[PATCH] Use alloc_percpu to allocate workqueues locally
[linux-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
31 /*
32  * The per-CPU workqueue (if single thread, we always use cpu 0's).
33  *
34  * The sequence counters are for flush_scheduled_work().  It wants to wait
35  * until until all currently-scheduled works are completed, but it doesn't
36  * want to be livelocked by new, incoming ones.  So it waits until
37  * remove_sequence is >= the insert_sequence which pertained when
38  * flush_scheduled_work() was called.
39  */
40 struct cpu_workqueue_struct {
41
42         spinlock_t lock;
43
44         long remove_sequence;   /* Least-recently added (next to run) */
45         long insert_sequence;   /* Next to add */
46
47         struct list_head worklist;
48         wait_queue_head_t more_work;
49         wait_queue_head_t work_done;
50
51         struct workqueue_struct *wq;
52         task_t *thread;
53
54         int run_depth;          /* Detect run_workqueue() recursion depth */
55 } ____cacheline_aligned;
56
57 /*
58  * The externally visible workqueue abstraction is an array of
59  * per-CPU workqueues:
60  */
61 struct workqueue_struct {
62         struct cpu_workqueue_struct *cpu_wq;
63         const char *name;
64         struct list_head list;  /* Empty if single thread */
65 };
66
67 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
68    threads to each one as cpus come/go. */
69 static DEFINE_SPINLOCK(workqueue_lock);
70 static LIST_HEAD(workqueues);
71
72 /* If it's single threaded, it isn't in the list of workqueues. */
73 static inline int is_single_threaded(struct workqueue_struct *wq)
74 {
75         return list_empty(&wq->list);
76 }
77
78 /* Preempt must be disabled. */
79 static void __queue_work(struct cpu_workqueue_struct *cwq,
80                          struct work_struct *work)
81 {
82         unsigned long flags;
83
84         spin_lock_irqsave(&cwq->lock, flags);
85         work->wq_data = cwq;
86         list_add_tail(&work->entry, &cwq->worklist);
87         cwq->insert_sequence++;
88         wake_up(&cwq->more_work);
89         spin_unlock_irqrestore(&cwq->lock, flags);
90 }
91
92 /*
93  * Queue work on a workqueue. Return non-zero if it was successfully
94  * added.
95  *
96  * We queue the work to the CPU it was submitted, but there is no
97  * guarantee that it will be processed by that CPU.
98  */
99 int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
100 {
101         int ret = 0, cpu = get_cpu();
102
103         if (!test_and_set_bit(0, &work->pending)) {
104                 if (unlikely(is_single_threaded(wq)))
105                         cpu = 0;
106                 BUG_ON(!list_empty(&work->entry));
107                 __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
108                 ret = 1;
109         }
110         put_cpu();
111         return ret;
112 }
113
114 static void delayed_work_timer_fn(unsigned long __data)
115 {
116         struct work_struct *work = (struct work_struct *)__data;
117         struct workqueue_struct *wq = work->wq_data;
118         int cpu = smp_processor_id();
119
120         if (unlikely(is_single_threaded(wq)))
121                 cpu = 0;
122
123         __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
124 }
125
126 int fastcall queue_delayed_work(struct workqueue_struct *wq,
127                         struct work_struct *work, unsigned long delay)
128 {
129         int ret = 0;
130         struct timer_list *timer = &work->timer;
131
132         if (!test_and_set_bit(0, &work->pending)) {
133                 BUG_ON(timer_pending(timer));
134                 BUG_ON(!list_empty(&work->entry));
135
136                 /* This stores wq for the moment, for the timer_fn */
137                 work->wq_data = wq;
138                 timer->expires = jiffies + delay;
139                 timer->data = (unsigned long)work;
140                 timer->function = delayed_work_timer_fn;
141                 add_timer(timer);
142                 ret = 1;
143         }
144         return ret;
145 }
146
147 static inline void run_workqueue(struct cpu_workqueue_struct *cwq)
148 {
149         unsigned long flags;
150
151         /*
152          * Keep taking off work from the queue until
153          * done.
154          */
155         spin_lock_irqsave(&cwq->lock, flags);
156         cwq->run_depth++;
157         if (cwq->run_depth > 3) {
158                 /* morton gets to eat his hat */
159                 printk("%s: recursion depth exceeded: %d\n",
160                         __FUNCTION__, cwq->run_depth);
161                 dump_stack();
162         }
163         while (!list_empty(&cwq->worklist)) {
164                 struct work_struct *work = list_entry(cwq->worklist.next,
165                                                 struct work_struct, entry);
166                 void (*f) (void *) = work->func;
167                 void *data = work->data;
168
169                 list_del_init(cwq->worklist.next);
170                 spin_unlock_irqrestore(&cwq->lock, flags);
171
172                 BUG_ON(work->wq_data != cwq);
173                 clear_bit(0, &work->pending);
174                 f(data);
175
176                 spin_lock_irqsave(&cwq->lock, flags);
177                 cwq->remove_sequence++;
178                 wake_up(&cwq->work_done);
179         }
180         cwq->run_depth--;
181         spin_unlock_irqrestore(&cwq->lock, flags);
182 }
183
184 static int worker_thread(void *__cwq)
185 {
186         struct cpu_workqueue_struct *cwq = __cwq;
187         DECLARE_WAITQUEUE(wait, current);
188         struct k_sigaction sa;
189         sigset_t blocked;
190
191         current->flags |= PF_NOFREEZE;
192
193         set_user_nice(current, -5);
194
195         /* Block and flush all signals */
196         sigfillset(&blocked);
197         sigprocmask(SIG_BLOCK, &blocked, NULL);
198         flush_signals(current);
199
200         /* SIG_IGN makes children autoreap: see do_notify_parent(). */
201         sa.sa.sa_handler = SIG_IGN;
202         sa.sa.sa_flags = 0;
203         siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
204         do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
205
206         set_current_state(TASK_INTERRUPTIBLE);
207         while (!kthread_should_stop()) {
208                 add_wait_queue(&cwq->more_work, &wait);
209                 if (list_empty(&cwq->worklist))
210                         schedule();
211                 else
212                         __set_current_state(TASK_RUNNING);
213                 remove_wait_queue(&cwq->more_work, &wait);
214
215                 if (!list_empty(&cwq->worklist))
216                         run_workqueue(cwq);
217                 set_current_state(TASK_INTERRUPTIBLE);
218         }
219         __set_current_state(TASK_RUNNING);
220         return 0;
221 }
222
223 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
224 {
225         if (cwq->thread == current) {
226                 /*
227                  * Probably keventd trying to flush its own queue. So simply run
228                  * it by hand rather than deadlocking.
229                  */
230                 run_workqueue(cwq);
231         } else {
232                 DEFINE_WAIT(wait);
233                 long sequence_needed;
234
235                 spin_lock_irq(&cwq->lock);
236                 sequence_needed = cwq->insert_sequence;
237
238                 while (sequence_needed - cwq->remove_sequence > 0) {
239                         prepare_to_wait(&cwq->work_done, &wait,
240                                         TASK_UNINTERRUPTIBLE);
241                         spin_unlock_irq(&cwq->lock);
242                         schedule();
243                         spin_lock_irq(&cwq->lock);
244                 }
245                 finish_wait(&cwq->work_done, &wait);
246                 spin_unlock_irq(&cwq->lock);
247         }
248 }
249
250 /*
251  * flush_workqueue - ensure that any scheduled work has run to completion.
252  *
253  * Forces execution of the workqueue and blocks until its completion.
254  * This is typically used in driver shutdown handlers.
255  *
256  * This function will sample each workqueue's current insert_sequence number and
257  * will sleep until the head sequence is greater than or equal to that.  This
258  * means that we sleep until all works which were queued on entry have been
259  * handled, but we are not livelocked by new incoming ones.
260  *
261  * This function used to run the workqueues itself.  Now we just wait for the
262  * helper threads to do it.
263  */
264 void fastcall flush_workqueue(struct workqueue_struct *wq)
265 {
266         might_sleep();
267
268         if (is_single_threaded(wq)) {
269                 /* Always use cpu 0's area. */
270                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, 0));
271         } else {
272                 int cpu;
273
274                 lock_cpu_hotplug();
275                 for_each_online_cpu(cpu)
276                         flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
277                 unlock_cpu_hotplug();
278         }
279 }
280
281 static struct task_struct *create_workqueue_thread(struct workqueue_struct *wq,
282                                                    int cpu)
283 {
284         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
285         struct task_struct *p;
286
287         spin_lock_init(&cwq->lock);
288         cwq->wq = wq;
289         cwq->thread = NULL;
290         cwq->insert_sequence = 0;
291         cwq->remove_sequence = 0;
292         INIT_LIST_HEAD(&cwq->worklist);
293         init_waitqueue_head(&cwq->more_work);
294         init_waitqueue_head(&cwq->work_done);
295
296         if (is_single_threaded(wq))
297                 p = kthread_create(worker_thread, cwq, "%s", wq->name);
298         else
299                 p = kthread_create(worker_thread, cwq, "%s/%d", wq->name, cpu);
300         if (IS_ERR(p))
301                 return NULL;
302         cwq->thread = p;
303         return p;
304 }
305
306 struct workqueue_struct *__create_workqueue(const char *name,
307                                             int singlethread)
308 {
309         int cpu, destroy = 0;
310         struct workqueue_struct *wq;
311         struct task_struct *p;
312
313         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
314         if (!wq)
315                 return NULL;
316
317         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
318         wq->name = name;
319         /* We don't need the distraction of CPUs appearing and vanishing. */
320         lock_cpu_hotplug();
321         if (singlethread) {
322                 INIT_LIST_HEAD(&wq->list);
323                 p = create_workqueue_thread(wq, 0);
324                 if (!p)
325                         destroy = 1;
326                 else
327                         wake_up_process(p);
328         } else {
329                 spin_lock(&workqueue_lock);
330                 list_add(&wq->list, &workqueues);
331                 spin_unlock(&workqueue_lock);
332                 for_each_online_cpu(cpu) {
333                         p = create_workqueue_thread(wq, cpu);
334                         if (p) {
335                                 kthread_bind(p, cpu);
336                                 wake_up_process(p);
337                         } else
338                                 destroy = 1;
339                 }
340         }
341         unlock_cpu_hotplug();
342
343         /*
344          * Was there any error during startup? If yes then clean up:
345          */
346         if (destroy) {
347                 destroy_workqueue(wq);
348                 wq = NULL;
349         }
350         return wq;
351 }
352
353 static void cleanup_workqueue_thread(struct workqueue_struct *wq, int cpu)
354 {
355         struct cpu_workqueue_struct *cwq;
356         unsigned long flags;
357         struct task_struct *p;
358
359         cwq = per_cpu_ptr(wq->cpu_wq, cpu);
360         spin_lock_irqsave(&cwq->lock, flags);
361         p = cwq->thread;
362         cwq->thread = NULL;
363         spin_unlock_irqrestore(&cwq->lock, flags);
364         if (p)
365                 kthread_stop(p);
366 }
367
368 void destroy_workqueue(struct workqueue_struct *wq)
369 {
370         int cpu;
371
372         flush_workqueue(wq);
373
374         /* We don't need the distraction of CPUs appearing and vanishing. */
375         lock_cpu_hotplug();
376         if (is_single_threaded(wq))
377                 cleanup_workqueue_thread(wq, 0);
378         else {
379                 for_each_online_cpu(cpu)
380                         cleanup_workqueue_thread(wq, cpu);
381                 spin_lock(&workqueue_lock);
382                 list_del(&wq->list);
383                 spin_unlock(&workqueue_lock);
384         }
385         unlock_cpu_hotplug();
386         free_percpu(wq->cpu_wq);
387         kfree(wq);
388 }
389
390 static struct workqueue_struct *keventd_wq;
391
392 int fastcall schedule_work(struct work_struct *work)
393 {
394         return queue_work(keventd_wq, work);
395 }
396
397 int fastcall schedule_delayed_work(struct work_struct *work, unsigned long delay)
398 {
399         return queue_delayed_work(keventd_wq, work, delay);
400 }
401
402 int schedule_delayed_work_on(int cpu,
403                         struct work_struct *work, unsigned long delay)
404 {
405         int ret = 0;
406         struct timer_list *timer = &work->timer;
407
408         if (!test_and_set_bit(0, &work->pending)) {
409                 BUG_ON(timer_pending(timer));
410                 BUG_ON(!list_empty(&work->entry));
411                 /* This stores keventd_wq for the moment, for the timer_fn */
412                 work->wq_data = keventd_wq;
413                 timer->expires = jiffies + delay;
414                 timer->data = (unsigned long)work;
415                 timer->function = delayed_work_timer_fn;
416                 add_timer_on(timer, cpu);
417                 ret = 1;
418         }
419         return ret;
420 }
421
422 void flush_scheduled_work(void)
423 {
424         flush_workqueue(keventd_wq);
425 }
426
427 /**
428  * cancel_rearming_delayed_workqueue - reliably kill off a delayed
429  *                      work whose handler rearms the delayed work.
430  * @wq:   the controlling workqueue structure
431  * @work: the delayed work struct
432  */
433 void cancel_rearming_delayed_workqueue(struct workqueue_struct *wq,
434                                        struct work_struct *work)
435 {
436         while (!cancel_delayed_work(work))
437                 flush_workqueue(wq);
438 }
439 EXPORT_SYMBOL(cancel_rearming_delayed_workqueue);
440
441 /**
442  * cancel_rearming_delayed_work - reliably kill off a delayed keventd
443  *                      work whose handler rearms the delayed work.
444  * @work: the delayed work struct
445  */
446 void cancel_rearming_delayed_work(struct work_struct *work)
447 {
448         cancel_rearming_delayed_workqueue(keventd_wq, work);
449 }
450 EXPORT_SYMBOL(cancel_rearming_delayed_work);
451
452 int keventd_up(void)
453 {
454         return keventd_wq != NULL;
455 }
456
457 int current_is_keventd(void)
458 {
459         struct cpu_workqueue_struct *cwq;
460         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
461         int ret = 0;
462
463         BUG_ON(!keventd_wq);
464
465         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
466         if (current == cwq->thread)
467                 ret = 1;
468
469         return ret;
470
471 }
472
473 #ifdef CONFIG_HOTPLUG_CPU
474 /* Take the work from this (downed) CPU. */
475 static void take_over_work(struct workqueue_struct *wq, unsigned int cpu)
476 {
477         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
478         LIST_HEAD(list);
479         struct work_struct *work;
480
481         spin_lock_irq(&cwq->lock);
482         list_splice_init(&cwq->worklist, &list);
483
484         while (!list_empty(&list)) {
485                 printk("Taking work for %s\n", wq->name);
486                 work = list_entry(list.next,struct work_struct,entry);
487                 list_del(&work->entry);
488                 __queue_work(per_cpu_ptr(wq->cpu_wq, smp_processor_id()), work);
489         }
490         spin_unlock_irq(&cwq->lock);
491 }
492
493 /* We're holding the cpucontrol mutex here */
494 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
495                                   unsigned long action,
496                                   void *hcpu)
497 {
498         unsigned int hotcpu = (unsigned long)hcpu;
499         struct workqueue_struct *wq;
500
501         switch (action) {
502         case CPU_UP_PREPARE:
503                 /* Create a new workqueue thread for it. */
504                 list_for_each_entry(wq, &workqueues, list) {
505                         if (!create_workqueue_thread(wq, hotcpu)) {
506                                 printk("workqueue for %i failed\n", hotcpu);
507                                 return NOTIFY_BAD;
508                         }
509                 }
510                 break;
511
512         case CPU_ONLINE:
513                 /* Kick off worker threads. */
514                 list_for_each_entry(wq, &workqueues, list) {
515                         struct cpu_workqueue_struct *cwq;
516
517                         cwq = per_cpu_ptr(wq->cpu_wq, hotcpu);
518                         kthread_bind(cwq->thread, hotcpu);
519                         wake_up_process(cwq->thread);
520                 }
521                 break;
522
523         case CPU_UP_CANCELED:
524                 list_for_each_entry(wq, &workqueues, list) {
525                         /* Unbind so it can run. */
526                         kthread_bind(per_cpu_ptr(wq->cpu_wq, hotcpu)->thread,
527                                      smp_processor_id());
528                         cleanup_workqueue_thread(wq, hotcpu);
529                 }
530                 break;
531
532         case CPU_DEAD:
533                 list_for_each_entry(wq, &workqueues, list)
534                         cleanup_workqueue_thread(wq, hotcpu);
535                 list_for_each_entry(wq, &workqueues, list)
536                         take_over_work(wq, hotcpu);
537                 break;
538         }
539
540         return NOTIFY_OK;
541 }
542 #endif
543
544 void init_workqueues(void)
545 {
546         hotcpu_notifier(workqueue_cpu_callback, 0);
547         keventd_wq = create_workqueue("events");
548         BUG_ON(!keventd_wq);
549 }
550
551 EXPORT_SYMBOL_GPL(__create_workqueue);
552 EXPORT_SYMBOL_GPL(queue_work);
553 EXPORT_SYMBOL_GPL(queue_delayed_work);
554 EXPORT_SYMBOL_GPL(flush_workqueue);
555 EXPORT_SYMBOL_GPL(destroy_workqueue);
556
557 EXPORT_SYMBOL(schedule_work);
558 EXPORT_SYMBOL(schedule_delayed_work);
559 EXPORT_SYMBOL(schedule_delayed_work_on);
560 EXPORT_SYMBOL(flush_scheduled_work);