request's retransmit timeout based on the transport's
538 * default timeout parameters. Used by transports that don't adjust
539 * the retransmit timeout based on round-trip time estimation.
541 void xprt_set_retrans_timeout_def(struct rpc_task *task)
543 task->tk_timeout = task->tk_rqstp->rq_timeout;
545 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
548 * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
549 * @task: task whose timeout is to be set
551 * Set a request's retransmit timeout using the RTT estimator.
553 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
555 int timer = task->tk_msg.rpc_proc->p_timer;
556 struct rpc_clnt *clnt = task->tk_client;
557 struct rpc_rtt *rtt = clnt->cl_rtt;
558 struct rpc_rqst *req = task->tk_rqstp;
559 unsigned long max_timeout = clnt->cl_timeout->to_maxval;
561 task->tk_timeout = rpc_calc_rto(rtt, timer);
562 task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
563 if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
564 task->tk_timeout = max_timeout;
566 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
568 static void xprt_reset_majortimeo(struct rpc_rqst *req)
570 const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
572 req->rq_majortimeo = req->rq_timeout;
573 if (to->to_exponential)
574 req->rq_majortimeo <<= to->to_retries;
576 req->rq_majortimeo += to->to_increment * to->to_retries;
577 if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
578 req->rq_majortimeo = to->to_maxval;
579 req->rq_majortimeo += jiffies;
583 * xprt_adjust_timeout - adjust timeout values for next retransmit
584 * @req: RPC request containing parameters to use for the adjustment
587 int xprt_adjust_timeout(struct rpc_rqst *req)
589 struct rpc_xprt *xprt = req->rq_xprt;
590 const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
593 if (time_before(jiffies, req->rq_majortimeo)) {
594 if (to->to_exponential)
595 req->rq_timeout <<= 1;
597 req->rq_timeout += to->to_increment;
598 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
599 req->rq_timeout = to->to_maxval;
602 req->rq_timeout = to->to_initval;
604 xprt_reset_majortimeo(req);
605 /* Reset the RTT counters == "slow start" */
606 spin_lock_bh(&xprt->transport_lock);
607 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
608 spin_unlock_bh(&xprt->transport_lock);
612 if (req->rq_timeout == 0) {
613 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
614 req->rq_timeout = 5 * HZ;
619 static void xprt_autoclose(struct work_struct *work)
621 struct rpc_xprt *xprt =
622 container_of(work, struct rpc_xprt, task_cleanup);
624 xprt->ops->close(xprt);
625 clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
626 xprt_release_write(xprt, NULL);
630 * xprt_disconnect_done - mark a transport as disconnected
631 * @xprt: transport to flag for disconnect
634 void xprt_disconnect_done(struct rpc_xprt *xprt)
636 dprintk("RPC: disconnected transport %p\n", xprt);
637 spin_lock_bh(&xprt->transport_lock);
638 xprt_clear_connected(xprt);
639 xprt_wake_pending_tasks(xprt, -EAGAIN);
640 spin_unlock_bh(&xprt->transport_lock);
642 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
645 * xprt_force_disconnect - force a transport to disconnect
646 * @xprt: transport to disconnect
649 void xprt_force_disconnect(struct rpc_xprt *xprt)
651 /* Don't race with the test_bit() in xprt_clear_locked() */
652 spin_lock_bh(&xprt->transport_lock);
653 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
654 /* Try to schedule an autoclose RPC call */
655 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
656 queue_work(rpciod_workqueue, &xprt->task_cleanup);
657 xprt_wake_pending_tasks(xprt, -EAGAIN);
658 spin_unlock_bh(&xprt->transport_lock);
662 * xprt_conditional_disconnect - force a transport to disconnect
663 * @xprt: transport to disconnect
664 * @cookie: 'connection cookie'
666 * This attempts to break the connection if and only if 'cookie' matches
667 * the current transport 'connection cookie'. It ensures that we don't
668 * try to break the connection more than once when we need to retransmit
669 * a batch of RPC requests.
672 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
674 /* Don't race with the test_bit() in xprt_clear_locked() */
675 spin_lock_bh(&xprt->transport_lock);
676 if (cookie != xprt->connect_cookie)
678 if (test_bit(XPRT_CLOSING, &xprt->state) || !xprt_connected(xprt))
680 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
681 /* Try to schedule an autoclose RPC call */
682 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
683 queue_work(rpciod_workqueue, &xprt->task_cleanup);
684 xprt_wake_pending_tasks(xprt, -EAGAIN);
686 spin_unlock_bh(&xprt->transport_lock);
690 xprt_init_autodisconnect(unsigned long data)
692 struct rpc_xprt *xprt = (struct rpc_xprt *)data;
694 spin_lock(&xprt->transport_lock);
695 if (!list_empty(&xprt->recv))
697 if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
699 spin_unlock(&xprt->transport_lock);
700 set_bit(XPRT_CONNECTION_CLOSE, &xprt->state);
701 queue_work(rpciod_workqueue, &xprt->task_cleanup);
704 spin_unlock(&xprt->transport_lock);
708 * xprt_connect - schedule a transport connect operation
709 * @task: RPC task that is requesting the connect
712 void xprt_connect(struct rpc_task *task)
714 struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
716 dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
717 xprt, (xprt_connected(xprt) ? "is" : "is not"));
719 if (!xprt_bound(xprt)) {
720 task->tk_status = -EAGAIN;
723 if (!xprt_lock_write(xprt, task))
726 if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state))
727 xprt->ops->close(xprt);
729 if (xprt_connected(xprt))
730 xprt_release_write(xprt, task);
732 task->tk_rqstp->rq_bytes_sent = 0;
733 task->tk_timeout = task->tk_rqstp->rq_timeout;
734 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
736 if (test_bit(XPRT_CLOSING, &xprt->state))
738 if (xprt_test_and_set_connecting(xprt))
740 xprt->stat.connect_start = jiffies;
741 xprt->ops->connect(xprt, task);
745 static void xprt_connect_status(struct rpc_task *task)
747 struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
749 if (task->tk_status == 0) {
750 xprt->stat.connect_count++;
751 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
752 dprintk("RPC: %5u xprt_connect_status: connection established\n",
757 switch (task->tk_status) {
759 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
762 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
763 "out\n", task->tk_pid);
766 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
767 "server %s\n", task->tk_pid, -task->tk_status,
769 xprt_release_write(xprt, task);
770 task->tk_status = -EIO;
775 * xprt_lookup_rqst - find an RPC request corresponding to an XID
776 * @xprt: transport on which the original request was transmitted
777 * @xid: RPC XID of incoming reply
780 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
782 struct rpc_rqst *entry;
784 list_for_each_entry(entry, &xprt->recv, rq_list)
785 if (entry->rq_xid == xid)
788 dprintk("RPC: xprt_lookup_rqst did not find xid %08x\n",
790 xprt->stat.bad_xids++;
793 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
795 static void xprt_update_rtt(struct rpc_task *task)
797 struct rpc_rqst *req = task->tk_rqstp;
798 struct rpc_rtt *rtt = task->tk_client->cl_rtt;
799 unsigned int timer = task->tk_msg.rpc_proc->p_timer;
800 long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt));
803 if (req->rq_ntrans == 1)
804 rpc_update_rtt(rtt, timer, m);
805 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
810 * xprt_complete_rqst - called when reply processing is complete
811 * @task: RPC request that recently completed
812 * @copied: actual number of bytes received from the transport
814 * Caller holds transport lock.
816 void xprt_complete_rqst(struct rpc_task *task, int copied)
818 struct rpc_rqst *req = task->tk_rqstp;
819 struct rpc_xprt *xprt = req->rq_xprt;
821 dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
822 task->tk_pid, ntohl(req->rq_xid), copied);
825 req->rq_rtt = ktime_sub(ktime_get(), req->rq_xtime);
826 if (xprt->ops->timer != NULL)
827 xprt_update_rtt(task);
829 list_del_init(&req->rq_list);
830 req->rq_private_buf.len = copied;
831 /* Ensure all writes are done before we update */
832 /* req->rq_reply_bytes_recvd */
834 req->rq_reply_bytes_recvd = copied;
835 rpc_wake_up_queued_task(&xprt->pending, task);
837 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
839 static void xprt_timer(struct rpc_task *task)
841 struct rpc_rqst *req = task->tk_rqstp;
842 struct rpc_xprt *xprt = req->rq_xprt;
844 if (task->tk_status != -ETIMEDOUT)
846 dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
848 spin_lock_bh(&xprt->transport_lock);
849 if (!req->rq_reply_bytes_recvd) {
850 if (xprt->ops->timer)
851 xprt->ops->timer(xprt, task);
854 spin_unlock_bh(&xprt->transport_lock);
857 static inline int xprt_has_timer(struct rpc_xprt *xprt)
859 return xprt->idle_timeout != 0;
863 * xprt_prepare_transmit - reserve the transport before sending a request
864 * @task: RPC task about to send a request
867 bool xprt_prepare_transmit(struct rpc_task *task)
869 struct rpc_rqst *req = task->tk_rqstp;
870 struct rpc_xprt *xprt = req->rq_xprt;
873 dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
875 spin_lock_bh(&xprt->transport_lock);
876 if (req->rq_reply_bytes_recvd && !req->rq_bytes_sent) {
877 task->tk_status = req->rq_reply_bytes_recvd;
880 if (!xprt->ops->reserve_xprt(xprt, task)) {
881 task->tk_status = -EAGAIN;
886 spin_unlock_bh(&xprt->transport_lock);
890 void xprt_end_transmit(struct rpc_task *task)
892 xprt_release_write(task->tk_rqstp->rq_xprt, task);
896 * xprt_transmit - send an RPC request on a transport
897 * @task: controlling RPC task
899 * We have to copy the iovec because sendmsg fiddles with its contents.
901 void xprt_transmit(struct rpc_task *task)
903 struct rpc_rqst *req = task->tk_rqstp;
904 struct rpc_xprt *xprt = req->rq_xprt;
907 dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
909 if (!req->rq_reply_bytes_recvd) {
910 if (list_empty(&req->rq_list) && rpc_reply_expected(task)) {
912 * Add to the list only if we're expecting a reply
914 spin_lock_bh(&xprt->transport_lock);
915 /* Update the softirq receive buffer */
916 memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
917 sizeof(req->rq_private_buf));
918 /* Add request to the receive list */
919 list_add_tail(&req->rq_list, &xprt->recv);
920 spin_unlock_bh(&xprt->transport_lock);
921 xprt_reset_majortimeo(req);
922 /* Turn off autodisconnect */
923 del_singleshot_timer_sync(&xprt->timer);
925 } else if (!req->rq_bytes_sent)
928 req->rq_xtime = ktime_get();
929 status = xprt->ops->send_request(task);
931 task->tk_status = status;
935 dprintk("RPC: %5u xmit complete\n", task->tk_pid);
936 task->tk_flags |= RPC_TASK_SENT;
937 spin_lock_bh(&xprt->transport_lock);
939 xprt->ops->set_retrans_timeout(task);
941 numreqs = atomic_read(&xprt->num_reqs);
942 if (numreqs > xprt->stat.max_slots)
943 xprt->stat.max_slots = numreqs;
945 xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
946 xprt->stat.bklog_u += xprt->backlog.qlen;
947 xprt->stat.sending_u += xprt->sending.qlen;
948 xprt->stat.pending_u += xprt->pending.qlen;
950 /* Don't race with disconnect */
951 if (!xprt_connected(xprt))
952 task->tk_status = -ENOTCONN;
955 * Sleep on the pending queue since
956 * we're expecting a reply.
958 if (!req->rq_reply_bytes_recvd && rpc_reply_expected(task))
959 rpc_sleep_on(&xprt->pending, task, xprt_timer);
960 req->rq_connect_cookie = xprt->connect_cookie;
962 spin_unlock_bh(&xprt->transport_lock);
965 static void xprt_add_backlog(struct rpc_xprt *xprt, struct rpc_task *task)
967 set_bit(XPRT_CONGESTED, &xprt->state);
968 rpc_sleep_on(&xprt->backlog, task, NULL);
971 static void xprt_wake_up_backlog(struct rpc_xprt *xprt)
973 if (rpc_wake_up_next(&xprt->backlog) == NULL)
974 clear_bit(XPRT_CONGESTED, &xprt->state);
977 static bool xprt_throttle_congested(struct rpc_xprt *xprt, struct rpc_task *task)
981 if (!test_bit(XPRT_CONGESTED, &xprt->state))
983 spin_lock(&xprt->reserve_lock);
984 if (test_bit(XPRT_CONGESTED, &xprt->state)) {
985 rpc_sleep_on(&xprt->backlog, task, NULL);
988 spin_unlock(&xprt->reserve_lock);
993 static struct rpc_rqst *xprt_dynamic_alloc_slot(struct rpc_xprt *xprt, gfp_t gfp_flags)
995 struct rpc_rqst *req = ERR_PTR(-EAGAIN);
997 if (!atomic_add_unless(&xprt->num_reqs, 1, xprt->max_reqs))
999 req = kzalloc(sizeof(struct rpc_rqst), gfp_flags);
1002 atomic_dec(&xprt->num_reqs);
1003 req = ERR_PTR(-ENOMEM);
1008 static bool xprt_dynamic_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1010 if (atomic_add_unless(&xprt->num_reqs, -1, xprt->min_reqs)) {
1017 void xprt_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1019 struct rpc_rqst *req;
1021 spin_lock(&xprt->reserve_lock);
1022 if (!list_empty(&xprt->free)) {
1023 req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
1024 list_del(&req->rq_list);
1027 req = xprt_dynamic_alloc_slot(xprt, GFP_NOWAIT|__GFP_NOWARN);
1030 switch (PTR_ERR(req)) {
1032 dprintk("RPC: dynamic allocation of request slot "
1033 "failed! Retrying\n");
1034 task->tk_status = -ENOMEM;
1037 xprt_add_backlog(xprt, task);
1038 dprintk("RPC: waiting for request slot\n");
1040 task->tk_status = -EAGAIN;
1042 spin_unlock(&xprt->reserve_lock);
1045 task->tk_status = 0;
1046 task->tk_rqstp = req;
1047 xprt_request_init(task, xprt);
1048 spin_unlock(&xprt->reserve_lock);
1050 EXPORT_SYMBOL_GPL(xprt_alloc_slot);
1052 void xprt_lock_and_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1054 /* Note: grabbing the xprt_lock_write() ensures that we throttle
1055 * new slot allocation if the transport is congested (i.e. when
1056 * reconnecting a stream transport or when out of socket write
1059 if (xprt_lock_write(xprt, task)) {
1060 xprt_alloc_slot(xprt, task);
1061 xprt_release_write(xprt, task);
1064 EXPORT_SYMBOL_GPL(xprt_lock_and_alloc_slot);
1066 static void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1068 spin_lock(&xprt->reserve_lock);
1069 if (!xprt_dynamic_free_slot(xprt, req)) {
1070 memset(req, 0, sizeof(*req)); /* mark unused */
1071 list_add(&req->rq_list, &xprt->free);
1073 xprt_wake_up_backlog(xprt);
1074 spin_unlock(&xprt->reserve_lock);
1077 static void xprt_free_all_slots(struct rpc_xprt *xprt)
1079 struct rpc_rqst *req;
1080 while (!list_empty(&xprt->free)) {
1081 req = list_first_entry(&xprt->free, struct rpc_rqst, rq_list);
1082 list_del(&req->rq_list);
1087 struct rpc_xprt *xprt_alloc(struct net *net, size_t size,
1088 unsigned int num_prealloc,
1089 unsigned int max_alloc)
1091 struct rpc_xprt *xprt;
1092 struct rpc_rqst *req;
1095 xprt = kzalloc(size, GFP_KERNEL);
1099 xprt_init(xprt, net);
1101 for (i = 0; i < num_prealloc; i++) {
1102 req = kzalloc(sizeof(struct rpc_rqst), GFP_KERNEL);
1105 list_add(&req->rq_list, &xprt->free);
1107 if (i < num_prealloc)
1109 if (max_alloc > num_prealloc)
1110 xprt->max_reqs = max_alloc;
1112 xprt->max_reqs = num_prealloc;
1113 xprt->min_reqs = num_prealloc;
1114 atomic_set(&xprt->num_reqs, num_prealloc);
1123 EXPORT_SYMBOL_GPL(xprt_alloc);
1125 void xprt_free(struct rpc_xprt *xprt)
1127 put_net(xprt->xprt_net);
1128 xprt_free_all_slots(xprt);
1131 EXPORT_SYMBOL_GPL(xprt_free);
1134 * xprt_reserve - allocate an RPC request slot
1135 * @task: RPC task requesting a slot allocation
1137 * If the transport is marked as being congested, or if no more
1138 * slots are available, place the task on the transport's
1141 void xprt_reserve(struct rpc_task *task)
1143 struct rpc_xprt *xprt;
1145 task->tk_status = 0;
1146 if (task->tk_rqstp != NULL)
1149 task->tk_timeout = 0;
1150 task->tk_status = -EAGAIN;
1152 xprt = rcu_dereference(task->tk_client->cl_xprt);
1153 if (!xprt_throttle_congested(xprt, task))
1154 xprt->ops->alloc_slot(xprt, task);
1159 * xprt_retry_reserve - allocate an RPC request slot
1160 * @task: RPC task requesting a slot allocation
1162 * If no more slots are available, place the task on the transport's
1164 * Note that the only difference with xprt_reserve is that we now
1165 * ignore the value of the XPRT_CONGESTED flag.
1167 void xprt_retry_reserve(struct rpc_task *task)
1169 struct rpc_xprt *xprt;
1171 task->tk_status = 0;
1172 if (task->tk_rqstp != NULL)
1175 task->tk_timeout = 0;
1176 task->tk_status = -EAGAIN;
1178 xprt = rcu_dereference(task->tk_client->cl_xprt);
1179 xprt->ops->alloc_slot(xprt, task);
1183 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
1185 return (__force __be32)xprt->xid++;
1188 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1190 xprt->xid = net_random();
1193 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1195 struct rpc_rqst *req = task->tk_rqstp;
1197 INIT_LIST_HEAD(&req->rq_list);
1198 req->rq_timeout = task->tk_client->cl_timeout->to_initval;
1199 req->rq_task = task;
1200 req->rq_xprt = xprt;
1201 req->rq_buffer = NULL;
1202 req->rq_xid = xprt_alloc_xid(xprt);
1203 req->rq_connect_cookie = xprt->connect_cookie - 1;
1204 req->rq_release_snd_buf = NULL;
1205 xprt_reset_majortimeo(req);
1206 dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
1207 req, ntohl(req->rq_xid));
1211 * xprt_release - release an RPC request slot
1212 * @task: task which is finished with the slot
1215 void xprt_release(struct rpc_task *task)
1217 struct rpc_xprt *xprt;
1218 struct rpc_rqst *req = task->tk_rqstp;
1221 if (task->tk_client) {
1223 xprt = rcu_dereference(task->tk_client->cl_xprt);
1224 if (xprt->snd_task == task)
1225 xprt_release_write(xprt, task);
1231 xprt = req->rq_xprt;
1232 if (task->tk_ops->rpc_count_stats != NULL)
1233 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
1234 else if (task->tk_client)
1235 rpc_count_iostats(task, task->tk_client->cl_metrics);
1236 spin_lock_bh(&xprt->transport_lock);
1237 xprt->ops->release_xprt(xprt, task);
1238 if (xprt->ops->release_request)
1239 xprt->ops->release_request(task);
1240 if (!list_empty(&req->rq_list))
1241 list_del(&req->rq_list);
1242 xprt->last_used = jiffies;
1243 if (list_empty(&xprt->recv) && xprt_has_timer(xprt))
1244 mod_timer(&xprt->timer,
1245 xprt->last_used + xprt->idle_timeout);
1246 spin_unlock_bh(&xprt->transport_lock);
1248 xprt->ops->buf_free(req->rq_buffer);
1249 if (req->rq_cred != NULL)
1250 put_rpccred(req->rq_cred);
1251 task->tk_rqstp = NULL;
1252 if (req->rq_release_snd_buf)
1253 req->rq_release_snd_buf(req);
1255 dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1256 if (likely(!bc_prealloc(req)))
1257 xprt_free_slot(xprt, req);
1259 xprt_free_bc_request(req);
1262 static void xprt_init(struct rpc_xprt *xprt, struct net *net)
1264 atomic_set(&xprt->count, 1);
1266 spin_lock_init(&xprt->transport_lock);
1267 spin_lock_init(&xprt->reserve_lock);
1269 INIT_LIST_HEAD(&xprt->free);
1270 INIT_LIST_HEAD(&xprt->recv);
1271 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1272 spin_lock_init(&xprt->bc_pa_lock);
1273 INIT_LIST_HEAD(&xprt->bc_pa_list);
1274 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1276 xprt->last_used = jiffies;
1277 xprt->cwnd = RPC_INITCWND;
1278 xprt->bind_index = 0;
1280 rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1281 rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1282 rpc_init_priority_wait_queue(&xprt->sending, "xprt_sending");
1283 rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1285 xprt_init_xid(xprt);
1287 xprt->xprt_net = get_net(net);
1291 * xprt_create_transport - create an RPC transport
1292 * @args: rpc transport creation arguments
1295 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1297 struct rpc_xprt *xprt;
1298 struct xprt_class *t;
1300 spin_lock(&xprt_list_lock);
1301 list_for_each_entry(t, &xprt_list, list) {
1302 if (t->ident == args->ident) {
1303 spin_unlock(&xprt_list_lock);
1307 spin_unlock(&xprt_list_lock);
1308 printk(KERN_ERR "RPC: transport (%d) not supported\n", args->ident);
1309 return ERR_PTR(-EIO);
1312 xprt = t->setup(args);
1314 dprintk("RPC: xprt_create_transport: failed, %ld\n",
1318 if (args->flags & XPRT_CREATE_NO_IDLE_TIMEOUT)
1319 xprt->idle_timeout = 0;
1320 INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1321 if (xprt_has_timer(xprt))
1322 setup_timer(&xprt->timer, xprt_init_autodisconnect,
1323 (unsigned long)xprt);
1325 init_timer(&xprt->timer);
1327 if (strlen(args->servername) > RPC_MAXNETNAMELEN) {
1329 return ERR_PTR(-EINVAL);
1331 xprt->servername = kstrdup(args->servername, GFP_KERNEL);
1332 if (xprt->servername == NULL) {
1334 return ERR_PTR(-ENOMEM);
1337 dprintk("RPC: created transport %p with %u slots\n", xprt,
1344 * xprt_destroy - destroy an RPC transport, killing off all requests.
1345 * @xprt: transport to destroy
1348 static void xprt_destroy(struct rpc_xprt *xprt)
1350 dprintk("RPC: destroying transport %p\n", xprt);
1351 del_timer_sync(&xprt->timer);
1353 rpc_destroy_wait_queue(&xprt->binding);
1354 rpc_destroy_wait_queue(&xprt->pending);
1355 rpc_destroy_wait_queue(&xprt->sending);
1356 rpc_destroy_wait_queue(&xprt->backlog);
1357 cancel_work_sync(&xprt->task_cleanup);
1358 kfree(xprt->servername);
1360 * Tear down transport state and free the rpc_xprt
1362 xprt->ops->destroy(xprt);
1366 * xprt_put - release a reference to an RPC transport.
1367 * @xprt: pointer to the transport
1370 void xprt_put(struct rpc_xprt *xprt)
1372 if (atomic_dec_and_test(&xprt->count))
1377 * xprt_get - return a reference to an RPC transport.
1378 * @xprt: pointer to the transport
1381 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1383 if (atomic_inc_not_zero(&xprt->count))