lockdep: Add an in_workqueue_context() lockdep-based test function
[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
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.
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 #include <linux/freezer.h>
33 #include <linux/kallsyms.h>
34 #include <linux/debug_locks.h>
35 #include <linux/lockdep.h>
36 #define CREATE_TRACE_POINTS
37 #include <trace/events/workqueue.h>
38
39 /*
40  * The per-CPU workqueue (if single thread, we always use the first
41  * possible cpu).
42  */
43 struct cpu_workqueue_struct {
44
45         spinlock_t lock;
46
47         struct list_head worklist;
48         wait_queue_head_t more_work;
49         struct work_struct *current_work;
50
51         struct workqueue_struct *wq;
52         struct task_struct *thread;
53 } ____cacheline_aligned;
54
55 /*
56  * The externally visible workqueue abstraction is an array of
57  * per-CPU workqueues:
58  */
59 struct workqueue_struct {
60         struct cpu_workqueue_struct *cpu_wq;
61         struct list_head list;
62         const char *name;
63         int singlethread;
64         int freezeable;         /* Freeze threads during suspend */
65         int rt;
66 #ifdef CONFIG_LOCKDEP
67         struct lockdep_map lockdep_map;
68 #endif
69 };
70
71 #ifdef CONFIG_LOCKDEP
72 /**
73  * in_workqueue_context() - in context of specified workqueue?
74  * @wq: the workqueue of interest
75  *
76  * Checks lockdep state to see if the current task is executing from
77  * within a workqueue item.  This function exists only if lockdep is
78  * enabled.
79  */
80 int in_workqueue_context(struct workqueue_struct *wq)
81 {
82         return lock_is_held(&wq->lockdep_map);
83 }
84 #endif
85
86 #ifdef CONFIG_DEBUG_OBJECTS_WORK
87
88 static struct debug_obj_descr work_debug_descr;
89
90 /*
91  * fixup_init is called when:
92  * - an active object is initialized
93  */
94 static int work_fixup_init(void *addr, enum debug_obj_state state)
95 {
96         struct work_struct *work = addr;
97
98         switch (state) {
99         case ODEBUG_STATE_ACTIVE:
100                 cancel_work_sync(work);
101                 debug_object_init(work, &work_debug_descr);
102                 return 1;
103         default:
104                 return 0;
105         }
106 }
107
108 /*
109  * fixup_activate is called when:
110  * - an active object is activated
111  * - an unknown object is activated (might be a statically initialized object)
112  */
113 static int work_fixup_activate(void *addr, enum debug_obj_state state)
114 {
115         struct work_struct *work = addr;
116
117         switch (state) {
118
119         case ODEBUG_STATE_NOTAVAILABLE:
120                 /*
121                  * This is not really a fixup. The work struct was
122                  * statically initialized. We just make sure that it
123                  * is tracked in the object tracker.
124                  */
125                 if (test_bit(WORK_STRUCT_STATIC, work_data_bits(work))) {
126                         debug_object_init(work, &work_debug_descr);
127                         debug_object_activate(work, &work_debug_descr);
128                         return 0;
129                 }
130                 WARN_ON_ONCE(1);
131                 return 0;
132
133         case ODEBUG_STATE_ACTIVE:
134                 WARN_ON(1);
135
136         default:
137                 return 0;
138         }
139 }
140
141 /*
142  * fixup_free is called when:
143  * - an active object is freed
144  */
145 static int work_fixup_free(void *addr, enum debug_obj_state state)
146 {
147         struct work_struct *work = addr;
148
149         switch (state) {
150         case ODEBUG_STATE_ACTIVE:
151                 cancel_work_sync(work);
152                 debug_object_free(work, &work_debug_descr);
153                 return 1;
154         default:
155                 return 0;
156         }
157 }
158
159 static struct debug_obj_descr work_debug_descr = {
160         .name           = "work_struct",
161         .fixup_init     = work_fixup_init,
162         .fixup_activate = work_fixup_activate,
163         .fixup_free     = work_fixup_free,
164 };
165
166 static inline void debug_work_activate(struct work_struct *work)
167 {
168         debug_object_activate(work, &work_debug_descr);
169 }
170
171 static inline void debug_work_deactivate(struct work_struct *work)
172 {
173         debug_object_deactivate(work, &work_debug_descr);
174 }
175
176 void __init_work(struct work_struct *work, int onstack)
177 {
178         if (onstack)
179                 debug_object_init_on_stack(work, &work_debug_descr);
180         else
181                 debug_object_init(work, &work_debug_descr);
182 }
183 EXPORT_SYMBOL_GPL(__init_work);
184
185 void destroy_work_on_stack(struct work_struct *work)
186 {
187         debug_object_free(work, &work_debug_descr);
188 }
189 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
190
191 #else
192 static inline void debug_work_activate(struct work_struct *work) { }
193 static inline void debug_work_deactivate(struct work_struct *work) { }
194 #endif
195
196 /* Serializes the accesses to the list of workqueues. */
197 static DEFINE_SPINLOCK(workqueue_lock);
198 static LIST_HEAD(workqueues);
199
200 static int singlethread_cpu __read_mostly;
201 static const struct cpumask *cpu_singlethread_map __read_mostly;
202 /*
203  * _cpu_down() first removes CPU from cpu_online_map, then CPU_DEAD
204  * flushes cwq->worklist. This means that flush_workqueue/wait_on_work
205  * which comes in between can't use for_each_online_cpu(). We could
206  * use cpu_possible_map, the cpumask below is more a documentation
207  * than optimization.
208  */
209 static cpumask_var_t cpu_populated_map __read_mostly;
210
211 /* If it's single threaded, it isn't in the list of workqueues. */
212 static inline int is_wq_single_threaded(struct workqueue_struct *wq)
213 {
214         return wq->singlethread;
215 }
216
217 static const struct cpumask *wq_cpu_map(struct workqueue_struct *wq)
218 {
219         return is_wq_single_threaded(wq)
220                 ? cpu_singlethread_map : cpu_populated_map;
221 }
222
223 static
224 struct cpu_workqueue_struct *wq_per_cpu(struct workqueue_struct *wq, int cpu)
225 {
226         if (unlikely(is_wq_single_threaded(wq)))
227                 cpu = singlethread_cpu;
228         return per_cpu_ptr(wq->cpu_wq, cpu);
229 }
230
231 /*
232  * Set the workqueue on which a work item is to be run
233  * - Must *only* be called if the pending flag is set
234  */
235 static inline void set_wq_data(struct work_struct *work,
236                                 struct cpu_workqueue_struct *cwq)
237 {
238         unsigned long new;
239
240         BUG_ON(!work_pending(work));
241
242         new = (unsigned long) cwq | (1UL << WORK_STRUCT_PENDING);
243         new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
244         atomic_long_set(&work->data, new);
245 }
246
247 /*
248  * Clear WORK_STRUCT_PENDING and the workqueue on which it was queued.
249  */
250 static inline void clear_wq_data(struct work_struct *work)
251 {
252         unsigned long flags = *work_data_bits(work) &
253                                 (1UL << WORK_STRUCT_STATIC);
254         atomic_long_set(&work->data, flags);
255 }
256
257 static inline
258 struct cpu_workqueue_struct *get_wq_data(struct work_struct *work)
259 {
260         return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
261 }
262
263 static void insert_work(struct cpu_workqueue_struct *cwq,
264                         struct work_struct *work, struct list_head *head)
265 {
266         trace_workqueue_insertion(cwq->thread, work);
267
268         set_wq_data(work, cwq);
269         /*
270          * Ensure that we get the right work->data if we see the
271          * result of list_add() below, see try_to_grab_pending().
272          */
273         smp_wmb();
274         list_add_tail(&work->entry, head);
275         wake_up(&cwq->more_work);
276 }
277
278 static void __queue_work(struct cpu_workqueue_struct *cwq,
279                          struct work_struct *work)
280 {
281         unsigned long flags;
282
283         debug_work_activate(work);
284         spin_lock_irqsave(&cwq->lock, flags);
285         insert_work(cwq, work, &cwq->worklist);
286         spin_unlock_irqrestore(&cwq->lock, flags);
287 }
288
289 /**
290  * queue_work - queue work on a workqueue
291  * @wq: workqueue to use
292  * @work: work to queue
293  *
294  * Returns 0 if @work was already on a queue, non-zero otherwise.
295  *
296  * We queue the work to the CPU on which it was submitted, but if the CPU dies
297  * it can be processed by another CPU.
298  */
299 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
300 {
301         int ret;
302
303         ret = queue_work_on(get_cpu(), wq, work);
304         put_cpu();
305
306         return ret;
307 }
308 EXPORT_SYMBOL_GPL(queue_work);
309
310 /**
311  * queue_work_on - queue work on specific cpu
312  * @cpu: CPU number to execute work on
313  * @wq: workqueue to use
314  * @work: work to queue
315  *
316  * Returns 0 if @work was already on a queue, non-zero otherwise.
317  *
318  * We queue the work to a specific CPU, the caller must ensure it
319  * can't go away.
320  */
321 int
322 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
323 {
324         int ret = 0;
325
326         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
327                 BUG_ON(!list_empty(&work->entry));
328                 __queue_work(wq_per_cpu(wq, cpu), work);
329                 ret = 1;
330         }
331         return ret;
332 }
333 EXPORT_SYMBOL_GPL(queue_work_on);
334
335 static void delayed_work_timer_fn(unsigned long __data)
336 {
337         struct delayed_work *dwork = (struct delayed_work *)__data;
338         struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
339         struct workqueue_struct *wq = cwq->wq;
340
341         __queue_work(wq_per_cpu(wq, smp_processor_id()), &dwork->work);
342 }
343
344 /**
345  * queue_delayed_work - queue work on a workqueue after delay
346  * @wq: workqueue to use
347  * @dwork: delayable work to queue
348  * @delay: number of jiffies to wait before queueing
349  *
350  * Returns 0 if @work was already on a queue, non-zero otherwise.
351  */
352 int queue_delayed_work(struct workqueue_struct *wq,
353                         struct delayed_work *dwork, unsigned long delay)
354 {
355         if (delay == 0)
356                 return queue_work(wq, &dwork->work);
357
358         return queue_delayed_work_on(-1, wq, dwork, delay);
359 }
360 EXPORT_SYMBOL_GPL(queue_delayed_work);
361
362 /**
363  * queue_delayed_work_on - queue work on specific CPU after delay
364  * @cpu: CPU number to execute work on
365  * @wq: workqueue to use
366  * @dwork: work to queue
367  * @delay: number of jiffies to wait before queueing
368  *
369  * Returns 0 if @work was already on a queue, non-zero otherwise.
370  */
371 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
372                         struct delayed_work *dwork, unsigned long delay)
373 {
374         int ret = 0;
375         struct timer_list *timer = &dwork->timer;
376         struct work_struct *work = &dwork->work;
377
378         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
379                 BUG_ON(timer_pending(timer));
380                 BUG_ON(!list_empty(&work->entry));
381
382                 timer_stats_timer_set_start_info(&dwork->timer);
383
384                 /* This stores cwq for the moment, for the timer_fn */
385                 set_wq_data(work, wq_per_cpu(wq, raw_smp_processor_id()));
386                 timer->expires = jiffies + delay;
387                 timer->data = (unsigned long)dwork;
388                 timer->function = delayed_work_timer_fn;
389
390                 if (unlikely(cpu >= 0))
391                         add_timer_on(timer, cpu);
392                 else
393                         add_timer(timer);
394                 ret = 1;
395         }
396         return ret;
397 }
398 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
399
400 static void run_workqueue(struct cpu_workqueue_struct *cwq)
401 {
402         spin_lock_irq(&cwq->lock);
403         while (!list_empty(&cwq->worklist)) {
404                 struct work_struct *work = list_entry(cwq->worklist.next,
405                                                 struct work_struct, entry);
406                 work_func_t f = work->func;
407 #ifdef CONFIG_LOCKDEP
408                 /*
409                  * It is permissible to free the struct work_struct
410                  * from inside the function that is called from it,
411                  * this we need to take into account for lockdep too.
412                  * To avoid bogus "held lock freed" warnings as well
413                  * as problems when looking into work->lockdep_map,
414                  * make a copy and use that here.
415                  */
416                 struct lockdep_map lockdep_map = work->lockdep_map;
417 #endif
418                 trace_workqueue_execution(cwq->thread, work);
419                 debug_work_deactivate(work);
420                 cwq->current_work = work;
421                 list_del_init(cwq->worklist.next);
422                 spin_unlock_irq(&cwq->lock);
423
424                 BUG_ON(get_wq_data(work) != cwq);
425                 work_clear_pending(work);
426                 lock_map_acquire(&cwq->wq->lockdep_map);
427                 lock_map_acquire(&lockdep_map);
428                 f(work);
429                 lock_map_release(&lockdep_map);
430                 lock_map_release(&cwq->wq->lockdep_map);
431
432                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
433                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
434                                         "%s/0x%08x/%d\n",
435                                         current->comm, preempt_count(),
436                                         task_pid_nr(current));
437                         printk(KERN_ERR "    last function: ");
438                         print_symbol("%s\n", (unsigned long)f);
439                         debug_show_held_locks(current);
440                         dump_stack();
441                 }
442
443                 spin_lock_irq(&cwq->lock);
444                 cwq->current_work = NULL;
445         }
446         spin_unlock_irq(&cwq->lock);
447 }
448
449 static int worker_thread(void *__cwq)
450 {
451         struct cpu_workqueue_struct *cwq = __cwq;
452         DEFINE_WAIT(wait);
453
454         if (cwq->wq->freezeable)
455                 set_freezable();
456
457         for (;;) {
458                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
459                 if (!freezing(current) &&
460                     !kthread_should_stop() &&
461                     list_empty(&cwq->worklist))
462                         schedule();
463                 finish_wait(&cwq->more_work, &wait);
464
465                 try_to_freeze();
466
467                 if (kthread_should_stop())
468                         break;
469
470                 run_workqueue(cwq);
471         }
472
473         return 0;
474 }
475
476 struct wq_barrier {
477         struct work_struct      work;
478         struct completion       done;
479 };
480
481 static void wq_barrier_func(struct work_struct *work)
482 {
483         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
484         complete(&barr->done);
485 }
486
487 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
488                         struct wq_barrier *barr, struct list_head *head)
489 {
490         /*
491          * debugobject calls are safe here even with cwq->lock locked
492          * as we know for sure that this will not trigger any of the
493          * checks and call back into the fixup functions where we
494          * might deadlock.
495          */
496         INIT_WORK_ON_STACK(&barr->work, wq_barrier_func);
497         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
498
499         init_completion(&barr->done);
500
501         debug_work_activate(&barr->work);
502         insert_work(cwq, &barr->work, head);
503 }
504
505 static int flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
506 {
507         int active = 0;
508         struct wq_barrier barr;
509
510         WARN_ON(cwq->thread == current);
511
512         spin_lock_irq(&cwq->lock);
513         if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
514                 insert_wq_barrier(cwq, &barr, &cwq->worklist);
515                 active = 1;
516         }
517         spin_unlock_irq(&cwq->lock);
518
519         if (active) {
520                 wait_for_completion(&barr.done);
521                 destroy_work_on_stack(&barr.work);
522         }
523
524         return active;
525 }
526
527 /**
528  * flush_workqueue - ensure that any scheduled work has run to completion.
529  * @wq: workqueue to flush
530  *
531  * Forces execution of the workqueue and blocks until its completion.
532  * This is typically used in driver shutdown handlers.
533  *
534  * We sleep until all works which were queued on entry have been handled,
535  * but we are not livelocked by new incoming ones.
536  *
537  * This function used to run the workqueues itself.  Now we just wait for the
538  * helper threads to do it.
539  */
540 void flush_workqueue(struct workqueue_struct *wq)
541 {
542         const struct cpumask *cpu_map = wq_cpu_map(wq);
543         int cpu;
544
545         might_sleep();
546         lock_map_acquire(&wq->lockdep_map);
547         lock_map_release(&wq->lockdep_map);
548         for_each_cpu(cpu, cpu_map)
549                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
550 }
551 EXPORT_SYMBOL_GPL(flush_workqueue);
552
553 /**
554  * flush_work - block until a work_struct's callback has terminated
555  * @work: the work which is to be flushed
556  *
557  * Returns false if @work has already terminated.
558  *
559  * It is expected that, prior to calling flush_work(), the caller has
560  * arranged for the work to not be requeued, otherwise it doesn't make
561  * sense to use this function.
562  */
563 int flush_work(struct work_struct *work)
564 {
565         struct cpu_workqueue_struct *cwq;
566         struct list_head *prev;
567         struct wq_barrier barr;
568
569         might_sleep();
570         cwq = get_wq_data(work);
571         if (!cwq)
572                 return 0;
573
574         lock_map_acquire(&cwq->wq->lockdep_map);
575         lock_map_release(&cwq->wq->lockdep_map);
576
577         prev = NULL;
578         spin_lock_irq(&cwq->lock);
579         if (!list_empty(&work->entry)) {
580                 /*
581                  * See the comment near try_to_grab_pending()->smp_rmb().
582                  * If it was re-queued under us we are not going to wait.
583                  */
584                 smp_rmb();
585                 if (unlikely(cwq != get_wq_data(work)))
586                         goto out;
587                 prev = &work->entry;
588         } else {
589                 if (cwq->current_work != work)
590                         goto out;
591                 prev = &cwq->worklist;
592         }
593         insert_wq_barrier(cwq, &barr, prev->next);
594 out:
595         spin_unlock_irq(&cwq->lock);
596         if (!prev)
597                 return 0;
598
599         wait_for_completion(&barr.done);
600         destroy_work_on_stack(&barr.work);
601         return 1;
602 }
603 EXPORT_SYMBOL_GPL(flush_work);
604
605 /*
606  * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
607  * so this work can't be re-armed in any way.
608  */
609 static int try_to_grab_pending(struct work_struct *work)
610 {
611         struct cpu_workqueue_struct *cwq;
612         int ret = -1;
613
614         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work)))
615                 return 0;
616
617         /*
618          * The queueing is in progress, or it is already queued. Try to
619          * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
620          */
621
622         cwq = get_wq_data(work);
623         if (!cwq)
624                 return ret;
625
626         spin_lock_irq(&cwq->lock);
627         if (!list_empty(&work->entry)) {
628                 /*
629                  * This work is queued, but perhaps we locked the wrong cwq.
630                  * In that case we must see the new value after rmb(), see
631                  * insert_work()->wmb().
632                  */
633                 smp_rmb();
634                 if (cwq == get_wq_data(work)) {
635                         debug_work_deactivate(work);
636                         list_del_init(&work->entry);
637                         ret = 1;
638                 }
639         }
640         spin_unlock_irq(&cwq->lock);
641
642         return ret;
643 }
644
645 static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq,
646                                 struct work_struct *work)
647 {
648         struct wq_barrier barr;
649         int running = 0;
650
651         spin_lock_irq(&cwq->lock);
652         if (unlikely(cwq->current_work == work)) {
653                 insert_wq_barrier(cwq, &barr, cwq->worklist.next);
654                 running = 1;
655         }
656         spin_unlock_irq(&cwq->lock);
657
658         if (unlikely(running)) {
659                 wait_for_completion(&barr.done);
660                 destroy_work_on_stack(&barr.work);
661         }
662 }
663
664 static void wait_on_work(struct work_struct *work)
665 {
666         struct cpu_workqueue_struct *cwq;
667         struct workqueue_struct *wq;
668         const struct cpumask *cpu_map;
669         int cpu;
670
671         might_sleep();
672
673         lock_map_acquire(&work->lockdep_map);
674         lock_map_release(&work->lockdep_map);
675
676         cwq = get_wq_data(work);
677         if (!cwq)
678                 return;
679
680         wq = cwq->wq;
681         cpu_map = wq_cpu_map(wq);
682
683         for_each_cpu(cpu, cpu_map)
684                 wait_on_cpu_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
685 }
686
687 static int __cancel_work_timer(struct work_struct *work,
688                                 struct timer_list* timer)
689 {
690         int ret;
691
692         do {
693                 ret = (timer && likely(del_timer(timer)));
694                 if (!ret)
695                         ret = try_to_grab_pending(work);
696                 wait_on_work(work);
697         } while (unlikely(ret < 0));
698
699         clear_wq_data(work);
700         return ret;
701 }
702
703 /**
704  * cancel_work_sync - block until a work_struct's callback has terminated
705  * @work: the work which is to be flushed
706  *
707  * Returns true if @work was pending.
708  *
709  * cancel_work_sync() will cancel the work if it is queued. If the work's
710  * callback appears to be running, cancel_work_sync() will block until it
711  * has completed.
712  *
713  * It is possible to use this function if the work re-queues itself. It can
714  * cancel the work even if it migrates to another workqueue, however in that
715  * case it only guarantees that work->func() has completed on the last queued
716  * workqueue.
717  *
718  * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
719  * pending, otherwise it goes into a busy-wait loop until the timer expires.
720  *
721  * The caller must ensure that workqueue_struct on which this work was last
722  * queued can't be destroyed before this function returns.
723  */
724 int cancel_work_sync(struct work_struct *work)
725 {
726         return __cancel_work_timer(work, NULL);
727 }
728 EXPORT_SYMBOL_GPL(cancel_work_sync);
729
730 /**
731  * cancel_delayed_work_sync - reliably kill off a delayed work.
732  * @dwork: the delayed work struct
733  *
734  * Returns true if @dwork was pending.
735  *
736  * It is possible to use this function if @dwork rearms itself via queue_work()
737  * or queue_delayed_work(). See also the comment for cancel_work_sync().
738  */
739 int cancel_delayed_work_sync(struct delayed_work *dwork)
740 {
741         return __cancel_work_timer(&dwork->work, &dwork->timer);
742 }
743 EXPORT_SYMBOL(cancel_delayed_work_sync);
744
745 static struct workqueue_struct *keventd_wq __read_mostly;
746
747 /**
748  * schedule_work - put work task in global workqueue
749  * @work: job to be done
750  *
751  * Returns zero if @work was already on the kernel-global workqueue and
752  * non-zero otherwise.
753  *
754  * This puts a job in the kernel-global workqueue if it was not already
755  * queued and leaves it in the same position on the kernel-global
756  * workqueue otherwise.
757  */
758 int schedule_work(struct work_struct *work)
759 {
760         return queue_work(keventd_wq, work);
761 }
762 EXPORT_SYMBOL(schedule_work);
763
764 /*
765  * schedule_work_on - put work task on a specific cpu
766  * @cpu: cpu to put the work task on
767  * @work: job to be done
768  *
769  * This puts a job on a specific cpu
770  */
771 int schedule_work_on(int cpu, struct work_struct *work)
772 {
773         return queue_work_on(cpu, keventd_wq, work);
774 }
775 EXPORT_SYMBOL(schedule_work_on);
776
777 /**
778  * schedule_delayed_work - put work task in global workqueue after delay
779  * @dwork: job to be done
780  * @delay: number of jiffies to wait or 0 for immediate execution
781  *
782  * After waiting for a given time this puts a job in the kernel-global
783  * workqueue.
784  */
785 int schedule_delayed_work(struct delayed_work *dwork,
786                                         unsigned long delay)
787 {
788         return queue_delayed_work(keventd_wq, dwork, delay);
789 }
790 EXPORT_SYMBOL(schedule_delayed_work);
791
792 /**
793  * flush_delayed_work - block until a dwork_struct's callback has terminated
794  * @dwork: the delayed work which is to be flushed
795  *
796  * Any timeout is cancelled, and any pending work is run immediately.
797  */
798 void flush_delayed_work(struct delayed_work *dwork)
799 {
800         if (del_timer_sync(&dwork->timer)) {
801                 struct cpu_workqueue_struct *cwq;
802                 cwq = wq_per_cpu(get_wq_data(&dwork->work)->wq, get_cpu());
803                 __queue_work(cwq, &dwork->work);
804                 put_cpu();
805         }
806         flush_work(&dwork->work);
807 }
808 EXPORT_SYMBOL(flush_delayed_work);
809
810 /**
811  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
812  * @cpu: cpu to use
813  * @dwork: job to be done
814  * @delay: number of jiffies to wait
815  *
816  * After waiting for a given time this puts a job in the kernel-global
817  * workqueue on the specified CPU.
818  */
819 int schedule_delayed_work_on(int cpu,
820                         struct delayed_work *dwork, unsigned long delay)
821 {
822         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
823 }
824 EXPORT_SYMBOL(schedule_delayed_work_on);
825
826 /**
827  * schedule_on_each_cpu - call a function on each online CPU from keventd
828  * @func: the function to call
829  *
830  * Returns zero on success.
831  * Returns -ve errno on failure.
832  *
833  * schedule_on_each_cpu() is very slow.
834  */
835 int schedule_on_each_cpu(work_func_t func)
836 {
837         int cpu;
838         int orig = -1;
839         struct work_struct *works;
840
841         works = alloc_percpu(struct work_struct);
842         if (!works)
843                 return -ENOMEM;
844
845         get_online_cpus();
846
847         /*
848          * When running in keventd don't schedule a work item on
849          * itself.  Can just call directly because the work queue is
850          * already bound.  This also is faster.
851          */
852         if (current_is_keventd())
853                 orig = raw_smp_processor_id();
854
855         for_each_online_cpu(cpu) {
856                 struct work_struct *work = per_cpu_ptr(works, cpu);
857
858                 INIT_WORK(work, func);
859                 if (cpu != orig)
860                         schedule_work_on(cpu, work);
861         }
862         if (orig >= 0)
863                 func(per_cpu_ptr(works, orig));
864
865         for_each_online_cpu(cpu)
866                 flush_work(per_cpu_ptr(works, cpu));
867
868         put_online_cpus();
869         free_percpu(works);
870         return 0;
871 }
872
873 /**
874  * flush_scheduled_work - ensure that any scheduled work has run to completion.
875  *
876  * Forces execution of the kernel-global workqueue and blocks until its
877  * completion.
878  *
879  * Think twice before calling this function!  It's very easy to get into
880  * trouble if you don't take great care.  Either of the following situations
881  * will lead to deadlock:
882  *
883  *      One of the work items currently on the workqueue needs to acquire
884  *      a lock held by your code or its caller.
885  *
886  *      Your code is running in the context of a work routine.
887  *
888  * They will be detected by lockdep when they occur, but the first might not
889  * occur very often.  It depends on what work items are on the workqueue and
890  * what locks they need, which you have no control over.
891  *
892  * In most situations flushing the entire workqueue is overkill; you merely
893  * need to know that a particular work item isn't queued and isn't running.
894  * In such cases you should use cancel_delayed_work_sync() or
895  * cancel_work_sync() instead.
896  */
897 void flush_scheduled_work(void)
898 {
899         flush_workqueue(keventd_wq);
900 }
901 EXPORT_SYMBOL(flush_scheduled_work);
902
903 /**
904  * execute_in_process_context - reliably execute the routine with user context
905  * @fn:         the function to execute
906  * @ew:         guaranteed storage for the execute work structure (must
907  *              be available when the work executes)
908  *
909  * Executes the function immediately if process context is available,
910  * otherwise schedules the function for delayed execution.
911  *
912  * Returns:     0 - function was executed
913  *              1 - function was scheduled for execution
914  */
915 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
916 {
917         if (!in_interrupt()) {
918                 fn(&ew->work);
919                 return 0;
920         }
921
922         INIT_WORK(&ew->work, fn);
923         schedule_work(&ew->work);
924
925         return 1;
926 }
927 EXPORT_SYMBOL_GPL(execute_in_process_context);
928
929 int keventd_up(void)
930 {
931         return keventd_wq != NULL;
932 }
933
934 int current_is_keventd(void)
935 {
936         struct cpu_workqueue_struct *cwq;
937         int cpu = raw_smp_processor_id(); /* preempt-safe: keventd is per-cpu */
938         int ret = 0;
939
940         BUG_ON(!keventd_wq);
941
942         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
943         if (current == cwq->thread)
944                 ret = 1;
945
946         return ret;
947
948 }
949
950 static struct cpu_workqueue_struct *
951 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
952 {
953         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
954
955         cwq->wq = wq;
956         spin_lock_init(&cwq->lock);
957         INIT_LIST_HEAD(&cwq->worklist);
958         init_waitqueue_head(&cwq->more_work);
959
960         return cwq;
961 }
962
963 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
964 {
965         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
966         struct workqueue_struct *wq = cwq->wq;
967         const char *fmt = is_wq_single_threaded(wq) ? "%s" : "%s/%d";
968         struct task_struct *p;
969
970         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
971         /*
972          * Nobody can add the work_struct to this cwq,
973          *      if (caller is __create_workqueue)
974          *              nobody should see this wq
975          *      else // caller is CPU_UP_PREPARE
976          *              cpu is not on cpu_online_map
977          * so we can abort safely.
978          */
979         if (IS_ERR(p))
980                 return PTR_ERR(p);
981         if (cwq->wq->rt)
982                 sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
983         cwq->thread = p;
984
985         trace_workqueue_creation(cwq->thread, cpu);
986
987         return 0;
988 }
989
990 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
991 {
992         struct task_struct *p = cwq->thread;
993
994         if (p != NULL) {
995                 if (cpu >= 0)
996                         kthread_bind(p, cpu);
997                 wake_up_process(p);
998         }
999 }
1000
1001 struct workqueue_struct *__create_workqueue_key(const char *name,
1002                                                 int singlethread,
1003                                                 int freezeable,
1004                                                 int rt,
1005                                                 struct lock_class_key *key,
1006                                                 const char *lock_name)
1007 {
1008         struct workqueue_struct *wq;
1009         struct cpu_workqueue_struct *cwq;
1010         int err = 0, cpu;
1011
1012         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
1013         if (!wq)
1014                 return NULL;
1015
1016         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
1017         if (!wq->cpu_wq) {
1018                 kfree(wq);
1019                 return NULL;
1020         }
1021
1022         wq->name = name;
1023         lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
1024         wq->singlethread = singlethread;
1025         wq->freezeable = freezeable;
1026         wq->rt = rt;
1027         INIT_LIST_HEAD(&wq->list);
1028
1029         if (singlethread) {
1030                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
1031                 err = create_workqueue_thread(cwq, singlethread_cpu);
1032                 start_workqueue_thread(cwq, -1);
1033         } else {
1034                 cpu_maps_update_begin();
1035                 /*
1036                  * We must place this wq on list even if the code below fails.
1037                  * cpu_down(cpu) can remove cpu from cpu_populated_map before
1038                  * destroy_workqueue() takes the lock, in that case we leak
1039                  * cwq[cpu]->thread.
1040                  */
1041                 spin_lock(&workqueue_lock);
1042                 list_add(&wq->list, &workqueues);
1043                 spin_unlock(&workqueue_lock);
1044                 /*
1045                  * We must initialize cwqs for each possible cpu even if we
1046                  * are going to call destroy_workqueue() finally. Otherwise
1047                  * cpu_up() can hit the uninitialized cwq once we drop the
1048                  * lock.
1049                  */
1050                 for_each_possible_cpu(cpu) {
1051                         cwq = init_cpu_workqueue(wq, cpu);
1052                         if (err || !cpu_online(cpu))
1053                                 continue;
1054                         err = create_workqueue_thread(cwq, cpu);
1055                         start_workqueue_thread(cwq, cpu);
1056                 }
1057                 cpu_maps_update_done();
1058         }
1059
1060         if (err) {
1061                 destroy_workqueue(wq);
1062                 wq = NULL;
1063         }
1064         return wq;
1065 }
1066 EXPORT_SYMBOL_GPL(__create_workqueue_key);
1067
1068 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq)
1069 {
1070         /*
1071          * Our caller is either destroy_workqueue() or CPU_POST_DEAD,
1072          * cpu_add_remove_lock protects cwq->thread.
1073          */
1074         if (cwq->thread == NULL)
1075                 return;
1076
1077         lock_map_acquire(&cwq->wq->lockdep_map);
1078         lock_map_release(&cwq->wq->lockdep_map);
1079
1080         flush_cpu_workqueue(cwq);
1081         /*
1082          * If the caller is CPU_POST_DEAD and cwq->worklist was not empty,
1083          * a concurrent flush_workqueue() can insert a barrier after us.
1084          * However, in that case run_workqueue() won't return and check
1085          * kthread_should_stop() until it flushes all work_struct's.
1086          * When ->worklist becomes empty it is safe to exit because no
1087          * more work_structs can be queued on this cwq: flush_workqueue
1088          * checks list_empty(), and a "normal" queue_work() can't use
1089          * a dead CPU.
1090          */
1091         trace_workqueue_destruction(cwq->thread);
1092         kthread_stop(cwq->thread);
1093         cwq->thread = NULL;
1094 }
1095
1096 /**
1097  * destroy_workqueue - safely terminate a workqueue
1098  * @wq: target workqueue
1099  *
1100  * Safely destroy a workqueue. All work currently pending will be done first.
1101  */
1102 void destroy_workqueue(struct workqueue_struct *wq)
1103 {
1104         const struct cpumask *cpu_map = wq_cpu_map(wq);
1105         int cpu;
1106
1107         cpu_maps_update_begin();
1108         spin_lock(&workqueue_lock);
1109         list_del(&wq->list);
1110         spin_unlock(&workqueue_lock);
1111
1112         for_each_cpu(cpu, cpu_map)
1113                 cleanup_workqueue_thread(per_cpu_ptr(wq->cpu_wq, cpu));
1114         cpu_maps_update_done();
1115
1116         free_percpu(wq->cpu_wq);
1117         kfree(wq);
1118 }
1119 EXPORT_SYMBOL_GPL(destroy_workqueue);
1120
1121 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
1122                                                 unsigned long action,
1123                                                 void *hcpu)
1124 {
1125         unsigned int cpu = (unsigned long)hcpu;
1126         struct cpu_workqueue_struct *cwq;
1127         struct workqueue_struct *wq;
1128         int err = 0;
1129
1130         action &= ~CPU_TASKS_FROZEN;
1131
1132         switch (action) {
1133         case CPU_UP_PREPARE:
1134                 cpumask_set_cpu(cpu, cpu_populated_map);
1135         }
1136 undo:
1137         list_for_each_entry(wq, &workqueues, list) {
1138                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
1139
1140                 switch (action) {
1141                 case CPU_UP_PREPARE:
1142                         err = create_workqueue_thread(cwq, cpu);
1143                         if (!err)
1144                                 break;
1145                         printk(KERN_ERR "workqueue [%s] for %i failed\n",
1146                                 wq->name, cpu);
1147                         action = CPU_UP_CANCELED;
1148                         err = -ENOMEM;
1149                         goto undo;
1150
1151                 case CPU_ONLINE:
1152                         start_workqueue_thread(cwq, cpu);
1153                         break;
1154
1155                 case CPU_UP_CANCELED:
1156                         start_workqueue_thread(cwq, -1);
1157                 case CPU_POST_DEAD:
1158                         cleanup_workqueue_thread(cwq);
1159                         break;
1160                 }
1161         }
1162
1163         switch (action) {
1164         case CPU_UP_CANCELED:
1165         case CPU_POST_DEAD:
1166                 cpumask_clear_cpu(cpu, cpu_populated_map);
1167         }
1168
1169         return notifier_from_errno(err);
1170 }
1171
1172 #ifdef CONFIG_SMP
1173
1174 struct work_for_cpu {
1175         struct completion completion;
1176         long (*fn)(void *);
1177         void *arg;
1178         long ret;
1179 };
1180
1181 static int do_work_for_cpu(void *_wfc)
1182 {
1183         struct work_for_cpu *wfc = _wfc;
1184         wfc->ret = wfc->fn(wfc->arg);
1185         complete(&wfc->completion);
1186         return 0;
1187 }
1188
1189 /**
1190  * work_on_cpu - run a function in user context on a particular cpu
1191  * @cpu: the cpu to run on
1192  * @fn: the function to run
1193  * @arg: the function arg
1194  *
1195  * This will return the value @fn returns.
1196  * It is up to the caller to ensure that the cpu doesn't go offline.
1197  * The caller must not hold any locks which would prevent @fn from completing.
1198  */
1199 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
1200 {
1201         struct task_struct *sub_thread;
1202         struct work_for_cpu wfc = {
1203                 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
1204                 .fn = fn,
1205                 .arg = arg,
1206         };
1207
1208         sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
1209         if (IS_ERR(sub_thread))
1210                 return PTR_ERR(sub_thread);
1211         kthread_bind(sub_thread, cpu);
1212         wake_up_process(sub_thread);
1213         wait_for_completion(&wfc.completion);
1214         return wfc.ret;
1215 }
1216 EXPORT_SYMBOL_GPL(work_on_cpu);
1217 #endif /* CONFIG_SMP */
1218
1219 void __init init_workqueues(void)
1220 {
1221         alloc_cpumask_var(&cpu_populated_map, GFP_KERNEL);
1222
1223         cpumask_copy(cpu_populated_map, cpu_online_mask);
1224         singlethread_cpu = cpumask_first(cpu_possible_mask);
1225         cpu_singlethread_map = cpumask_of(singlethread_cpu);
1226         hotcpu_notifier(workqueue_cpu_callback, 0);
1227         keventd_wq = create_workqueue("events");
1228         BUG_ON(!keventd_wq);
1229 }