[IRDA] sem2mutex: drivers/net/irda
[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
22 #include <linux/sunrpc/clnt.h>
23 #include <linux/sunrpc/xprt.h>
24
25 #ifdef RPC_DEBUG
26 #define RPCDBG_FACILITY         RPCDBG_SCHED
27 #define RPC_TASK_MAGIC_ID       0xf00baa
28 static int                      rpc_task_id;
29 #endif
30
31 /*
32  * RPC slabs and memory pools
33  */
34 #define RPC_BUFFER_MAXSIZE      (2048)
35 #define RPC_BUFFER_POOLSIZE     (8)
36 #define RPC_TASK_POOLSIZE       (8)
37 static kmem_cache_t     *rpc_task_slabp __read_mostly;
38 static kmem_cache_t     *rpc_buffer_slabp __read_mostly;
39 static mempool_t        *rpc_task_mempool __read_mostly;
40 static mempool_t        *rpc_buffer_mempool __read_mostly;
41
42 static void                     __rpc_default_timer(struct rpc_task *task);
43 static void                     rpciod_killall(void);
44 static void                     rpc_async_schedule(void *);
45
46 /*
47  * RPC tasks that create another task (e.g. for contacting the portmapper)
48  * will wait on this queue for their child's completion
49  */
50 static RPC_WAITQ(childq, "childq");
51
52 /*
53  * RPC tasks sit here while waiting for conditions to improve.
54  */
55 static RPC_WAITQ(delay_queue, "delayq");
56
57 /*
58  * All RPC tasks are linked into this list
59  */
60 static LIST_HEAD(all_tasks);
61
62 /*
63  * rpciod-related stuff
64  */
65 static DECLARE_MUTEX(rpciod_sema);
66 static unsigned int             rpciod_users;
67 static struct workqueue_struct *rpciod_workqueue;
68
69 /*
70  * Spinlock for other critical sections of code.
71  */
72 static DEFINE_SPINLOCK(rpc_sched_lock);
73
74 /*
75  * Disable the timer for a given RPC task. Should be called with
76  * queue->lock and bh_disabled in order to avoid races within
77  * rpc_run_timer().
78  */
79 static inline void
80 __rpc_disable_timer(struct rpc_task *task)
81 {
82         dprintk("RPC: %4d disabling timer\n", task->tk_pid);
83         task->tk_timeout_fn = NULL;
84         task->tk_timeout = 0;
85 }
86
87 /*
88  * Run a timeout function.
89  * We use the callback in order to allow __rpc_wake_up_task()
90  * and friends to disable the timer synchronously on SMP systems
91  * without calling del_timer_sync(). The latter could cause a
92  * deadlock if called while we're holding spinlocks...
93  */
94 static void rpc_run_timer(struct rpc_task *task)
95 {
96         void (*callback)(struct rpc_task *);
97
98         callback = task->tk_timeout_fn;
99         task->tk_timeout_fn = NULL;
100         if (callback && RPC_IS_QUEUED(task)) {
101                 dprintk("RPC: %4d running timer\n", task->tk_pid);
102                 callback(task);
103         }
104         smp_mb__before_clear_bit();
105         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
106         smp_mb__after_clear_bit();
107 }
108
109 /*
110  * Set up a timer for the current task.
111  */
112 static inline void
113 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
114 {
115         if (!task->tk_timeout)
116                 return;
117
118         dprintk("RPC: %4d setting alarm for %lu ms\n",
119                         task->tk_pid, task->tk_timeout * 1000 / HZ);
120
121         if (timer)
122                 task->tk_timeout_fn = timer;
123         else
124                 task->tk_timeout_fn = __rpc_default_timer;
125         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
126         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
127 }
128
129 /*
130  * Delete any timer for the current task. Because we use del_timer_sync(),
131  * this function should never be called while holding queue->lock.
132  */
133 static void
134 rpc_delete_timer(struct rpc_task *task)
135 {
136         if (RPC_IS_QUEUED(task))
137                 return;
138         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
139                 del_singleshot_timer_sync(&task->tk_timer);
140                 dprintk("RPC: %4d deleting timer\n", task->tk_pid);
141         }
142 }
143
144 /*
145  * Add new request to a priority queue.
146  */
147 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
148 {
149         struct list_head *q;
150         struct rpc_task *t;
151
152         INIT_LIST_HEAD(&task->u.tk_wait.links);
153         q = &queue->tasks[task->tk_priority];
154         if (unlikely(task->tk_priority > queue->maxpriority))
155                 q = &queue->tasks[queue->maxpriority];
156         list_for_each_entry(t, q, u.tk_wait.list) {
157                 if (t->tk_cookie == task->tk_cookie) {
158                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
159                         return;
160                 }
161         }
162         list_add_tail(&task->u.tk_wait.list, q);
163 }
164
165 /*
166  * Add new request to wait queue.
167  *
168  * Swapper tasks always get inserted at the head of the queue.
169  * This should avoid many nasty memory deadlocks and hopefully
170  * improve overall performance.
171  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
172  */
173 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
174 {
175         BUG_ON (RPC_IS_QUEUED(task));
176
177         if (RPC_IS_PRIORITY(queue))
178                 __rpc_add_wait_queue_priority(queue, task);
179         else if (RPC_IS_SWAPPER(task))
180                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
181         else
182                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
183         task->u.tk_wait.rpc_waitq = queue;
184         rpc_set_queued(task);
185
186         dprintk("RPC: %4d added to queue %p \"%s\"\n",
187                                 task->tk_pid, queue, rpc_qname(queue));
188 }
189
190 /*
191  * Remove request from a priority queue.
192  */
193 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
194 {
195         struct rpc_task *t;
196
197         if (!list_empty(&task->u.tk_wait.links)) {
198                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
199                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
200                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
201         }
202         list_del(&task->u.tk_wait.list);
203 }
204
205 /*
206  * Remove request from queue.
207  * Note: must be called with spin lock held.
208  */
209 static void __rpc_remove_wait_queue(struct rpc_task *task)
210 {
211         struct rpc_wait_queue *queue;
212         queue = task->u.tk_wait.rpc_waitq;
213
214         if (RPC_IS_PRIORITY(queue))
215                 __rpc_remove_wait_queue_priority(task);
216         else
217                 list_del(&task->u.tk_wait.list);
218         dprintk("RPC: %4d removed from queue %p \"%s\"\n",
219                                 task->tk_pid, queue, rpc_qname(queue));
220 }
221
222 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
223 {
224         queue->priority = priority;
225         queue->count = 1 << (priority * 2);
226 }
227
228 static inline void rpc_set_waitqueue_cookie(struct rpc_wait_queue *queue, unsigned long cookie)
229 {
230         queue->cookie = cookie;
231         queue->nr = RPC_BATCH_COUNT;
232 }
233
234 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
235 {
236         rpc_set_waitqueue_priority(queue, queue->maxpriority);
237         rpc_set_waitqueue_cookie(queue, 0);
238 }
239
240 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, int maxprio)
241 {
242         int i;
243
244         spin_lock_init(&queue->lock);
245         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
246                 INIT_LIST_HEAD(&queue->tasks[i]);
247         queue->maxpriority = maxprio;
248         rpc_reset_waitqueue_priority(queue);
249 #ifdef RPC_DEBUG
250         queue->name = qname;
251 #endif
252 }
253
254 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
255 {
256         __rpc_init_priority_wait_queue(queue, qname, RPC_PRIORITY_HIGH);
257 }
258
259 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
260 {
261         __rpc_init_priority_wait_queue(queue, qname, 0);
262 }
263 EXPORT_SYMBOL(rpc_init_wait_queue);
264
265 static int rpc_wait_bit_interruptible(void *word)
266 {
267         if (signal_pending(current))
268                 return -ERESTARTSYS;
269         schedule();
270         return 0;
271 }
272
273 /*
274  * Mark an RPC call as having completed by clearing the 'active' bit
275  */
276 static inline void rpc_mark_complete_task(struct rpc_task *task)
277 {
278         rpc_clear_active(task);
279         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
280 }
281
282 /*
283  * Allow callers to wait for completion of an RPC call
284  */
285 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
286 {
287         if (action == NULL)
288                 action = rpc_wait_bit_interruptible;
289         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
290                         action, TASK_INTERRUPTIBLE);
291 }
292 EXPORT_SYMBOL(__rpc_wait_for_completion_task);
293
294 /*
295  * Make an RPC task runnable.
296  *
297  * Note: If the task is ASYNC, this must be called with 
298  * the spinlock held to protect the wait queue operation.
299  */
300 static void rpc_make_runnable(struct rpc_task *task)
301 {
302         int do_ret;
303
304         BUG_ON(task->tk_timeout_fn);
305         do_ret = rpc_test_and_set_running(task);
306         rpc_clear_queued(task);
307         if (do_ret)
308                 return;
309         if (RPC_IS_ASYNC(task)) {
310                 int status;
311
312                 INIT_WORK(&task->u.tk_work, rpc_async_schedule, (void *)task);
313                 status = queue_work(task->tk_workqueue, &task->u.tk_work);
314                 if (status < 0) {
315                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
316                         task->tk_status = status;
317                         return;
318                 }
319         } else
320                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
321 }
322
323 /*
324  * Place a newly initialized task on the workqueue.
325  */
326 static inline void
327 rpc_schedule_run(struct rpc_task *task)
328 {
329         rpc_set_active(task);
330         rpc_make_runnable(task);
331 }
332
333 /*
334  * Prepare for sleeping on a wait queue.
335  * By always appending tasks to the list we ensure FIFO behavior.
336  * NB: An RPC task will only receive interrupt-driven events as long
337  * as it's on a wait queue.
338  */
339 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
340                         rpc_action action, rpc_action timer)
341 {
342         dprintk("RPC: %4d sleep_on(queue \"%s\" time %ld)\n", task->tk_pid,
343                                 rpc_qname(q), jiffies);
344
345         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
346                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
347                 return;
348         }
349
350         /* Mark the task as being activated if so needed */
351         rpc_set_active(task);
352
353         __rpc_add_wait_queue(q, task);
354
355         BUG_ON(task->tk_callback != NULL);
356         task->tk_callback = action;
357         __rpc_add_timer(task, timer);
358 }
359
360 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
361                                 rpc_action action, rpc_action timer)
362 {
363         /*
364          * Protect the queue operations.
365          */
366         spin_lock_bh(&q->lock);
367         __rpc_sleep_on(q, task, action, timer);
368         spin_unlock_bh(&q->lock);
369 }
370
371 /**
372  * __rpc_do_wake_up_task - wake up a single rpc_task
373  * @task: task to be woken up
374  *
375  * Caller must hold queue->lock, and have cleared the task queued flag.
376  */
377 static void __rpc_do_wake_up_task(struct rpc_task *task)
378 {
379         dprintk("RPC: %4d __rpc_wake_up_task (now %ld)\n", task->tk_pid, jiffies);
380
381 #ifdef RPC_DEBUG
382         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
383 #endif
384         /* Has the task been executed yet? If not, we cannot wake it up! */
385         if (!RPC_IS_ACTIVATED(task)) {
386                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
387                 return;
388         }
389
390         __rpc_disable_timer(task);
391         __rpc_remove_wait_queue(task);
392
393         rpc_make_runnable(task);
394
395         dprintk("RPC:      __rpc_wake_up_task done\n");
396 }
397
398 /*
399  * Wake up the specified task
400  */
401 static void __rpc_wake_up_task(struct rpc_task *task)
402 {
403         if (rpc_start_wakeup(task)) {
404                 if (RPC_IS_QUEUED(task))
405                         __rpc_do_wake_up_task(task);
406                 rpc_finish_wakeup(task);
407         }
408 }
409
410 /*
411  * Default timeout handler if none specified by user
412  */
413 static void
414 __rpc_default_timer(struct rpc_task *task)
415 {
416         dprintk("RPC: %d timeout (default timer)\n", task->tk_pid);
417         task->tk_status = -ETIMEDOUT;
418         rpc_wake_up_task(task);
419 }
420
421 /*
422  * Wake up the specified task
423  */
424 void rpc_wake_up_task(struct rpc_task *task)
425 {
426         if (rpc_start_wakeup(task)) {
427                 if (RPC_IS_QUEUED(task)) {
428                         struct rpc_wait_queue *queue = task->u.tk_wait.rpc_waitq;
429
430                         spin_lock_bh(&queue->lock);
431                         __rpc_do_wake_up_task(task);
432                         spin_unlock_bh(&queue->lock);
433                 }
434                 rpc_finish_wakeup(task);
435         }
436 }
437
438 /*
439  * Wake up the next task on a priority queue.
440  */
441 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
442 {
443         struct list_head *q;
444         struct rpc_task *task;
445
446         /*
447          * Service a batch of tasks from a single cookie.
448          */
449         q = &queue->tasks[queue->priority];
450         if (!list_empty(q)) {
451                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
452                 if (queue->cookie == task->tk_cookie) {
453                         if (--queue->nr)
454                                 goto out;
455                         list_move_tail(&task->u.tk_wait.list, q);
456                 }
457                 /*
458                  * Check if we need to switch queues.
459                  */
460                 if (--queue->count)
461                         goto new_cookie;
462         }
463
464         /*
465          * Service the next queue.
466          */
467         do {
468                 if (q == &queue->tasks[0])
469                         q = &queue->tasks[queue->maxpriority];
470                 else
471                         q = q - 1;
472                 if (!list_empty(q)) {
473                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
474                         goto new_queue;
475                 }
476         } while (q != &queue->tasks[queue->priority]);
477
478         rpc_reset_waitqueue_priority(queue);
479         return NULL;
480
481 new_queue:
482         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
483 new_cookie:
484         rpc_set_waitqueue_cookie(queue, task->tk_cookie);
485 out:
486         __rpc_wake_up_task(task);
487         return task;
488 }
489
490 /*
491  * Wake up the next task on the wait queue.
492  */
493 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
494 {
495         struct rpc_task *task = NULL;
496
497         dprintk("RPC:      wake_up_next(%p \"%s\")\n", queue, rpc_qname(queue));
498         spin_lock_bh(&queue->lock);
499         if (RPC_IS_PRIORITY(queue))
500                 task = __rpc_wake_up_next_priority(queue);
501         else {
502                 task_for_first(task, &queue->tasks[0])
503                         __rpc_wake_up_task(task);
504         }
505         spin_unlock_bh(&queue->lock);
506
507         return task;
508 }
509
510 /**
511  * rpc_wake_up - wake up all rpc_tasks
512  * @queue: rpc_wait_queue on which the tasks are sleeping
513  *
514  * Grabs queue->lock
515  */
516 void rpc_wake_up(struct rpc_wait_queue *queue)
517 {
518         struct rpc_task *task, *next;
519         struct list_head *head;
520
521         spin_lock_bh(&queue->lock);
522         head = &queue->tasks[queue->maxpriority];
523         for (;;) {
524                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
525                         __rpc_wake_up_task(task);
526                 if (head == &queue->tasks[0])
527                         break;
528                 head--;
529         }
530         spin_unlock_bh(&queue->lock);
531 }
532
533 /**
534  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
535  * @queue: rpc_wait_queue on which the tasks are sleeping
536  * @status: status value to set
537  *
538  * Grabs queue->lock
539  */
540 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
541 {
542         struct rpc_task *task, *next;
543         struct list_head *head;
544
545         spin_lock_bh(&queue->lock);
546         head = &queue->tasks[queue->maxpriority];
547         for (;;) {
548                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
549                         task->tk_status = status;
550                         __rpc_wake_up_task(task);
551                 }
552                 if (head == &queue->tasks[0])
553                         break;
554                 head--;
555         }
556         spin_unlock_bh(&queue->lock);
557 }
558
559 /*
560  * Run a task at a later time
561  */
562 static void     __rpc_atrun(struct rpc_task *);
563 void
564 rpc_delay(struct rpc_task *task, unsigned long delay)
565 {
566         task->tk_timeout = delay;
567         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
568 }
569
570 static void
571 __rpc_atrun(struct rpc_task *task)
572 {
573         task->tk_status = 0;
574         rpc_wake_up_task(task);
575 }
576
577 /*
578  * Helper to call task->tk_ops->rpc_call_prepare
579  */
580 static void rpc_prepare_task(struct rpc_task *task)
581 {
582         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
583 }
584
585 /*
586  * Helper that calls task->tk_ops->rpc_call_done if it exists
587  */
588 void rpc_exit_task(struct rpc_task *task)
589 {
590         task->tk_action = NULL;
591         if (task->tk_ops->rpc_call_done != NULL) {
592                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
593                 if (task->tk_action != NULL) {
594                         WARN_ON(RPC_ASSASSINATED(task));
595                         /* Always release the RPC slot and buffer memory */
596                         xprt_release(task);
597                 }
598         }
599 }
600 EXPORT_SYMBOL(rpc_exit_task);
601
602 /*
603  * This is the RPC `scheduler' (or rather, the finite state machine).
604  */
605 static int __rpc_execute(struct rpc_task *task)
606 {
607         int             status = 0;
608
609         dprintk("RPC: %4d rpc_execute flgs %x\n",
610                                 task->tk_pid, task->tk_flags);
611
612         BUG_ON(RPC_IS_QUEUED(task));
613
614         for (;;) {
615                 /*
616                  * Garbage collection of pending timers...
617                  */
618                 rpc_delete_timer(task);
619
620                 /*
621                  * Execute any pending callback.
622                  */
623                 if (RPC_DO_CALLBACK(task)) {
624                         /* Define a callback save pointer */
625                         void (*save_callback)(struct rpc_task *);
626         
627                         /* 
628                          * If a callback exists, save it, reset it,
629                          * call it.
630                          * The save is needed to stop from resetting
631                          * another callback set within the callback handler
632                          * - Dave
633                          */
634                         save_callback=task->tk_callback;
635                         task->tk_callback=NULL;
636                         lock_kernel();
637                         save_callback(task);
638                         unlock_kernel();
639                 }
640
641                 /*
642                  * Perform the next FSM step.
643                  * tk_action may be NULL when the task has been killed
644                  * by someone else.
645                  */
646                 if (!RPC_IS_QUEUED(task)) {
647                         if (task->tk_action == NULL)
648                                 break;
649                         lock_kernel();
650                         task->tk_action(task);
651                         unlock_kernel();
652                 }
653
654                 /*
655                  * Lockless check for whether task is sleeping or not.
656                  */
657                 if (!RPC_IS_QUEUED(task))
658                         continue;
659                 rpc_clear_running(task);
660                 if (RPC_IS_ASYNC(task)) {
661                         /* Careful! we may have raced... */
662                         if (RPC_IS_QUEUED(task))
663                                 return 0;
664                         if (rpc_test_and_set_running(task))
665                                 return 0;
666                         continue;
667                 }
668
669                 /* sync task: sleep here */
670                 dprintk("RPC: %4d sync task going to sleep\n", task->tk_pid);
671                 /* Note: Caller should be using rpc_clnt_sigmask() */
672                 status = out_of_line_wait_on_bit(&task->tk_runstate,
673                                 RPC_TASK_QUEUED, rpc_wait_bit_interruptible,
674                                 TASK_INTERRUPTIBLE);
675                 if (status == -ERESTARTSYS) {
676                         /*
677                          * When a sync task receives a signal, it exits with
678                          * -ERESTARTSYS. In order to catch any callbacks that
679                          * clean up after sleeping on some queue, we don't
680                          * break the loop here, but go around once more.
681                          */
682                         dprintk("RPC: %4d got signal\n", task->tk_pid);
683                         task->tk_flags |= RPC_TASK_KILLED;
684                         rpc_exit(task, -ERESTARTSYS);
685                         rpc_wake_up_task(task);
686                 }
687                 rpc_set_running(task);
688                 dprintk("RPC: %4d sync task resuming\n", task->tk_pid);
689         }
690
691         dprintk("RPC: %4d, return %d, status %d\n", task->tk_pid, status, task->tk_status);
692         /* Wake up anyone who is waiting for task completion */
693         rpc_mark_complete_task(task);
694         /* Release all resources associated with the task */
695         rpc_release_task(task);
696         return status;
697 }
698
699 /*
700  * User-visible entry point to the scheduler.
701  *
702  * This may be called recursively if e.g. an async NFS task updates
703  * the attributes and finds that dirty pages must be flushed.
704  * NOTE: Upon exit of this function the task is guaranteed to be
705  *       released. In particular note that tk_release() will have
706  *       been called, so your task memory may have been freed.
707  */
708 int
709 rpc_execute(struct rpc_task *task)
710 {
711         rpc_set_active(task);
712         rpc_set_running(task);
713         return __rpc_execute(task);
714 }
715
716 static void rpc_async_schedule(void *arg)
717 {
718         __rpc_execute((struct rpc_task *)arg);
719 }
720
721 /**
722  * rpc_malloc - allocate an RPC buffer
723  * @task: RPC task that will use this buffer
724  * @size: requested byte size
725  *
726  * We try to ensure that some NFS reads and writes can always proceed
727  * by using a mempool when allocating 'small' buffers.
728  * In order to avoid memory starvation triggering more writebacks of
729  * NFS requests, we use GFP_NOFS rather than GFP_KERNEL.
730  */
731 void * rpc_malloc(struct rpc_task *task, size_t size)
732 {
733         struct rpc_rqst *req = task->tk_rqstp;
734         gfp_t   gfp;
735
736         if (task->tk_flags & RPC_TASK_SWAPPER)
737                 gfp = GFP_ATOMIC;
738         else
739                 gfp = GFP_NOFS;
740
741         if (size > RPC_BUFFER_MAXSIZE) {
742                 req->rq_buffer = kmalloc(size, gfp);
743                 if (req->rq_buffer)
744                         req->rq_bufsize = size;
745         } else {
746                 req->rq_buffer = mempool_alloc(rpc_buffer_mempool, gfp);
747                 if (req->rq_buffer)
748                         req->rq_bufsize = RPC_BUFFER_MAXSIZE;
749         }
750         return req->rq_buffer;
751 }
752
753 /**
754  * rpc_free - free buffer allocated via rpc_malloc
755  * @task: RPC task with a buffer to be freed
756  *
757  */
758 void rpc_free(struct rpc_task *task)
759 {
760         struct rpc_rqst *req = task->tk_rqstp;
761
762         if (req->rq_buffer) {
763                 if (req->rq_bufsize == RPC_BUFFER_MAXSIZE)
764                         mempool_free(req->rq_buffer, rpc_buffer_mempool);
765                 else
766                         kfree(req->rq_buffer);
767                 req->rq_buffer = NULL;
768                 req->rq_bufsize = 0;
769         }
770 }
771
772 /*
773  * Creation and deletion of RPC task structures
774  */
775 void rpc_init_task(struct rpc_task *task, struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
776 {
777         memset(task, 0, sizeof(*task));
778         init_timer(&task->tk_timer);
779         task->tk_timer.data     = (unsigned long) task;
780         task->tk_timer.function = (void (*)(unsigned long)) rpc_run_timer;
781         atomic_set(&task->tk_count, 1);
782         task->tk_client = clnt;
783         task->tk_flags  = flags;
784         task->tk_ops = tk_ops;
785         if (tk_ops->rpc_call_prepare != NULL)
786                 task->tk_action = rpc_prepare_task;
787         task->tk_calldata = calldata;
788
789         /* Initialize retry counters */
790         task->tk_garb_retry = 2;
791         task->tk_cred_retry = 2;
792
793         task->tk_priority = RPC_PRIORITY_NORMAL;
794         task->tk_cookie = (unsigned long)current;
795
796         /* Initialize workqueue for async tasks */
797         task->tk_workqueue = rpciod_workqueue;
798
799         if (clnt) {
800                 atomic_inc(&clnt->cl_users);
801                 if (clnt->cl_softrtry)
802                         task->tk_flags |= RPC_TASK_SOFT;
803                 if (!clnt->cl_intr)
804                         task->tk_flags |= RPC_TASK_NOINTR;
805         }
806
807 #ifdef RPC_DEBUG
808         task->tk_magic = RPC_TASK_MAGIC_ID;
809         task->tk_pid = rpc_task_id++;
810 #endif
811         /* Add to global list of all tasks */
812         spin_lock(&rpc_sched_lock);
813         list_add_tail(&task->tk_task, &all_tasks);
814         spin_unlock(&rpc_sched_lock);
815
816         BUG_ON(task->tk_ops == NULL);
817
818         dprintk("RPC: %4d new task procpid %d\n", task->tk_pid,
819                                 current->pid);
820 }
821
822 static struct rpc_task *
823 rpc_alloc_task(void)
824 {
825         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
826 }
827
828 static void rpc_free_task(struct rpc_task *task)
829 {
830         dprintk("RPC: %4d freeing task\n", task->tk_pid);
831         mempool_free(task, rpc_task_mempool);
832 }
833
834 /*
835  * Create a new task for the specified client.  We have to
836  * clean up after an allocation failure, as the client may
837  * have specified "oneshot".
838  */
839 struct rpc_task *rpc_new_task(struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
840 {
841         struct rpc_task *task;
842
843         task = rpc_alloc_task();
844         if (!task)
845                 goto cleanup;
846
847         rpc_init_task(task, clnt, flags, tk_ops, calldata);
848
849         dprintk("RPC: %4d allocated task\n", task->tk_pid);
850         task->tk_flags |= RPC_TASK_DYNAMIC;
851 out:
852         return task;
853
854 cleanup:
855         /* Check whether to release the client */
856         if (clnt) {
857                 printk("rpc_new_task: failed, users=%d, oneshot=%d\n",
858                         atomic_read(&clnt->cl_users), clnt->cl_oneshot);
859                 atomic_inc(&clnt->cl_users); /* pretend we were used ... */
860                 rpc_release_client(clnt);
861         }
862         goto out;
863 }
864
865 void rpc_release_task(struct rpc_task *task)
866 {
867         const struct rpc_call_ops *tk_ops = task->tk_ops;
868         void *calldata = task->tk_calldata;
869
870 #ifdef RPC_DEBUG
871         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
872 #endif
873         if (!atomic_dec_and_test(&task->tk_count))
874                 return;
875         dprintk("RPC: %4d release task\n", task->tk_pid);
876
877         /* Remove from global task list */
878         spin_lock(&rpc_sched_lock);
879         list_del(&task->tk_task);
880         spin_unlock(&rpc_sched_lock);
881
882         BUG_ON (RPC_IS_QUEUED(task));
883
884         /* Synchronously delete any running timer */
885         rpc_delete_timer(task);
886
887         /* Release resources */
888         if (task->tk_rqstp)
889                 xprt_release(task);
890         if (task->tk_msg.rpc_cred)
891                 rpcauth_unbindcred(task);
892         if (task->tk_client) {
893                 rpc_release_client(task->tk_client);
894                 task->tk_client = NULL;
895         }
896
897 #ifdef RPC_DEBUG
898         task->tk_magic = 0;
899 #endif
900         if (task->tk_flags & RPC_TASK_DYNAMIC)
901                 rpc_free_task(task);
902         if (tk_ops->rpc_release)
903                 tk_ops->rpc_release(calldata);
904 }
905
906 /**
907  * rpc_run_task - Allocate a new RPC task, then run rpc_execute against it
908  * @clnt: pointer to RPC client
909  * @flags: RPC flags
910  * @ops: RPC call ops
911  * @data: user call data
912  */
913 struct rpc_task *rpc_run_task(struct rpc_clnt *clnt, int flags,
914                                         const struct rpc_call_ops *ops,
915                                         void *data)
916 {
917         struct rpc_task *task;
918         task = rpc_new_task(clnt, flags, ops, data);
919         if (task == NULL)
920                 return ERR_PTR(-ENOMEM);
921         atomic_inc(&task->tk_count);
922         rpc_execute(task);
923         return task;
924 }
925 EXPORT_SYMBOL(rpc_run_task);
926
927 /**
928  * rpc_find_parent - find the parent of a child task.
929  * @child: child task
930  * @parent: parent task
931  *
932  * Checks that the parent task is still sleeping on the
933  * queue 'childq'. If so returns a pointer to the parent.
934  * Upon failure returns NULL.
935  *
936  * Caller must hold childq.lock
937  */
938 static inline struct rpc_task *rpc_find_parent(struct rpc_task *child, struct rpc_task *parent)
939 {
940         struct rpc_task *task;
941         struct list_head *le;
942
943         task_for_each(task, le, &childq.tasks[0])
944                 if (task == parent)
945                         return parent;
946
947         return NULL;
948 }
949
950 static void rpc_child_exit(struct rpc_task *child, void *calldata)
951 {
952         struct rpc_task *parent;
953
954         spin_lock_bh(&childq.lock);
955         if ((parent = rpc_find_parent(child, calldata)) != NULL) {
956                 parent->tk_status = child->tk_status;
957                 __rpc_wake_up_task(parent);
958         }
959         spin_unlock_bh(&childq.lock);
960 }
961
962 static const struct rpc_call_ops rpc_child_ops = {
963         .rpc_call_done = rpc_child_exit,
964 };
965
966 /*
967  * Note: rpc_new_task releases the client after a failure.
968  */
969 struct rpc_task *
970 rpc_new_child(struct rpc_clnt *clnt, struct rpc_task *parent)
971 {
972         struct rpc_task *task;
973
974         task = rpc_new_task(clnt, RPC_TASK_ASYNC | RPC_TASK_CHILD, &rpc_child_ops, parent);
975         if (!task)
976                 goto fail;
977         return task;
978
979 fail:
980         parent->tk_status = -ENOMEM;
981         return NULL;
982 }
983
984 void rpc_run_child(struct rpc_task *task, struct rpc_task *child, rpc_action func)
985 {
986         spin_lock_bh(&childq.lock);
987         /* N.B. Is it possible for the child to have already finished? */
988         __rpc_sleep_on(&childq, task, func, NULL);
989         rpc_schedule_run(child);
990         spin_unlock_bh(&childq.lock);
991 }
992
993 /*
994  * Kill all tasks for the given client.
995  * XXX: kill their descendants as well?
996  */
997 void rpc_killall_tasks(struct rpc_clnt *clnt)
998 {
999         struct rpc_task *rovr;
1000         struct list_head *le;
1001
1002         dprintk("RPC:      killing all tasks for client %p\n", clnt);
1003
1004         /*
1005          * Spin lock all_tasks to prevent changes...
1006          */
1007         spin_lock(&rpc_sched_lock);
1008         alltask_for_each(rovr, le, &all_tasks) {
1009                 if (! RPC_IS_ACTIVATED(rovr))
1010                         continue;
1011                 if (!clnt || rovr->tk_client == clnt) {
1012                         rovr->tk_flags |= RPC_TASK_KILLED;
1013                         rpc_exit(rovr, -EIO);
1014                         rpc_wake_up_task(rovr);
1015                 }
1016         }
1017         spin_unlock(&rpc_sched_lock);
1018 }
1019
1020 static DECLARE_MUTEX_LOCKED(rpciod_running);
1021
1022 static void rpciod_killall(void)
1023 {
1024         unsigned long flags;
1025
1026         while (!list_empty(&all_tasks)) {
1027                 clear_thread_flag(TIF_SIGPENDING);
1028                 rpc_killall_tasks(NULL);
1029                 flush_workqueue(rpciod_workqueue);
1030                 if (!list_empty(&all_tasks)) {
1031                         dprintk("rpciod_killall: waiting for tasks to exit\n");
1032                         yield();
1033                 }
1034         }
1035
1036         spin_lock_irqsave(&current->sighand->siglock, flags);
1037         recalc_sigpending();
1038         spin_unlock_irqrestore(&current->sighand->siglock, flags);
1039 }
1040
1041 /*
1042  * Start up the rpciod process if it's not already running.
1043  */
1044 int
1045 rpciod_up(void)
1046 {
1047         struct workqueue_struct *wq;
1048         int error = 0;
1049
1050         down(&rpciod_sema);
1051         dprintk("rpciod_up: users %d\n", rpciod_users);
1052         rpciod_users++;
1053         if (rpciod_workqueue)
1054                 goto out;
1055         /*
1056          * If there's no pid, we should be the first user.
1057          */
1058         if (rpciod_users > 1)
1059                 printk(KERN_WARNING "rpciod_up: no workqueue, %d users??\n", rpciod_users);
1060         /*
1061          * Create the rpciod thread and wait for it to start.
1062          */
1063         error = -ENOMEM;
1064         wq = create_workqueue("rpciod");
1065         if (wq == NULL) {
1066                 printk(KERN_WARNING "rpciod_up: create workqueue failed, error=%d\n", error);
1067                 rpciod_users--;
1068                 goto out;
1069         }
1070         rpciod_workqueue = wq;
1071         error = 0;
1072 out:
1073         up(&rpciod_sema);
1074         return error;
1075 }
1076
1077 void
1078 rpciod_down(void)
1079 {
1080         down(&rpciod_sema);
1081         dprintk("rpciod_down sema %d\n", rpciod_users);
1082         if (rpciod_users) {
1083                 if (--rpciod_users)
1084                         goto out;
1085         } else
1086                 printk(KERN_WARNING "rpciod_down: no users??\n");
1087
1088         if (!rpciod_workqueue) {
1089                 dprintk("rpciod_down: Nothing to do!\n");
1090                 goto out;
1091         }
1092         rpciod_killall();
1093
1094         destroy_workqueue(rpciod_workqueue);
1095         rpciod_workqueue = NULL;
1096  out:
1097         up(&rpciod_sema);
1098 }
1099
1100 #ifdef RPC_DEBUG
1101 void rpc_show_tasks(void)
1102 {
1103         struct list_head *le;
1104         struct rpc_task *t;
1105
1106         spin_lock(&rpc_sched_lock);
1107         if (list_empty(&all_tasks)) {
1108                 spin_unlock(&rpc_sched_lock);
1109                 return;
1110         }
1111         printk("-pid- proc flgs status -client- -prog- --rqstp- -timeout "
1112                 "-rpcwait -action- ---ops--\n");
1113         alltask_for_each(t, le, &all_tasks) {
1114                 const char *rpc_waitq = "none";
1115
1116                 if (RPC_IS_QUEUED(t))
1117                         rpc_waitq = rpc_qname(t->u.tk_wait.rpc_waitq);
1118
1119                 printk("%05d %04d %04x %06d %8p %6d %8p %08ld %8s %8p %8p\n",
1120                         t->tk_pid,
1121                         (t->tk_msg.rpc_proc ? t->tk_msg.rpc_proc->p_proc : -1),
1122                         t->tk_flags, t->tk_status,
1123                         t->tk_client,
1124                         (t->tk_client ? t->tk_client->cl_prog : 0),
1125                         t->tk_rqstp, t->tk_timeout,
1126                         rpc_waitq,
1127                         t->tk_action, t->tk_ops);
1128         }
1129         spin_unlock(&rpc_sched_lock);
1130 }
1131 #endif
1132
1133 void
1134 rpc_destroy_mempool(void)
1135 {
1136         if (rpc_buffer_mempool)
1137                 mempool_destroy(rpc_buffer_mempool);
1138         if (rpc_task_mempool)
1139                 mempool_destroy(rpc_task_mempool);
1140         if (rpc_task_slabp && kmem_cache_destroy(rpc_task_slabp))
1141                 printk(KERN_INFO "rpc_task: not all structures were freed\n");
1142         if (rpc_buffer_slabp && kmem_cache_destroy(rpc_buffer_slabp))
1143                 printk(KERN_INFO "rpc_buffers: not all structures were freed\n");
1144 }
1145
1146 int
1147 rpc_init_mempool(void)
1148 {
1149         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1150                                              sizeof(struct rpc_task),
1151                                              0, SLAB_HWCACHE_ALIGN,
1152                                              NULL, NULL);
1153         if (!rpc_task_slabp)
1154                 goto err_nomem;
1155         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1156                                              RPC_BUFFER_MAXSIZE,
1157                                              0, SLAB_HWCACHE_ALIGN,
1158                                              NULL, NULL);
1159         if (!rpc_buffer_slabp)
1160                 goto err_nomem;
1161         rpc_task_mempool = mempool_create(RPC_TASK_POOLSIZE,
1162                                             mempool_alloc_slab,
1163                                             mempool_free_slab,
1164                                             rpc_task_slabp);
1165         if (!rpc_task_mempool)
1166                 goto err_nomem;
1167         rpc_buffer_mempool = mempool_create(RPC_BUFFER_POOLSIZE,
1168                                             mempool_alloc_slab,
1169                                             mempool_free_slab,
1170                                             rpc_buffer_slabp);
1171         if (!rpc_buffer_mempool)
1172                 goto err_nomem;
1173         return 0;
1174 err_nomem:
1175         rpc_destroy_mempool();
1176         return -ENOMEM;
1177 }