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->sock_lock);
110 retval = __xprt_lock_write(xprt, task);
111 spin_unlock_bh(&xprt->sock_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->sock_lock);
165 __xprt_release_write(xprt, task);
166 spin_unlock_bh(&xprt->sock_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);
231 * Reset the major timeout value
233 static void xprt_reset_majortimeo(struct rpc_rqst *req)
235 struct rpc_timeout *to = &req->rq_xprt->timeout;
237 req->rq_majortimeo = req->rq_timeout;
238 if (to->to_exponential)
239 req->rq_majortimeo <<= to->to_retries;
241 req->rq_majortimeo += to->to_increment * to->to_retries;
242 if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
243 req->rq_majortimeo = to->to_maxval;
244 req->rq_majortimeo += jiffies;
248 * Adjust timeout values etc for next retransmit
250 int xprt_adjust_timeout(struct rpc_rqst *req)
252 struct rpc_xprt *xprt = req->rq_xprt;
253 struct rpc_timeout *to = &xprt->timeout;
256 if (time_before(jiffies, req->rq_majortimeo)) {
257 if (to->to_exponential)
258 req->rq_timeout <<= 1;
260 req->rq_timeout += to->to_increment;
261 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
262 req->rq_timeout = to->to_maxval;
264 pprintk("RPC: %lu retrans\n", jiffies);
266 req->rq_timeout = to->to_initval;
268 xprt_reset_majortimeo(req);
269 /* Reset the RTT counters == "slow start" */
270 spin_lock_bh(&xprt->sock_lock);
271 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
272 spin_unlock_bh(&xprt->sock_lock);
273 pprintk("RPC: %lu timeout\n", jiffies);
277 if (req->rq_timeout == 0) {
278 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
279 req->rq_timeout = 5 * HZ;
285 xprt_socket_autoclose(void *args)
287 struct rpc_xprt *xprt = (struct rpc_xprt *)args;
289 xprt_disconnect(xprt);
290 xprt->ops->close(xprt);
291 xprt_release_write(xprt, NULL);
295 * Mark a transport as disconnected
297 void xprt_disconnect(struct rpc_xprt *xprt)
299 dprintk("RPC: disconnected transport %p\n", xprt);
300 spin_lock_bh(&xprt->sock_lock);
301 xprt_clear_connected(xprt);
302 rpc_wake_up_status(&xprt->pending, -ENOTCONN);
303 spin_unlock_bh(&xprt->sock_lock);
307 * Used to allow disconnection when we've been idle
310 xprt_init_autodisconnect(unsigned long data)
312 struct rpc_xprt *xprt = (struct rpc_xprt *)data;
314 spin_lock(&xprt->sock_lock);
315 if (!list_empty(&xprt->recv) || xprt->shutdown)
317 if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
319 spin_unlock(&xprt->sock_lock);
320 /* Let keventd close the socket */
321 if (test_bit(XPRT_CONNECTING, &xprt->sockstate) != 0)
322 xprt_release_write(xprt, NULL);
324 schedule_work(&xprt->task_cleanup);
327 spin_unlock(&xprt->sock_lock);
331 * Attempt to connect a TCP socket.
334 void xprt_connect(struct rpc_task *task)
336 struct rpc_xprt *xprt = task->tk_xprt;
338 dprintk("RPC: %4d xprt_connect xprt %p %s connected\n", task->tk_pid,
339 xprt, (xprt_connected(xprt) ? "is" : "is not"));
341 if (xprt->shutdown) {
342 task->tk_status = -EIO;
345 if (!xprt->addr.sin_port) {
346 task->tk_status = -EIO;
349 if (!xprt_lock_write(xprt, task))
351 if (xprt_connected(xprt))
352 xprt_release_write(xprt, task);
355 task->tk_rqstp->rq_bytes_sent = 0;
357 task->tk_timeout = RPC_CONNECT_TIMEOUT;
358 rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
359 xprt->ops->connect(task);
365 * We arrive here when awoken from waiting on connection establishment.
368 xprt_connect_status(struct rpc_task *task)
370 struct rpc_xprt *xprt = task->tk_xprt;
372 if (task->tk_status >= 0) {
373 dprintk("RPC: %4d xprt_connect_status: connection established\n",
378 switch (task->tk_status) {
381 dprintk("RPC: %4d xprt_connect_status: server %s refused connection\n",
382 task->tk_pid, task->tk_client->cl_server);
385 dprintk("RPC: %4d xprt_connect_status: connection broken\n",
389 dprintk("RPC: %4d xprt_connect_status: connect attempt timed out\n",
393 dprintk("RPC: %4d xprt_connect_status: error %d connecting to server %s\n",
394 task->tk_pid, -task->tk_status, task->tk_client->cl_server);
395 xprt_release_write(xprt, task);
396 task->tk_status = -EIO;
400 /* if soft mounted, just cause this RPC to fail */
401 if (RPC_IS_SOFT(task)) {
402 xprt_release_write(xprt, task);
403 task->tk_status = -EIO;
408 * Look up the RPC request corresponding to a reply, and then lock it.
410 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
412 struct list_head *pos;
413 struct rpc_rqst *req = NULL;
415 list_for_each(pos, &xprt->recv) {
416 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
417 if (entry->rq_xid == xid) {
426 * Complete reply received.
427 * The TCP code relies on us to remove the request from xprt->pending.
429 void xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
431 struct rpc_task *task = req->rq_task;
432 struct rpc_clnt *clnt = task->tk_client;
434 /* Adjust congestion window */
436 unsigned timer = task->tk_msg.rpc_proc->p_timer;
437 xprt_adjust_cwnd(xprt, copied);
438 __xprt_put_cong(xprt, req);
440 if (req->rq_ntrans == 1)
441 rpc_update_rtt(clnt->cl_rtt, timer,
442 (long)jiffies - req->rq_xtime);
443 rpc_set_timeo(clnt->cl_rtt, timer, req->rq_ntrans - 1);
448 /* Profile only reads for now */
450 static unsigned long nextstat;
451 static unsigned long pkt_rtt, pkt_len, pkt_cnt;
454 pkt_len += req->rq_slen + copied;
455 pkt_rtt += jiffies - req->rq_xtime;
456 if (time_before(nextstat, jiffies)) {
457 printk("RPC: %lu %ld cwnd\n", jiffies, xprt->cwnd);
458 printk("RPC: %ld %ld %ld %ld stat\n",
459 jiffies, pkt_cnt, pkt_len, pkt_rtt);
460 pkt_rtt = pkt_len = pkt_cnt = 0;
461 nextstat = jiffies + 5 * HZ;
466 dprintk("RPC: %4d has input (%d bytes)\n", task->tk_pid, copied);
467 list_del_init(&req->rq_list);
468 req->rq_received = req->rq_private_buf.len = copied;
470 /* ... and wake up the process. */
471 rpc_wake_up_task(task);
476 * RPC receive timeout handler.
479 xprt_timer(struct rpc_task *task)
481 struct rpc_rqst *req = task->tk_rqstp;
482 struct rpc_xprt *xprt = req->rq_xprt;
484 spin_lock(&xprt->sock_lock);
485 if (req->rq_received)
488 xprt_adjust_cwnd(req->rq_xprt, -ETIMEDOUT);
489 __xprt_put_cong(xprt, req);
491 dprintk("RPC: %4d xprt_timer (%s request)\n",
492 task->tk_pid, req ? "pending" : "backlogged");
494 task->tk_status = -ETIMEDOUT;
496 task->tk_timeout = 0;
497 rpc_wake_up_task(task);
498 spin_unlock(&xprt->sock_lock);
502 * Place the actual RPC call.
503 * We have to copy the iovec because sendmsg fiddles with its contents.
506 xprt_prepare_transmit(struct rpc_task *task)
508 struct rpc_rqst *req = task->tk_rqstp;
509 struct rpc_xprt *xprt = req->rq_xprt;
512 dprintk("RPC: %4d xprt_prepare_transmit\n", task->tk_pid);
517 spin_lock_bh(&xprt->sock_lock);
518 if (req->rq_received && !req->rq_bytes_sent) {
519 err = req->rq_received;
522 if (!__xprt_lock_write(xprt, task)) {
527 if (!xprt_connected(xprt)) {
532 spin_unlock_bh(&xprt->sock_lock);
537 xprt_transmit(struct rpc_task *task)
539 struct rpc_clnt *clnt = task->tk_client;
540 struct rpc_rqst *req = task->tk_rqstp;
541 struct rpc_xprt *xprt = req->rq_xprt;
544 dprintk("RPC: %4d xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
547 if (!req->rq_received) {
548 if (list_empty(&req->rq_list)) {
549 spin_lock_bh(&xprt->sock_lock);
550 /* Update the softirq receive buffer */
551 memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
552 sizeof(req->rq_private_buf));
553 /* Add request to the receive list */
554 list_add_tail(&req->rq_list, &xprt->recv);
555 spin_unlock_bh(&xprt->sock_lock);
556 xprt_reset_majortimeo(req);
557 /* Turn off autodisconnect */
558 del_singleshot_timer_sync(&xprt->timer);
560 } else if (!req->rq_bytes_sent)
563 status = xprt->ops->send_request(task);
567 /* Note: at this point, task->tk_sleeping has not yet been set,
568 * hence there is no danger of the waking up task being put on
569 * schedq, and being picked up by a parallel run of rpciod().
571 task->tk_status = status;
575 task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
576 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
582 xprt_disconnect(xprt);
584 xprt_release_write(xprt, task);
587 dprintk("RPC: %4d xmit complete\n", task->tk_pid);
588 /* Set the task's receive timeout value */
589 spin_lock_bh(&xprt->sock_lock);
591 int timer = task->tk_msg.rpc_proc->p_timer;
592 task->tk_timeout = rpc_calc_rto(clnt->cl_rtt, timer);
593 task->tk_timeout <<= rpc_ntimeo(clnt->cl_rtt, timer) + req->rq_retries;
594 if (task->tk_timeout > xprt->timeout.to_maxval || task->tk_timeout == 0)
595 task->tk_timeout = xprt->timeout.to_maxval;
597 task->tk_timeout = req->rq_timeout;
598 /* Don't race with disconnect */
599 if (!xprt_connected(xprt))
600 task->tk_status = -ENOTCONN;
601 else if (!req->rq_received)
602 rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
603 __xprt_release_write(xprt, task);
604 spin_unlock_bh(&xprt->sock_lock);
608 * Reserve an RPC call slot.
611 do_xprt_reserve(struct rpc_task *task)
613 struct rpc_xprt *xprt = task->tk_xprt;
618 if (!list_empty(&xprt->free)) {
619 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
620 list_del_init(&req->rq_list);
621 task->tk_rqstp = req;
622 xprt_request_init(task, xprt);
625 dprintk("RPC: waiting for request slot\n");
626 task->tk_status = -EAGAIN;
627 task->tk_timeout = 0;
628 rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
632 xprt_reserve(struct rpc_task *task)
634 struct rpc_xprt *xprt = task->tk_xprt;
636 task->tk_status = -EIO;
637 if (!xprt->shutdown) {
638 spin_lock(&xprt->xprt_lock);
639 do_xprt_reserve(task);
640 spin_unlock(&xprt->xprt_lock);
645 * Allocate a 'unique' XID
647 static inline u32 xprt_alloc_xid(struct rpc_xprt *xprt)
652 static inline void xprt_init_xid(struct rpc_xprt *xprt)
654 get_random_bytes(&xprt->xid, sizeof(xprt->xid));
658 * Initialize RPC request
661 xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
663 struct rpc_rqst *req = task->tk_rqstp;
665 req->rq_timeout = xprt->timeout.to_initval;
668 req->rq_xid = xprt_alloc_xid(xprt);
669 dprintk("RPC: %4d reserved req %p xid %08x\n", task->tk_pid,
670 req, ntohl(req->rq_xid));
674 * Release an RPC call slot
677 xprt_release(struct rpc_task *task)
679 struct rpc_xprt *xprt = task->tk_xprt;
680 struct rpc_rqst *req;
682 if (!(req = task->tk_rqstp))
684 spin_lock_bh(&xprt->sock_lock);
685 __xprt_release_write(xprt, task);
686 __xprt_put_cong(xprt, req);
687 if (!list_empty(&req->rq_list))
688 list_del(&req->rq_list);
689 xprt->last_used = jiffies;
690 if (list_empty(&xprt->recv) && !xprt->shutdown)
691 mod_timer(&xprt->timer,
692 xprt->last_used + RPC_IDLE_DISCONNECT_TIMEOUT);
693 spin_unlock_bh(&xprt->sock_lock);
694 task->tk_rqstp = NULL;
695 memset(req, 0, sizeof(*req)); /* mark unused */
697 dprintk("RPC: %4d release request %p\n", task->tk_pid, req);
699 spin_lock(&xprt->xprt_lock);
700 list_add(&req->rq_list, &xprt->free);
701 xprt_clear_backlog(xprt);
702 spin_unlock(&xprt->xprt_lock);
706 * Set constant timeout
709 xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
712 to->to_increment = incr;
713 to->to_maxval = to->to_initval + (incr * retr);
714 to->to_retries = retr;
715 to->to_exponential = 0;
719 * Initialize an RPC client
721 static struct rpc_xprt *
722 xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
725 struct rpc_xprt *xprt;
726 struct rpc_rqst *req;
728 if ((xprt = kmalloc(sizeof(struct rpc_xprt), GFP_KERNEL)) == NULL)
729 return ERR_PTR(-ENOMEM);
730 memset(xprt, 0, sizeof(*xprt)); /* Nnnngh! */
736 result = xs_setup_udp(xprt, to);
739 result = xs_setup_tcp(xprt, to);
742 printk(KERN_ERR "RPC: unrecognized transport protocol: %d\n",
749 return ERR_PTR(result);
752 spin_lock_init(&xprt->sock_lock);
753 spin_lock_init(&xprt->xprt_lock);
754 init_waitqueue_head(&xprt->cong_wait);
756 INIT_LIST_HEAD(&xprt->free);
757 INIT_LIST_HEAD(&xprt->recv);
758 INIT_WORK(&xprt->task_cleanup, xprt_socket_autoclose, xprt);
759 init_timer(&xprt->timer);
760 xprt->timer.function = xprt_init_autodisconnect;
761 xprt->timer.data = (unsigned long) xprt;
762 xprt->last_used = jiffies;
764 rpc_init_wait_queue(&xprt->pending, "xprt_pending");
765 rpc_init_wait_queue(&xprt->sending, "xprt_sending");
766 rpc_init_wait_queue(&xprt->resend, "xprt_resend");
767 rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
769 /* initialize free list */
770 for (req = &xprt->slot[xprt->max_reqs-1]; req >= &xprt->slot[0]; req--)
771 list_add(&req->rq_list, &xprt->free);
775 dprintk("RPC: created transport %p with %u slots\n", xprt,
782 * Create an RPC client transport given the protocol and peer address.
785 xprt_create_proto(int proto, struct sockaddr_in *sap, struct rpc_timeout *to)
787 struct rpc_xprt *xprt;
789 xprt = xprt_setup(proto, sap, to);
791 dprintk("RPC: xprt_create_proto failed\n");
793 dprintk("RPC: xprt_create_proto created xprt %p\n", xprt);
798 * Prepare for transport shutdown.
801 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);
813 * Clear the xprt backlog queue
816 xprt_clear_backlog(struct rpc_xprt *xprt) {
817 rpc_wake_up_next(&xprt->backlog);
818 wake_up(&xprt->cong_wait);
823 * Destroy an RPC transport, killing off all requests.
826 xprt_destroy(struct rpc_xprt *xprt)
828 dprintk("RPC: destroying transport %p\n", xprt);
830 xprt->ops->destroy(xprt);