xref: /openbmc/linux/net/sunrpc/xprt.c (revision e868d61272caa648214046a096e5a6bfc068dc8c)
1 /*
2  *  linux/net/sunrpc/xprt.c
3  *
4  *  This is a generic RPC call interface supporting congestion avoidance,
5  *  and asynchronous calls.
6  *
7  *  The interface works like this:
8  *
9  *  -	When a process places a call, it allocates a request slot if
10  *	one is available. Otherwise, it sleeps on the backlog queue
11  *	(xprt_reserve).
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, 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 transport. 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
23  *	of -ETIMEDOUT.
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
28  *	again.
29  *
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.
33  *
34  *  Copyright (C) 1995-1997, Olaf Kirch <okir@monad.swb.de>
35  *
36  *  Transport switch API copyright (C) 2005, Chuck Lever <cel@netapp.com>
37  */
38 
39 #include <linux/module.h>
40 
41 #include <linux/types.h>
42 #include <linux/interrupt.h>
43 #include <linux/workqueue.h>
44 #include <linux/net.h>
45 
46 #include <linux/sunrpc/clnt.h>
47 #include <linux/sunrpc/metrics.h>
48 
49 /*
50  * Local variables
51  */
52 
53 #ifdef RPC_DEBUG
54 # define RPCDBG_FACILITY	RPCDBG_XPRT
55 #endif
56 
57 /*
58  * Local functions
59  */
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 *);
64 
65 /*
66  * The transport code maintains an estimate on the maximum number of out-
67  * standing RPC requests, using a smoothed version of the congestion
68  * avoidance implemented in 44BSD. This is basically the Van Jacobson
69  * congestion algorithm: If a retransmit occurs, the congestion window is
70  * halved; otherwise, it is incremented by 1/cwnd when
71  *
72  *	-	a reply is received and
73  *	-	a full number of requests are outstanding and
74  *	-	the congestion window hasn't been updated recently.
75  */
76 #define RPC_CWNDSHIFT		(8U)
77 #define RPC_CWNDSCALE		(1U << RPC_CWNDSHIFT)
78 #define RPC_INITCWND		RPC_CWNDSCALE
79 #define RPC_MAXCWND(xprt)	((xprt)->max_reqs << RPC_CWNDSHIFT)
80 
81 #define RPCXPRT_CONGESTED(xprt) ((xprt)->cong >= (xprt)->cwnd)
82 
83 /**
84  * xprt_reserve_xprt - serialize write access to transports
85  * @task: task that is requesting access to the transport
86  *
87  * This prevents mixing the payload of separate requests, and prevents
88  * transport connects from colliding with writes.  No congestion control
89  * is provided.
90  */
91 int xprt_reserve_xprt(struct rpc_task *task)
92 {
93 	struct rpc_xprt	*xprt = task->tk_xprt;
94 	struct rpc_rqst *req = task->tk_rqstp;
95 
96 	if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
97 		if (task == xprt->snd_task)
98 			return 1;
99 		if (task == NULL)
100 			return 0;
101 		goto out_sleep;
102 	}
103 	xprt->snd_task = task;
104 	if (req) {
105 		req->rq_bytes_sent = 0;
106 		req->rq_ntrans++;
107 	}
108 	return 1;
109 
110 out_sleep:
111 	dprintk("RPC: %5u failed to lock transport %p\n",
112 			task->tk_pid, xprt);
113 	task->tk_timeout = 0;
114 	task->tk_status = -EAGAIN;
115 	if (req && req->rq_ntrans)
116 		rpc_sleep_on(&xprt->resend, task, NULL, NULL);
117 	else
118 		rpc_sleep_on(&xprt->sending, task, NULL, NULL);
119 	return 0;
120 }
121 
122 static void xprt_clear_locked(struct rpc_xprt *xprt)
123 {
124 	xprt->snd_task = NULL;
125 	if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state) || xprt->shutdown) {
126 		smp_mb__before_clear_bit();
127 		clear_bit(XPRT_LOCKED, &xprt->state);
128 		smp_mb__after_clear_bit();
129 	} else
130 		schedule_work(&xprt->task_cleanup);
131 }
132 
133 /*
134  * xprt_reserve_xprt_cong - serialize write access to transports
135  * @task: task that is requesting access to the transport
136  *
137  * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
138  * integrated into the decision of whether a request is allowed to be
139  * woken up and given access to the transport.
140  */
141 int xprt_reserve_xprt_cong(struct rpc_task *task)
142 {
143 	struct rpc_xprt	*xprt = task->tk_xprt;
144 	struct rpc_rqst *req = task->tk_rqstp;
145 
146 	if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
147 		if (task == xprt->snd_task)
148 			return 1;
149 		goto out_sleep;
150 	}
151 	if (__xprt_get_cong(xprt, task)) {
152 		xprt->snd_task = task;
153 		if (req) {
154 			req->rq_bytes_sent = 0;
155 			req->rq_ntrans++;
156 		}
157 		return 1;
158 	}
159 	xprt_clear_locked(xprt);
160 out_sleep:
161 	dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
162 	task->tk_timeout = 0;
163 	task->tk_status = -EAGAIN;
164 	if (req && req->rq_ntrans)
165 		rpc_sleep_on(&xprt->resend, task, NULL, NULL);
166 	else
167 		rpc_sleep_on(&xprt->sending, task, NULL, NULL);
168 	return 0;
169 }
170 
171 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
172 {
173 	int retval;
174 
175 	spin_lock_bh(&xprt->transport_lock);
176 	retval = xprt->ops->reserve_xprt(task);
177 	spin_unlock_bh(&xprt->transport_lock);
178 	return retval;
179 }
180 
181 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
182 {
183 	struct rpc_task *task;
184 	struct rpc_rqst *req;
185 
186 	if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
187 		return;
188 
189 	task = rpc_wake_up_next(&xprt->resend);
190 	if (!task) {
191 		task = rpc_wake_up_next(&xprt->sending);
192 		if (!task)
193 			goto out_unlock;
194 	}
195 
196 	req = task->tk_rqstp;
197 	xprt->snd_task = task;
198 	if (req) {
199 		req->rq_bytes_sent = 0;
200 		req->rq_ntrans++;
201 	}
202 	return;
203 
204 out_unlock:
205 	xprt_clear_locked(xprt);
206 }
207 
208 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
209 {
210 	struct rpc_task *task;
211 
212 	if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
213 		return;
214 	if (RPCXPRT_CONGESTED(xprt))
215 		goto out_unlock;
216 	task = rpc_wake_up_next(&xprt->resend);
217 	if (!task) {
218 		task = rpc_wake_up_next(&xprt->sending);
219 		if (!task)
220 			goto out_unlock;
221 	}
222 	if (__xprt_get_cong(xprt, task)) {
223 		struct rpc_rqst *req = task->tk_rqstp;
224 		xprt->snd_task = task;
225 		if (req) {
226 			req->rq_bytes_sent = 0;
227 			req->rq_ntrans++;
228 		}
229 		return;
230 	}
231 out_unlock:
232 	xprt_clear_locked(xprt);
233 }
234 
235 /**
236  * xprt_release_xprt - allow other requests to use a transport
237  * @xprt: transport with other tasks potentially waiting
238  * @task: task that is releasing access to the transport
239  *
240  * Note that "task" can be NULL.  No congestion control is provided.
241  */
242 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
243 {
244 	if (xprt->snd_task == task) {
245 		xprt_clear_locked(xprt);
246 		__xprt_lock_write_next(xprt);
247 	}
248 }
249 
250 /**
251  * xprt_release_xprt_cong - allow other requests to use a transport
252  * @xprt: transport with other tasks potentially waiting
253  * @task: task that is releasing access to the transport
254  *
255  * Note that "task" can be NULL.  Another task is awoken to use the
256  * transport if the transport's congestion window allows it.
257  */
258 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
259 {
260 	if (xprt->snd_task == task) {
261 		xprt_clear_locked(xprt);
262 		__xprt_lock_write_next_cong(xprt);
263 	}
264 }
265 
266 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
267 {
268 	spin_lock_bh(&xprt->transport_lock);
269 	xprt->ops->release_xprt(xprt, task);
270 	spin_unlock_bh(&xprt->transport_lock);
271 }
272 
273 /*
274  * Van Jacobson congestion avoidance. Check if the congestion window
275  * overflowed. Put the task to sleep if this is the case.
276  */
277 static int
278 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
279 {
280 	struct rpc_rqst *req = task->tk_rqstp;
281 
282 	if (req->rq_cong)
283 		return 1;
284 	dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
285 			task->tk_pid, xprt->cong, xprt->cwnd);
286 	if (RPCXPRT_CONGESTED(xprt))
287 		return 0;
288 	req->rq_cong = 1;
289 	xprt->cong += RPC_CWNDSCALE;
290 	return 1;
291 }
292 
293 /*
294  * Adjust the congestion window, and wake up the next task
295  * that has been sleeping due to congestion
296  */
297 static void
298 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
299 {
300 	if (!req->rq_cong)
301 		return;
302 	req->rq_cong = 0;
303 	xprt->cong -= RPC_CWNDSCALE;
304 	__xprt_lock_write_next_cong(xprt);
305 }
306 
307 /**
308  * xprt_release_rqst_cong - housekeeping when request is complete
309  * @task: RPC request that recently completed
310  *
311  * Useful for transports that require congestion control.
312  */
313 void xprt_release_rqst_cong(struct rpc_task *task)
314 {
315 	__xprt_put_cong(task->tk_xprt, task->tk_rqstp);
316 }
317 
318 /**
319  * xprt_adjust_cwnd - adjust transport congestion window
320  * @task: recently completed RPC request used to adjust window
321  * @result: result code of completed RPC request
322  *
323  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
324  */
325 void xprt_adjust_cwnd(struct rpc_task *task, int result)
326 {
327 	struct rpc_rqst *req = task->tk_rqstp;
328 	struct rpc_xprt *xprt = task->tk_xprt;
329 	unsigned long cwnd = xprt->cwnd;
330 
331 	if (result >= 0 && cwnd <= xprt->cong) {
332 		/* The (cwnd >> 1) term makes sure
333 		 * the result gets rounded properly. */
334 		cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
335 		if (cwnd > RPC_MAXCWND(xprt))
336 			cwnd = RPC_MAXCWND(xprt);
337 		__xprt_lock_write_next_cong(xprt);
338 	} else if (result == -ETIMEDOUT) {
339 		cwnd >>= 1;
340 		if (cwnd < RPC_CWNDSCALE)
341 			cwnd = RPC_CWNDSCALE;
342 	}
343 	dprintk("RPC:       cong %ld, cwnd was %ld, now %ld\n",
344 			xprt->cong, xprt->cwnd, cwnd);
345 	xprt->cwnd = cwnd;
346 	__xprt_put_cong(xprt, req);
347 }
348 
349 /**
350  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
351  * @xprt: transport with waiting tasks
352  * @status: result code to plant in each task before waking it
353  *
354  */
355 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
356 {
357 	if (status < 0)
358 		rpc_wake_up_status(&xprt->pending, status);
359 	else
360 		rpc_wake_up(&xprt->pending);
361 }
362 
363 /**
364  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
365  * @task: task to be put to sleep
366  *
367  */
368 void xprt_wait_for_buffer_space(struct rpc_task *task)
369 {
370 	struct rpc_rqst *req = task->tk_rqstp;
371 	struct rpc_xprt *xprt = req->rq_xprt;
372 
373 	task->tk_timeout = req->rq_timeout;
374 	rpc_sleep_on(&xprt->pending, task, NULL, NULL);
375 }
376 
377 /**
378  * xprt_write_space - wake the task waiting for transport output buffer space
379  * @xprt: transport with waiting tasks
380  *
381  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
382  */
383 void xprt_write_space(struct rpc_xprt *xprt)
384 {
385 	if (unlikely(xprt->shutdown))
386 		return;
387 
388 	spin_lock_bh(&xprt->transport_lock);
389 	if (xprt->snd_task) {
390 		dprintk("RPC:       write space: waking waiting task on "
391 				"xprt %p\n", xprt);
392 		rpc_wake_up_task(xprt->snd_task);
393 	}
394 	spin_unlock_bh(&xprt->transport_lock);
395 }
396 
397 /**
398  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
399  * @task: task whose timeout is to be set
400  *
401  * Set a request's retransmit timeout based on the transport's
402  * default timeout parameters.  Used by transports that don't adjust
403  * the retransmit timeout based on round-trip time estimation.
404  */
405 void xprt_set_retrans_timeout_def(struct rpc_task *task)
406 {
407 	task->tk_timeout = task->tk_rqstp->rq_timeout;
408 }
409 
410 /*
411  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
412  * @task: task whose timeout is to be set
413  *
414  * Set a request's retransmit timeout using the RTT estimator.
415  */
416 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
417 {
418 	int timer = task->tk_msg.rpc_proc->p_timer;
419 	struct rpc_rtt *rtt = task->tk_client->cl_rtt;
420 	struct rpc_rqst *req = task->tk_rqstp;
421 	unsigned long max_timeout = req->rq_xprt->timeout.to_maxval;
422 
423 	task->tk_timeout = rpc_calc_rto(rtt, timer);
424 	task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
425 	if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
426 		task->tk_timeout = max_timeout;
427 }
428 
429 static void xprt_reset_majortimeo(struct rpc_rqst *req)
430 {
431 	struct rpc_timeout *to = &req->rq_xprt->timeout;
432 
433 	req->rq_majortimeo = req->rq_timeout;
434 	if (to->to_exponential)
435 		req->rq_majortimeo <<= to->to_retries;
436 	else
437 		req->rq_majortimeo += to->to_increment * to->to_retries;
438 	if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
439 		req->rq_majortimeo = to->to_maxval;
440 	req->rq_majortimeo += jiffies;
441 }
442 
443 /**
444  * xprt_adjust_timeout - adjust timeout values for next retransmit
445  * @req: RPC request containing parameters to use for the adjustment
446  *
447  */
448 int xprt_adjust_timeout(struct rpc_rqst *req)
449 {
450 	struct rpc_xprt *xprt = req->rq_xprt;
451 	struct rpc_timeout *to = &xprt->timeout;
452 	int status = 0;
453 
454 	if (time_before(jiffies, req->rq_majortimeo)) {
455 		if (to->to_exponential)
456 			req->rq_timeout <<= 1;
457 		else
458 			req->rq_timeout += to->to_increment;
459 		if (to->to_maxval && req->rq_timeout >= to->to_maxval)
460 			req->rq_timeout = to->to_maxval;
461 		req->rq_retries++;
462 	} else {
463 		req->rq_timeout = to->to_initval;
464 		req->rq_retries = 0;
465 		xprt_reset_majortimeo(req);
466 		/* Reset the RTT counters == "slow start" */
467 		spin_lock_bh(&xprt->transport_lock);
468 		rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
469 		spin_unlock_bh(&xprt->transport_lock);
470 		status = -ETIMEDOUT;
471 	}
472 
473 	if (req->rq_timeout == 0) {
474 		printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
475 		req->rq_timeout = 5 * HZ;
476 	}
477 	return status;
478 }
479 
480 static void xprt_autoclose(struct work_struct *work)
481 {
482 	struct rpc_xprt *xprt =
483 		container_of(work, struct rpc_xprt, task_cleanup);
484 
485 	xprt_disconnect(xprt);
486 	xprt->ops->close(xprt);
487 	xprt_release_write(xprt, NULL);
488 }
489 
490 /**
491  * xprt_disconnect - mark a transport as disconnected
492  * @xprt: transport to flag for disconnect
493  *
494  */
495 void xprt_disconnect(struct rpc_xprt *xprt)
496 {
497 	dprintk("RPC:       disconnected transport %p\n", xprt);
498 	spin_lock_bh(&xprt->transport_lock);
499 	xprt_clear_connected(xprt);
500 	xprt_wake_pending_tasks(xprt, -ENOTCONN);
501 	spin_unlock_bh(&xprt->transport_lock);
502 }
503 
504 static void
505 xprt_init_autodisconnect(unsigned long data)
506 {
507 	struct rpc_xprt *xprt = (struct rpc_xprt *)data;
508 
509 	spin_lock(&xprt->transport_lock);
510 	if (!list_empty(&xprt->recv) || xprt->shutdown)
511 		goto out_abort;
512 	if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
513 		goto out_abort;
514 	spin_unlock(&xprt->transport_lock);
515 	if (xprt_connecting(xprt))
516 		xprt_release_write(xprt, NULL);
517 	else
518 		schedule_work(&xprt->task_cleanup);
519 	return;
520 out_abort:
521 	spin_unlock(&xprt->transport_lock);
522 }
523 
524 /**
525  * xprt_connect - schedule a transport connect operation
526  * @task: RPC task that is requesting the connect
527  *
528  */
529 void xprt_connect(struct rpc_task *task)
530 {
531 	struct rpc_xprt	*xprt = task->tk_xprt;
532 
533 	dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
534 			xprt, (xprt_connected(xprt) ? "is" : "is not"));
535 
536 	if (!xprt_bound(xprt)) {
537 		task->tk_status = -EIO;
538 		return;
539 	}
540 	if (!xprt_lock_write(xprt, task))
541 		return;
542 	if (xprt_connected(xprt))
543 		xprt_release_write(xprt, task);
544 	else {
545 		if (task->tk_rqstp)
546 			task->tk_rqstp->rq_bytes_sent = 0;
547 
548 		task->tk_timeout = xprt->connect_timeout;
549 		rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
550 		xprt->stat.connect_start = jiffies;
551 		xprt->ops->connect(task);
552 	}
553 	return;
554 }
555 
556 static void xprt_connect_status(struct rpc_task *task)
557 {
558 	struct rpc_xprt	*xprt = task->tk_xprt;
559 
560 	if (task->tk_status >= 0) {
561 		xprt->stat.connect_count++;
562 		xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
563 		dprintk("RPC: %5u xprt_connect_status: connection established\n",
564 				task->tk_pid);
565 		return;
566 	}
567 
568 	switch (task->tk_status) {
569 	case -ECONNREFUSED:
570 	case -ECONNRESET:
571 		dprintk("RPC: %5u xprt_connect_status: server %s refused "
572 				"connection\n", task->tk_pid,
573 				task->tk_client->cl_server);
574 		break;
575 	case -ENOTCONN:
576 		dprintk("RPC: %5u xprt_connect_status: connection broken\n",
577 				task->tk_pid);
578 		break;
579 	case -ETIMEDOUT:
580 		dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
581 				"out\n", task->tk_pid);
582 		break;
583 	default:
584 		dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
585 				"server %s\n", task->tk_pid, -task->tk_status,
586 				task->tk_client->cl_server);
587 		xprt_release_write(xprt, task);
588 		task->tk_status = -EIO;
589 	}
590 }
591 
592 /**
593  * xprt_lookup_rqst - find an RPC request corresponding to an XID
594  * @xprt: transport on which the original request was transmitted
595  * @xid: RPC XID of incoming reply
596  *
597  */
598 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
599 {
600 	struct list_head *pos;
601 
602 	list_for_each(pos, &xprt->recv) {
603 		struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
604 		if (entry->rq_xid == xid)
605 			return entry;
606 	}
607 
608 	dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
609 			ntohl(xid));
610 	xprt->stat.bad_xids++;
611 	return NULL;
612 }
613 
614 /**
615  * xprt_update_rtt - update an RPC client's RTT state after receiving a reply
616  * @task: RPC request that recently completed
617  *
618  */
619 void xprt_update_rtt(struct rpc_task *task)
620 {
621 	struct rpc_rqst *req = task->tk_rqstp;
622 	struct rpc_rtt *rtt = task->tk_client->cl_rtt;
623 	unsigned timer = task->tk_msg.rpc_proc->p_timer;
624 
625 	if (timer) {
626 		if (req->rq_ntrans == 1)
627 			rpc_update_rtt(rtt, timer,
628 					(long)jiffies - req->rq_xtime);
629 		rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
630 	}
631 }
632 
633 /**
634  * xprt_complete_rqst - called when reply processing is complete
635  * @task: RPC request that recently completed
636  * @copied: actual number of bytes received from the transport
637  *
638  * Caller holds transport lock.
639  */
640 void xprt_complete_rqst(struct rpc_task *task, int copied)
641 {
642 	struct rpc_rqst *req = task->tk_rqstp;
643 
644 	dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
645 			task->tk_pid, ntohl(req->rq_xid), copied);
646 
647 	task->tk_xprt->stat.recvs++;
648 	task->tk_rtt = (long)jiffies - req->rq_xtime;
649 
650 	list_del_init(&req->rq_list);
651 	/* Ensure all writes are done before we update req->rq_received */
652 	smp_wmb();
653 	req->rq_received = req->rq_private_buf.len = copied;
654 	rpc_wake_up_task(task);
655 }
656 
657 static void xprt_timer(struct rpc_task *task)
658 {
659 	struct rpc_rqst *req = task->tk_rqstp;
660 	struct rpc_xprt *xprt = req->rq_xprt;
661 
662 	dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
663 
664 	spin_lock(&xprt->transport_lock);
665 	if (!req->rq_received) {
666 		if (xprt->ops->timer)
667 			xprt->ops->timer(task);
668 		task->tk_status = -ETIMEDOUT;
669 	}
670 	task->tk_timeout = 0;
671 	rpc_wake_up_task(task);
672 	spin_unlock(&xprt->transport_lock);
673 }
674 
675 /**
676  * xprt_prepare_transmit - reserve the transport before sending a request
677  * @task: RPC task about to send a request
678  *
679  */
680 int xprt_prepare_transmit(struct rpc_task *task)
681 {
682 	struct rpc_rqst	*req = task->tk_rqstp;
683 	struct rpc_xprt	*xprt = req->rq_xprt;
684 	int err = 0;
685 
686 	dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
687 
688 	spin_lock_bh(&xprt->transport_lock);
689 	if (req->rq_received && !req->rq_bytes_sent) {
690 		err = req->rq_received;
691 		goto out_unlock;
692 	}
693 	if (!xprt->ops->reserve_xprt(task)) {
694 		err = -EAGAIN;
695 		goto out_unlock;
696 	}
697 
698 	if (!xprt_connected(xprt)) {
699 		err = -ENOTCONN;
700 		goto out_unlock;
701 	}
702 out_unlock:
703 	spin_unlock_bh(&xprt->transport_lock);
704 	return err;
705 }
706 
707 void xprt_end_transmit(struct rpc_task *task)
708 {
709 	xprt_release_write(task->tk_xprt, task);
710 }
711 
712 /**
713  * xprt_transmit - send an RPC request on a transport
714  * @task: controlling RPC task
715  *
716  * We have to copy the iovec because sendmsg fiddles with its contents.
717  */
718 void xprt_transmit(struct rpc_task *task)
719 {
720 	struct rpc_rqst	*req = task->tk_rqstp;
721 	struct rpc_xprt	*xprt = req->rq_xprt;
722 	int status;
723 
724 	dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
725 
726 	if (!req->rq_received) {
727 		if (list_empty(&req->rq_list)) {
728 			spin_lock_bh(&xprt->transport_lock);
729 			/* Update the softirq receive buffer */
730 			memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
731 					sizeof(req->rq_private_buf));
732 			/* Add request to the receive list */
733 			list_add_tail(&req->rq_list, &xprt->recv);
734 			spin_unlock_bh(&xprt->transport_lock);
735 			xprt_reset_majortimeo(req);
736 			/* Turn off autodisconnect */
737 			del_singleshot_timer_sync(&xprt->timer);
738 		}
739 	} else if (!req->rq_bytes_sent)
740 		return;
741 
742 	status = xprt->ops->send_request(task);
743 	if (status == 0) {
744 		dprintk("RPC: %5u xmit complete\n", task->tk_pid);
745 		spin_lock_bh(&xprt->transport_lock);
746 
747 		xprt->ops->set_retrans_timeout(task);
748 
749 		xprt->stat.sends++;
750 		xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
751 		xprt->stat.bklog_u += xprt->backlog.qlen;
752 
753 		/* Don't race with disconnect */
754 		if (!xprt_connected(xprt))
755 			task->tk_status = -ENOTCONN;
756 		else if (!req->rq_received)
757 			rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
758 		spin_unlock_bh(&xprt->transport_lock);
759 		return;
760 	}
761 
762 	/* Note: at this point, task->tk_sleeping has not yet been set,
763 	 *	 hence there is no danger of the waking up task being put on
764 	 *	 schedq, and being picked up by a parallel run of rpciod().
765 	 */
766 	task->tk_status = status;
767 	if (status == -ECONNREFUSED)
768 		rpc_sleep_on(&xprt->sending, task, NULL, NULL);
769 }
770 
771 static inline void do_xprt_reserve(struct rpc_task *task)
772 {
773 	struct rpc_xprt	*xprt = task->tk_xprt;
774 
775 	task->tk_status = 0;
776 	if (task->tk_rqstp)
777 		return;
778 	if (!list_empty(&xprt->free)) {
779 		struct rpc_rqst	*req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
780 		list_del_init(&req->rq_list);
781 		task->tk_rqstp = req;
782 		xprt_request_init(task, xprt);
783 		return;
784 	}
785 	dprintk("RPC:       waiting for request slot\n");
786 	task->tk_status = -EAGAIN;
787 	task->tk_timeout = 0;
788 	rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
789 }
790 
791 /**
792  * xprt_reserve - allocate an RPC request slot
793  * @task: RPC task requesting a slot allocation
794  *
795  * If no more slots are available, place the task on the transport's
796  * backlog queue.
797  */
798 void xprt_reserve(struct rpc_task *task)
799 {
800 	struct rpc_xprt	*xprt = task->tk_xprt;
801 
802 	task->tk_status = -EIO;
803 	spin_lock(&xprt->reserve_lock);
804 	do_xprt_reserve(task);
805 	spin_unlock(&xprt->reserve_lock);
806 }
807 
808 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
809 {
810 	return xprt->xid++;
811 }
812 
813 static inline void xprt_init_xid(struct rpc_xprt *xprt)
814 {
815 	xprt->xid = net_random();
816 }
817 
818 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
819 {
820 	struct rpc_rqst	*req = task->tk_rqstp;
821 
822 	req->rq_timeout = xprt->timeout.to_initval;
823 	req->rq_task	= task;
824 	req->rq_xprt    = xprt;
825 	req->rq_buffer  = NULL;
826 	req->rq_xid     = xprt_alloc_xid(xprt);
827 	req->rq_release_snd_buf = NULL;
828 	xprt_reset_majortimeo(req);
829 	dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
830 			req, ntohl(req->rq_xid));
831 }
832 
833 /**
834  * xprt_release - release an RPC request slot
835  * @task: task which is finished with the slot
836  *
837  */
838 void xprt_release(struct rpc_task *task)
839 {
840 	struct rpc_xprt	*xprt = task->tk_xprt;
841 	struct rpc_rqst	*req;
842 
843 	if (!(req = task->tk_rqstp))
844 		return;
845 	rpc_count_iostats(task);
846 	spin_lock_bh(&xprt->transport_lock);
847 	xprt->ops->release_xprt(xprt, task);
848 	if (xprt->ops->release_request)
849 		xprt->ops->release_request(task);
850 	if (!list_empty(&req->rq_list))
851 		list_del(&req->rq_list);
852 	xprt->last_used = jiffies;
853 	if (list_empty(&xprt->recv))
854 		mod_timer(&xprt->timer,
855 				xprt->last_used + xprt->idle_timeout);
856 	spin_unlock_bh(&xprt->transport_lock);
857 	xprt->ops->buf_free(req->rq_buffer);
858 	task->tk_rqstp = NULL;
859 	if (req->rq_release_snd_buf)
860 		req->rq_release_snd_buf(req);
861 	memset(req, 0, sizeof(*req));	/* mark unused */
862 
863 	dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
864 
865 	spin_lock(&xprt->reserve_lock);
866 	list_add(&req->rq_list, &xprt->free);
867 	rpc_wake_up_next(&xprt->backlog);
868 	spin_unlock(&xprt->reserve_lock);
869 }
870 
871 /**
872  * xprt_set_timeout - set constant RPC timeout
873  * @to: RPC timeout parameters to set up
874  * @retr: number of retries
875  * @incr: amount of increase after each retry
876  *
877  */
878 void xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
879 {
880 	to->to_initval   =
881 	to->to_increment = incr;
882 	to->to_maxval    = to->to_initval + (incr * retr);
883 	to->to_retries   = retr;
884 	to->to_exponential = 0;
885 }
886 
887 /**
888  * xprt_create_transport - create an RPC transport
889  * @proto: requested transport protocol
890  * @ap: remote peer address
891  * @size: length of address
892  * @to: timeout parameters
893  *
894  */
895 struct rpc_xprt *xprt_create_transport(int proto, struct sockaddr *ap, size_t size, struct rpc_timeout *to)
896 {
897 	struct rpc_xprt	*xprt;
898 	struct rpc_rqst	*req;
899 
900 	switch (proto) {
901 	case IPPROTO_UDP:
902 		xprt = xs_setup_udp(ap, size, to);
903 		break;
904 	case IPPROTO_TCP:
905 		xprt = xs_setup_tcp(ap, size, to);
906 		break;
907 	default:
908 		printk(KERN_ERR "RPC: unrecognized transport protocol: %d\n",
909 				proto);
910 		return ERR_PTR(-EIO);
911 	}
912 	if (IS_ERR(xprt)) {
913 		dprintk("RPC:       xprt_create_transport: failed, %ld\n",
914 				-PTR_ERR(xprt));
915 		return xprt;
916 	}
917 
918 	kref_init(&xprt->kref);
919 	spin_lock_init(&xprt->transport_lock);
920 	spin_lock_init(&xprt->reserve_lock);
921 
922 	INIT_LIST_HEAD(&xprt->free);
923 	INIT_LIST_HEAD(&xprt->recv);
924 	INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
925 	init_timer(&xprt->timer);
926 	xprt->timer.function = xprt_init_autodisconnect;
927 	xprt->timer.data = (unsigned long) xprt;
928 	xprt->last_used = jiffies;
929 	xprt->cwnd = RPC_INITCWND;
930 	xprt->bind_index = 0;
931 
932 	rpc_init_wait_queue(&xprt->binding, "xprt_binding");
933 	rpc_init_wait_queue(&xprt->pending, "xprt_pending");
934 	rpc_init_wait_queue(&xprt->sending, "xprt_sending");
935 	rpc_init_wait_queue(&xprt->resend, "xprt_resend");
936 	rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
937 
938 	/* initialize free list */
939 	for (req = &xprt->slot[xprt->max_reqs-1]; req >= &xprt->slot[0]; req--)
940 		list_add(&req->rq_list, &xprt->free);
941 
942 	xprt_init_xid(xprt);
943 
944 	dprintk("RPC:       created transport %p with %u slots\n", xprt,
945 			xprt->max_reqs);
946 
947 	return xprt;
948 }
949 
950 /**
951  * xprt_destroy - destroy an RPC transport, killing off all requests.
952  * @kref: kref for the transport to destroy
953  *
954  */
955 static void xprt_destroy(struct kref *kref)
956 {
957 	struct rpc_xprt *xprt = container_of(kref, struct rpc_xprt, kref);
958 
959 	dprintk("RPC:       destroying transport %p\n", xprt);
960 	xprt->shutdown = 1;
961 	del_timer_sync(&xprt->timer);
962 
963 	/*
964 	 * Tear down transport state and free the rpc_xprt
965 	 */
966 	xprt->ops->destroy(xprt);
967 }
968 
969 /**
970  * xprt_put - release a reference to an RPC transport.
971  * @xprt: pointer to the transport
972  *
973  */
974 void xprt_put(struct rpc_xprt *xprt)
975 {
976 	kref_put(&xprt->kref, xprt_destroy);
977 }
978 
979 /**
980  * xprt_get - return a reference to an RPC transport.
981  * @xprt: pointer to the transport
982  *
983  */
984 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
985 {
986 	kref_get(&xprt->kref);
987 	return xprt;
988 }
989