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_call().
14 * - xprt_call transmits the message and installs the caller on the
15 * socket's wait list. At the same time, it installs a timer that
16 * is run after the packet's timeout has expired.
17 * - When a packet arrives, the data_ready handler walks the list of
18 * pending requests for that socket. If a matching XID is found, the
19 * caller is woken up, and the timer removed.
20 * - When no reply arrives within the timeout interval, the timer is
21 * fired by the kernel and runs xprt_timer(). It either adjusts the
22 * timeout values (minor timeout) or wakes up the caller with a status
24 * - When the caller receives a notification from RPC that a reply arrived,
25 * it should release the RPC slot, and process the reply.
26 * If the call timed out, it may choose to retry the operation by
27 * adjusting the initial timeout value, and simply calling rpc_call
30 * Support for async RPC is done through a set of RPC-specific scheduling
31 * primitives that `transparently' work for processes as well as async
32 * tasks that rely on callbacks.
34 * Copyright (C) 1995-1997, Olaf Kirch <okir@monad.swb.de>
37 #include <linux/module.h>
39 #include <linux/types.h>
40 #include <linux/interrupt.h>
41 #include <linux/workqueue.h>
42 #include <linux/random.h>
44 #include <linux/sunrpc/clnt.h>
51 # undef RPC_DEBUG_DATA
52 # define RPCDBG_FACILITY RPCDBG_XPRT
55 #define XPRT_MAX_BACKOFF (8)
60 static void xprt_request_init(struct rpc_task *, struct rpc_xprt *);
61 static inline void do_xprt_reserve(struct rpc_task *);
62 static void xprt_connect_status(struct rpc_task *task);
63 static int __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
65 static int xprt_clear_backlog(struct rpc_xprt *xprt);
68 * Serialize write access to sockets, in order to prevent different
69 * requests from interfering with each other.
70 * Also prevents TCP socket connects from colliding with writes.
73 __xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
75 struct rpc_rqst *req = task->tk_rqstp;
77 if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate)) {
78 if (task == xprt->snd_task)
82 if (xprt->nocong || __xprt_get_cong(xprt, task)) {
83 xprt->snd_task = task;
85 req->rq_bytes_sent = 0;
90 smp_mb__before_clear_bit();
91 clear_bit(XPRT_LOCKED, &xprt->sockstate);
92 smp_mb__after_clear_bit();
94 dprintk("RPC: %4d failed to lock socket %p\n", task->tk_pid, xprt);
96 task->tk_status = -EAGAIN;
97 if (req && req->rq_ntrans)
98 rpc_sleep_on(&xprt->resend, task, NULL, NULL);
100 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
105 xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
109 spin_lock_bh(&xprt->transport_lock);
110 retval = __xprt_lock_write(xprt, task);
111 spin_unlock_bh(&xprt->transport_lock);
117 __xprt_lock_write_next(struct rpc_xprt *xprt)
119 struct rpc_task *task;
121 if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
123 if (!xprt->nocong && RPCXPRT_CONGESTED(xprt))
125 task = rpc_wake_up_next(&xprt->resend);
127 task = rpc_wake_up_next(&xprt->sending);
131 if (xprt->nocong || __xprt_get_cong(xprt, task)) {
132 struct rpc_rqst *req = task->tk_rqstp;
133 xprt->snd_task = task;
135 req->rq_bytes_sent = 0;
141 smp_mb__before_clear_bit();
142 clear_bit(XPRT_LOCKED, &xprt->sockstate);
143 smp_mb__after_clear_bit();
147 * Releases the socket for use by other requests.
150 __xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
152 if (xprt->snd_task == task) {
153 xprt->snd_task = NULL;
154 smp_mb__before_clear_bit();
155 clear_bit(XPRT_LOCKED, &xprt->sockstate);
156 smp_mb__after_clear_bit();
157 __xprt_lock_write_next(xprt);
162 xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
164 spin_lock_bh(&xprt->transport_lock);
165 __xprt_release_write(xprt, task);
166 spin_unlock_bh(&xprt->transport_lock);
170 * Van Jacobson congestion avoidance. Check if the congestion window
171 * overflowed. Put the task to sleep if this is the case.
174 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
176 struct rpc_rqst *req = task->tk_rqstp;
180 dprintk("RPC: %4d xprt_cwnd_limited cong = %ld cwnd = %ld\n",
181 task->tk_pid, xprt->cong, xprt->cwnd);
182 if (RPCXPRT_CONGESTED(xprt))
185 xprt->cong += RPC_CWNDSCALE;
190 * Adjust the congestion window, and wake up the next task
191 * that has been sleeping due to congestion
194 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
199 xprt->cong -= RPC_CWNDSCALE;
200 __xprt_lock_write_next(xprt);
204 * Adjust RPC congestion window
205 * We use a time-smoothed congestion estimator to avoid heavy oscillation.
208 xprt_adjust_cwnd(struct rpc_xprt *xprt, int result)
213 if (result >= 0 && cwnd <= xprt->cong) {
214 /* The (cwnd >> 1) term makes sure
215 * the result gets rounded properly. */
216 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
217 if (cwnd > RPC_MAXCWND(xprt))
218 cwnd = RPC_MAXCWND(xprt);
219 __xprt_lock_write_next(xprt);
220 } else if (result == -ETIMEDOUT) {
222 if (cwnd < RPC_CWNDSCALE)
223 cwnd = RPC_CWNDSCALE;
225 dprintk("RPC: cong %ld, cwnd was %ld, now %ld\n",
226 xprt->cong, xprt->cwnd, cwnd);
230 static void xprt_reset_majortimeo(struct rpc_rqst *req)
232 struct rpc_timeout *to = &req->rq_xprt->timeout;
234 req->rq_majortimeo = req->rq_timeout;
235 if (to->to_exponential)
236 req->rq_majortimeo <<= to->to_retries;
238 req->rq_majortimeo += to->to_increment * to->to_retries;
239 if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
240 req->rq_majortimeo = to->to_maxval;
241 req->rq_majortimeo += jiffies;
245 * xprt_adjust_timeout - adjust timeout values for next retransmit
246 * @req: RPC request containing parameters to use for the adjustment
249 int xprt_adjust_timeout(struct rpc_rqst *req)
251 struct rpc_xprt *xprt = req->rq_xprt;
252 struct rpc_timeout *to = &xprt->timeout;
255 if (time_before(jiffies, req->rq_majortimeo)) {
256 if (to->to_exponential)
257 req->rq_timeout <<= 1;
259 req->rq_timeout += to->to_increment;
260 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
261 req->rq_timeout = to->to_maxval;
263 pprintk("RPC: %lu retrans\n", jiffies);
265 req->rq_timeout = to->to_initval;
267 xprt_reset_majortimeo(req);
268 /* Reset the RTT counters == "slow start" */
269 spin_lock_bh(&xprt->transport_lock);
270 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
271 spin_unlock_bh(&xprt->transport_lock);
272 pprintk("RPC: %lu timeout\n", jiffies);
276 if (req->rq_timeout == 0) {
277 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
278 req->rq_timeout = 5 * HZ;
284 xprt_socket_autoclose(void *args)
286 struct rpc_xprt *xprt = (struct rpc_xprt *)args;
288 xprt_disconnect(xprt);
289 xprt->ops->close(xprt);
290 xprt_release_write(xprt, NULL);
294 * xprt_disconnect - mark a transport as disconnected
295 * @xprt: transport to flag for disconnect
298 void xprt_disconnect(struct rpc_xprt *xprt)
300 dprintk("RPC: disconnected transport %p\n", xprt);
301 spin_lock_bh(&xprt->transport_lock);
302 xprt_clear_connected(xprt);
303 rpc_wake_up_status(&xprt->pending, -ENOTCONN);
304 spin_unlock_bh(&xprt->transport_lock);
308 xprt_init_autodisconnect(unsigned long data)
310 struct rpc_xprt *xprt = (struct rpc_xprt *)data;
312 spin_lock(&xprt->transport_lock);
313 if (!list_empty(&xprt->recv) || xprt->shutdown)
315 if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
317 spin_unlock(&xprt->transport_lock);
318 /* Let keventd close the socket */
319 if (test_bit(XPRT_CONNECTING, &xprt->sockstate) != 0)
320 xprt_release_write(xprt, NULL);
322 schedule_work(&xprt->task_cleanup);
325 spin_unlock(&xprt->transport_lock);
329 * xprt_connect - schedule a transport connect operation
330 * @task: RPC task that is requesting the connect
333 void xprt_connect(struct rpc_task *task)
335 struct rpc_xprt *xprt = task->tk_xprt;
337 dprintk("RPC: %4d xprt_connect xprt %p %s connected\n", task->tk_pid,
338 xprt, (xprt_connected(xprt) ? "is" : "is not"));
340 if (xprt->shutdown) {
341 task->tk_status = -EIO;
344 if (!xprt->addr.sin_port) {
345 task->tk_status = -EIO;
348 if (!xprt_lock_write(xprt, task))
350 if (xprt_connected(xprt))
351 xprt_release_write(xprt, task);
354 task->tk_rqstp->rq_bytes_sent = 0;
356 task->tk_timeout = RPC_CONNECT_TIMEOUT;
357 rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
358 xprt->ops->connect(task);
363 static void xprt_connect_status(struct rpc_task *task)
365 struct rpc_xprt *xprt = task->tk_xprt;
367 if (task->tk_status >= 0) {
368 dprintk("RPC: %4d xprt_connect_status: connection established\n",
373 switch (task->tk_status) {
376 dprintk("RPC: %4d xprt_connect_status: server %s refused connection\n",
377 task->tk_pid, task->tk_client->cl_server);
380 dprintk("RPC: %4d xprt_connect_status: connection broken\n",
384 dprintk("RPC: %4d xprt_connect_status: connect attempt timed out\n",
388 dprintk("RPC: %4d xprt_connect_status: error %d connecting to server %s\n",
389 task->tk_pid, -task->tk_status, task->tk_client->cl_server);
390 xprt_release_write(xprt, task);
391 task->tk_status = -EIO;
395 /* if soft mounted, just cause this RPC to fail */
396 if (RPC_IS_SOFT(task)) {
397 xprt_release_write(xprt, task);
398 task->tk_status = -EIO;
403 * xprt_lookup_rqst - find an RPC request corresponding to an XID
404 * @xprt: transport on which the original request was transmitted
405 * @xid: RPC XID of incoming reply
408 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
410 struct list_head *pos;
411 struct rpc_rqst *req = NULL;
413 list_for_each(pos, &xprt->recv) {
414 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
415 if (entry->rq_xid == xid) {
424 * xprt_complete_rqst - called when reply processing is complete
425 * @xprt: controlling transport
426 * @req: RPC request that just completed
427 * @copied: actual number of bytes received from the transport
430 void xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
432 struct rpc_task *task = req->rq_task;
433 struct rpc_clnt *clnt = task->tk_client;
435 /* Adjust congestion window */
437 unsigned timer = task->tk_msg.rpc_proc->p_timer;
438 xprt_adjust_cwnd(xprt, copied);
439 __xprt_put_cong(xprt, req);
441 if (req->rq_ntrans == 1)
442 rpc_update_rtt(clnt->cl_rtt, timer,
443 (long)jiffies - req->rq_xtime);
444 rpc_set_timeo(clnt->cl_rtt, timer, req->rq_ntrans - 1);
449 /* Profile only reads for now */
451 static unsigned long nextstat;
452 static unsigned long pkt_rtt, pkt_len, pkt_cnt;
455 pkt_len += req->rq_slen + copied;
456 pkt_rtt += jiffies - req->rq_xtime;
457 if (time_before(nextstat, jiffies)) {
458 printk("RPC: %lu %ld cwnd\n", jiffies, xprt->cwnd);
459 printk("RPC: %ld %ld %ld %ld stat\n",
460 jiffies, pkt_cnt, pkt_len, pkt_rtt);
461 pkt_rtt = pkt_len = pkt_cnt = 0;
462 nextstat = jiffies + 5 * HZ;
467 dprintk("RPC: %4d has input (%d bytes)\n", task->tk_pid, copied);
468 list_del_init(&req->rq_list);
469 req->rq_received = req->rq_private_buf.len = copied;
471 /* ... and wake up the process. */
472 rpc_wake_up_task(task);
477 * RPC receive timeout handler.
480 xprt_timer(struct rpc_task *task)
482 struct rpc_rqst *req = task->tk_rqstp;
483 struct rpc_xprt *xprt = req->rq_xprt;
485 spin_lock(&xprt->transport_lock);
486 if (req->rq_received)
489 xprt_adjust_cwnd(req->rq_xprt, -ETIMEDOUT);
490 __xprt_put_cong(xprt, req);
492 dprintk("RPC: %4d xprt_timer (%s request)\n",
493 task->tk_pid, req ? "pending" : "backlogged");
495 task->tk_status = -ETIMEDOUT;
497 task->tk_timeout = 0;
498 rpc_wake_up_task(task);
499 spin_unlock(&xprt->transport_lock);
503 * xprt_prepare_transmit - reserve the transport before sending a request
504 * @task: RPC task about to send a request
507 int xprt_prepare_transmit(struct rpc_task *task)
509 struct rpc_rqst *req = task->tk_rqstp;
510 struct rpc_xprt *xprt = req->rq_xprt;
513 dprintk("RPC: %4d xprt_prepare_transmit\n", task->tk_pid);
518 spin_lock_bh(&xprt->transport_lock);
519 if (req->rq_received && !req->rq_bytes_sent) {
520 err = req->rq_received;
523 if (!__xprt_lock_write(xprt, task)) {
528 if (!xprt_connected(xprt)) {
533 spin_unlock_bh(&xprt->transport_lock);
538 * xprt_transmit - send an RPC request on a transport
539 * @task: controlling RPC task
541 * We have to copy the iovec because sendmsg fiddles with its contents.
543 void xprt_transmit(struct rpc_task *task)
545 struct rpc_clnt *clnt = task->tk_client;
546 struct rpc_rqst *req = task->tk_rqstp;
547 struct rpc_xprt *xprt = req->rq_xprt;
550 dprintk("RPC: %4d xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
553 if (!req->rq_received) {
554 if (list_empty(&req->rq_list)) {
555 spin_lock_bh(&xprt->transport_lock);
556 /* Update the softirq receive buffer */
557 memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
558 sizeof(req->rq_private_buf));
559 /* Add request to the receive list */
560 list_add_tail(&req->rq_list, &xprt->recv);
561 spin_unlock_bh(&xprt->transport_lock);
562 xprt_reset_majortimeo(req);
563 /* Turn off autodisconnect */
564 del_singleshot_timer_sync(&xprt->timer);
566 } else if (!req->rq_bytes_sent)
569 status = xprt->ops->send_request(task);
573 /* Note: at this point, task->tk_sleeping has not yet been set,
574 * hence there is no danger of the waking up task being put on
575 * schedq, and being picked up by a parallel run of rpciod().
577 task->tk_status = status;
581 task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
582 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
588 xprt_disconnect(xprt);
590 xprt_release_write(xprt, task);
593 dprintk("RPC: %4d xmit complete\n", task->tk_pid);
594 /* Set the task's receive timeout value */
595 spin_lock_bh(&xprt->transport_lock);
597 int timer = task->tk_msg.rpc_proc->p_timer;
598 task->tk_timeout = rpc_calc_rto(clnt->cl_rtt, timer);
599 task->tk_timeout <<= rpc_ntimeo(clnt->cl_rtt, timer) + req->rq_retries;
600 if (task->tk_timeout > xprt->timeout.to_maxval || task->tk_timeout == 0)
601 task->tk_timeout = xprt->timeout.to_maxval;
603 task->tk_timeout = req->rq_timeout;
604 /* Don't race with disconnect */
605 if (!xprt_connected(xprt))
606 task->tk_status = -ENOTCONN;
607 else if (!req->rq_received)
608 rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
609 __xprt_release_write(xprt, task);
610 spin_unlock_bh(&xprt->transport_lock);
613 static inline void do_xprt_reserve(struct rpc_task *task)
615 struct rpc_xprt *xprt = task->tk_xprt;
620 if (!list_empty(&xprt->free)) {
621 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
622 list_del_init(&req->rq_list);
623 task->tk_rqstp = req;
624 xprt_request_init(task, xprt);
627 dprintk("RPC: waiting for request slot\n");
628 task->tk_status = -EAGAIN;
629 task->tk_timeout = 0;
630 rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
634 * xprt_reserve - allocate an RPC request slot
635 * @task: RPC task requesting a slot allocation
637 * If no more slots are available, place the task on the transport's
640 void xprt_reserve(struct rpc_task *task)
642 struct rpc_xprt *xprt = task->tk_xprt;
644 task->tk_status = -EIO;
645 if (!xprt->shutdown) {
646 spin_lock(&xprt->xprt_lock);
647 do_xprt_reserve(task);
648 spin_unlock(&xprt->xprt_lock);
652 static inline u32 xprt_alloc_xid(struct rpc_xprt *xprt)
657 static inline void xprt_init_xid(struct rpc_xprt *xprt)
659 get_random_bytes(&xprt->xid, sizeof(xprt->xid));
662 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
664 struct rpc_rqst *req = task->tk_rqstp;
666 req->rq_timeout = xprt->timeout.to_initval;
669 req->rq_xid = xprt_alloc_xid(xprt);
670 dprintk("RPC: %4d reserved req %p xid %08x\n", task->tk_pid,
671 req, ntohl(req->rq_xid));
675 * xprt_release - release an RPC request slot
676 * @task: task which is finished with the slot
679 void xprt_release(struct rpc_task *task)
681 struct rpc_xprt *xprt = task->tk_xprt;
682 struct rpc_rqst *req;
684 if (!(req = task->tk_rqstp))
686 spin_lock_bh(&xprt->transport_lock);
687 __xprt_release_write(xprt, task);
688 __xprt_put_cong(xprt, req);
689 if (!list_empty(&req->rq_list))
690 list_del(&req->rq_list);
691 xprt->last_used = jiffies;
692 if (list_empty(&xprt->recv) && !xprt->shutdown)
693 mod_timer(&xprt->timer,
694 xprt->last_used + RPC_IDLE_DISCONNECT_TIMEOUT);
695 spin_unlock_bh(&xprt->transport_lock);
696 task->tk_rqstp = NULL;
697 memset(req, 0, sizeof(*req)); /* mark unused */
699 dprintk("RPC: %4d release request %p\n", task->tk_pid, req);
701 spin_lock(&xprt->xprt_lock);
702 list_add(&req->rq_list, &xprt->free);
703 xprt_clear_backlog(xprt);
704 spin_unlock(&xprt->xprt_lock);
708 * xprt_set_timeout - set constant RPC timeout
709 * @to: RPC timeout parameters to set up
710 * @retr: number of retries
711 * @incr: amount of increase after each retry
714 void xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
717 to->to_increment = incr;
718 to->to_maxval = to->to_initval + (incr * retr);
719 to->to_retries = retr;
720 to->to_exponential = 0;
723 static struct rpc_xprt *xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
726 struct rpc_xprt *xprt;
727 struct rpc_rqst *req;
729 if ((xprt = kmalloc(sizeof(struct rpc_xprt), GFP_KERNEL)) == NULL)
730 return ERR_PTR(-ENOMEM);
731 memset(xprt, 0, sizeof(*xprt)); /* Nnnngh! */
737 result = xs_setup_udp(xprt, to);
740 result = xs_setup_tcp(xprt, to);
743 printk(KERN_ERR "RPC: unrecognized transport protocol: %d\n",
750 return ERR_PTR(result);
753 spin_lock_init(&xprt->transport_lock);
754 spin_lock_init(&xprt->xprt_lock);
755 init_waitqueue_head(&xprt->cong_wait);
757 INIT_LIST_HEAD(&xprt->free);
758 INIT_LIST_HEAD(&xprt->recv);
759 INIT_WORK(&xprt->task_cleanup, xprt_socket_autoclose, xprt);
760 init_timer(&xprt->timer);
761 xprt->timer.function = xprt_init_autodisconnect;
762 xprt->timer.data = (unsigned long) xprt;
763 xprt->last_used = jiffies;
765 rpc_init_wait_queue(&xprt->pending, "xprt_pending");
766 rpc_init_wait_queue(&xprt->sending, "xprt_sending");
767 rpc_init_wait_queue(&xprt->resend, "xprt_resend");
768 rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
770 /* initialize free list */
771 for (req = &xprt->slot[xprt->max_reqs-1]; req >= &xprt->slot[0]; req--)
772 list_add(&req->rq_list, &xprt->free);
776 dprintk("RPC: created transport %p with %u slots\n", xprt,
783 * xprt_create_proto - create an RPC client transport
784 * @proto: requested transport protocol
785 * @sap: remote peer's address
786 * @to: timeout parameters for new transport
789 struct rpc_xprt *xprt_create_proto(int proto, struct sockaddr_in *sap, struct rpc_timeout *to)
791 struct rpc_xprt *xprt;
793 xprt = xprt_setup(proto, sap, to);
795 dprintk("RPC: xprt_create_proto failed\n");
797 dprintk("RPC: xprt_create_proto created xprt %p\n", xprt);
801 static void xprt_shutdown(struct rpc_xprt *xprt)
804 rpc_wake_up(&xprt->sending);
805 rpc_wake_up(&xprt->resend);
806 rpc_wake_up(&xprt->pending);
807 rpc_wake_up(&xprt->backlog);
808 wake_up(&xprt->cong_wait);
809 del_timer_sync(&xprt->timer);
812 static int xprt_clear_backlog(struct rpc_xprt *xprt) {
813 rpc_wake_up_next(&xprt->backlog);
814 wake_up(&xprt->cong_wait);
819 * xprt_destroy - destroy an RPC transport, killing off all requests.
820 * @xprt: transport to destroy
823 int xprt_destroy(struct rpc_xprt *xprt)
825 dprintk("RPC: destroying transport %p\n", xprt);
827 xprt->ops->destroy(xprt);