Fix a second potential rpc_wakeup race...
[safe/jmp/linux-2.6] / 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
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 sit here while waiting for conditions to improve.
48  */
49 static RPC_WAITQ(delay_queue, "delayq");
50
51 /*
52  * All RPC tasks are linked into this list
53  */
54 static LIST_HEAD(all_tasks);
55
56 /*
57  * rpciod-related stuff
58  */
59 static DEFINE_MUTEX(rpciod_mutex);
60 static unsigned int             rpciod_users;
61 struct workqueue_struct *rpciod_workqueue;
62
63 /*
64  * Spinlock for other critical sections of code.
65  */
66 static DEFINE_SPINLOCK(rpc_sched_lock);
67
68 /*
69  * Disable the timer for a given RPC task. Should be called with
70  * queue->lock and bh_disabled in order to avoid races within
71  * rpc_run_timer().
72  */
73 static inline void
74 __rpc_disable_timer(struct rpc_task *task)
75 {
76         dprintk("RPC: %4d disabling timer\n", task->tk_pid);
77         task->tk_timeout_fn = NULL;
78         task->tk_timeout = 0;
79 }
80
81 /*
82  * Run a timeout function.
83  * We use the callback in order to allow __rpc_wake_up_task()
84  * and friends to disable the timer synchronously on SMP systems
85  * without calling del_timer_sync(). The latter could cause a
86  * deadlock if called while we're holding spinlocks...
87  */
88 static void rpc_run_timer(struct rpc_task *task)
89 {
90         void (*callback)(struct rpc_task *);
91
92         callback = task->tk_timeout_fn;
93         task->tk_timeout_fn = NULL;
94         if (callback && RPC_IS_QUEUED(task)) {
95                 dprintk("RPC: %4d running timer\n", task->tk_pid);
96                 callback(task);
97         }
98         smp_mb__before_clear_bit();
99         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
100         smp_mb__after_clear_bit();
101 }
102
103 /*
104  * Set up a timer for the current task.
105  */
106 static inline void
107 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
108 {
109         if (!task->tk_timeout)
110                 return;
111
112         dprintk("RPC: %4d setting alarm for %lu ms\n",
113                         task->tk_pid, task->tk_timeout * 1000 / HZ);
114
115         if (timer)
116                 task->tk_timeout_fn = timer;
117         else
118                 task->tk_timeout_fn = __rpc_default_timer;
119         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
120         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
121 }
122
123 /*
124  * Delete any timer for the current task. Because we use del_timer_sync(),
125  * this function should never be called while holding queue->lock.
126  */
127 static void
128 rpc_delete_timer(struct rpc_task *task)
129 {
130         if (RPC_IS_QUEUED(task))
131                 return;
132         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
133                 del_singleshot_timer_sync(&task->tk_timer);
134                 dprintk("RPC: %4d deleting timer\n", task->tk_pid);
135         }
136 }
137
138 /*
139  * Add new request to a priority queue.
140  */
141 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
142 {
143         struct list_head *q;
144         struct rpc_task *t;
145
146         INIT_LIST_HEAD(&task->u.tk_wait.links);
147         q = &queue->tasks[task->tk_priority];
148         if (unlikely(task->tk_priority > queue->maxpriority))
149                 q = &queue->tasks[queue->maxpriority];
150         list_for_each_entry(t, q, u.tk_wait.list) {
151                 if (t->tk_cookie == task->tk_cookie) {
152                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
153                         return;
154                 }
155         }
156         list_add_tail(&task->u.tk_wait.list, q);
157 }
158
159 /*
160  * Add new request to wait queue.
161  *
162  * Swapper tasks always get inserted at the head of the queue.
163  * This should avoid many nasty memory deadlocks and hopefully
164  * improve overall performance.
165  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
166  */
167 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
168 {
169         BUG_ON (RPC_IS_QUEUED(task));
170
171         if (RPC_IS_PRIORITY(queue))
172                 __rpc_add_wait_queue_priority(queue, task);
173         else if (RPC_IS_SWAPPER(task))
174                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
175         else
176                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
177         task->u.tk_wait.rpc_waitq = queue;
178         queue->qlen++;
179         rpc_set_queued(task);
180
181         dprintk("RPC: %4d added to queue %p \"%s\"\n",
182                                 task->tk_pid, queue, rpc_qname(queue));
183 }
184
185 /*
186  * Remove request from a priority queue.
187  */
188 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
189 {
190         struct rpc_task *t;
191
192         if (!list_empty(&task->u.tk_wait.links)) {
193                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
194                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
195                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
196         }
197         list_del(&task->u.tk_wait.list);
198 }
199
200 /*
201  * Remove request from queue.
202  * Note: must be called with spin lock held.
203  */
204 static void __rpc_remove_wait_queue(struct rpc_task *task)
205 {
206         struct rpc_wait_queue *queue;
207         queue = task->u.tk_wait.rpc_waitq;
208
209         if (RPC_IS_PRIORITY(queue))
210                 __rpc_remove_wait_queue_priority(task);
211         else
212                 list_del(&task->u.tk_wait.list);
213         queue->qlen--;
214         dprintk("RPC: %4d removed from queue %p \"%s\"\n",
215                                 task->tk_pid, queue, rpc_qname(queue));
216 }
217
218 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
219 {
220         queue->priority = priority;
221         queue->count = 1 << (priority * 2);
222 }
223
224 static inline void rpc_set_waitqueue_cookie(struct rpc_wait_queue *queue, unsigned long cookie)
225 {
226         queue->cookie = cookie;
227         queue->nr = RPC_BATCH_COUNT;
228 }
229
230 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
231 {
232         rpc_set_waitqueue_priority(queue, queue->maxpriority);
233         rpc_set_waitqueue_cookie(queue, 0);
234 }
235
236 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, int maxprio)
237 {
238         int i;
239
240         spin_lock_init(&queue->lock);
241         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
242                 INIT_LIST_HEAD(&queue->tasks[i]);
243         queue->maxpriority = maxprio;
244         rpc_reset_waitqueue_priority(queue);
245 #ifdef RPC_DEBUG
246         queue->name = qname;
247 #endif
248 }
249
250 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
251 {
252         __rpc_init_priority_wait_queue(queue, qname, RPC_PRIORITY_HIGH);
253 }
254
255 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
256 {
257         __rpc_init_priority_wait_queue(queue, qname, 0);
258 }
259 EXPORT_SYMBOL(rpc_init_wait_queue);
260
261 static int rpc_wait_bit_interruptible(void *word)
262 {
263         if (signal_pending(current))
264                 return -ERESTARTSYS;
265         schedule();
266         return 0;
267 }
268
269 static void rpc_set_active(struct rpc_task *task)
270 {
271         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
272                 return;
273         spin_lock(&rpc_sched_lock);
274 #ifdef RPC_DEBUG
275         task->tk_magic = RPC_TASK_MAGIC_ID;
276         task->tk_pid = rpc_task_id++;
277 #endif
278         /* Add to global list of all tasks */
279         list_add_tail(&task->tk_task, &all_tasks);
280         spin_unlock(&rpc_sched_lock);
281 }
282
283 /*
284  * Mark an RPC call as having completed by clearing the 'active' bit
285  */
286 static void rpc_mark_complete_task(struct rpc_task *task)
287 {
288         smp_mb__before_clear_bit();
289         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
290         smp_mb__after_clear_bit();
291         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
292 }
293
294 /*
295  * Allow callers to wait for completion of an RPC call
296  */
297 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
298 {
299         if (action == NULL)
300                 action = rpc_wait_bit_interruptible;
301         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
302                         action, TASK_INTERRUPTIBLE);
303 }
304 EXPORT_SYMBOL(__rpc_wait_for_completion_task);
305
306 /*
307  * Make an RPC task runnable.
308  *
309  * Note: If the task is ASYNC, this must be called with 
310  * the spinlock held to protect the wait queue operation.
311  */
312 static void rpc_make_runnable(struct rpc_task *task)
313 {
314         BUG_ON(task->tk_timeout_fn);
315         rpc_clear_queued(task);
316         if (rpc_test_and_set_running(task))
317                 return;
318         /* We might have raced */
319         if (RPC_IS_QUEUED(task)) {
320                 rpc_clear_running(task);
321                 return;
322         }
323         if (RPC_IS_ASYNC(task)) {
324                 int status;
325
326                 INIT_WORK(&task->u.tk_work, rpc_async_schedule, (void *)task);
327                 status = queue_work(task->tk_workqueue, &task->u.tk_work);
328                 if (status < 0) {
329                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
330                         task->tk_status = status;
331                         return;
332                 }
333         } else
334                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
335 }
336
337 /*
338  * Prepare for sleeping on a wait queue.
339  * By always appending tasks to the list we ensure FIFO behavior.
340  * NB: An RPC task will only receive interrupt-driven events as long
341  * as it's on a wait queue.
342  */
343 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
344                         rpc_action action, rpc_action timer)
345 {
346         dprintk("RPC: %4d sleep_on(queue \"%s\" time %ld)\n", task->tk_pid,
347                                 rpc_qname(q), jiffies);
348
349         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
350                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
351                 return;
352         }
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         /* Mark the task as being activated if so needed */
365         rpc_set_active(task);
366
367         /*
368          * Protect the queue operations.
369          */
370         spin_lock_bh(&q->lock);
371         __rpc_sleep_on(q, task, action, timer);
372         spin_unlock_bh(&q->lock);
373 }
374
375 /**
376  * __rpc_do_wake_up_task - wake up a single rpc_task
377  * @task: task to be woken up
378  *
379  * Caller must hold queue->lock, and have cleared the task queued flag.
380  */
381 static void __rpc_do_wake_up_task(struct rpc_task *task)
382 {
383         dprintk("RPC: %4d __rpc_wake_up_task (now %ld)\n", task->tk_pid, jiffies);
384
385 #ifdef RPC_DEBUG
386         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
387 #endif
388         /* Has the task been executed yet? If not, we cannot wake it up! */
389         if (!RPC_IS_ACTIVATED(task)) {
390                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
391                 return;
392         }
393
394         __rpc_disable_timer(task);
395         __rpc_remove_wait_queue(task);
396
397         rpc_make_runnable(task);
398
399         dprintk("RPC:      __rpc_wake_up_task done\n");
400 }
401
402 /*
403  * Wake up the specified task
404  */
405 static void __rpc_wake_up_task(struct rpc_task *task)
406 {
407         if (rpc_start_wakeup(task)) {
408                 if (RPC_IS_QUEUED(task))
409                         __rpc_do_wake_up_task(task);
410                 rpc_finish_wakeup(task);
411         }
412 }
413
414 /*
415  * Default timeout handler if none specified by user
416  */
417 static void
418 __rpc_default_timer(struct rpc_task *task)
419 {
420         dprintk("RPC: %d timeout (default timer)\n", task->tk_pid);
421         task->tk_status = -ETIMEDOUT;
422         rpc_wake_up_task(task);
423 }
424
425 /*
426  * Wake up the specified task
427  */
428 void rpc_wake_up_task(struct rpc_task *task)
429 {
430         if (rpc_start_wakeup(task)) {
431                 if (RPC_IS_QUEUED(task)) {
432                         struct rpc_wait_queue *queue = task->u.tk_wait.rpc_waitq;
433
434                         spin_lock_bh(&queue->lock);
435                         __rpc_do_wake_up_task(task);
436                         spin_unlock_bh(&queue->lock);
437                 }
438                 rpc_finish_wakeup(task);
439         }
440 }
441
442 /*
443  * Wake up the next task on a priority queue.
444  */
445 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
446 {
447         struct list_head *q;
448         struct rpc_task *task;
449
450         /*
451          * Service a batch of tasks from a single cookie.
452          */
453         q = &queue->tasks[queue->priority];
454         if (!list_empty(q)) {
455                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
456                 if (queue->cookie == task->tk_cookie) {
457                         if (--queue->nr)
458                                 goto out;
459                         list_move_tail(&task->u.tk_wait.list, q);
460                 }
461                 /*
462                  * Check if we need to switch queues.
463                  */
464                 if (--queue->count)
465                         goto new_cookie;
466         }
467
468         /*
469          * Service the next queue.
470          */
471         do {
472                 if (q == &queue->tasks[0])
473                         q = &queue->tasks[queue->maxpriority];
474                 else
475                         q = q - 1;
476                 if (!list_empty(q)) {
477                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
478                         goto new_queue;
479                 }
480         } while (q != &queue->tasks[queue->priority]);
481
482         rpc_reset_waitqueue_priority(queue);
483         return NULL;
484
485 new_queue:
486         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
487 new_cookie:
488         rpc_set_waitqueue_cookie(queue, task->tk_cookie);
489 out:
490         __rpc_wake_up_task(task);
491         return task;
492 }
493
494 /*
495  * Wake up the next task on the wait queue.
496  */
497 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
498 {
499         struct rpc_task *task = NULL;
500
501         dprintk("RPC:      wake_up_next(%p \"%s\")\n", queue, rpc_qname(queue));
502         spin_lock_bh(&queue->lock);
503         if (RPC_IS_PRIORITY(queue))
504                 task = __rpc_wake_up_next_priority(queue);
505         else {
506                 task_for_first(task, &queue->tasks[0])
507                         __rpc_wake_up_task(task);
508         }
509         spin_unlock_bh(&queue->lock);
510
511         return task;
512 }
513
514 /**
515  * rpc_wake_up - wake up all rpc_tasks
516  * @queue: rpc_wait_queue on which the tasks are sleeping
517  *
518  * Grabs queue->lock
519  */
520 void rpc_wake_up(struct rpc_wait_queue *queue)
521 {
522         struct rpc_task *task, *next;
523         struct list_head *head;
524
525         spin_lock_bh(&queue->lock);
526         head = &queue->tasks[queue->maxpriority];
527         for (;;) {
528                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
529                         __rpc_wake_up_task(task);
530                 if (head == &queue->tasks[0])
531                         break;
532                 head--;
533         }
534         spin_unlock_bh(&queue->lock);
535 }
536
537 /**
538  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
539  * @queue: rpc_wait_queue on which the tasks are sleeping
540  * @status: status value to set
541  *
542  * Grabs queue->lock
543  */
544 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
545 {
546         struct rpc_task *task, *next;
547         struct list_head *head;
548
549         spin_lock_bh(&queue->lock);
550         head = &queue->tasks[queue->maxpriority];
551         for (;;) {
552                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
553                         task->tk_status = status;
554                         __rpc_wake_up_task(task);
555                 }
556                 if (head == &queue->tasks[0])
557                         break;
558                 head--;
559         }
560         spin_unlock_bh(&queue->lock);
561 }
562
563 static void __rpc_atrun(struct rpc_task *task)
564 {
565         rpc_wake_up_task(task);
566 }
567
568 /*
569  * Run a task at a later time
570  */
571 void rpc_delay(struct rpc_task *task, unsigned long delay)
572 {
573         task->tk_timeout = delay;
574         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
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         /* Release all resources associated with the task */
693         rpc_release_task(task);
694         return status;
695 }
696
697 /*
698  * User-visible entry point to the scheduler.
699  *
700  * This may be called recursively if e.g. an async NFS task updates
701  * the attributes and finds that dirty pages must be flushed.
702  * NOTE: Upon exit of this function the task is guaranteed to be
703  *       released. In particular note that tk_release() will have
704  *       been called, so your task memory may have been freed.
705  */
706 int
707 rpc_execute(struct rpc_task *task)
708 {
709         rpc_set_active(task);
710         rpc_set_running(task);
711         return __rpc_execute(task);
712 }
713
714 static void rpc_async_schedule(void *arg)
715 {
716         __rpc_execute((struct rpc_task *)arg);
717 }
718
719 /**
720  * rpc_malloc - allocate an RPC buffer
721  * @task: RPC task that will use this buffer
722  * @size: requested byte size
723  *
724  * We try to ensure that some NFS reads and writes can always proceed
725  * by using a mempool when allocating 'small' buffers.
726  * In order to avoid memory starvation triggering more writebacks of
727  * NFS requests, we use GFP_NOFS rather than GFP_KERNEL.
728  */
729 void * rpc_malloc(struct rpc_task *task, size_t size)
730 {
731         struct rpc_rqst *req = task->tk_rqstp;
732         gfp_t   gfp;
733
734         if (task->tk_flags & RPC_TASK_SWAPPER)
735                 gfp = GFP_ATOMIC;
736         else
737                 gfp = GFP_NOFS;
738
739         if (size > RPC_BUFFER_MAXSIZE) {
740                 req->rq_buffer = kmalloc(size, gfp);
741                 if (req->rq_buffer)
742                         req->rq_bufsize = size;
743         } else {
744                 req->rq_buffer = mempool_alloc(rpc_buffer_mempool, gfp);
745                 if (req->rq_buffer)
746                         req->rq_bufsize = RPC_BUFFER_MAXSIZE;
747         }
748         return req->rq_buffer;
749 }
750
751 /**
752  * rpc_free - free buffer allocated via rpc_malloc
753  * @task: RPC task with a buffer to be freed
754  *
755  */
756 void rpc_free(struct rpc_task *task)
757 {
758         struct rpc_rqst *req = task->tk_rqstp;
759
760         if (req->rq_buffer) {
761                 if (req->rq_bufsize == RPC_BUFFER_MAXSIZE)
762                         mempool_free(req->rq_buffer, rpc_buffer_mempool);
763                 else
764                         kfree(req->rq_buffer);
765                 req->rq_buffer = NULL;
766                 req->rq_bufsize = 0;
767         }
768 }
769
770 /*
771  * Creation and deletion of RPC task structures
772  */
773 void rpc_init_task(struct rpc_task *task, struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
774 {
775         memset(task, 0, sizeof(*task));
776         init_timer(&task->tk_timer);
777         task->tk_timer.data     = (unsigned long) task;
778         task->tk_timer.function = (void (*)(unsigned long)) rpc_run_timer;
779         atomic_set(&task->tk_count, 1);
780         task->tk_client = clnt;
781         task->tk_flags  = flags;
782         task->tk_ops = tk_ops;
783         if (tk_ops->rpc_call_prepare != NULL)
784                 task->tk_action = rpc_prepare_task;
785         task->tk_calldata = calldata;
786
787         /* Initialize retry counters */
788         task->tk_garb_retry = 2;
789         task->tk_cred_retry = 2;
790
791         task->tk_priority = RPC_PRIORITY_NORMAL;
792         task->tk_cookie = (unsigned long)current;
793
794         /* Initialize workqueue for async tasks */
795         task->tk_workqueue = rpciod_workqueue;
796
797         if (clnt) {
798                 atomic_inc(&clnt->cl_users);
799                 if (clnt->cl_softrtry)
800                         task->tk_flags |= RPC_TASK_SOFT;
801                 if (!clnt->cl_intr)
802                         task->tk_flags |= RPC_TASK_NOINTR;
803         }
804
805         BUG_ON(task->tk_ops == NULL);
806
807         /* starting timestamp */
808         task->tk_start = jiffies;
809
810         dprintk("RPC: %4d new task procpid %d\n", task->tk_pid,
811                                 current->pid);
812 }
813
814 static struct rpc_task *
815 rpc_alloc_task(void)
816 {
817         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
818 }
819
820 static void rpc_free_task(struct rpc_task *task)
821 {
822         dprintk("RPC: %4d freeing task\n", task->tk_pid);
823         mempool_free(task, rpc_task_mempool);
824 }
825
826 /*
827  * Create a new task for the specified client.  We have to
828  * clean up after an allocation failure, as the client may
829  * have specified "oneshot".
830  */
831 struct rpc_task *rpc_new_task(struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
832 {
833         struct rpc_task *task;
834
835         task = rpc_alloc_task();
836         if (!task)
837                 goto cleanup;
838
839         rpc_init_task(task, clnt, flags, tk_ops, calldata);
840
841         dprintk("RPC: %4d allocated task\n", task->tk_pid);
842         task->tk_flags |= RPC_TASK_DYNAMIC;
843 out:
844         return task;
845
846 cleanup:
847         /* Check whether to release the client */
848         if (clnt) {
849                 printk("rpc_new_task: failed, users=%d, oneshot=%d\n",
850                         atomic_read(&clnt->cl_users), clnt->cl_oneshot);
851                 atomic_inc(&clnt->cl_users); /* pretend we were used ... */
852                 rpc_release_client(clnt);
853         }
854         goto out;
855 }
856
857
858 void rpc_put_task(struct rpc_task *task)
859 {
860         const struct rpc_call_ops *tk_ops = task->tk_ops;
861         void *calldata = task->tk_calldata;
862
863         if (!atomic_dec_and_test(&task->tk_count))
864                 return;
865         /* Release resources */
866         if (task->tk_rqstp)
867                 xprt_release(task);
868         if (task->tk_msg.rpc_cred)
869                 rpcauth_unbindcred(task);
870         if (task->tk_client) {
871                 rpc_release_client(task->tk_client);
872                 task->tk_client = NULL;
873         }
874         if (task->tk_flags & RPC_TASK_DYNAMIC)
875                 rpc_free_task(task);
876         if (tk_ops->rpc_release)
877                 tk_ops->rpc_release(calldata);
878 }
879 EXPORT_SYMBOL(rpc_put_task);
880
881 void rpc_release_task(struct rpc_task *task)
882 {
883 #ifdef RPC_DEBUG
884         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
885 #endif
886         dprintk("RPC: %4d release task\n", task->tk_pid);
887
888         /* Remove from global task list */
889         spin_lock(&rpc_sched_lock);
890         list_del(&task->tk_task);
891         spin_unlock(&rpc_sched_lock);
892
893         BUG_ON (RPC_IS_QUEUED(task));
894
895         /* Synchronously delete any running timer */
896         rpc_delete_timer(task);
897
898 #ifdef RPC_DEBUG
899         task->tk_magic = 0;
900 #endif
901         /* Wake up anyone who is waiting for task completion */
902         rpc_mark_complete_task(task);
903
904         rpc_put_task(task);
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                 if (ops->rpc_release != NULL)
922                         ops->rpc_release(data);
923                 return ERR_PTR(-ENOMEM);
924         }
925         atomic_inc(&task->tk_count);
926         rpc_execute(task);
927         return task;
928 }
929 EXPORT_SYMBOL(rpc_run_task);
930
931 /*
932  * Kill all tasks for the given client.
933  * XXX: kill their descendants as well?
934  */
935 void rpc_killall_tasks(struct rpc_clnt *clnt)
936 {
937         struct rpc_task *rovr;
938         struct list_head *le;
939
940         dprintk("RPC:      killing all tasks for client %p\n", clnt);
941
942         /*
943          * Spin lock all_tasks to prevent changes...
944          */
945         spin_lock(&rpc_sched_lock);
946         alltask_for_each(rovr, le, &all_tasks) {
947                 if (! RPC_IS_ACTIVATED(rovr))
948                         continue;
949                 if (!clnt || rovr->tk_client == clnt) {
950                         rovr->tk_flags |= RPC_TASK_KILLED;
951                         rpc_exit(rovr, -EIO);
952                         rpc_wake_up_task(rovr);
953                 }
954         }
955         spin_unlock(&rpc_sched_lock);
956 }
957
958 static DECLARE_MUTEX_LOCKED(rpciod_running);
959
960 static void rpciod_killall(void)
961 {
962         unsigned long flags;
963
964         while (!list_empty(&all_tasks)) {
965                 clear_thread_flag(TIF_SIGPENDING);
966                 rpc_killall_tasks(NULL);
967                 flush_workqueue(rpciod_workqueue);
968                 if (!list_empty(&all_tasks)) {
969                         dprintk("rpciod_killall: waiting for tasks to exit\n");
970                         yield();
971                 }
972         }
973
974         spin_lock_irqsave(&current->sighand->siglock, flags);
975         recalc_sigpending();
976         spin_unlock_irqrestore(&current->sighand->siglock, flags);
977 }
978
979 /*
980  * Start up the rpciod process if it's not already running.
981  */
982 int
983 rpciod_up(void)
984 {
985         struct workqueue_struct *wq;
986         int error = 0;
987
988         mutex_lock(&rpciod_mutex);
989         dprintk("rpciod_up: users %d\n", rpciod_users);
990         rpciod_users++;
991         if (rpciod_workqueue)
992                 goto out;
993         /*
994          * If there's no pid, we should be the first user.
995          */
996         if (rpciod_users > 1)
997                 printk(KERN_WARNING "rpciod_up: no workqueue, %d users??\n", rpciod_users);
998         /*
999          * Create the rpciod thread and wait for it to start.
1000          */
1001         error = -ENOMEM;
1002         wq = create_workqueue("rpciod");
1003         if (wq == NULL) {
1004                 printk(KERN_WARNING "rpciod_up: create workqueue failed, error=%d\n", error);
1005                 rpciod_users--;
1006                 goto out;
1007         }
1008         rpciod_workqueue = wq;
1009         error = 0;
1010 out:
1011         mutex_unlock(&rpciod_mutex);
1012         return error;
1013 }
1014
1015 void
1016 rpciod_down(void)
1017 {
1018         mutex_lock(&rpciod_mutex);
1019         dprintk("rpciod_down sema %d\n", rpciod_users);
1020         if (rpciod_users) {
1021                 if (--rpciod_users)
1022                         goto out;
1023         } else
1024                 printk(KERN_WARNING "rpciod_down: no users??\n");
1025
1026         if (!rpciod_workqueue) {
1027                 dprintk("rpciod_down: Nothing to do!\n");
1028                 goto out;
1029         }
1030         rpciod_killall();
1031
1032         destroy_workqueue(rpciod_workqueue);
1033         rpciod_workqueue = NULL;
1034  out:
1035         mutex_unlock(&rpciod_mutex);
1036 }
1037
1038 #ifdef RPC_DEBUG
1039 void rpc_show_tasks(void)
1040 {
1041         struct list_head *le;
1042         struct rpc_task *t;
1043
1044         spin_lock(&rpc_sched_lock);
1045         if (list_empty(&all_tasks)) {
1046                 spin_unlock(&rpc_sched_lock);
1047                 return;
1048         }
1049         printk("-pid- proc flgs status -client- -prog- --rqstp- -timeout "
1050                 "-rpcwait -action- ---ops--\n");
1051         alltask_for_each(t, le, &all_tasks) {
1052                 const char *rpc_waitq = "none";
1053
1054                 if (RPC_IS_QUEUED(t))
1055                         rpc_waitq = rpc_qname(t->u.tk_wait.rpc_waitq);
1056
1057                 printk("%05d %04d %04x %06d %8p %6d %8p %08ld %8s %8p %8p\n",
1058                         t->tk_pid,
1059                         (t->tk_msg.rpc_proc ? t->tk_msg.rpc_proc->p_proc : -1),
1060                         t->tk_flags, t->tk_status,
1061                         t->tk_client,
1062                         (t->tk_client ? t->tk_client->cl_prog : 0),
1063                         t->tk_rqstp, t->tk_timeout,
1064                         rpc_waitq,
1065                         t->tk_action, t->tk_ops);
1066         }
1067         spin_unlock(&rpc_sched_lock);
1068 }
1069 #endif
1070
1071 void
1072 rpc_destroy_mempool(void)
1073 {
1074         if (rpc_buffer_mempool)
1075                 mempool_destroy(rpc_buffer_mempool);
1076         if (rpc_task_mempool)
1077                 mempool_destroy(rpc_task_mempool);
1078         if (rpc_task_slabp)
1079                 kmem_cache_destroy(rpc_task_slabp);
1080         if (rpc_buffer_slabp)
1081                 kmem_cache_destroy(rpc_buffer_slabp);
1082 }
1083
1084 int
1085 rpc_init_mempool(void)
1086 {
1087         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1088                                              sizeof(struct rpc_task),
1089                                              0, SLAB_HWCACHE_ALIGN,
1090                                              NULL, NULL);
1091         if (!rpc_task_slabp)
1092                 goto err_nomem;
1093         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1094                                              RPC_BUFFER_MAXSIZE,
1095                                              0, SLAB_HWCACHE_ALIGN,
1096                                              NULL, NULL);
1097         if (!rpc_buffer_slabp)
1098                 goto err_nomem;
1099         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1100                                                     rpc_task_slabp);
1101         if (!rpc_task_mempool)
1102                 goto err_nomem;
1103         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1104                                                       rpc_buffer_slabp);
1105         if (!rpc_buffer_mempool)
1106                 goto err_nomem;
1107         return 0;
1108 err_nomem:
1109         rpc_destroy_mempool();
1110         return -ENOMEM;
1111 }