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