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) {
361 xprt_clear_locked(xprt);
362 __xprt_lock_write_next(xprt);
365 EXPORT_SYMBOL_GPL(xprt_release_xprt);
368 * xprt_release_xprt_cong - allow other requests to use a transport
369 * @xprt: transport with other tasks potentially waiting
370 * @task: task that is releasing access to the transport
372 * Note that "task" can be NULL. Another task is awoken to use the
373 * transport if the transport's congestion window allows it.
375 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
377 if (xprt->snd_task == task) {
378 xprt_clear_locked(xprt);
379 __xprt_lock_write_next_cong(xprt);
382 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
384 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
386 spin_lock_bh(&xprt->transport_lock);
387 xprt->ops->release_xprt(xprt, task);
388 spin_unlock_bh(&xprt->transport_lock);
392 * Van Jacobson congestion avoidance. Check if the congestion window
393 * overflowed. Put the task to sleep if this is the case.
396 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
398 struct rpc_rqst *req = task->tk_rqstp;
402 dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
403 task->tk_pid, xprt->cong, xprt->cwnd);
404 if (RPCXPRT_CONGESTED(xprt))
407 xprt->cong += RPC_CWNDSCALE;
412 * Adjust the congestion window, and wake up the next task
413 * that has been sleeping due to congestion
416 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
421 xprt->cong -= RPC_CWNDSCALE;
422 __xprt_lock_write_next_cong(xprt);
426 * xprt_release_rqst_cong - housekeeping when request is complete
427 * @task: RPC request that recently completed
429 * Useful for transports that require congestion control.
431 void xprt_release_rqst_cong(struct rpc_task *task)
433 __xprt_put_cong(task->tk_xprt, task->tk_rqstp);
435 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
438 * xprt_adjust_cwnd - adjust transport congestion window
439 * @task: recently completed RPC request used to adjust window
440 * @result: result code of completed RPC request
442 * We use a time-smoothed congestion estimator to avoid heavy oscillation.
444 void xprt_adjust_cwnd(struct rpc_task *task, int result)
446 struct rpc_rqst *req = task->tk_rqstp;
447 struct rpc_xprt *xprt = task->tk_xprt;
448 unsigned long cwnd = xprt->cwnd;
450 if (result >= 0 && cwnd <= xprt->cong) {
451 /* The (cwnd >> 1) term makes sure
452 * the result gets rounded properly. */
453 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
454 if (cwnd > RPC_MAXCWND(xprt))
455 cwnd = RPC_MAXCWND(xprt);
456 __xprt_lock_write_next_cong(xprt);
457 } else if (result == -ETIMEDOUT) {
459 if (cwnd < RPC_CWNDSCALE)
460 cwnd = RPC_CWNDSCALE;
462 dprintk("RPC: cong %ld, cwnd was %ld, now %ld\n",
463 xprt->cong, xprt->cwnd, cwnd);
465 __xprt_put_cong(xprt, req);
467 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
470 * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
471 * @xprt: transport with waiting tasks
472 * @status: result code to plant in each task before waking it
475 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
478 rpc_wake_up_status(&xprt->pending, status);
480 rpc_wake_up(&xprt->pending);
482 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
485 * xprt_wait_for_buffer_space - wait for transport output buffer to clear
486 * @task: task to be put to sleep
487 * @action: function pointer to be executed after wait
489 * Note that we only set the timer for the case of RPC_IS_SOFT(), since
490 * we don't in general want to force a socket disconnection due to
491 * an incomplete RPC call transmission.
493 void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action)
495 struct rpc_rqst *req = task->tk_rqstp;
496 struct rpc_xprt *xprt = req->rq_xprt;
498 task->tk_timeout = RPC_IS_SOFT(task) ? req->rq_timeout : 0;
499 rpc_sleep_on(&xprt->pending, task, action);
501 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
504 * xprt_write_space - wake the task waiting for transport output buffer space
505 * @xprt: transport with waiting tasks
507 * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
509 void xprt_write_space(struct rpc_xprt *xprt)
511 spin_lock_bh(&xprt->transport_lock);
512 if (xprt->snd_task) {
513 dprintk("RPC: write space: waking waiting task on "
515 rpc_wake_up_queued_task(&xprt->pending, xprt->snd_task);
517 spin_unlock_bh(&xprt->transport_lock);
519 EXPORT_SYMBOL_GPL(xprt_write_space);
522 * xprt_set_retrans_timeout_def - set a request's retransmit timeout
523 * @task: task whose timeout is to be set
525 * Set a request's retransmit timeout based on the transport's
526 * default timeout parameters. Used by transports that don't adjust
527 * the retransmit timeout based on round-trip time estimation.
529 void xprt_set_retrans_timeout_def(struct rpc_task *task)
531 task->tk_timeout = task->tk_rqstp->rq_timeout;
533 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
536 * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
537 * @task: task whose timeout is to be set
539 * Set a request's retransmit timeout using the RTT estimator.
541 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
543 int timer = task->tk_msg.rpc_proc->p_timer;
544 struct rpc_clnt *clnt = task->tk_client;
545 struct rpc_rtt *rtt = clnt->cl_rtt;
546 struct rpc_rqst *req = task->tk_rqstp;
547 unsigned long max_timeout = clnt->cl_timeout->to_maxval;
549 task->tk_timeout = rpc_calc_rto(rtt, timer);
550 task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
551 if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
552 task->tk_timeout = max_timeout;
554 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
556 static void xprt_reset_majortimeo(struct rpc_rqst *req)
558 const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
560 req->rq_majortimeo = req->rq_timeout;
561 if (to->to_exponential)
562 req->rq_majortimeo <<= to->to_retries;
564 req->rq_majortimeo += to->to_increment * to->to_retries;
565 if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
566 req->rq_majortimeo = to->to_maxval;
567 req->rq_majortimeo += jiffies;
571 * xprt_adjust_timeout - adjust timeout values for next retransmit
572 * @req: RPC request containing parameters to use for the adjustment
575 int xprt_adjust_timeout(struct rpc_rqst *req)
577 struct rpc_xprt *xprt = req->rq_xprt;
578 const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
581 if (time_before(jiffies, req->rq_majortimeo)) {
582 if (to->to_exponential)
583 req->rq_timeout <<= 1;
585 req->rq_timeout += to->to_increment;
586 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
587 req->rq_timeout = to->to_maxval;
590 req->rq_timeout = to->to_initval;
592 xprt_reset_majortimeo(req);
593 /* Reset the RTT counters == "slow start" */
594 spin_lock_bh(&xprt->transport_lock);
595 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
596 spin_unlock_bh(&xprt->transport_lock);
600 if (req->rq_timeout == 0) {
601 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
602 req->rq_timeout = 5 * HZ;
607 static void xprt_autoclose(struct work_struct *work)
609 struct rpc_xprt *xprt =
610 container_of(work, struct rpc_xprt, task_cleanup);
612 xprt->ops->close(xprt);
613 clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
614 xprt_release_write(xprt, NULL);
618 * xprt_disconnect_done - mark a transport as disconnected
619 * @xprt: transport to flag for disconnect
622 void xprt_disconnect_done(struct rpc_xprt *xprt)
624 dprintk("RPC: disconnected transport %p\n", xprt);
625 spin_lock_bh(&xprt->transport_lock);
626 xprt_clear_connected(xprt);
627 xprt_wake_pending_tasks(xprt, -EAGAIN);
628 spin_unlock_bh(&xprt->transport_lock);
630 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
633 * xprt_force_disconnect - force a transport to disconnect
634 * @xprt: transport to disconnect
637 void xprt_force_disconnect(struct rpc_xprt *xprt)
639 /* Don't race with the test_bit() in xprt_clear_locked() */
640 spin_lock_bh(&xprt->transport_lock);
641 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
642 /* Try to schedule an autoclose RPC call */
643 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
644 queue_work(rpciod_workqueue, &xprt->task_cleanup);
645 xprt_wake_pending_tasks(xprt, -EAGAIN);
646 spin_unlock_bh(&xprt->transport_lock);
650 * xprt_conditional_disconnect - force a transport to disconnect
651 * @xprt: transport to disconnect
652 * @cookie: 'connection cookie'
654 * This attempts to break the connection if and only if 'cookie' matches
655 * the current transport 'connection cookie'. It ensures that we don't
656 * try to break the connection more than once when we need to retransmit
657 * a batch of RPC requests.
660 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
662 /* Don't race with the test_bit() in xprt_clear_locked() */
663 spin_lock_bh(&xprt->transport_lock);
664 if (cookie != xprt->connect_cookie)
666 if (test_bit(XPRT_CLOSING, &xprt->state) || !xprt_connected(xprt))
668 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
669 /* Try to schedule an autoclose RPC call */
670 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
671 queue_work(rpciod_workqueue, &xprt->task_cleanup);
672 xprt_wake_pending_tasks(xprt, -EAGAIN);
674 spin_unlock_bh(&xprt->transport_lock);
678 xprt_init_autodisconnect(unsigned long data)
680 struct rpc_xprt *xprt = (struct rpc_xprt *)data;
682 spin_lock(&xprt->transport_lock);
683 if (!list_empty(&xprt->recv))
685 if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
687 spin_unlock(&xprt->transport_lock);
688 set_bit(XPRT_CONNECTION_CLOSE, &xprt->state);
689 queue_work(rpciod_workqueue, &xprt->task_cleanup);
692 spin_unlock(&xprt->transport_lock);
696 * xprt_connect - schedule a transport connect operation
697 * @task: RPC task that is requesting the connect
700 void xprt_connect(struct rpc_task *task)
702 struct rpc_xprt *xprt = task->tk_xprt;
704 dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
705 xprt, (xprt_connected(xprt) ? "is" : "is not"));
707 if (!xprt_bound(xprt)) {
708 task->tk_status = -EAGAIN;
711 if (!xprt_lock_write(xprt, task))
714 if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state))
715 xprt->ops->close(xprt);
717 if (xprt_connected(xprt))
718 xprt_release_write(xprt, task);
720 task->tk_rqstp->rq_bytes_sent = 0;
721 task->tk_timeout = task->tk_rqstp->rq_timeout;
722 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
724 if (test_bit(XPRT_CLOSING, &xprt->state))
726 if (xprt_test_and_set_connecting(xprt))
728 xprt->stat.connect_start = jiffies;
729 xprt->ops->connect(task);
733 static void xprt_connect_status(struct rpc_task *task)
735 struct rpc_xprt *xprt = task->tk_xprt;
737 if (task->tk_status == 0) {
738 xprt->stat.connect_count++;
739 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
740 dprintk("RPC: %5u xprt_connect_status: connection established\n",
745 switch (task->tk_status) {
747 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
750 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
751 "out\n", task->tk_pid);
754 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
755 "server %s\n", task->tk_pid, -task->tk_status,
757 xprt_release_write(xprt, task);
758 task->tk_status = -EIO;
763 * xprt_lookup_rqst - find an RPC request corresponding to an XID
764 * @xprt: transport on which the original request was transmitted
765 * @xid: RPC XID of incoming reply
768 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
770 struct rpc_rqst *entry;
772 list_for_each_entry(entry, &xprt->recv, rq_list)
773 if (entry->rq_xid == xid)
776 dprintk("RPC: xprt_lookup_rqst did not find xid %08x\n",
778 xprt->stat.bad_xids++;
781 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
783 static void xprt_update_rtt(struct rpc_task *task)
785 struct rpc_rqst *req = task->tk_rqstp;
786 struct rpc_rtt *rtt = task->tk_client->cl_rtt;
787 unsigned int timer = task->tk_msg.rpc_proc->p_timer;
788 long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt));
791 if (req->rq_ntrans == 1)
792 rpc_update_rtt(rtt, timer, m);
793 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
798 * xprt_complete_rqst - called when reply processing is complete
799 * @task: RPC request that recently completed
800 * @copied: actual number of bytes received from the transport
802 * Caller holds transport lock.
804 void xprt_complete_rqst(struct rpc_task *task, int copied)
806 struct rpc_rqst *req = task->tk_rqstp;
807 struct rpc_xprt *xprt = req->rq_xprt;
809 dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
810 task->tk_pid, ntohl(req->rq_xid), copied);
813 req->rq_rtt = ktime_sub(ktime_get(), req->rq_xtime);
814 if (xprt->ops->timer != NULL)
815 xprt_update_rtt(task);
817 list_del_init(&req->rq_list);
818 req->rq_private_buf.len = copied;
819 /* Ensure all writes are done before we update */
820 /* req->rq_reply_bytes_recvd */
822 req->rq_reply_bytes_recvd = copied;
823 rpc_wake_up_queued_task(&xprt->pending, task);
825 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
827 static void xprt_timer(struct rpc_task *task)
829 struct rpc_rqst *req = task->tk_rqstp;
830 struct rpc_xprt *xprt = req->rq_xprt;
832 if (task->tk_status != -ETIMEDOUT)
834 dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
836 spin_lock_bh(&xprt->transport_lock);
837 if (!req->rq_reply_bytes_recvd) {
838 if (xprt->ops->timer)
839 xprt->ops->timer(task);
842 spin_unlock_bh(&xprt->transport_lock);
845 static inline int xprt_has_timer(struct rpc_xprt *xprt)
847 return xprt->idle_timeout != 0;
851 * xprt_prepare_transmit - reserve the transport before sending a request
852 * @task: RPC task about to send a request
855 int xprt_prepare_transmit(struct rpc_task *task)
857 struct rpc_rqst *req = task->tk_rqstp;
858 struct rpc_xprt *xprt = req->rq_xprt;
861 dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
863 spin_lock_bh(&xprt->transport_lock);
864 if (req->rq_reply_bytes_recvd && !req->rq_bytes_sent) {
865 err = req->rq_reply_bytes_recvd;
868 if (!xprt->ops->reserve_xprt(xprt, task))
871 spin_unlock_bh(&xprt->transport_lock);
875 void xprt_end_transmit(struct rpc_task *task)
877 xprt_release_write(task->tk_rqstp->rq_xprt, task);
881 * xprt_transmit - send an RPC request on a transport
882 * @task: controlling RPC task
884 * We have to copy the iovec because sendmsg fiddles with its contents.
886 void xprt_transmit(struct rpc_task *task)
888 struct rpc_rqst *req = task->tk_rqstp;
889 struct rpc_xprt *xprt = req->rq_xprt;
892 dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
894 if (!req->rq_reply_bytes_recvd) {
895 if (list_empty(&req->rq_list) && rpc_reply_expected(task)) {
897 * Add to the list only if we're expecting a reply
899 spin_lock_bh(&xprt->transport_lock);
900 /* Update the softirq receive buffer */
901 memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
902 sizeof(req->rq_private_buf));
903 /* Add request to the receive list */
904 list_add_tail(&req->rq_list, &xprt->recv);
905 spin_unlock_bh(&xprt->transport_lock);
906 xprt_reset_majortimeo(req);
907 /* Turn off autodisconnect */
908 del_singleshot_timer_sync(&xprt->timer);
910 } else if (!req->rq_bytes_sent)
913 req->rq_connect_cookie = xprt->connect_cookie;
914 req->rq_xtime = ktime_get();
915 status = xprt->ops->send_request(task);
917 task->tk_status = status;
921 dprintk("RPC: %5u xmit complete\n", task->tk_pid);
922 task->tk_flags |= RPC_TASK_SENT;
923 spin_lock_bh(&xprt->transport_lock);
925 xprt->ops->set_retrans_timeout(task);
927 numreqs = atomic_read(&xprt->num_reqs);
928 if (numreqs > xprt->stat.max_slots)
929 xprt->stat.max_slots = numreqs;
931 xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
932 xprt->stat.bklog_u += xprt->backlog.qlen;
933 xprt->stat.sending_u += xprt->sending.qlen;
934 xprt->stat.pending_u += xprt->pending.qlen;
936 /* Don't race with disconnect */
937 if (!xprt_connected(xprt))
938 task->tk_status = -ENOTCONN;
939 else if (!req->rq_reply_bytes_recvd && rpc_reply_expected(task)) {
941 * Sleep on the pending queue since
942 * we're expecting a reply.
944 rpc_sleep_on(&xprt->pending, task, xprt_timer);
946 spin_unlock_bh(&xprt->transport_lock);
949 static struct rpc_rqst *xprt_dynamic_alloc_slot(struct rpc_xprt *xprt, gfp_t gfp_flags)
951 struct rpc_rqst *req = ERR_PTR(-EAGAIN);
953 if (!atomic_add_unless(&xprt->num_reqs, 1, xprt->max_reqs))
955 req = kzalloc(sizeof(struct rpc_rqst), gfp_flags);
958 atomic_dec(&xprt->num_reqs);
959 req = ERR_PTR(-ENOMEM);
964 static bool xprt_dynamic_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
966 if (atomic_add_unless(&xprt->num_reqs, -1, xprt->min_reqs)) {
973 void xprt_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
975 struct rpc_rqst *req;
977 spin_lock(&xprt->reserve_lock);
978 if (!list_empty(&xprt->free)) {
979 req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
980 list_del(&req->rq_list);
983 req = xprt_dynamic_alloc_slot(xprt, GFP_NOWAIT|__GFP_NOWARN);
986 switch (PTR_ERR(req)) {
988 dprintk("RPC: dynamic allocation of request slot "
989 "failed! Retrying\n");
990 task->tk_status = -ENOMEM;
993 rpc_sleep_on(&xprt->backlog, task, NULL);
994 dprintk("RPC: waiting for request slot\n");
996 task->tk_status = -EAGAIN;
998 spin_unlock(&xprt->reserve_lock);
1001 task->tk_status = 0;
1002 task->tk_rqstp = req;
1003 xprt_request_init(task, xprt);
1004 spin_unlock(&xprt->reserve_lock);
1006 EXPORT_SYMBOL_GPL(xprt_alloc_slot);
1008 void xprt_lock_and_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1010 /* Note: grabbing the xprt_lock_write() ensures that we throttle
1011 * new slot allocation if the transport is congested (i.e. when
1012 * reconnecting a stream transport or when out of socket write
1015 if (xprt_lock_write(xprt, task)) {
1016 xprt_alloc_slot(xprt, task);
1017 xprt_release_write(xprt, task);
1020 EXPORT_SYMBOL_GPL(xprt_lock_and_alloc_slot);
1022 static void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1024 spin_lock(&xprt->reserve_lock);
1025 if (!xprt_dynamic_free_slot(xprt, req)) {
1026 memset(req, 0, sizeof(*req)); /* mark unused */
1027 list_add(&req->rq_list, &xprt->free);
1029 rpc_wake_up_next(&xprt->backlog);
1030 spin_unlock(&xprt->reserve_lock);
1033 static void xprt_free_all_slots(struct rpc_xprt *xprt)
1035 struct rpc_rqst *req;
1036 while (!list_empty(&xprt->free)) {
1037 req = list_first_entry(&xprt->free, struct rpc_rqst, rq_list);
1038 list_del(&req->rq_list);
1043 struct rpc_xprt *xprt_alloc(struct net *net, size_t size,
1044 unsigned int num_prealloc,
1045 unsigned int max_alloc)
1047 struct rpc_xprt *xprt;
1048 struct rpc_rqst *req;
1051 xprt = kzalloc(size, GFP_KERNEL);
1055 xprt_init(xprt, net);
1057 for (i = 0; i < num_prealloc; i++) {
1058 req = kzalloc(sizeof(struct rpc_rqst), GFP_KERNEL);
1061 list_add(&req->rq_list, &xprt->free);
1063 if (i < num_prealloc)
1065 if (max_alloc > num_prealloc)
1066 xprt->max_reqs = max_alloc;
1068 xprt->max_reqs = num_prealloc;
1069 xprt->min_reqs = num_prealloc;
1070 atomic_set(&xprt->num_reqs, num_prealloc);
1079 EXPORT_SYMBOL_GPL(xprt_alloc);
1081 void xprt_free(struct rpc_xprt *xprt)
1083 put_net(xprt->xprt_net);
1084 xprt_free_all_slots(xprt);
1087 EXPORT_SYMBOL_GPL(xprt_free);
1090 * xprt_reserve - allocate an RPC request slot
1091 * @task: RPC task requesting a slot allocation
1093 * If no more slots are available, place the task on the transport's
1096 void xprt_reserve(struct rpc_task *task)
1098 struct rpc_xprt *xprt = task->tk_xprt;
1100 task->tk_status = 0;
1101 if (task->tk_rqstp != NULL)
1104 task->tk_timeout = 0;
1105 task->tk_status = -EAGAIN;
1106 xprt->ops->alloc_slot(xprt, task);
1109 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
1111 return (__force __be32)xprt->xid++;
1114 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1116 xprt->xid = net_random();
1119 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1121 struct rpc_rqst *req = task->tk_rqstp;
1123 INIT_LIST_HEAD(&req->rq_list);
1124 req->rq_timeout = task->tk_client->cl_timeout->to_initval;
1125 req->rq_task = task;
1126 req->rq_xprt = xprt;
1127 req->rq_buffer = NULL;
1128 req->rq_xid = xprt_alloc_xid(xprt);
1129 req->rq_release_snd_buf = NULL;
1130 xprt_reset_majortimeo(req);
1131 dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
1132 req, ntohl(req->rq_xid));
1136 * xprt_release - release an RPC request slot
1137 * @task: task which is finished with the slot
1140 void xprt_release(struct rpc_task *task)
1142 struct rpc_xprt *xprt;
1143 struct rpc_rqst *req = task->tk_rqstp;
1146 if (task->tk_client) {
1148 xprt = rcu_dereference(task->tk_client->cl_xprt);
1149 if (xprt->snd_task == task)
1150 xprt_release_write(xprt, task);
1156 xprt = req->rq_xprt;
1157 if (task->tk_ops->rpc_count_stats != NULL)
1158 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
1159 else if (task->tk_client)
1160 rpc_count_iostats(task, task->tk_client->cl_metrics);
1161 spin_lock_bh(&xprt->transport_lock);
1162 xprt->ops->release_xprt(xprt, task);
1163 if (xprt->ops->release_request)
1164 xprt->ops->release_request(task);
1165 if (!list_empty(&req->rq_list))
1166 list_del(&req->rq_list);
1167 xprt->last_used = jiffies;
1168 if (list_empty(&xprt->recv) && xprt_has_timer(xprt))
1169 mod_timer(&xprt->timer,
1170 xprt->last_used + xprt->idle_timeout);
1171 spin_unlock_bh(&xprt->transport_lock);
1173 xprt->ops->buf_free(req->rq_buffer);
1174 if (req->rq_cred != NULL)
1175 put_rpccred(req->rq_cred);
1176 task->tk_rqstp = NULL;
1177 if (req->rq_release_snd_buf)
1178 req->rq_release_snd_buf(req);
1180 dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1181 if (likely(!bc_prealloc(req)))
1182 xprt_free_slot(xprt, req);
1184 xprt_free_bc_request(req);
1187 static void xprt_init(struct rpc_xprt *xprt, struct net *net)
1189 atomic_set(&xprt->count, 1);
1191 spin_lock_init(&xprt->transport_lock);
1192 spin_lock_init(&xprt->reserve_lock);
1194 INIT_LIST_HEAD(&xprt->free);
1195 INIT_LIST_HEAD(&xprt->recv);
1196 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1197 spin_lock_init(&xprt->bc_pa_lock);
1198 INIT_LIST_HEAD(&xprt->bc_pa_list);
1199 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1201 xprt->last_used = jiffies;
1202 xprt->cwnd = RPC_INITCWND;
1203 xprt->bind_index = 0;
1205 rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1206 rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1207 rpc_init_priority_wait_queue(&xprt->sending, "xprt_sending");
1208 rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1210 xprt_init_xid(xprt);
1212 xprt->xprt_net = get_net(net);
1216 * xprt_create_transport - create an RPC transport
1217 * @args: rpc transport creation arguments
1220 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1222 struct rpc_xprt *xprt;
1223 struct xprt_class *t;
1225 spin_lock(&xprt_list_lock);
1226 list_for_each_entry(t, &xprt_list, list) {
1227 if (t->ident == args->ident) {
1228 spin_unlock(&xprt_list_lock);
1232 spin_unlock(&xprt_list_lock);
1233 printk(KERN_ERR "RPC: transport (%d) not supported\n", args->ident);
1234 return ERR_PTR(-EIO);
1237 xprt = t->setup(args);
1239 dprintk("RPC: xprt_create_transport: failed, %ld\n",
1243 INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1244 if (xprt_has_timer(xprt))
1245 setup_timer(&xprt->timer, xprt_init_autodisconnect,
1246 (unsigned long)xprt);
1248 init_timer(&xprt->timer);
1250 if (strlen(args->servername) > RPC_MAXNETNAMELEN) {
1252 return ERR_PTR(-EINVAL);
1254 xprt->servername = kstrdup(args->servername, GFP_KERNEL);
1255 if (xprt->servername == NULL) {
1257 return ERR_PTR(-ENOMEM);
1260 dprintk("RPC: created transport %p with %u slots\n", xprt,
1267 * xprt_destroy - destroy an RPC transport, killing off all requests.
1268 * @xprt: transport to destroy
1271 static void xprt_destroy(struct rpc_xprt *xprt)
1273 dprintk("RPC: destroying transport %p\n", xprt);
1274 del_timer_sync(&xprt->timer);
1276 rpc_destroy_wait_queue(&xprt->binding);
1277 rpc_destroy_wait_queue(&xprt->pending);
1278 rpc_destroy_wait_queue(&xprt->sending);
1279 rpc_destroy_wait_queue(&xprt->backlog);
1280 cancel_work_sync(&xprt->task_cleanup);
1281 kfree(xprt->servername);
1283 * Tear down transport state and free the rpc_xprt
1285 xprt->ops->destroy(xprt);
1289 * xprt_put - release a reference to an RPC transport.
1290 * @xprt: pointer to the transport
1293 void xprt_put(struct rpc_xprt *xprt)
1295 if (atomic_dec_and_test(&xprt->count))
1300 * xprt_get - return a reference to an RPC transport.
1301 * @xprt: pointer to the transport
1304 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1306 if (atomic_inc_not_zero(&xprt->count))