Merge tag 'jg-20061012-00' of git://electric-eye.fr.zoreil.com/home/romieu/linux...
[sfrench/cifs-2.6.git] / kernel / workqueue.c
1 /*
2  * linux/kernel/workqueue.c
3  *
4  * Generic mechanism for defining kernel helper threads for running
5  * arbitrary tasks in process context.
6  *
7  * Started by Ingo Molnar, Copyright (C) 2002
8  *
9  * Derived from the taskqueue/keventd code by:
10  *
11  *   David Woodhouse <dwmw2@infradead.org>
12  *   Andrew Morton <andrewm@uow.edu.au>
13  *   Kai Petzke <wpp@marie.physik.tu-berlin.de>
14  *   Theodore Ts'o <tytso@mit.edu>
15  *
16  * Made to use alloc_percpu by Christoph Lameter <clameter@sgi.com>.
17  */
18
19 #include <linux/module.h>
20 #include <linux/kernel.h>
21 #include <linux/sched.h>
22 #include <linux/init.h>
23 #include <linux/signal.h>
24 #include <linux/completion.h>
25 #include <linux/workqueue.h>
26 #include <linux/slab.h>
27 #include <linux/cpu.h>
28 #include <linux/notifier.h>
29 #include <linux/kthread.h>
30 #include <linux/hardirq.h>
31 #include <linux/mempolicy.h>
32
33 /*
34  * The per-CPU workqueue (if single thread, we always use the first
35  * possible cpu).
36  *
37  * The sequence counters are for flush_scheduled_work().  It wants to wait
38  * until all currently-scheduled works are completed, but it doesn't
39  * want to be livelocked by new, incoming ones.  So it waits until
40  * remove_sequence is >= the insert_sequence which pertained when
41  * flush_scheduled_work() was called.
42  */
43 struct cpu_workqueue_struct {
44
45         spinlock_t lock;
46
47         long remove_sequence;   /* Least-recently added (next to run) */
48         long insert_sequence;   /* Next to add */
49
50         struct list_head worklist;
51         wait_queue_head_t more_work;
52         wait_queue_head_t work_done;
53
54         struct workqueue_struct *wq;
55         struct task_struct *thread;
56
57         int run_depth;          /* Detect run_workqueue() recursion depth */
58 } ____cacheline_aligned;
59
60 /*
61  * The externally visible workqueue abstraction is an array of
62  * per-CPU workqueues:
63  */
64 struct workqueue_struct {
65         struct cpu_workqueue_struct *cpu_wq;
66         const char *name;
67         struct list_head list;  /* Empty if single thread */
68 };
69
70 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
71    threads to each one as cpus come/go. */
72 static DEFINE_MUTEX(workqueue_mutex);
73 static LIST_HEAD(workqueues);
74
75 static int singlethread_cpu;
76
77 /* If it's single threaded, it isn't in the list of workqueues. */
78 static inline int is_single_threaded(struct workqueue_struct *wq)
79 {
80         return list_empty(&wq->list);
81 }
82
83 /* Preempt must be disabled. */
84 static void __queue_work(struct cpu_workqueue_struct *cwq,
85                          struct work_struct *work)
86 {
87         unsigned long flags;
88
89         spin_lock_irqsave(&cwq->lock, flags);
90         work->wq_data = cwq;
91         list_add_tail(&work->entry, &cwq->worklist);
92         cwq->insert_sequence++;
93         wake_up(&cwq->more_work);
94         spin_unlock_irqrestore(&cwq->lock, flags);
95 }
96
97 /**
98  * queue_work - queue work on a workqueue
99  * @wq: workqueue to use
100  * @work: work to queue
101  *
102  * Returns non-zero if it was successfully added.
103  *
104  * We queue the work to the CPU it was submitted, but there is no
105  * guarantee that it will be processed by that CPU.
106  */
107 int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
108 {
109         int ret = 0, cpu = get_cpu();
110
111         if (!test_and_set_bit(0, &work->pending)) {
112                 if (unlikely(is_single_threaded(wq)))
113                         cpu = singlethread_cpu;
114                 BUG_ON(!list_empty(&work->entry));
115                 __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
116                 ret = 1;
117         }
118         put_cpu();
119         return ret;
120 }
121 EXPORT_SYMBOL_GPL(queue_work);
122
123 static void delayed_work_timer_fn(unsigned long __data)
124 {
125         struct work_struct *work = (struct work_struct *)__data;
126         struct workqueue_struct *wq = work->wq_data;
127         int cpu = smp_processor_id();
128
129         if (unlikely(is_single_threaded(wq)))
130                 cpu = singlethread_cpu;
131
132         __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
133 }
134
135 /**
136  * queue_delayed_work - queue work on a workqueue after delay
137  * @wq: workqueue to use
138  * @work: work to queue
139  * @delay: number of jiffies to wait before queueing
140  *
141  * Returns non-zero if it was successfully added.
142  */
143 int fastcall queue_delayed_work(struct workqueue_struct *wq,
144                         struct work_struct *work, unsigned long delay)
145 {
146         int ret = 0;
147         struct timer_list *timer = &work->timer;
148
149         if (!test_and_set_bit(0, &work->pending)) {
150                 BUG_ON(timer_pending(timer));
151                 BUG_ON(!list_empty(&work->entry));
152
153                 /* This stores wq for the moment, for the timer_fn */
154                 work->wq_data = wq;
155                 timer->expires = jiffies + delay;
156                 timer->data = (unsigned long)work;
157                 timer->function = delayed_work_timer_fn;
158                 add_timer(timer);
159                 ret = 1;
160         }
161         return ret;
162 }
163 EXPORT_SYMBOL_GPL(queue_delayed_work);
164
165 /**
166  * queue_delayed_work_on - queue work on specific CPU after delay
167  * @cpu: CPU number to execute work on
168  * @wq: workqueue to use
169  * @work: work to queue
170  * @delay: number of jiffies to wait before queueing
171  *
172  * Returns non-zero if it was successfully added.
173  */
174 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
175                         struct work_struct *work, unsigned long delay)
176 {
177         int ret = 0;
178         struct timer_list *timer = &work->timer;
179
180         if (!test_and_set_bit(0, &work->pending)) {
181                 BUG_ON(timer_pending(timer));
182                 BUG_ON(!list_empty(&work->entry));
183
184                 /* This stores wq for the moment, for the timer_fn */
185                 work->wq_data = wq;
186                 timer->expires = jiffies + delay;
187                 timer->data = (unsigned long)work;
188                 timer->function = delayed_work_timer_fn;
189                 add_timer_on(timer, cpu);
190                 ret = 1;
191         }
192         return ret;
193 }
194 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
195
196 static void run_workqueue(struct cpu_workqueue_struct *cwq)
197 {
198         unsigned long flags;
199
200         /*
201          * Keep taking off work from the queue until
202          * done.
203          */
204         spin_lock_irqsave(&cwq->lock, flags);
205         cwq->run_depth++;
206         if (cwq->run_depth > 3) {
207                 /* morton gets to eat his hat */
208                 printk("%s: recursion depth exceeded: %d\n",
209                         __FUNCTION__, cwq->run_depth);
210                 dump_stack();
211         }
212         while (!list_empty(&cwq->worklist)) {
213                 struct work_struct *work = list_entry(cwq->worklist.next,
214                                                 struct work_struct, entry);
215                 void (*f) (void *) = work->func;
216                 void *data = work->data;
217
218                 list_del_init(cwq->worklist.next);
219                 spin_unlock_irqrestore(&cwq->lock, flags);
220
221                 BUG_ON(work->wq_data != cwq);
222                 clear_bit(0, &work->pending);
223                 f(data);
224
225                 spin_lock_irqsave(&cwq->lock, flags);
226                 cwq->remove_sequence++;
227                 wake_up(&cwq->work_done);
228         }
229         cwq->run_depth--;
230         spin_unlock_irqrestore(&cwq->lock, flags);
231 }
232
233 static int worker_thread(void *__cwq)
234 {
235         struct cpu_workqueue_struct *cwq = __cwq;
236         DECLARE_WAITQUEUE(wait, current);
237         struct k_sigaction sa;
238         sigset_t blocked;
239
240         current->flags |= PF_NOFREEZE;
241
242         set_user_nice(current, -5);
243
244         /* Block and flush all signals */
245         sigfillset(&blocked);
246         sigprocmask(SIG_BLOCK, &blocked, NULL);
247         flush_signals(current);
248
249         /*
250          * We inherited MPOL_INTERLEAVE from the booting kernel.
251          * Set MPOL_DEFAULT to insure node local allocations.
252          */
253         numa_default_policy();
254
255         /* SIG_IGN makes children autoreap: see do_notify_parent(). */
256         sa.sa.sa_handler = SIG_IGN;
257         sa.sa.sa_flags = 0;
258         siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
259         do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
260
261         set_current_state(TASK_INTERRUPTIBLE);
262         while (!kthread_should_stop()) {
263                 add_wait_queue(&cwq->more_work, &wait);
264                 if (list_empty(&cwq->worklist))
265                         schedule();
266                 else
267                         __set_current_state(TASK_RUNNING);
268                 remove_wait_queue(&cwq->more_work, &wait);
269
270                 if (!list_empty(&cwq->worklist))
271                         run_workqueue(cwq);
272                 set_current_state(TASK_INTERRUPTIBLE);
273         }
274         __set_current_state(TASK_RUNNING);
275         return 0;
276 }
277
278 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
279 {
280         if (cwq->thread == current) {
281                 /*
282                  * Probably keventd trying to flush its own queue. So simply run
283                  * it by hand rather than deadlocking.
284                  */
285                 run_workqueue(cwq);
286         } else {
287                 DEFINE_WAIT(wait);
288                 long sequence_needed;
289
290                 spin_lock_irq(&cwq->lock);
291                 sequence_needed = cwq->insert_sequence;
292
293                 while (sequence_needed - cwq->remove_sequence > 0) {
294                         prepare_to_wait(&cwq->work_done, &wait,
295                                         TASK_UNINTERRUPTIBLE);
296                         spin_unlock_irq(&cwq->lock);
297                         schedule();
298                         spin_lock_irq(&cwq->lock);
299                 }
300                 finish_wait(&cwq->work_done, &wait);
301                 spin_unlock_irq(&cwq->lock);
302         }
303 }
304
305 /**
306  * flush_workqueue - ensure that any scheduled work has run to completion.
307  * @wq: workqueue to flush
308  *
309  * Forces execution of the workqueue and blocks until its completion.
310  * This is typically used in driver shutdown handlers.
311  *
312  * This function will sample each workqueue's current insert_sequence number and
313  * will sleep until the head sequence is greater than or equal to that.  This
314  * means that we sleep until all works which were queued on entry have been
315  * handled, but we are not livelocked by new incoming ones.
316  *
317  * This function used to run the workqueues itself.  Now we just wait for the
318  * helper threads to do it.
319  */
320 void fastcall flush_workqueue(struct workqueue_struct *wq)
321 {
322         might_sleep();
323
324         if (is_single_threaded(wq)) {
325                 /* Always use first cpu's area. */
326                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, singlethread_cpu));
327         } else {
328                 int cpu;
329
330                 mutex_lock(&workqueue_mutex);
331                 for_each_online_cpu(cpu)
332                         flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
333                 mutex_unlock(&workqueue_mutex);
334         }
335 }
336 EXPORT_SYMBOL_GPL(flush_workqueue);
337
338 static struct task_struct *create_workqueue_thread(struct workqueue_struct *wq,
339                                                    int cpu)
340 {
341         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
342         struct task_struct *p;
343
344         spin_lock_init(&cwq->lock);
345         cwq->wq = wq;
346         cwq->thread = NULL;
347         cwq->insert_sequence = 0;
348         cwq->remove_sequence = 0;
349         INIT_LIST_HEAD(&cwq->worklist);
350         init_waitqueue_head(&cwq->more_work);
351         init_waitqueue_head(&cwq->work_done);
352
353         if (is_single_threaded(wq))
354                 p = kthread_create(worker_thread, cwq, "%s", wq->name);
355         else
356                 p = kthread_create(worker_thread, cwq, "%s/%d", wq->name, cpu);
357         if (IS_ERR(p))
358                 return NULL;
359         cwq->thread = p;
360         return p;
361 }
362
363 struct workqueue_struct *__create_workqueue(const char *name,
364                                             int singlethread)
365 {
366         int cpu, destroy = 0;
367         struct workqueue_struct *wq;
368         struct task_struct *p;
369
370         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
371         if (!wq)
372                 return NULL;
373
374         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
375         if (!wq->cpu_wq) {
376                 kfree(wq);
377                 return NULL;
378         }
379
380         wq->name = name;
381         mutex_lock(&workqueue_mutex);
382         if (singlethread) {
383                 INIT_LIST_HEAD(&wq->list);
384                 p = create_workqueue_thread(wq, singlethread_cpu);
385                 if (!p)
386                         destroy = 1;
387                 else
388                         wake_up_process(p);
389         } else {
390                 list_add(&wq->list, &workqueues);
391                 for_each_online_cpu(cpu) {
392                         p = create_workqueue_thread(wq, cpu);
393                         if (p) {
394                                 kthread_bind(p, cpu);
395                                 wake_up_process(p);
396                         } else
397                                 destroy = 1;
398                 }
399         }
400         mutex_unlock(&workqueue_mutex);
401
402         /*
403          * Was there any error during startup? If yes then clean up:
404          */
405         if (destroy) {
406                 destroy_workqueue(wq);
407                 wq = NULL;
408         }
409         return wq;
410 }
411 EXPORT_SYMBOL_GPL(__create_workqueue);
412
413 static void cleanup_workqueue_thread(struct workqueue_struct *wq, int cpu)
414 {
415         struct cpu_workqueue_struct *cwq;
416         unsigned long flags;
417         struct task_struct *p;
418
419         cwq = per_cpu_ptr(wq->cpu_wq, cpu);
420         spin_lock_irqsave(&cwq->lock, flags);
421         p = cwq->thread;
422         cwq->thread = NULL;
423         spin_unlock_irqrestore(&cwq->lock, flags);
424         if (p)
425                 kthread_stop(p);
426 }
427
428 /**
429  * destroy_workqueue - safely terminate a workqueue
430  * @wq: target workqueue
431  *
432  * Safely destroy a workqueue. All work currently pending will be done first.
433  */
434 void destroy_workqueue(struct workqueue_struct *wq)
435 {
436         int cpu;
437
438         flush_workqueue(wq);
439
440         /* We don't need the distraction of CPUs appearing and vanishing. */
441         mutex_lock(&workqueue_mutex);
442         if (is_single_threaded(wq))
443                 cleanup_workqueue_thread(wq, singlethread_cpu);
444         else {
445                 for_each_online_cpu(cpu)
446                         cleanup_workqueue_thread(wq, cpu);
447                 list_del(&wq->list);
448         }
449         mutex_unlock(&workqueue_mutex);
450         free_percpu(wq->cpu_wq);
451         kfree(wq);
452 }
453 EXPORT_SYMBOL_GPL(destroy_workqueue);
454
455 static struct workqueue_struct *keventd_wq;
456
457 /**
458  * schedule_work - put work task in global workqueue
459  * @work: job to be done
460  *
461  * This puts a job in the kernel-global workqueue.
462  */
463 int fastcall schedule_work(struct work_struct *work)
464 {
465         return queue_work(keventd_wq, work);
466 }
467 EXPORT_SYMBOL(schedule_work);
468
469 /**
470  * schedule_delayed_work - put work task in global workqueue after delay
471  * @work: job to be done
472  * @delay: number of jiffies to wait
473  *
474  * After waiting for a given time this puts a job in the kernel-global
475  * workqueue.
476  */
477 int fastcall schedule_delayed_work(struct work_struct *work, unsigned long delay)
478 {
479         return queue_delayed_work(keventd_wq, work, delay);
480 }
481 EXPORT_SYMBOL(schedule_delayed_work);
482
483 /**
484  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
485  * @cpu: cpu to use
486  * @work: job to be done
487  * @delay: number of jiffies to wait
488  *
489  * After waiting for a given time this puts a job in the kernel-global
490  * workqueue on the specified CPU.
491  */
492 int schedule_delayed_work_on(int cpu,
493                         struct work_struct *work, unsigned long delay)
494 {
495         return queue_delayed_work_on(cpu, keventd_wq, work, delay);
496 }
497 EXPORT_SYMBOL(schedule_delayed_work_on);
498
499 /**
500  * schedule_on_each_cpu - call a function on each online CPU from keventd
501  * @func: the function to call
502  * @info: a pointer to pass to func()
503  *
504  * Returns zero on success.
505  * Returns -ve errno on failure.
506  *
507  * Appears to be racy against CPU hotplug.
508  *
509  * schedule_on_each_cpu() is very slow.
510  */
511 int schedule_on_each_cpu(void (*func)(void *info), void *info)
512 {
513         int cpu;
514         struct work_struct *works;
515
516         works = alloc_percpu(struct work_struct);
517         if (!works)
518                 return -ENOMEM;
519
520         mutex_lock(&workqueue_mutex);
521         for_each_online_cpu(cpu) {
522                 INIT_WORK(per_cpu_ptr(works, cpu), func, info);
523                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu),
524                                 per_cpu_ptr(works, cpu));
525         }
526         mutex_unlock(&workqueue_mutex);
527         flush_workqueue(keventd_wq);
528         free_percpu(works);
529         return 0;
530 }
531
532 void flush_scheduled_work(void)
533 {
534         flush_workqueue(keventd_wq);
535 }
536 EXPORT_SYMBOL(flush_scheduled_work);
537
538 /**
539  * cancel_rearming_delayed_workqueue - reliably kill off a delayed
540  *                      work whose handler rearms the delayed work.
541  * @wq:   the controlling workqueue structure
542  * @work: the delayed work struct
543  */
544 void cancel_rearming_delayed_workqueue(struct workqueue_struct *wq,
545                                        struct work_struct *work)
546 {
547         while (!cancel_delayed_work(work))
548                 flush_workqueue(wq);
549 }
550 EXPORT_SYMBOL(cancel_rearming_delayed_workqueue);
551
552 /**
553  * cancel_rearming_delayed_work - reliably kill off a delayed keventd
554  *                      work whose handler rearms the delayed work.
555  * @work: the delayed work struct
556  */
557 void cancel_rearming_delayed_work(struct work_struct *work)
558 {
559         cancel_rearming_delayed_workqueue(keventd_wq, work);
560 }
561 EXPORT_SYMBOL(cancel_rearming_delayed_work);
562
563 /**
564  * execute_in_process_context - reliably execute the routine with user context
565  * @fn:         the function to execute
566  * @data:       data to pass to the function
567  * @ew:         guaranteed storage for the execute work structure (must
568  *              be available when the work executes)
569  *
570  * Executes the function immediately if process context is available,
571  * otherwise schedules the function for delayed execution.
572  *
573  * Returns:     0 - function was executed
574  *              1 - function was scheduled for execution
575  */
576 int execute_in_process_context(void (*fn)(void *data), void *data,
577                                struct execute_work *ew)
578 {
579         if (!in_interrupt()) {
580                 fn(data);
581                 return 0;
582         }
583
584         INIT_WORK(&ew->work, fn, data);
585         schedule_work(&ew->work);
586
587         return 1;
588 }
589 EXPORT_SYMBOL_GPL(execute_in_process_context);
590
591 int keventd_up(void)
592 {
593         return keventd_wq != NULL;
594 }
595
596 int current_is_keventd(void)
597 {
598         struct cpu_workqueue_struct *cwq;
599         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
600         int ret = 0;
601
602         BUG_ON(!keventd_wq);
603
604         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
605         if (current == cwq->thread)
606                 ret = 1;
607
608         return ret;
609
610 }
611
612 #ifdef CONFIG_HOTPLUG_CPU
613 /* Take the work from this (downed) CPU. */
614 static void take_over_work(struct workqueue_struct *wq, unsigned int cpu)
615 {
616         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
617         struct list_head list;
618         struct work_struct *work;
619
620         spin_lock_irq(&cwq->lock);
621         list_replace_init(&cwq->worklist, &list);
622
623         while (!list_empty(&list)) {
624                 printk("Taking work for %s\n", wq->name);
625                 work = list_entry(list.next,struct work_struct,entry);
626                 list_del(&work->entry);
627                 __queue_work(per_cpu_ptr(wq->cpu_wq, smp_processor_id()), work);
628         }
629         spin_unlock_irq(&cwq->lock);
630 }
631
632 /* We're holding the cpucontrol mutex here */
633 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
634                                   unsigned long action,
635                                   void *hcpu)
636 {
637         unsigned int hotcpu = (unsigned long)hcpu;
638         struct workqueue_struct *wq;
639
640         switch (action) {
641         case CPU_UP_PREPARE:
642                 mutex_lock(&workqueue_mutex);
643                 /* Create a new workqueue thread for it. */
644                 list_for_each_entry(wq, &workqueues, list) {
645                         if (!create_workqueue_thread(wq, hotcpu)) {
646                                 printk("workqueue for %i failed\n", hotcpu);
647                                 return NOTIFY_BAD;
648                         }
649                 }
650                 break;
651
652         case CPU_ONLINE:
653                 /* Kick off worker threads. */
654                 list_for_each_entry(wq, &workqueues, list) {
655                         struct cpu_workqueue_struct *cwq;
656
657                         cwq = per_cpu_ptr(wq->cpu_wq, hotcpu);
658                         kthread_bind(cwq->thread, hotcpu);
659                         wake_up_process(cwq->thread);
660                 }
661                 mutex_unlock(&workqueue_mutex);
662                 break;
663
664         case CPU_UP_CANCELED:
665                 list_for_each_entry(wq, &workqueues, list) {
666                         if (!per_cpu_ptr(wq->cpu_wq, hotcpu)->thread)
667                                 continue;
668                         /* Unbind so it can run. */
669                         kthread_bind(per_cpu_ptr(wq->cpu_wq, hotcpu)->thread,
670                                      any_online_cpu(cpu_online_map));
671                         cleanup_workqueue_thread(wq, hotcpu);
672                 }
673                 mutex_unlock(&workqueue_mutex);
674                 break;
675
676         case CPU_DOWN_PREPARE:
677                 mutex_lock(&workqueue_mutex);
678                 break;
679
680         case CPU_DOWN_FAILED:
681                 mutex_unlock(&workqueue_mutex);
682                 break;
683
684         case CPU_DEAD:
685                 list_for_each_entry(wq, &workqueues, list)
686                         cleanup_workqueue_thread(wq, hotcpu);
687                 list_for_each_entry(wq, &workqueues, list)
688                         take_over_work(wq, hotcpu);
689                 mutex_unlock(&workqueue_mutex);
690                 break;
691         }
692
693         return NOTIFY_OK;
694 }
695 #endif
696
697 void init_workqueues(void)
698 {
699         singlethread_cpu = first_cpu(cpu_possible_map);
700         hotcpu_notifier(workqueue_cpu_callback, 0);
701         keventd_wq = create_workqueue("events");
702         BUG_ON(!keventd_wq);
703 }
704