dff07795bd16af0dbbb5ff1cb51e4729a28ef3ab
[linux-2.6.git] / net / sunrpc / sched.c
1 /*
2  * linux/net/sunrpc/sched.c
3  *
4  * Scheduling for synchronous and asynchronous RPC requests.
5  *
6  * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
7  * 
8  * TCP NFS related read + write fixes
9  * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
10  */
11
12 #include <linux/module.h>
13
14 #include <linux/sched.h>
15 #include <linux/interrupt.h>
16 #include <linux/slab.h>
17 #include <linux/mempool.h>
18 #include <linux/smp.h>
19 #include <linux/smp_lock.h>
20 #include <linux/spinlock.h>
21 #include <linux/mutex.h>
22
23 #include <linux/sunrpc/clnt.h>
24 #include <linux/sunrpc/xprt.h>
25
26 #ifdef RPC_DEBUG
27 #define RPCDBG_FACILITY         RPCDBG_SCHED
28 #define RPC_TASK_MAGIC_ID       0xf00baa
29 static int                      rpc_task_id;
30 #endif
31
32 /*
33  * RPC slabs and memory pools
34  */
35 #define RPC_BUFFER_MAXSIZE      (2048)
36 #define RPC_BUFFER_POOLSIZE     (8)
37 #define RPC_TASK_POOLSIZE       (8)
38 static kmem_cache_t     *rpc_task_slabp __read_mostly;
39 static kmem_cache_t     *rpc_buffer_slabp __read_mostly;
40 static mempool_t        *rpc_task_mempool __read_mostly;
41 static mempool_t        *rpc_buffer_mempool __read_mostly;
42
43 static void                     __rpc_default_timer(struct rpc_task *task);
44 static void                     rpciod_killall(void);
45 static void                     rpc_async_schedule(void *);
46
47 /*
48  * RPC tasks that create another task (e.g. for contacting the portmapper)
49  * will wait on this queue for their child's completion
50  */
51 static RPC_WAITQ(childq, "childq");
52
53 /*
54  * RPC tasks sit here while waiting for conditions to improve.
55  */
56 static RPC_WAITQ(delay_queue, "delayq");
57
58 /*
59  * All RPC tasks are linked into this list
60  */
61 static LIST_HEAD(all_tasks);
62
63 /*
64  * rpciod-related stuff
65  */
66 static DEFINE_MUTEX(rpciod_mutex);
67 static unsigned int             rpciod_users;
68 static struct workqueue_struct *rpciod_workqueue;
69
70 /*
71  * Spinlock for other critical sections of code.
72  */
73 static DEFINE_SPINLOCK(rpc_sched_lock);
74
75 /*
76  * Disable the timer for a given RPC task. Should be called with
77  * queue->lock and bh_disabled in order to avoid races within
78  * rpc_run_timer().
79  */
80 static inline void
81 __rpc_disable_timer(struct rpc_task *task)
82 {
83         dprintk("RPC: %4d disabling timer\n", task->tk_pid);
84         task->tk_timeout_fn = NULL;
85         task->tk_timeout = 0;
86 }
87
88 /*
89  * Run a timeout function.
90  * We use the callback in order to allow __rpc_wake_up_task()
91  * and friends to disable the timer synchronously on SMP systems
92  * without calling del_timer_sync(). The latter could cause a
93  * deadlock if called while we're holding spinlocks...
94  */
95 static void rpc_run_timer(struct rpc_task *task)
96 {
97         void (*callback)(struct rpc_task *);
98
99         callback = task->tk_timeout_fn;
100         task->tk_timeout_fn = NULL;
101         if (callback && RPC_IS_QUEUED(task)) {
102                 dprintk("RPC: %4d running timer\n", task->tk_pid);
103                 callback(task);
104         }
105         smp_mb__before_clear_bit();
106         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
107         smp_mb__after_clear_bit();
108 }
109
110 /*
111  * Set up a timer for the current task.
112  */
113 static inline void
114 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
115 {
116         if (!task->tk_timeout)
117                 return;
118
119         dprintk("RPC: %4d setting alarm for %lu ms\n",
120                         task->tk_pid, task->tk_timeout * 1000 / HZ);
121
122         if (timer)
123                 task->tk_timeout_fn = timer;
124         else
125                 task->tk_timeout_fn = __rpc_default_timer;
126         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
127         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
128 }
129
130 /*
131  * Delete any timer for the current task. Because we use del_timer_sync(),
132  * this function should never be called while holding queue->lock.
133  */
134 static void
135 rpc_delete_timer(struct rpc_task *task)
136 {
137         if (RPC_IS_QUEUED(task))
138                 return;
139         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
140                 del_singleshot_timer_sync(&task->tk_timer);
141                 dprintk("RPC: %4d deleting timer\n", task->tk_pid);
142         }
143 }
144
145 /*
146  * Add new request to a priority queue.
147  */
148 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
149 {
150         struct list_head *q;
151         struct rpc_task *t;
152
153         INIT_LIST_HEAD(&task->u.tk_wait.links);
154         q = &queue->tasks[task->tk_priority];
155         if (unlikely(task->tk_priority > queue->maxpriority))
156                 q = &queue->tasks[queue->maxpriority];
157         list_for_each_entry(t, q, u.tk_wait.list) {
158                 if (t->tk_cookie == task->tk_cookie) {
159                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
160                         return;
161                 }
162         }
163         list_add_tail(&task->u.tk_wait.list, q);
164 }
165
166 /*
167  * Add new request to wait queue.
168  *
169  * Swapper tasks always get inserted at the head of the queue.
170  * This should avoid many nasty memory deadlocks and hopefully
171  * improve overall performance.
172  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
173  */
174 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
175 {
176         BUG_ON (RPC_IS_QUEUED(task));
177
178         if (RPC_IS_PRIORITY(queue))
179                 __rpc_add_wait_queue_priority(queue, task);
180         else if (RPC_IS_SWAPPER(task))
181                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
182         else
183                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
184         task->u.tk_wait.rpc_waitq = queue;
185         rpc_set_queued(task);
186
187         dprintk("RPC: %4d added to queue %p \"%s\"\n",
188                                 task->tk_pid, queue, rpc_qname(queue));
189 }
190
191 /*
192  * Remove request from a priority queue.
193  */
194 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
195 {
196         struct rpc_task *t;
197
198         if (!list_empty(&task->u.tk_wait.links)) {
199                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
200                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
201                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
202         }
203         list_del(&task->u.tk_wait.list);
204 }
205
206 /*
207  * Remove request from queue.
208  * Note: must be called with spin lock held.
209  */
210 static void __rpc_remove_wait_queue(struct rpc_task *task)
211 {
212         struct rpc_wait_queue *queue;
213         queue = task->u.tk_wait.rpc_waitq;
214
215         if (RPC_IS_PRIORITY(queue))
216                 __rpc_remove_wait_queue_priority(task);
217         else
218                 list_del(&task->u.tk_wait.list);
219         dprintk("RPC: %4d removed from queue %p \"%s\"\n",
220                                 task->tk_pid, queue, rpc_qname(queue));
221 }
222
223 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
224 {
225         queue->priority = priority;
226         queue->count = 1 << (priority * 2);
227 }
228
229 static inline void rpc_set_waitqueue_cookie(struct rpc_wait_queue *queue, unsigned long cookie)
230 {
231         queue->cookie = cookie;
232         queue->nr = RPC_BATCH_COUNT;
233 }
234
235 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
236 {
237         rpc_set_waitqueue_priority(queue, queue->maxpriority);
238         rpc_set_waitqueue_cookie(queue, 0);
239 }
240
241 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, int maxprio)
242 {
243         int i;
244
245         spin_lock_init(&queue->lock);
246         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
247                 INIT_LIST_HEAD(&queue->tasks[i]);
248         queue->maxpriority = maxprio;
249         rpc_reset_waitqueue_priority(queue);
250 #ifdef RPC_DEBUG
251         queue->name = qname;
252 #endif
253 }
254
255 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
256 {
257         __rpc_init_priority_wait_queue(queue, qname, RPC_PRIORITY_HIGH);
258 }
259
260 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
261 {
262         __rpc_init_priority_wait_queue(queue, qname, 0);
263 }
264 EXPORT_SYMBOL(rpc_init_wait_queue);
265
266 static int rpc_wait_bit_interruptible(void *word)
267 {
268         if (signal_pending(current))
269                 return -ERESTARTSYS;
270         schedule();
271         return 0;
272 }
273
274 /*
275  * Mark an RPC call as having completed by clearing the 'active' bit
276  */
277 static inline void rpc_mark_complete_task(struct rpc_task *task)
278 {
279         rpc_clear_active(task);
280         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
281 }
282
283 /*
284  * Allow callers to wait for completion of an RPC call
285  */
286 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
287 {
288         if (action == NULL)
289                 action = rpc_wait_bit_interruptible;
290         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
291                         action, TASK_INTERRUPTIBLE);
292 }
293 EXPORT_SYMBOL(__rpc_wait_for_completion_task);
294
295 /*
296  * Make an RPC task runnable.
297  *
298  * Note: If the task is ASYNC, this must be called with 
299  * the spinlock held to protect the wait queue operation.
300  */
301 static void rpc_make_runnable(struct rpc_task *task)
302 {
303         int do_ret;
304
305         BUG_ON(task->tk_timeout_fn);
306         do_ret = rpc_test_and_set_running(task);
307         rpc_clear_queued(task);
308         if (do_ret)
309                 return;
310         if (RPC_IS_ASYNC(task)) {
311                 int status;
312
313                 INIT_WORK(&task->u.tk_work, rpc_async_schedule, (void *)task);
314                 status = queue_work(task->tk_workqueue, &task->u.tk_work);
315                 if (status < 0) {
316                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
317                         task->tk_status = status;
318                         return;
319                 }
320         } else
321                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
322 }
323
324 /*
325  * Place a newly initialized task on the workqueue.
326  */
327 static inline void
328 rpc_schedule_run(struct rpc_task *task)
329 {
330         rpc_set_active(task);
331         rpc_make_runnable(task);
332 }
333
334 /*
335  * Prepare for sleeping on a wait queue.
336  * By always appending tasks to the list we ensure FIFO behavior.
337  * NB: An RPC task will only receive interrupt-driven events as long
338  * as it's on a wait queue.
339  */
340 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
341                         rpc_action action, rpc_action timer)
342 {
343         dprintk("RPC: %4d sleep_on(queue \"%s\" time %ld)\n", task->tk_pid,
344                                 rpc_qname(q), jiffies);
345
346         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
347                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
348                 return;
349         }
350
351         /* Mark the task as being activated if so needed */
352         rpc_set_active(task);
353
354         __rpc_add_wait_queue(q, task);
355
356         BUG_ON(task->tk_callback != NULL);
357         task->tk_callback = action;
358         __rpc_add_timer(task, timer);
359 }
360
361 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
362                                 rpc_action action, rpc_action timer)
363 {
364         /*
365          * Protect the queue operations.
366          */
367         spin_lock_bh(&q->lock);
368         __rpc_sleep_on(q, task, action, timer);
369         spin_unlock_bh(&q->lock);
370 }
371
372 /**
373  * __rpc_do_wake_up_task - wake up a single rpc_task
374  * @task: task to be woken up
375  *
376  * Caller must hold queue->lock, and have cleared the task queued flag.
377  */
378 static void __rpc_do_wake_up_task(struct rpc_task *task)
379 {
380         dprintk("RPC: %4d __rpc_wake_up_task (now %ld)\n", task->tk_pid, jiffies);
381
382 #ifdef RPC_DEBUG
383         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
384 #endif
385         /* Has the task been executed yet? If not, we cannot wake it up! */
386         if (!RPC_IS_ACTIVATED(task)) {
387                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
388                 return;
389         }
390
391         __rpc_disable_timer(task);
392         __rpc_remove_wait_queue(task);
393
394         rpc_make_runnable(task);
395
396         dprintk("RPC:      __rpc_wake_up_task done\n");
397 }
398
399 /*
400  * Wake up the specified task
401  */
402 static void __rpc_wake_up_task(struct rpc_task *task)
403 {
404         if (rpc_start_wakeup(task)) {
405                 if (RPC_IS_QUEUED(task))
406                         __rpc_do_wake_up_task(task);
407                 rpc_finish_wakeup(task);
408         }
409 }
410
411 /*
412  * Default timeout handler if none specified by user
413  */
414 static void
415 __rpc_default_timer(struct rpc_task *task)
416 {
417         dprintk("RPC: %d timeout (default timer)\n", task->tk_pid);
418         task->tk_status = -ETIMEDOUT;
419         rpc_wake_up_task(task);
420 }
421
422 /*
423  * Wake up the specified task
424  */
425 void rpc_wake_up_task(struct rpc_task *task)
426 {
427         if (rpc_start_wakeup(task)) {
428                 if (RPC_IS_QUEUED(task)) {
429                         struct rpc_wait_queue *queue = task->u.tk_wait.rpc_waitq;
430
431                         spin_lock_bh(&queue->lock);
432                         __rpc_do_wake_up_task(task);
433                         spin_unlock_bh(&queue->lock);
434                 }
435                 rpc_finish_wakeup(task);
436         }
437 }
438
439 /*
440  * Wake up the next task on a priority queue.
441  */
442 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
443 {
444         struct list_head *q;
445         struct rpc_task *task;
446
447         /*
448          * Service a batch of tasks from a single cookie.
449          */
450         q = &queue->tasks[queue->priority];
451         if (!list_empty(q)) {
452                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
453                 if (queue->cookie == task->tk_cookie) {
454                         if (--queue->nr)
455                                 goto out;
456                         list_move_tail(&task->u.tk_wait.list, q);
457                 }
458                 /*
459                  * Check if we need to switch queues.
460                  */
461                 if (--queue->count)
462                         goto new_cookie;
463         }
464
465         /*
466          * Service the next queue.
467          */
468         do {
469                 if (q == &queue->tasks[0])
470                         q = &queue->tasks[queue->maxpriority];
471                 else
472                         q = q - 1;
473                 if (!list_empty(q)) {
474                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
475                         goto new_queue;
476                 }
477         } while (q != &queue->tasks[queue->priority]);
478
479         rpc_reset_waitqueue_priority(queue);
480         return NULL;
481
482 new_queue:
483         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
484 new_cookie:
485         rpc_set_waitqueue_cookie(queue, task->tk_cookie);
486 out:
487         __rpc_wake_up_task(task);
488         return task;
489 }
490
491 /*
492  * Wake up the next task on the wait queue.
493  */
494 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
495 {
496         struct rpc_task *task = NULL;
497
498         dprintk("RPC:      wake_up_next(%p \"%s\")\n", queue, rpc_qname(queue));
499         spin_lock_bh(&queue->lock);
500         if (RPC_IS_PRIORITY(queue))
501                 task = __rpc_wake_up_next_priority(queue);
502         else {
503                 task_for_first(task, &queue->tasks[0])
504                         __rpc_wake_up_task(task);
505         }
506         spin_unlock_bh(&queue->lock);
507
508         return task;
509 }
510
511 /**
512  * rpc_wake_up - wake up all rpc_tasks
513  * @queue: rpc_wait_queue on which the tasks are sleeping
514  *
515  * Grabs queue->lock
516  */
517 void rpc_wake_up(struct rpc_wait_queue *queue)
518 {
519         struct rpc_task *task, *next;
520         struct list_head *head;
521
522         spin_lock_bh(&queue->lock);
523         head = &queue->tasks[queue->maxpriority];
524         for (;;) {
525                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
526                         __rpc_wake_up_task(task);
527                 if (head == &queue->tasks[0])
528                         break;
529                 head--;
530         }
531         spin_unlock_bh(&queue->lock);
532 }
533
534 /**
535  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
536  * @queue: rpc_wait_queue on which the tasks are sleeping
537  * @status: status value to set
538  *
539  * Grabs queue->lock
540  */
541 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
542 {
543         struct rpc_task *task, *next;
544         struct list_head *head;
545
546         spin_lock_bh(&queue->lock);
547         head = &queue->tasks[queue->maxpriority];
548         for (;;) {
549                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
550                         task->tk_status = status;
551                         __rpc_wake_up_task(task);
552                 }
553                 if (head == &queue->tasks[0])
554                         break;
555                 head--;
556         }
557         spin_unlock_bh(&queue->lock);
558 }
559
560 /*
561  * Run a task at a later time
562  */
563 static void     __rpc_atrun(struct rpc_task *);
564 void
565 rpc_delay(struct rpc_task *task, unsigned long delay)
566 {
567         task->tk_timeout = delay;
568         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
569 }
570
571 static void
572 __rpc_atrun(struct rpc_task *task)
573 {
574         task->tk_status = 0;
575         rpc_wake_up_task(task);
576 }
577
578 /*
579  * Helper to call task->tk_ops->rpc_call_prepare
580  */
581 static void rpc_prepare_task(struct rpc_task *task)
582 {
583         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
584 }
585
586 /*
587  * Helper that calls task->tk_ops->rpc_call_done if it exists
588  */
589 void rpc_exit_task(struct rpc_task *task)
590 {
591         task->tk_action = NULL;
592         if (task->tk_ops->rpc_call_done != NULL) {
593                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
594                 if (task->tk_action != NULL) {
595                         WARN_ON(RPC_ASSASSINATED(task));
596                         /* Always release the RPC slot and buffer memory */
597                         xprt_release(task);
598                 }
599         }
600 }
601 EXPORT_SYMBOL(rpc_exit_task);
602
603 /*
604  * This is the RPC `scheduler' (or rather, the finite state machine).
605  */
606 static int __rpc_execute(struct rpc_task *task)
607 {
608         int             status = 0;
609
610         dprintk("RPC: %4d rpc_execute flgs %x\n",
611                                 task->tk_pid, task->tk_flags);
612
613         BUG_ON(RPC_IS_QUEUED(task));
614
615         for (;;) {
616                 /*
617                  * Garbage collection of pending timers...
618                  */
619                 rpc_delete_timer(task);
620
621                 /*
622                  * Execute any pending callback.
623                  */
624                 if (RPC_DO_CALLBACK(task)) {
625                         /* Define a callback save pointer */
626                         void (*save_callback)(struct rpc_task *);
627         
628                         /* 
629                          * If a callback exists, save it, reset it,
630                          * call it.
631                          * The save is needed to stop from resetting
632                          * another callback set within the callback handler
633                          * - Dave
634                          */
635                         save_callback=task->tk_callback;
636                         task->tk_callback=NULL;
637                         lock_kernel();
638                         save_callback(task);
639                         unlock_kernel();
640                 }
641
642                 /*
643                  * Perform the next FSM step.
644                  * tk_action may be NULL when the task has been killed
645                  * by someone else.
646                  */
647                 if (!RPC_IS_QUEUED(task)) {
648                         if (task->tk_action == NULL)
649                                 break;
650                         lock_kernel();
651                         task->tk_action(task);
652                         unlock_kernel();
653                 }
654
655                 /*
656                  * Lockless check for whether task is sleeping or not.
657                  */
658                 if (!RPC_IS_QUEUED(task))
659                         continue;
660                 rpc_clear_running(task);
661                 if (RPC_IS_ASYNC(task)) {
662                         /* Careful! we may have raced... */
663                         if (RPC_IS_QUEUED(task))
664                                 return 0;
665                         if (rpc_test_and_set_running(task))
666                                 return 0;
667                         continue;
668                 }
669
670                 /* sync task: sleep here */
671                 dprintk("RPC: %4d sync task going to sleep\n", task->tk_pid);
672                 /* Note: Caller should be using rpc_clnt_sigmask() */
673                 status = out_of_line_wait_on_bit(&task->tk_runstate,
674                                 RPC_TASK_QUEUED, rpc_wait_bit_interruptible,
675                                 TASK_INTERRUPTIBLE);
676                 if (status == -ERESTARTSYS) {
677                         /*
678                          * When a sync task receives a signal, it exits with
679                          * -ERESTARTSYS. In order to catch any callbacks that
680                          * clean up after sleeping on some queue, we don't
681                          * break the loop here, but go around once more.
682                          */
683                         dprintk("RPC: %4d got signal\n", task->tk_pid);
684                         task->tk_flags |= RPC_TASK_KILLED;
685                         rpc_exit(task, -ERESTARTSYS);
686                         rpc_wake_up_task(task);
687                 }
688                 rpc_set_running(task);
689                 dprintk("RPC: %4d sync task resuming\n", task->tk_pid);
690         }
691
692         dprintk("RPC: %4d, return %d, status %d\n", task->tk_pid, status, task->tk_status);
693         /* Wake up anyone who is waiting for task completion */
694         rpc_mark_complete_task(task);
695         /* Release all resources associated with the task */
696         rpc_release_task(task);
697         return status;
698 }
699
700 /*
701  * User-visible entry point to the scheduler.
702  *
703  * This may be called recursively if e.g. an async NFS task updates
704  * the attributes and finds that dirty pages must be flushed.
705  * NOTE: Upon exit of this function the task is guaranteed to be
706  *       released. In particular note that tk_release() will have
707  *       been called, so your task memory may have been freed.
708  */
709 int
710 rpc_execute(struct rpc_task *task)
711 {
712         rpc_set_active(task);
713         rpc_set_running(task);
714         return __rpc_execute(task);
715 }
716
717 static void rpc_async_schedule(void *arg)
718 {
719         __rpc_execute((struct rpc_task *)arg);
720 }
721
722 /**
723  * rpc_malloc - allocate an RPC buffer
724  * @task: RPC task that will use this buffer
725  * @size: requested byte size
726  *
727  * We try to ensure that some NFS reads and writes can always proceed
728  * by using a mempool when allocating 'small' buffers.
729  * In order to avoid memory starvation triggering more writebacks of
730  * NFS requests, we use GFP_NOFS rather than GFP_KERNEL.
731  */
732 void * rpc_malloc(struct rpc_task *task, size_t size)
733 {
734         struct rpc_rqst *req = task->tk_rqstp;
735         gfp_t   gfp;
736
737         if (task->tk_flags & RPC_TASK_SWAPPER)
738                 gfp = GFP_ATOMIC;
739         else
740                 gfp = GFP_NOFS;
741
742         if (size > RPC_BUFFER_MAXSIZE) {
743                 req->rq_buffer = kmalloc(size, gfp);
744                 if (req->rq_buffer)
745                         req->rq_bufsize = size;
746         } else {
747                 req->rq_buffer = mempool_alloc(rpc_buffer_mempool, gfp);
748                 if (req->rq_buffer)
749                         req->rq_bufsize = RPC_BUFFER_MAXSIZE;
750         }
751         return req->rq_buffer;
752 }
753
754 /**
755  * rpc_free - free buffer allocated via rpc_malloc
756  * @task: RPC task with a buffer to be freed
757  *
758  */
759 void rpc_free(struct rpc_task *task)
760 {
761         struct rpc_rqst *req = task->tk_rqstp;
762
763         if (req->rq_buffer) {
764                 if (req->rq_bufsize == RPC_BUFFER_MAXSIZE)
765                         mempool_free(req->rq_buffer, rpc_buffer_mempool);
766                 else
767                         kfree(req->rq_buffer);
768                 req->rq_buffer = NULL;
769                 req->rq_bufsize = 0;
770         }
771 }
772
773 /*
774  * Creation and deletion of RPC task structures
775  */
776 void rpc_init_task(struct rpc_task *task, struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
777 {
778         memset(task, 0, sizeof(*task));
779         init_timer(&task->tk_timer);
780         task->tk_timer.data     = (unsigned long) task;
781         task->tk_timer.function = (void (*)(unsigned long)) rpc_run_timer;
782         atomic_set(&task->tk_count, 1);
783         task->tk_client = clnt;
784         task->tk_flags  = flags;
785         task->tk_ops = tk_ops;
786         if (tk_ops->rpc_call_prepare != NULL)
787                 task->tk_action = rpc_prepare_task;
788         task->tk_calldata = calldata;
789
790         /* Initialize retry counters */
791         task->tk_garb_retry = 2;
792         task->tk_cred_retry = 2;
793
794         task->tk_priority = RPC_PRIORITY_NORMAL;
795         task->tk_cookie = (unsigned long)current;
796
797         /* Initialize workqueue for async tasks */
798         task->tk_workqueue = rpciod_workqueue;
799
800         if (clnt) {
801                 atomic_inc(&clnt->cl_users);
802                 if (clnt->cl_softrtry)
803                         task->tk_flags |= RPC_TASK_SOFT;
804                 if (!clnt->cl_intr)
805                         task->tk_flags |= RPC_TASK_NOINTR;
806         }
807
808 #ifdef RPC_DEBUG
809         task->tk_magic = RPC_TASK_MAGIC_ID;
810         task->tk_pid = rpc_task_id++;
811 #endif
812         /* Add to global list of all tasks */
813         spin_lock(&rpc_sched_lock);
814         list_add_tail(&task->tk_task, &all_tasks);
815         spin_unlock(&rpc_sched_lock);
816
817         BUG_ON(task->tk_ops == NULL);
818
819         dprintk("RPC: %4d new task procpid %d\n", task->tk_pid,
820                                 current->pid);
821 }
822
823 static struct rpc_task *
824 rpc_alloc_task(void)
825 {
826         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
827 }
828
829 static void rpc_free_task(struct rpc_task *task)
830 {
831         dprintk("RPC: %4d freeing task\n", task->tk_pid);
832         mempool_free(task, rpc_task_mempool);
833 }
834
835 /*
836  * Create a new task for the specified client.  We have to
837  * clean up after an allocation failure, as the client may
838  * have specified "oneshot".
839  */
840 struct rpc_task *rpc_new_task(struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
841 {
842         struct rpc_task *task;
843
844         task = rpc_alloc_task();
845         if (!task)
846                 goto cleanup;
847
848         rpc_init_task(task, clnt, flags, tk_ops, calldata);
849
850         dprintk("RPC: %4d allocated task\n", task->tk_pid);
851         task->tk_flags |= RPC_TASK_DYNAMIC;
852 out:
853         return task;
854
855 cleanup:
856         /* Check whether to release the client */
857         if (clnt) {
858                 printk("rpc_new_task: failed, users=%d, oneshot=%d\n",
859                         atomic_read(&clnt->cl_users), clnt->cl_oneshot);
860                 atomic_inc(&clnt->cl_users); /* pretend we were used ... */
861                 rpc_release_client(clnt);
862         }
863         goto out;
864 }
865
866 void rpc_release_task(struct rpc_task *task)
867 {
868         const struct rpc_call_ops *tk_ops = task->tk_ops;
869         void *calldata = task->tk_calldata;
870
871 #ifdef RPC_DEBUG
872         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
873 #endif
874         if (!atomic_dec_and_test(&task->tk_count))
875                 return;
876         dprintk("RPC: %4d release task\n", task->tk_pid);
877
878         /* Remove from global task list */
879         spin_lock(&rpc_sched_lock);
880         list_del(&task->tk_task);
881         spin_unlock(&rpc_sched_lock);
882
883         BUG_ON (RPC_IS_QUEUED(task));
884
885         /* Synchronously delete any running timer */
886         rpc_delete_timer(task);
887
888         /* Release resources */
889         if (task->tk_rqstp)
890                 xprt_release(task);
891         if (task->tk_msg.rpc_cred)
892                 rpcauth_unbindcred(task);
893         if (task->tk_client) {
894                 rpc_release_client(task->tk_client);
895                 task->tk_client = NULL;
896         }
897
898 #ifdef RPC_DEBUG
899         task->tk_magic = 0;
900 #endif
901         if (task->tk_flags & RPC_TASK_DYNAMIC)
902                 rpc_free_task(task);
903         if (tk_ops->rpc_release)
904                 tk_ops->rpc_release(calldata);
905 }
906
907 /**
908  * rpc_run_task - Allocate a new RPC task, then run rpc_execute against it
909  * @clnt: pointer to RPC client
910  * @flags: RPC flags
911  * @ops: RPC call ops
912  * @data: user call data
913  */
914 struct rpc_task *rpc_run_task(struct rpc_clnt *clnt, int flags,
915                                         const struct rpc_call_ops *ops,
916                                         void *data)
917 {
918         struct rpc_task *task;
919         task = rpc_new_task(clnt, flags, ops, data);
920         if (task == NULL)
921                 return ERR_PTR(-ENOMEM);
922         atomic_inc(&task->tk_count);
923         rpc_execute(task);
924         return task;
925 }
926 EXPORT_SYMBOL(rpc_run_task);
927
928 /**
929  * rpc_find_parent - find the parent of a child task.
930  * @child: child task
931  * @parent: parent task
932  *
933  * Checks that the parent task is still sleeping on the
934  * queue 'childq'. If so returns a pointer to the parent.
935  * Upon failure returns NULL.
936  *
937  * Caller must hold childq.lock
938  */
939 static inline struct rpc_task *rpc_find_parent(struct rpc_task *child, struct rpc_task *parent)
940 {
941         struct rpc_task *task;
942         struct list_head *le;
943
944         task_for_each(task, le, &childq.tasks[0])
945                 if (task == parent)
946                         return parent;
947
948         return NULL;
949 }
950
951 static void rpc_child_exit(struct rpc_task *child, void *calldata)
952 {
953         struct rpc_task *parent;
954
955         spin_lock_bh(&childq.lock);
956         if ((parent = rpc_find_parent(child, calldata)) != NULL) {
957                 parent->tk_status = child->tk_status;
958                 __rpc_wake_up_task(parent);
959         }
960         spin_unlock_bh(&childq.lock);
961 }
962
963 static const struct rpc_call_ops rpc_child_ops = {
964         .rpc_call_done = rpc_child_exit,
965 };
966
967 /*
968  * Note: rpc_new_task releases the client after a failure.
969  */
970 struct rpc_task *
971 rpc_new_child(struct rpc_clnt *clnt, struct rpc_task *parent)
972 {
973         struct rpc_task *task;
974
975         task = rpc_new_task(clnt, RPC_TASK_ASYNC | RPC_TASK_CHILD, &rpc_child_ops, parent);
976         if (!task)
977                 goto fail;
978         return task;
979
980 fail:
981         parent->tk_status = -ENOMEM;
982         return NULL;
983 }
984
985 void rpc_run_child(struct rpc_task *task, struct rpc_task *child, rpc_action func)
986 {
987         spin_lock_bh(&childq.lock);
988         /* N.B. Is it possible for the child to have already finished? */
989         __rpc_sleep_on(&childq, task, func, NULL);
990         rpc_schedule_run(child);
991         spin_unlock_bh(&childq.lock);
992 }
993
994 /*
995  * Kill all tasks for the given client.
996  * XXX: kill their descendants as well?
997  */
998 void rpc_killall_tasks(struct rpc_clnt *clnt)
999 {
1000         struct rpc_task *rovr;
1001         struct list_head *le;
1002
1003         dprintk("RPC:      killing all tasks for client %p\n", clnt);
1004
1005         /*
1006          * Spin lock all_tasks to prevent changes...
1007          */
1008         spin_lock(&rpc_sched_lock);
1009         alltask_for_each(rovr, le, &all_tasks) {
1010                 if (! RPC_IS_ACTIVATED(rovr))
1011                         continue;
1012                 if (!clnt || rovr->tk_client == clnt) {
1013                         rovr->tk_flags |= RPC_TASK_KILLED;
1014                         rpc_exit(rovr, -EIO);
1015                         rpc_wake_up_task(rovr);
1016                 }
1017         }
1018         spin_unlock(&rpc_sched_lock);
1019 }
1020
1021 static DECLARE_MUTEX_LOCKED(rpciod_running);
1022
1023 static void rpciod_killall(void)
1024 {
1025         unsigned long flags;
1026
1027         while (!list_empty(&all_tasks)) {
1028                 clear_thread_flag(TIF_SIGPENDING);
1029                 rpc_killall_tasks(NULL);
1030                 flush_workqueue(rpciod_workqueue);
1031                 if (!list_empty(&all_tasks)) {
1032                         dprintk("rpciod_killall: waiting for tasks to exit\n");
1033                         yield();
1034                 }
1035         }
1036
1037         spin_lock_irqsave(&current->sighand->siglock, flags);
1038         recalc_sigpending();
1039         spin_unlock_irqrestore(&current->sighand->siglock, flags);
1040 }
1041
1042 /*
1043  * Start up the rpciod process if it's not already running.
1044  */
1045 int
1046 rpciod_up(void)
1047 {
1048         struct workqueue_struct *wq;
1049         int error = 0;
1050
1051         mutex_lock(&rpciod_mutex);
1052         dprintk("rpciod_up: users %d\n", rpciod_users);
1053         rpciod_users++;
1054         if (rpciod_workqueue)
1055                 goto out;
1056         /*
1057          * If there's no pid, we should be the first user.
1058          */
1059         if (rpciod_users > 1)
1060                 printk(KERN_WARNING "rpciod_up: no workqueue, %d users??\n", rpciod_users);
1061         /*
1062          * Create the rpciod thread and wait for it to start.
1063          */
1064         error = -ENOMEM;
1065         wq = create_workqueue("rpciod");
1066         if (wq == NULL) {
1067                 printk(KERN_WARNING "rpciod_up: create workqueue failed, error=%d\n", error);
1068                 rpciod_users--;
1069                 goto out;
1070         }
1071         rpciod_workqueue = wq;
1072         error = 0;
1073 out:
1074         mutex_unlock(&rpciod_mutex);
1075         return error;
1076 }
1077
1078 void
1079 rpciod_down(void)
1080 {
1081         mutex_lock(&rpciod_mutex);
1082         dprintk("rpciod_down sema %d\n", rpciod_users);
1083         if (rpciod_users) {
1084                 if (--rpciod_users)
1085                         goto out;
1086         } else
1087                 printk(KERN_WARNING "rpciod_down: no users??\n");
1088
1089         if (!rpciod_workqueue) {
1090                 dprintk("rpciod_down: Nothing to do!\n");
1091                 goto out;
1092         }
1093         rpciod_killall();
1094
1095         destroy_workqueue(rpciod_workqueue);
1096         rpciod_workqueue = NULL;
1097  out:
1098         mutex_unlock(&rpciod_mutex);
1099 }
1100
1101 #ifdef RPC_DEBUG
1102 void rpc_show_tasks(void)
1103 {
1104         struct list_head *le;
1105         struct rpc_task *t;
1106
1107         spin_lock(&rpc_sched_lock);
1108         if (list_empty(&all_tasks)) {
1109                 spin_unlock(&rpc_sched_lock);
1110                 return;
1111         }
1112         printk("-pid- proc flgs status -client- -prog- --rqstp- -timeout "
1113                 "-rpcwait -action- ---ops--\n");
1114         alltask_for_each(t, le, &all_tasks) {
1115                 const char *rpc_waitq = "none";
1116
1117                 if (RPC_IS_QUEUED(t))
1118                         rpc_waitq = rpc_qname(t->u.tk_wait.rpc_waitq);
1119
1120                 printk("%05d %04d %04x %06d %8p %6d %8p %08ld %8s %8p %8p\n",
1121                         t->tk_pid,
1122                         (t->tk_msg.rpc_proc ? t->tk_msg.rpc_proc->p_proc : -1),
1123                         t->tk_flags, t->tk_status,
1124                         t->tk_client,
1125                         (t->tk_client ? t->tk_client->cl_prog : 0),
1126                         t->tk_rqstp, t->tk_timeout,
1127                         rpc_waitq,
1128                         t->tk_action, t->tk_ops);
1129         }
1130         spin_unlock(&rpc_sched_lock);
1131 }
1132 #endif
1133
1134 void
1135 rpc_destroy_mempool(void)
1136 {
1137         if (rpc_buffer_mempool)
1138                 mempool_destroy(rpc_buffer_mempool);
1139         if (rpc_task_mempool)
1140                 mempool_destroy(rpc_task_mempool);
1141         if (rpc_task_slabp && kmem_cache_destroy(rpc_task_slabp))
1142                 printk(KERN_INFO "rpc_task: not all structures were freed\n");
1143         if (rpc_buffer_slabp && kmem_cache_destroy(rpc_buffer_slabp))
1144                 printk(KERN_INFO "rpc_buffers: not all structures were freed\n");
1145 }
1146
1147 int
1148 rpc_init_mempool(void)
1149 {
1150         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1151                                              sizeof(struct rpc_task),
1152                                              0, SLAB_HWCACHE_ALIGN,
1153                                              NULL, NULL);
1154         if (!rpc_task_slabp)
1155                 goto err_nomem;
1156         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1157                                              RPC_BUFFER_MAXSIZE,
1158                                              0, SLAB_HWCACHE_ALIGN,
1159                                              NULL, NULL);
1160         if (!rpc_buffer_slabp)
1161                 goto err_nomem;
1162         rpc_task_mempool = mempool_create(RPC_TASK_POOLSIZE,
1163                                             mempool_alloc_slab,
1164                                             mempool_free_slab,
1165                                             rpc_task_slabp);
1166         if (!rpc_task_mempool)
1167                 goto err_nomem;
1168         rpc_buffer_mempool = mempool_create(RPC_BUFFER_POOLSIZE,
1169                                             mempool_alloc_slab,
1170                                             mempool_free_slab,
1171                                             rpc_buffer_slabp);
1172         if (!rpc_buffer_mempool)
1173                 goto err_nomem;
1174         return 0;
1175 err_nomem:
1176         rpc_destroy_mempool();
1177         return -ENOMEM;
1178 }