Merge git://git.kernel.org/pub/scm/linux/kernel/git/kyle/parisc-2.6
[linux-2.6] / net / sunrpc / xprt.c
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 static DEFINE_SPINLOCK(xprt_list_lock);
66 static LIST_HEAD(xprt_list);
67
68 /*
69  * The transport code maintains an estimate on the maximum number of out-
70  * standing RPC requests, using a smoothed version of the congestion
71  * avoidance implemented in 44BSD. This is basically the Van Jacobson
72  * congestion algorithm: If a retransmit occurs, the congestion window is
73  * halved; otherwise, it is incremented by 1/cwnd when
74  *
75  *      -       a reply is received and
76  *      -       a full number of requests are outstanding and
77  *      -       the congestion window hasn't been updated recently.
78  */
79 #define RPC_CWNDSHIFT           (8U)
80 #define RPC_CWNDSCALE           (1U << RPC_CWNDSHIFT)
81 #define RPC_INITCWND            RPC_CWNDSCALE
82 #define RPC_MAXCWND(xprt)       ((xprt)->max_reqs << RPC_CWNDSHIFT)
83
84 #define RPCXPRT_CONGESTED(xprt) ((xprt)->cong >= (xprt)->cwnd)
85
86 /**
87  * xprt_register_transport - register a transport implementation
88  * @transport: transport to register
89  *
90  * If a transport implementation is loaded as a kernel module, it can
91  * call this interface to make itself known to the RPC client.
92  *
93  * Returns:
94  * 0:           transport successfully registered
95  * -EEXIST:     transport already registered
96  * -EINVAL:     transport module being unloaded
97  */
98 int xprt_register_transport(struct xprt_class *transport)
99 {
100         struct xprt_class *t;
101         int result;
102
103         result = -EEXIST;
104         spin_lock(&xprt_list_lock);
105         list_for_each_entry(t, &xprt_list, list) {
106                 /* don't register the same transport class twice */
107                 if (t->ident == transport->ident)
108                         goto out;
109         }
110
111         list_add_tail(&transport->list, &xprt_list);
112         printk(KERN_INFO "RPC: Registered %s transport module.\n",
113                transport->name);
114         result = 0;
115
116 out:
117         spin_unlock(&xprt_list_lock);
118         return result;
119 }
120 EXPORT_SYMBOL_GPL(xprt_register_transport);
121
122 /**
123  * xprt_unregister_transport - unregister a transport implementation
124  * @transport: transport to unregister
125  *
126  * Returns:
127  * 0:           transport successfully unregistered
128  * -ENOENT:     transport never registered
129  */
130 int xprt_unregister_transport(struct xprt_class *transport)
131 {
132         struct xprt_class *t;
133         int result;
134
135         result = 0;
136         spin_lock(&xprt_list_lock);
137         list_for_each_entry(t, &xprt_list, list) {
138                 if (t == transport) {
139                         printk(KERN_INFO
140                                 "RPC: Unregistered %s transport module.\n",
141                                 transport->name);
142                         list_del_init(&transport->list);
143                         goto out;
144                 }
145         }
146         result = -ENOENT;
147
148 out:
149         spin_unlock(&xprt_list_lock);
150         return result;
151 }
152 EXPORT_SYMBOL_GPL(xprt_unregister_transport);
153
154 /**
155  * xprt_load_transport - load a transport implementation
156  * @transport_name: transport to load
157  *
158  * Returns:
159  * 0:           transport successfully loaded
160  * -ENOENT:     transport module not available
161  */
162 int xprt_load_transport(const char *transport_name)
163 {
164         struct xprt_class *t;
165         char module_name[sizeof t->name + 5];
166         int result;
167
168         result = 0;
169         spin_lock(&xprt_list_lock);
170         list_for_each_entry(t, &xprt_list, list) {
171                 if (strcmp(t->name, transport_name) == 0) {
172                         spin_unlock(&xprt_list_lock);
173                         goto out;
174                 }
175         }
176         spin_unlock(&xprt_list_lock);
177         strcpy(module_name, "xprt");
178         strncat(module_name, transport_name, sizeof t->name);
179         result = request_module(module_name);
180 out:
181         return result;
182 }
183 EXPORT_SYMBOL_GPL(xprt_load_transport);
184
185 /**
186  * xprt_reserve_xprt - serialize write access to transports
187  * @task: task that is requesting access to the transport
188  *
189  * This prevents mixing the payload of separate requests, and prevents
190  * transport connects from colliding with writes.  No congestion control
191  * is provided.
192  */
193 int xprt_reserve_xprt(struct rpc_task *task)
194 {
195         struct rpc_xprt *xprt = task->tk_xprt;
196         struct rpc_rqst *req = task->tk_rqstp;
197
198         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
199                 if (task == xprt->snd_task)
200                         return 1;
201                 if (task == NULL)
202                         return 0;
203                 goto out_sleep;
204         }
205         xprt->snd_task = task;
206         if (req) {
207                 req->rq_bytes_sent = 0;
208                 req->rq_ntrans++;
209         }
210         return 1;
211
212 out_sleep:
213         dprintk("RPC: %5u failed to lock transport %p\n",
214                         task->tk_pid, xprt);
215         task->tk_timeout = 0;
216         task->tk_status = -EAGAIN;
217         if (req && req->rq_ntrans)
218                 rpc_sleep_on(&xprt->resend, task, NULL);
219         else
220                 rpc_sleep_on(&xprt->sending, task, NULL);
221         return 0;
222 }
223 EXPORT_SYMBOL_GPL(xprt_reserve_xprt);
224
225 static void xprt_clear_locked(struct rpc_xprt *xprt)
226 {
227         xprt->snd_task = NULL;
228         if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state) || xprt->shutdown) {
229                 smp_mb__before_clear_bit();
230                 clear_bit(XPRT_LOCKED, &xprt->state);
231                 smp_mb__after_clear_bit();
232         } else
233                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
234 }
235
236 /*
237  * xprt_reserve_xprt_cong - serialize write access to transports
238  * @task: task that is requesting access to the transport
239  *
240  * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
241  * integrated into the decision of whether a request is allowed to be
242  * woken up and given access to the transport.
243  */
244 int xprt_reserve_xprt_cong(struct rpc_task *task)
245 {
246         struct rpc_xprt *xprt = task->tk_xprt;
247         struct rpc_rqst *req = task->tk_rqstp;
248
249         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
250                 if (task == xprt->snd_task)
251                         return 1;
252                 goto out_sleep;
253         }
254         if (__xprt_get_cong(xprt, task)) {
255                 xprt->snd_task = task;
256                 if (req) {
257                         req->rq_bytes_sent = 0;
258                         req->rq_ntrans++;
259                 }
260                 return 1;
261         }
262         xprt_clear_locked(xprt);
263 out_sleep:
264         dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
265         task->tk_timeout = 0;
266         task->tk_status = -EAGAIN;
267         if (req && req->rq_ntrans)
268                 rpc_sleep_on(&xprt->resend, task, NULL);
269         else
270                 rpc_sleep_on(&xprt->sending, task, NULL);
271         return 0;
272 }
273 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
274
275 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
276 {
277         int retval;
278
279         spin_lock_bh(&xprt->transport_lock);
280         retval = xprt->ops->reserve_xprt(task);
281         spin_unlock_bh(&xprt->transport_lock);
282         return retval;
283 }
284
285 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
286 {
287         struct rpc_task *task;
288         struct rpc_rqst *req;
289
290         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
291                 return;
292
293         task = rpc_wake_up_next(&xprt->resend);
294         if (!task) {
295                 task = rpc_wake_up_next(&xprt->sending);
296                 if (!task)
297                         goto out_unlock;
298         }
299
300         req = task->tk_rqstp;
301         xprt->snd_task = task;
302         if (req) {
303                 req->rq_bytes_sent = 0;
304                 req->rq_ntrans++;
305         }
306         return;
307
308 out_unlock:
309         xprt_clear_locked(xprt);
310 }
311
312 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
313 {
314         struct rpc_task *task;
315
316         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
317                 return;
318         if (RPCXPRT_CONGESTED(xprt))
319                 goto out_unlock;
320         task = rpc_wake_up_next(&xprt->resend);
321         if (!task) {
322                 task = rpc_wake_up_next(&xprt->sending);
323                 if (!task)
324                         goto out_unlock;
325         }
326         if (__xprt_get_cong(xprt, task)) {
327                 struct rpc_rqst *req = task->tk_rqstp;
328                 xprt->snd_task = task;
329                 if (req) {
330                         req->rq_bytes_sent = 0;
331                         req->rq_ntrans++;
332                 }
333                 return;
334         }
335 out_unlock:
336         xprt_clear_locked(xprt);
337 }
338
339 /**
340  * xprt_release_xprt - allow other requests to use a transport
341  * @xprt: transport with other tasks potentially waiting
342  * @task: task that is releasing access to the transport
343  *
344  * Note that "task" can be NULL.  No congestion control is provided.
345  */
346 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
347 {
348         if (xprt->snd_task == task) {
349                 xprt_clear_locked(xprt);
350                 __xprt_lock_write_next(xprt);
351         }
352 }
353 EXPORT_SYMBOL_GPL(xprt_release_xprt);
354
355 /**
356  * xprt_release_xprt_cong - allow other requests to use a transport
357  * @xprt: transport with other tasks potentially waiting
358  * @task: task that is releasing access to the transport
359  *
360  * Note that "task" can be NULL.  Another task is awoken to use the
361  * transport if the transport's congestion window allows it.
362  */
363 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
364 {
365         if (xprt->snd_task == task) {
366                 xprt_clear_locked(xprt);
367                 __xprt_lock_write_next_cong(xprt);
368         }
369 }
370 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
371
372 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
373 {
374         spin_lock_bh(&xprt->transport_lock);
375         xprt->ops->release_xprt(xprt, task);
376         spin_unlock_bh(&xprt->transport_lock);
377 }
378
379 /*
380  * Van Jacobson congestion avoidance. Check if the congestion window
381  * overflowed. Put the task to sleep if this is the case.
382  */
383 static int
384 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
385 {
386         struct rpc_rqst *req = task->tk_rqstp;
387
388         if (req->rq_cong)
389                 return 1;
390         dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
391                         task->tk_pid, xprt->cong, xprt->cwnd);
392         if (RPCXPRT_CONGESTED(xprt))
393                 return 0;
394         req->rq_cong = 1;
395         xprt->cong += RPC_CWNDSCALE;
396         return 1;
397 }
398
399 /*
400  * Adjust the congestion window, and wake up the next task
401  * that has been sleeping due to congestion
402  */
403 static void
404 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
405 {
406         if (!req->rq_cong)
407                 return;
408         req->rq_cong = 0;
409         xprt->cong -= RPC_CWNDSCALE;
410         __xprt_lock_write_next_cong(xprt);
411 }
412
413 /**
414  * xprt_release_rqst_cong - housekeeping when request is complete
415  * @task: RPC request that recently completed
416  *
417  * Useful for transports that require congestion control.
418  */
419 void xprt_release_rqst_cong(struct rpc_task *task)
420 {
421         __xprt_put_cong(task->tk_xprt, task->tk_rqstp);
422 }
423 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
424
425 /**
426  * xprt_adjust_cwnd - adjust transport congestion window
427  * @task: recently completed RPC request used to adjust window
428  * @result: result code of completed RPC request
429  *
430  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
431  */
432 void xprt_adjust_cwnd(struct rpc_task *task, int result)
433 {
434         struct rpc_rqst *req = task->tk_rqstp;
435         struct rpc_xprt *xprt = task->tk_xprt;
436         unsigned long cwnd = xprt->cwnd;
437
438         if (result >= 0 && cwnd <= xprt->cong) {
439                 /* The (cwnd >> 1) term makes sure
440                  * the result gets rounded properly. */
441                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
442                 if (cwnd > RPC_MAXCWND(xprt))
443                         cwnd = RPC_MAXCWND(xprt);
444                 __xprt_lock_write_next_cong(xprt);
445         } else if (result == -ETIMEDOUT) {
446                 cwnd >>= 1;
447                 if (cwnd < RPC_CWNDSCALE)
448                         cwnd = RPC_CWNDSCALE;
449         }
450         dprintk("RPC:       cong %ld, cwnd was %ld, now %ld\n",
451                         xprt->cong, xprt->cwnd, cwnd);
452         xprt->cwnd = cwnd;
453         __xprt_put_cong(xprt, req);
454 }
455 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
456
457 /**
458  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
459  * @xprt: transport with waiting tasks
460  * @status: result code to plant in each task before waking it
461  *
462  */
463 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
464 {
465         if (status < 0)
466                 rpc_wake_up_status(&xprt->pending, status);
467         else
468                 rpc_wake_up(&xprt->pending);
469 }
470 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
471
472 /**
473  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
474  * @task: task to be put to sleep
475  * @action: function pointer to be executed after wait
476  */
477 void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action)
478 {
479         struct rpc_rqst *req = task->tk_rqstp;
480         struct rpc_xprt *xprt = req->rq_xprt;
481
482         task->tk_timeout = req->rq_timeout;
483         rpc_sleep_on(&xprt->pending, task, action);
484 }
485 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
486
487 /**
488  * xprt_write_space - wake the task waiting for transport output buffer space
489  * @xprt: transport with waiting tasks
490  *
491  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
492  */
493 void xprt_write_space(struct rpc_xprt *xprt)
494 {
495         if (unlikely(xprt->shutdown))
496                 return;
497
498         spin_lock_bh(&xprt->transport_lock);
499         if (xprt->snd_task) {
500                 dprintk("RPC:       write space: waking waiting task on "
501                                 "xprt %p\n", xprt);
502                 rpc_wake_up_queued_task(&xprt->pending, xprt->snd_task);
503         }
504         spin_unlock_bh(&xprt->transport_lock);
505 }
506 EXPORT_SYMBOL_GPL(xprt_write_space);
507
508 /**
509  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
510  * @task: task whose timeout is to be set
511  *
512  * Set a request's retransmit timeout based on the transport's
513  * default timeout parameters.  Used by transports that don't adjust
514  * the retransmit timeout based on round-trip time estimation.
515  */
516 void xprt_set_retrans_timeout_def(struct rpc_task *task)
517 {
518         task->tk_timeout = task->tk_rqstp->rq_timeout;
519 }
520 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
521
522 /*
523  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
524  * @task: task whose timeout is to be set
525  *
526  * Set a request's retransmit timeout using the RTT estimator.
527  */
528 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
529 {
530         int timer = task->tk_msg.rpc_proc->p_timer;
531         struct rpc_clnt *clnt = task->tk_client;
532         struct rpc_rtt *rtt = clnt->cl_rtt;
533         struct rpc_rqst *req = task->tk_rqstp;
534         unsigned long max_timeout = clnt->cl_timeout->to_maxval;
535
536         task->tk_timeout = rpc_calc_rto(rtt, timer);
537         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
538         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
539                 task->tk_timeout = max_timeout;
540 }
541 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
542
543 static void xprt_reset_majortimeo(struct rpc_rqst *req)
544 {
545         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
546
547         req->rq_majortimeo = req->rq_timeout;
548         if (to->to_exponential)
549                 req->rq_majortimeo <<= to->to_retries;
550         else
551                 req->rq_majortimeo += to->to_increment * to->to_retries;
552         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
553                 req->rq_majortimeo = to->to_maxval;
554         req->rq_majortimeo += jiffies;
555 }
556
557 /**
558  * xprt_adjust_timeout - adjust timeout values for next retransmit
559  * @req: RPC request containing parameters to use for the adjustment
560  *
561  */
562 int xprt_adjust_timeout(struct rpc_rqst *req)
563 {
564         struct rpc_xprt *xprt = req->rq_xprt;
565         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
566         int status = 0;
567
568         if (time_before(jiffies, req->rq_majortimeo)) {
569                 if (to->to_exponential)
570                         req->rq_timeout <<= 1;
571                 else
572                         req->rq_timeout += to->to_increment;
573                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
574                         req->rq_timeout = to->to_maxval;
575                 req->rq_retries++;
576         } else {
577                 req->rq_timeout = to->to_initval;
578                 req->rq_retries = 0;
579                 xprt_reset_majortimeo(req);
580                 /* Reset the RTT counters == "slow start" */
581                 spin_lock_bh(&xprt->transport_lock);
582                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
583                 spin_unlock_bh(&xprt->transport_lock);
584                 status = -ETIMEDOUT;
585         }
586
587         if (req->rq_timeout == 0) {
588                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
589                 req->rq_timeout = 5 * HZ;
590         }
591         return status;
592 }
593
594 static void xprt_autoclose(struct work_struct *work)
595 {
596         struct rpc_xprt *xprt =
597                 container_of(work, struct rpc_xprt, task_cleanup);
598
599         xprt->ops->close(xprt);
600         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
601         xprt_release_write(xprt, NULL);
602 }
603
604 /**
605  * xprt_disconnect_done - mark a transport as disconnected
606  * @xprt: transport to flag for disconnect
607  *
608  */
609 void xprt_disconnect_done(struct rpc_xprt *xprt)
610 {
611         dprintk("RPC:       disconnected transport %p\n", xprt);
612         spin_lock_bh(&xprt->transport_lock);
613         xprt_clear_connected(xprt);
614         xprt_wake_pending_tasks(xprt, -EAGAIN);
615         spin_unlock_bh(&xprt->transport_lock);
616 }
617 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
618
619 /**
620  * xprt_force_disconnect - force a transport to disconnect
621  * @xprt: transport to disconnect
622  *
623  */
624 void xprt_force_disconnect(struct rpc_xprt *xprt)
625 {
626         /* Don't race with the test_bit() in xprt_clear_locked() */
627         spin_lock_bh(&xprt->transport_lock);
628         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
629         /* Try to schedule an autoclose RPC call */
630         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
631                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
632         xprt_wake_pending_tasks(xprt, -EAGAIN);
633         spin_unlock_bh(&xprt->transport_lock);
634 }
635
636 /**
637  * xprt_conditional_disconnect - force a transport to disconnect
638  * @xprt: transport to disconnect
639  * @cookie: 'connection cookie'
640  *
641  * This attempts to break the connection if and only if 'cookie' matches
642  * the current transport 'connection cookie'. It ensures that we don't
643  * try to break the connection more than once when we need to retransmit
644  * a batch of RPC requests.
645  *
646  */
647 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
648 {
649         /* Don't race with the test_bit() in xprt_clear_locked() */
650         spin_lock_bh(&xprt->transport_lock);
651         if (cookie != xprt->connect_cookie)
652                 goto out;
653         if (test_bit(XPRT_CLOSING, &xprt->state) || !xprt_connected(xprt))
654                 goto out;
655         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
656         /* Try to schedule an autoclose RPC call */
657         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
658                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
659         xprt_wake_pending_tasks(xprt, -EAGAIN);
660 out:
661         spin_unlock_bh(&xprt->transport_lock);
662 }
663
664 static void
665 xprt_init_autodisconnect(unsigned long data)
666 {
667         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
668
669         spin_lock(&xprt->transport_lock);
670         if (!list_empty(&xprt->recv) || xprt->shutdown)
671                 goto out_abort;
672         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
673                 goto out_abort;
674         spin_unlock(&xprt->transport_lock);
675         if (xprt_connecting(xprt))
676                 xprt_release_write(xprt, NULL);
677         else
678                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
679         return;
680 out_abort:
681         spin_unlock(&xprt->transport_lock);
682 }
683
684 /**
685  * xprt_connect - schedule a transport connect operation
686  * @task: RPC task that is requesting the connect
687  *
688  */
689 void xprt_connect(struct rpc_task *task)
690 {
691         struct rpc_xprt *xprt = task->tk_xprt;
692
693         dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
694                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
695
696         if (!xprt_bound(xprt)) {
697                 task->tk_status = -EAGAIN;
698                 return;
699         }
700         if (!xprt_lock_write(xprt, task))
701                 return;
702         if (xprt_connected(xprt))
703                 xprt_release_write(xprt, task);
704         else {
705                 if (task->tk_rqstp)
706                         task->tk_rqstp->rq_bytes_sent = 0;
707
708                 task->tk_timeout = xprt->connect_timeout;
709                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
710                 xprt->stat.connect_start = jiffies;
711                 xprt->ops->connect(task);
712         }
713         return;
714 }
715
716 static void xprt_connect_status(struct rpc_task *task)
717 {
718         struct rpc_xprt *xprt = task->tk_xprt;
719
720         if (task->tk_status == 0) {
721                 xprt->stat.connect_count++;
722                 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
723                 dprintk("RPC: %5u xprt_connect_status: connection established\n",
724                                 task->tk_pid);
725                 return;
726         }
727
728         switch (task->tk_status) {
729         case -EAGAIN:
730                 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
731                 break;
732         case -ETIMEDOUT:
733                 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
734                                 "out\n", task->tk_pid);
735                 break;
736         default:
737                 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
738                                 "server %s\n", task->tk_pid, -task->tk_status,
739                                 task->tk_client->cl_server);
740                 xprt_release_write(xprt, task);
741                 task->tk_status = -EIO;
742         }
743 }
744
745 /**
746  * xprt_lookup_rqst - find an RPC request corresponding to an XID
747  * @xprt: transport on which the original request was transmitted
748  * @xid: RPC XID of incoming reply
749  *
750  */
751 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
752 {
753         struct list_head *pos;
754
755         list_for_each(pos, &xprt->recv) {
756                 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
757                 if (entry->rq_xid == xid)
758                         return entry;
759         }
760
761         dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
762                         ntohl(xid));
763         xprt->stat.bad_xids++;
764         return NULL;
765 }
766 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
767
768 /**
769  * xprt_update_rtt - update an RPC client's RTT state after receiving a reply
770  * @task: RPC request that recently completed
771  *
772  */
773 void xprt_update_rtt(struct rpc_task *task)
774 {
775         struct rpc_rqst *req = task->tk_rqstp;
776         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
777         unsigned timer = task->tk_msg.rpc_proc->p_timer;
778
779         if (timer) {
780                 if (req->rq_ntrans == 1)
781                         rpc_update_rtt(rtt, timer,
782                                         (long)jiffies - req->rq_xtime);
783                 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
784         }
785 }
786 EXPORT_SYMBOL_GPL(xprt_update_rtt);
787
788 /**
789  * xprt_complete_rqst - called when reply processing is complete
790  * @task: RPC request that recently completed
791  * @copied: actual number of bytes received from the transport
792  *
793  * Caller holds transport lock.
794  */
795 void xprt_complete_rqst(struct rpc_task *task, int copied)
796 {
797         struct rpc_rqst *req = task->tk_rqstp;
798         struct rpc_xprt *xprt = req->rq_xprt;
799
800         dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
801                         task->tk_pid, ntohl(req->rq_xid), copied);
802
803         xprt->stat.recvs++;
804         task->tk_rtt = (long)jiffies - req->rq_xtime;
805
806         list_del_init(&req->rq_list);
807         req->rq_private_buf.len = copied;
808         /* Ensure all writes are done before we update req->rq_received */
809         smp_wmb();
810         req->rq_received = copied;
811         rpc_wake_up_queued_task(&xprt->pending, task);
812 }
813 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
814
815 static void xprt_timer(struct rpc_task *task)
816 {
817         struct rpc_rqst *req = task->tk_rqstp;
818         struct rpc_xprt *xprt = req->rq_xprt;
819
820         if (task->tk_status != -ETIMEDOUT)
821                 return;
822         dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
823
824         spin_lock_bh(&xprt->transport_lock);
825         if (!req->rq_received) {
826                 if (xprt->ops->timer)
827                         xprt->ops->timer(task);
828         } else
829                 task->tk_status = 0;
830         spin_unlock_bh(&xprt->transport_lock);
831 }
832
833 /**
834  * xprt_prepare_transmit - reserve the transport before sending a request
835  * @task: RPC task about to send a request
836  *
837  */
838 int xprt_prepare_transmit(struct rpc_task *task)
839 {
840         struct rpc_rqst *req = task->tk_rqstp;
841         struct rpc_xprt *xprt = req->rq_xprt;
842         int err = 0;
843
844         dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
845
846         spin_lock_bh(&xprt->transport_lock);
847         if (req->rq_received && !req->rq_bytes_sent) {
848                 err = req->rq_received;
849                 goto out_unlock;
850         }
851         if (!xprt->ops->reserve_xprt(task))
852                 err = -EAGAIN;
853 out_unlock:
854         spin_unlock_bh(&xprt->transport_lock);
855         return err;
856 }
857
858 void xprt_end_transmit(struct rpc_task *task)
859 {
860         xprt_release_write(task->tk_xprt, task);
861 }
862
863 /**
864  * xprt_transmit - send an RPC request on a transport
865  * @task: controlling RPC task
866  *
867  * We have to copy the iovec because sendmsg fiddles with its contents.
868  */
869 void xprt_transmit(struct rpc_task *task)
870 {
871         struct rpc_rqst *req = task->tk_rqstp;
872         struct rpc_xprt *xprt = req->rq_xprt;
873         int status;
874
875         dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
876
877         if (!req->rq_received) {
878                 if (list_empty(&req->rq_list)) {
879                         spin_lock_bh(&xprt->transport_lock);
880                         /* Update the softirq receive buffer */
881                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
882                                         sizeof(req->rq_private_buf));
883                         /* Add request to the receive list */
884                         list_add_tail(&req->rq_list, &xprt->recv);
885                         spin_unlock_bh(&xprt->transport_lock);
886                         xprt_reset_majortimeo(req);
887                         /* Turn off autodisconnect */
888                         del_singleshot_timer_sync(&xprt->timer);
889                 }
890         } else if (!req->rq_bytes_sent)
891                 return;
892
893         req->rq_connect_cookie = xprt->connect_cookie;
894         req->rq_xtime = jiffies;
895         status = xprt->ops->send_request(task);
896         if (status != 0) {
897                 task->tk_status = status;
898                 return;
899         }
900
901         dprintk("RPC: %5u xmit complete\n", task->tk_pid);
902         spin_lock_bh(&xprt->transport_lock);
903
904         xprt->ops->set_retrans_timeout(task);
905
906         xprt->stat.sends++;
907         xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
908         xprt->stat.bklog_u += xprt->backlog.qlen;
909
910         /* Don't race with disconnect */
911         if (!xprt_connected(xprt))
912                 task->tk_status = -ENOTCONN;
913         else if (!req->rq_received)
914                 rpc_sleep_on(&xprt->pending, task, xprt_timer);
915         spin_unlock_bh(&xprt->transport_lock);
916 }
917
918 static inline void do_xprt_reserve(struct rpc_task *task)
919 {
920         struct rpc_xprt *xprt = task->tk_xprt;
921
922         task->tk_status = 0;
923         if (task->tk_rqstp)
924                 return;
925         if (!list_empty(&xprt->free)) {
926                 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
927                 list_del_init(&req->rq_list);
928                 task->tk_rqstp = req;
929                 xprt_request_init(task, xprt);
930                 return;
931         }
932         dprintk("RPC:       waiting for request slot\n");
933         task->tk_status = -EAGAIN;
934         task->tk_timeout = 0;
935         rpc_sleep_on(&xprt->backlog, task, NULL);
936 }
937
938 /**
939  * xprt_reserve - allocate an RPC request slot
940  * @task: RPC task requesting a slot allocation
941  *
942  * If no more slots are available, place the task on the transport's
943  * backlog queue.
944  */
945 void xprt_reserve(struct rpc_task *task)
946 {
947         struct rpc_xprt *xprt = task->tk_xprt;
948
949         task->tk_status = -EIO;
950         spin_lock(&xprt->reserve_lock);
951         do_xprt_reserve(task);
952         spin_unlock(&xprt->reserve_lock);
953 }
954
955 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
956 {
957         return xprt->xid++;
958 }
959
960 static inline void xprt_init_xid(struct rpc_xprt *xprt)
961 {
962         xprt->xid = net_random();
963 }
964
965 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
966 {
967         struct rpc_rqst *req = task->tk_rqstp;
968
969         req->rq_timeout = task->tk_client->cl_timeout->to_initval;
970         req->rq_task    = task;
971         req->rq_xprt    = xprt;
972         req->rq_buffer  = NULL;
973         req->rq_xid     = xprt_alloc_xid(xprt);
974         req->rq_release_snd_buf = NULL;
975         xprt_reset_majortimeo(req);
976         dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
977                         req, ntohl(req->rq_xid));
978 }
979
980 /**
981  * xprt_release - release an RPC request slot
982  * @task: task which is finished with the slot
983  *
984  */
985 void xprt_release(struct rpc_task *task)
986 {
987         struct rpc_xprt *xprt = task->tk_xprt;
988         struct rpc_rqst *req;
989
990         if (!(req = task->tk_rqstp))
991                 return;
992         rpc_count_iostats(task);
993         spin_lock_bh(&xprt->transport_lock);
994         xprt->ops->release_xprt(xprt, task);
995         if (xprt->ops->release_request)
996                 xprt->ops->release_request(task);
997         if (!list_empty(&req->rq_list))
998                 list_del(&req->rq_list);
999         xprt->last_used = jiffies;
1000         if (list_empty(&xprt->recv))
1001                 mod_timer(&xprt->timer,
1002                                 xprt->last_used + xprt->idle_timeout);
1003         spin_unlock_bh(&xprt->transport_lock);
1004         xprt->ops->buf_free(req->rq_buffer);
1005         task->tk_rqstp = NULL;
1006         if (req->rq_release_snd_buf)
1007                 req->rq_release_snd_buf(req);
1008         memset(req, 0, sizeof(*req));   /* mark unused */
1009
1010         dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1011
1012         spin_lock(&xprt->reserve_lock);
1013         list_add(&req->rq_list, &xprt->free);
1014         rpc_wake_up_next(&xprt->backlog);
1015         spin_unlock(&xprt->reserve_lock);
1016 }
1017
1018 /**
1019  * xprt_create_transport - create an RPC transport
1020  * @args: rpc transport creation arguments
1021  *
1022  */
1023 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1024 {
1025         struct rpc_xprt *xprt;
1026         struct rpc_rqst *req;
1027         struct xprt_class *t;
1028
1029         spin_lock(&xprt_list_lock);
1030         list_for_each_entry(t, &xprt_list, list) {
1031                 if (t->ident == args->ident) {
1032                         spin_unlock(&xprt_list_lock);
1033                         goto found;
1034                 }
1035         }
1036         spin_unlock(&xprt_list_lock);
1037         printk(KERN_ERR "RPC: transport (%d) not supported\n", args->ident);
1038         return ERR_PTR(-EIO);
1039
1040 found:
1041         xprt = t->setup(args);
1042         if (IS_ERR(xprt)) {
1043                 dprintk("RPC:       xprt_create_transport: failed, %ld\n",
1044                                 -PTR_ERR(xprt));
1045                 return xprt;
1046         }
1047
1048         kref_init(&xprt->kref);
1049         spin_lock_init(&xprt->transport_lock);
1050         spin_lock_init(&xprt->reserve_lock);
1051
1052         INIT_LIST_HEAD(&xprt->free);
1053         INIT_LIST_HEAD(&xprt->recv);
1054         INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1055         setup_timer(&xprt->timer, xprt_init_autodisconnect,
1056                         (unsigned long)xprt);
1057         xprt->last_used = jiffies;
1058         xprt->cwnd = RPC_INITCWND;
1059         xprt->bind_index = 0;
1060
1061         rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1062         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1063         rpc_init_wait_queue(&xprt->sending, "xprt_sending");
1064         rpc_init_wait_queue(&xprt->resend, "xprt_resend");
1065         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1066
1067         /* initialize free list */
1068         for (req = &xprt->slot[xprt->max_reqs-1]; req >= &xprt->slot[0]; req--)
1069                 list_add(&req->rq_list, &xprt->free);
1070
1071         xprt_init_xid(xprt);
1072
1073         dprintk("RPC:       created transport %p with %u slots\n", xprt,
1074                         xprt->max_reqs);
1075
1076         return xprt;
1077 }
1078
1079 /**
1080  * xprt_destroy - destroy an RPC transport, killing off all requests.
1081  * @kref: kref for the transport to destroy
1082  *
1083  */
1084 static void xprt_destroy(struct kref *kref)
1085 {
1086         struct rpc_xprt *xprt = container_of(kref, struct rpc_xprt, kref);
1087
1088         dprintk("RPC:       destroying transport %p\n", xprt);
1089         xprt->shutdown = 1;
1090         del_timer_sync(&xprt->timer);
1091
1092         rpc_destroy_wait_queue(&xprt->binding);
1093         rpc_destroy_wait_queue(&xprt->pending);
1094         rpc_destroy_wait_queue(&xprt->sending);
1095         rpc_destroy_wait_queue(&xprt->resend);
1096         rpc_destroy_wait_queue(&xprt->backlog);
1097         /*
1098          * Tear down transport state and free the rpc_xprt
1099          */
1100         xprt->ops->destroy(xprt);
1101 }
1102
1103 /**
1104  * xprt_put - release a reference to an RPC transport.
1105  * @xprt: pointer to the transport
1106  *
1107  */
1108 void xprt_put(struct rpc_xprt *xprt)
1109 {
1110         kref_put(&xprt->kref, xprt_destroy);
1111 }
1112
1113 /**
1114  * xprt_get - return a reference to an RPC transport.
1115  * @xprt: pointer to the transport
1116  *
1117  */
1118 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1119 {
1120         kref_get(&xprt->kref);
1121         return xprt;
1122 }