workqueue: merge feature parameters into flags
[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  * Structure fields follow one of the following exclusion rules.
41  *
42  * I: Set during initialization and read-only afterwards.
43  *
44  * L: cwq->lock protected.  Access with cwq->lock held.
45  *
46  * W: workqueue_lock protected.
47  */
48
49 /*
50  * The per-CPU workqueue (if single thread, we always use the first
51  * possible cpu).
52  */
53 struct cpu_workqueue_struct {
54
55         spinlock_t lock;
56
57         struct list_head worklist;
58         wait_queue_head_t more_work;
59         struct work_struct *current_work;
60
61         struct workqueue_struct *wq;            /* I: the owning workqueue */
62         struct task_struct      *thread;
63 } ____cacheline_aligned;
64
65 /*
66  * The externally visible workqueue abstraction is an array of
67  * per-CPU workqueues:
68  */
69 struct workqueue_struct {
70         unsigned int            flags;          /* I: WQ_* flags */
71         struct cpu_workqueue_struct *cpu_wq;    /* I: cwq's */
72         struct list_head        list;           /* W: list of all workqueues */
73         const char              *name;          /* I: workqueue name */
74 #ifdef CONFIG_LOCKDEP
75         struct lockdep_map      lockdep_map;
76 #endif
77 };
78
79 #ifdef CONFIG_DEBUG_OBJECTS_WORK
80
81 static struct debug_obj_descr work_debug_descr;
82
83 /*
84  * fixup_init is called when:
85  * - an active object is initialized
86  */
87 static int work_fixup_init(void *addr, enum debug_obj_state state)
88 {
89         struct work_struct *work = addr;
90
91         switch (state) {
92         case ODEBUG_STATE_ACTIVE:
93                 cancel_work_sync(work);
94                 debug_object_init(work, &work_debug_descr);
95                 return 1;
96         default:
97                 return 0;
98         }
99 }
100
101 /*
102  * fixup_activate is called when:
103  * - an active object is activated
104  * - an unknown object is activated (might be a statically initialized object)
105  */
106 static int work_fixup_activate(void *addr, enum debug_obj_state state)
107 {
108         struct work_struct *work = addr;
109
110         switch (state) {
111
112         case ODEBUG_STATE_NOTAVAILABLE:
113                 /*
114                  * This is not really a fixup. The work struct was
115                  * statically initialized. We just make sure that it
116                  * is tracked in the object tracker.
117                  */
118                 if (test_bit(WORK_STRUCT_STATIC, work_data_bits(work))) {
119                         debug_object_init(work, &work_debug_descr);
120                         debug_object_activate(work, &work_debug_descr);
121                         return 0;
122                 }
123                 WARN_ON_ONCE(1);
124                 return 0;
125
126         case ODEBUG_STATE_ACTIVE:
127                 WARN_ON(1);
128
129         default:
130                 return 0;
131         }
132 }
133
134 /*
135  * fixup_free is called when:
136  * - an active object is freed
137  */
138 static int work_fixup_free(void *addr, enum debug_obj_state state)
139 {
140         struct work_struct *work = addr;
141
142         switch (state) {
143         case ODEBUG_STATE_ACTIVE:
144                 cancel_work_sync(work);
145                 debug_object_free(work, &work_debug_descr);
146                 return 1;
147         default:
148                 return 0;
149         }
150 }
151
152 static struct debug_obj_descr work_debug_descr = {
153         .name           = "work_struct",
154         .fixup_init     = work_fixup_init,
155         .fixup_activate = work_fixup_activate,
156         .fixup_free     = work_fixup_free,
157 };
158
159 static inline void debug_work_activate(struct work_struct *work)
160 {
161         debug_object_activate(work, &work_debug_descr);
162 }
163
164 static inline void debug_work_deactivate(struct work_struct *work)
165 {
166         debug_object_deactivate(work, &work_debug_descr);
167 }
168
169 void __init_work(struct work_struct *work, int onstack)
170 {
171         if (onstack)
172                 debug_object_init_on_stack(work, &work_debug_descr);
173         else
174                 debug_object_init(work, &work_debug_descr);
175 }
176 EXPORT_SYMBOL_GPL(__init_work);
177
178 void destroy_work_on_stack(struct work_struct *work)
179 {
180         debug_object_free(work, &work_debug_descr);
181 }
182 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
183
184 #else
185 static inline void debug_work_activate(struct work_struct *work) { }
186 static inline void debug_work_deactivate(struct work_struct *work) { }
187 #endif
188
189 /* Serializes the accesses to the list of workqueues. */
190 static DEFINE_SPINLOCK(workqueue_lock);
191 static LIST_HEAD(workqueues);
192
193 static int singlethread_cpu __read_mostly;
194 static const struct cpumask *cpu_singlethread_map __read_mostly;
195 /*
196  * _cpu_down() first removes CPU from cpu_online_map, then CPU_DEAD
197  * flushes cwq->worklist. This means that flush_workqueue/wait_on_work
198  * which comes in between can't use for_each_online_cpu(). We could
199  * use cpu_possible_map, the cpumask below is more a documentation
200  * than optimization.
201  */
202 static cpumask_var_t cpu_populated_map __read_mostly;
203
204 /* If it's single threaded, it isn't in the list of workqueues. */
205 static inline bool is_wq_single_threaded(struct workqueue_struct *wq)
206 {
207         return wq->flags & WQ_SINGLE_THREAD;
208 }
209
210 static const struct cpumask *wq_cpu_map(struct workqueue_struct *wq)
211 {
212         return is_wq_single_threaded(wq)
213                 ? cpu_singlethread_map : cpu_populated_map;
214 }
215
216 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
217                                             struct workqueue_struct *wq)
218 {
219         if (unlikely(is_wq_single_threaded(wq)))
220                 cpu = singlethread_cpu;
221         return per_cpu_ptr(wq->cpu_wq, cpu);
222 }
223
224 /*
225  * Set the workqueue on which a work item is to be run
226  * - Must *only* be called if the pending flag is set
227  */
228 static inline void set_wq_data(struct work_struct *work,
229                                struct cpu_workqueue_struct *cwq,
230                                unsigned long extra_flags)
231 {
232         BUG_ON(!work_pending(work));
233
234         atomic_long_set(&work->data, (unsigned long)cwq | work_static(work) |
235                         (1UL << WORK_STRUCT_PENDING) | extra_flags);
236 }
237
238 /*
239  * Clear WORK_STRUCT_PENDING and the workqueue on which it was queued.
240  */
241 static inline void clear_wq_data(struct work_struct *work)
242 {
243         atomic_long_set(&work->data, work_static(work));
244 }
245
246 static inline
247 struct cpu_workqueue_struct *get_wq_data(struct work_struct *work)
248 {
249         return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
250 }
251
252 /**
253  * insert_work - insert a work into cwq
254  * @cwq: cwq @work belongs to
255  * @work: work to insert
256  * @head: insertion point
257  * @extra_flags: extra WORK_STRUCT_* flags to set
258  *
259  * Insert @work into @cwq after @head.
260  *
261  * CONTEXT:
262  * spin_lock_irq(cwq->lock).
263  */
264 static void insert_work(struct cpu_workqueue_struct *cwq,
265                         struct work_struct *work, struct list_head *head,
266                         unsigned int extra_flags)
267 {
268         trace_workqueue_insertion(cwq->thread, work);
269
270         /* we own @work, set data and link */
271         set_wq_data(work, cwq, extra_flags);
272
273         /*
274          * Ensure that we get the right work->data if we see the
275          * result of list_add() below, see try_to_grab_pending().
276          */
277         smp_wmb();
278
279         list_add_tail(&work->entry, head);
280         wake_up(&cwq->more_work);
281 }
282
283 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
284                          struct work_struct *work)
285 {
286         struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
287         unsigned long flags;
288
289         debug_work_activate(work);
290         spin_lock_irqsave(&cwq->lock, flags);
291         BUG_ON(!list_empty(&work->entry));
292         insert_work(cwq, work, &cwq->worklist, 0);
293         spin_unlock_irqrestore(&cwq->lock, flags);
294 }
295
296 /**
297  * queue_work - queue work on a workqueue
298  * @wq: workqueue to use
299  * @work: work to queue
300  *
301  * Returns 0 if @work was already on a queue, non-zero otherwise.
302  *
303  * We queue the work to the CPU on which it was submitted, but if the CPU dies
304  * it can be processed by another CPU.
305  */
306 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
307 {
308         int ret;
309
310         ret = queue_work_on(get_cpu(), wq, work);
311         put_cpu();
312
313         return ret;
314 }
315 EXPORT_SYMBOL_GPL(queue_work);
316
317 /**
318  * queue_work_on - queue work on specific cpu
319  * @cpu: CPU number to execute work on
320  * @wq: workqueue to use
321  * @work: work to queue
322  *
323  * Returns 0 if @work was already on a queue, non-zero otherwise.
324  *
325  * We queue the work to a specific CPU, the caller must ensure it
326  * can't go away.
327  */
328 int
329 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
330 {
331         int ret = 0;
332
333         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
334                 __queue_work(cpu, wq, work);
335                 ret = 1;
336         }
337         return ret;
338 }
339 EXPORT_SYMBOL_GPL(queue_work_on);
340
341 static void delayed_work_timer_fn(unsigned long __data)
342 {
343         struct delayed_work *dwork = (struct delayed_work *)__data;
344         struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
345
346         __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
347 }
348
349 /**
350  * queue_delayed_work - queue work on a workqueue after delay
351  * @wq: workqueue to use
352  * @dwork: delayable work to queue
353  * @delay: number of jiffies to wait before queueing
354  *
355  * Returns 0 if @work was already on a queue, non-zero otherwise.
356  */
357 int queue_delayed_work(struct workqueue_struct *wq,
358                         struct delayed_work *dwork, unsigned long delay)
359 {
360         if (delay == 0)
361                 return queue_work(wq, &dwork->work);
362
363         return queue_delayed_work_on(-1, wq, dwork, delay);
364 }
365 EXPORT_SYMBOL_GPL(queue_delayed_work);
366
367 /**
368  * queue_delayed_work_on - queue work on specific CPU after delay
369  * @cpu: CPU number to execute work on
370  * @wq: workqueue to use
371  * @dwork: work to queue
372  * @delay: number of jiffies to wait before queueing
373  *
374  * Returns 0 if @work was already on a queue, non-zero otherwise.
375  */
376 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
377                         struct delayed_work *dwork, unsigned long delay)
378 {
379         int ret = 0;
380         struct timer_list *timer = &dwork->timer;
381         struct work_struct *work = &dwork->work;
382
383         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
384                 BUG_ON(timer_pending(timer));
385                 BUG_ON(!list_empty(&work->entry));
386
387                 timer_stats_timer_set_start_info(&dwork->timer);
388
389                 /* This stores cwq for the moment, for the timer_fn */
390                 set_wq_data(work, get_cwq(raw_smp_processor_id(), wq), 0);
391                 timer->expires = jiffies + delay;
392                 timer->data = (unsigned long)dwork;
393                 timer->function = delayed_work_timer_fn;
394
395                 if (unlikely(cpu >= 0))
396                         add_timer_on(timer, cpu);
397                 else
398                         add_timer(timer);
399                 ret = 1;
400         }
401         return ret;
402 }
403 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
404
405 static void run_workqueue(struct cpu_workqueue_struct *cwq)
406 {
407         spin_lock_irq(&cwq->lock);
408         while (!list_empty(&cwq->worklist)) {
409                 struct work_struct *work = list_entry(cwq->worklist.next,
410                                                 struct work_struct, entry);
411                 work_func_t f = work->func;
412 #ifdef CONFIG_LOCKDEP
413                 /*
414                  * It is permissible to free the struct work_struct
415                  * from inside the function that is called from it,
416                  * this we need to take into account for lockdep too.
417                  * To avoid bogus "held lock freed" warnings as well
418                  * as problems when looking into work->lockdep_map,
419                  * make a copy and use that here.
420                  */
421                 struct lockdep_map lockdep_map = work->lockdep_map;
422 #endif
423                 trace_workqueue_execution(cwq->thread, work);
424                 debug_work_deactivate(work);
425                 cwq->current_work = work;
426                 list_del_init(cwq->worklist.next);
427                 spin_unlock_irq(&cwq->lock);
428
429                 BUG_ON(get_wq_data(work) != cwq);
430                 work_clear_pending(work);
431                 lock_map_acquire(&cwq->wq->lockdep_map);
432                 lock_map_acquire(&lockdep_map);
433                 f(work);
434                 lock_map_release(&lockdep_map);
435                 lock_map_release(&cwq->wq->lockdep_map);
436
437                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
438                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
439                                         "%s/0x%08x/%d\n",
440                                         current->comm, preempt_count(),
441                                         task_pid_nr(current));
442                         printk(KERN_ERR "    last function: ");
443                         print_symbol("%s\n", (unsigned long)f);
444                         debug_show_held_locks(current);
445                         dump_stack();
446                 }
447
448                 spin_lock_irq(&cwq->lock);
449                 cwq->current_work = NULL;
450         }
451         spin_unlock_irq(&cwq->lock);
452 }
453
454 /**
455  * worker_thread - the worker thread function
456  * @__cwq: cwq to serve
457  *
458  * The cwq worker thread function.
459  */
460 static int worker_thread(void *__cwq)
461 {
462         struct cpu_workqueue_struct *cwq = __cwq;
463         DEFINE_WAIT(wait);
464
465         if (cwq->wq->flags & WQ_FREEZEABLE)
466                 set_freezable();
467
468         for (;;) {
469                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
470                 if (!freezing(current) &&
471                     !kthread_should_stop() &&
472                     list_empty(&cwq->worklist))
473                         schedule();
474                 finish_wait(&cwq->more_work, &wait);
475
476                 try_to_freeze();
477
478                 if (kthread_should_stop())
479                         break;
480
481                 run_workqueue(cwq);
482         }
483
484         return 0;
485 }
486
487 struct wq_barrier {
488         struct work_struct      work;
489         struct completion       done;
490 };
491
492 static void wq_barrier_func(struct work_struct *work)
493 {
494         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
495         complete(&barr->done);
496 }
497
498 /**
499  * insert_wq_barrier - insert a barrier work
500  * @cwq: cwq to insert barrier into
501  * @barr: wq_barrier to insert
502  * @head: insertion point
503  *
504  * Insert barrier @barr into @cwq before @head.
505  *
506  * CONTEXT:
507  * spin_lock_irq(cwq->lock).
508  */
509 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
510                         struct wq_barrier *barr, struct list_head *head)
511 {
512         /*
513          * debugobject calls are safe here even with cwq->lock locked
514          * as we know for sure that this will not trigger any of the
515          * checks and call back into the fixup functions where we
516          * might deadlock.
517          */
518         INIT_WORK_ON_STACK(&barr->work, wq_barrier_func);
519         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
520         init_completion(&barr->done);
521
522         debug_work_activate(&barr->work);
523         insert_work(cwq, &barr->work, head, 0);
524 }
525
526 static int flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
527 {
528         int active = 0;
529         struct wq_barrier barr;
530
531         WARN_ON(cwq->thread == current);
532
533         spin_lock_irq(&cwq->lock);
534         if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
535                 insert_wq_barrier(cwq, &barr, &cwq->worklist);
536                 active = 1;
537         }
538         spin_unlock_irq(&cwq->lock);
539
540         if (active) {
541                 wait_for_completion(&barr.done);
542                 destroy_work_on_stack(&barr.work);
543         }
544
545         return active;
546 }
547
548 /**
549  * flush_workqueue - ensure that any scheduled work has run to completion.
550  * @wq: workqueue to flush
551  *
552  * Forces execution of the workqueue and blocks until its completion.
553  * This is typically used in driver shutdown handlers.
554  *
555  * We sleep until all works which were queued on entry have been handled,
556  * but we are not livelocked by new incoming ones.
557  */
558 void flush_workqueue(struct workqueue_struct *wq)
559 {
560         const struct cpumask *cpu_map = wq_cpu_map(wq);
561         int cpu;
562
563         might_sleep();
564         lock_map_acquire(&wq->lockdep_map);
565         lock_map_release(&wq->lockdep_map);
566         for_each_cpu(cpu, cpu_map)
567                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
568 }
569 EXPORT_SYMBOL_GPL(flush_workqueue);
570
571 /**
572  * flush_work - block until a work_struct's callback has terminated
573  * @work: the work which is to be flushed
574  *
575  * Returns false if @work has already terminated.
576  *
577  * It is expected that, prior to calling flush_work(), the caller has
578  * arranged for the work to not be requeued, otherwise it doesn't make
579  * sense to use this function.
580  */
581 int flush_work(struct work_struct *work)
582 {
583         struct cpu_workqueue_struct *cwq;
584         struct list_head *prev;
585         struct wq_barrier barr;
586
587         might_sleep();
588         cwq = get_wq_data(work);
589         if (!cwq)
590                 return 0;
591
592         lock_map_acquire(&cwq->wq->lockdep_map);
593         lock_map_release(&cwq->wq->lockdep_map);
594
595         spin_lock_irq(&cwq->lock);
596         if (!list_empty(&work->entry)) {
597                 /*
598                  * See the comment near try_to_grab_pending()->smp_rmb().
599                  * If it was re-queued under us we are not going to wait.
600                  */
601                 smp_rmb();
602                 if (unlikely(cwq != get_wq_data(work)))
603                         goto already_gone;
604                 prev = &work->entry;
605         } else {
606                 if (cwq->current_work != work)
607                         goto already_gone;
608                 prev = &cwq->worklist;
609         }
610         insert_wq_barrier(cwq, &barr, prev->next);
611
612         spin_unlock_irq(&cwq->lock);
613         wait_for_completion(&barr.done);
614         destroy_work_on_stack(&barr.work);
615         return 1;
616 already_gone:
617         spin_unlock_irq(&cwq->lock);
618         return 0;
619 }
620 EXPORT_SYMBOL_GPL(flush_work);
621
622 /*
623  * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
624  * so this work can't be re-armed in any way.
625  */
626 static int try_to_grab_pending(struct work_struct *work)
627 {
628         struct cpu_workqueue_struct *cwq;
629         int ret = -1;
630
631         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work)))
632                 return 0;
633
634         /*
635          * The queueing is in progress, or it is already queued. Try to
636          * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
637          */
638
639         cwq = get_wq_data(work);
640         if (!cwq)
641                 return ret;
642
643         spin_lock_irq(&cwq->lock);
644         if (!list_empty(&work->entry)) {
645                 /*
646                  * This work is queued, but perhaps we locked the wrong cwq.
647                  * In that case we must see the new value after rmb(), see
648                  * insert_work()->wmb().
649                  */
650                 smp_rmb();
651                 if (cwq == get_wq_data(work)) {
652                         debug_work_deactivate(work);
653                         list_del_init(&work->entry);
654                         ret = 1;
655                 }
656         }
657         spin_unlock_irq(&cwq->lock);
658
659         return ret;
660 }
661
662 static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq,
663                                 struct work_struct *work)
664 {
665         struct wq_barrier barr;
666         int running = 0;
667
668         spin_lock_irq(&cwq->lock);
669         if (unlikely(cwq->current_work == work)) {
670                 insert_wq_barrier(cwq, &barr, cwq->worklist.next);
671                 running = 1;
672         }
673         spin_unlock_irq(&cwq->lock);
674
675         if (unlikely(running)) {
676                 wait_for_completion(&barr.done);
677                 destroy_work_on_stack(&barr.work);
678         }
679 }
680
681 static void wait_on_work(struct work_struct *work)
682 {
683         struct cpu_workqueue_struct *cwq;
684         struct workqueue_struct *wq;
685         const struct cpumask *cpu_map;
686         int cpu;
687
688         might_sleep();
689
690         lock_map_acquire(&work->lockdep_map);
691         lock_map_release(&work->lockdep_map);
692
693         cwq = get_wq_data(work);
694         if (!cwq)
695                 return;
696
697         wq = cwq->wq;
698         cpu_map = wq_cpu_map(wq);
699
700         for_each_cpu(cpu, cpu_map)
701                 wait_on_cpu_work(get_cwq(cpu, wq), work);
702 }
703
704 static int __cancel_work_timer(struct work_struct *work,
705                                 struct timer_list* timer)
706 {
707         int ret;
708
709         do {
710                 ret = (timer && likely(del_timer(timer)));
711                 if (!ret)
712                         ret = try_to_grab_pending(work);
713                 wait_on_work(work);
714         } while (unlikely(ret < 0));
715
716         clear_wq_data(work);
717         return ret;
718 }
719
720 /**
721  * cancel_work_sync - block until a work_struct's callback has terminated
722  * @work: the work which is to be flushed
723  *
724  * Returns true if @work was pending.
725  *
726  * cancel_work_sync() will cancel the work if it is queued. If the work's
727  * callback appears to be running, cancel_work_sync() will block until it
728  * has completed.
729  *
730  * It is possible to use this function if the work re-queues itself. It can
731  * cancel the work even if it migrates to another workqueue, however in that
732  * case it only guarantees that work->func() has completed on the last queued
733  * workqueue.
734  *
735  * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
736  * pending, otherwise it goes into a busy-wait loop until the timer expires.
737  *
738  * The caller must ensure that workqueue_struct on which this work was last
739  * queued can't be destroyed before this function returns.
740  */
741 int cancel_work_sync(struct work_struct *work)
742 {
743         return __cancel_work_timer(work, NULL);
744 }
745 EXPORT_SYMBOL_GPL(cancel_work_sync);
746
747 /**
748  * cancel_delayed_work_sync - reliably kill off a delayed work.
749  * @dwork: the delayed work struct
750  *
751  * Returns true if @dwork was pending.
752  *
753  * It is possible to use this function if @dwork rearms itself via queue_work()
754  * or queue_delayed_work(). See also the comment for cancel_work_sync().
755  */
756 int cancel_delayed_work_sync(struct delayed_work *dwork)
757 {
758         return __cancel_work_timer(&dwork->work, &dwork->timer);
759 }
760 EXPORT_SYMBOL(cancel_delayed_work_sync);
761
762 static struct workqueue_struct *keventd_wq __read_mostly;
763
764 /**
765  * schedule_work - put work task in global workqueue
766  * @work: job to be done
767  *
768  * Returns zero if @work was already on the kernel-global workqueue and
769  * non-zero otherwise.
770  *
771  * This puts a job in the kernel-global workqueue if it was not already
772  * queued and leaves it in the same position on the kernel-global
773  * workqueue otherwise.
774  */
775 int schedule_work(struct work_struct *work)
776 {
777         return queue_work(keventd_wq, work);
778 }
779 EXPORT_SYMBOL(schedule_work);
780
781 /*
782  * schedule_work_on - put work task on a specific cpu
783  * @cpu: cpu to put the work task on
784  * @work: job to be done
785  *
786  * This puts a job on a specific cpu
787  */
788 int schedule_work_on(int cpu, struct work_struct *work)
789 {
790         return queue_work_on(cpu, keventd_wq, work);
791 }
792 EXPORT_SYMBOL(schedule_work_on);
793
794 /**
795  * schedule_delayed_work - put work task in global workqueue after delay
796  * @dwork: job to be done
797  * @delay: number of jiffies to wait or 0 for immediate execution
798  *
799  * After waiting for a given time this puts a job in the kernel-global
800  * workqueue.
801  */
802 int schedule_delayed_work(struct delayed_work *dwork,
803                                         unsigned long delay)
804 {
805         return queue_delayed_work(keventd_wq, dwork, delay);
806 }
807 EXPORT_SYMBOL(schedule_delayed_work);
808
809 /**
810  * flush_delayed_work - block until a dwork_struct's callback has terminated
811  * @dwork: the delayed work which is to be flushed
812  *
813  * Any timeout is cancelled, and any pending work is run immediately.
814  */
815 void flush_delayed_work(struct delayed_work *dwork)
816 {
817         if (del_timer_sync(&dwork->timer)) {
818                 __queue_work(get_cpu(), get_wq_data(&dwork->work)->wq,
819                              &dwork->work);
820                 put_cpu();
821         }
822         flush_work(&dwork->work);
823 }
824 EXPORT_SYMBOL(flush_delayed_work);
825
826 /**
827  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
828  * @cpu: cpu to use
829  * @dwork: job to be done
830  * @delay: number of jiffies to wait
831  *
832  * After waiting for a given time this puts a job in the kernel-global
833  * workqueue on the specified CPU.
834  */
835 int schedule_delayed_work_on(int cpu,
836                         struct delayed_work *dwork, unsigned long delay)
837 {
838         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
839 }
840 EXPORT_SYMBOL(schedule_delayed_work_on);
841
842 /**
843  * schedule_on_each_cpu - call a function on each online CPU from keventd
844  * @func: the function to call
845  *
846  * Returns zero on success.
847  * Returns -ve errno on failure.
848  *
849  * schedule_on_each_cpu() is very slow.
850  */
851 int schedule_on_each_cpu(work_func_t func)
852 {
853         int cpu;
854         int orig = -1;
855         struct work_struct *works;
856
857         works = alloc_percpu(struct work_struct);
858         if (!works)
859                 return -ENOMEM;
860
861         get_online_cpus();
862
863         /*
864          * When running in keventd don't schedule a work item on
865          * itself.  Can just call directly because the work queue is
866          * already bound.  This also is faster.
867          */
868         if (current_is_keventd())
869                 orig = raw_smp_processor_id();
870
871         for_each_online_cpu(cpu) {
872                 struct work_struct *work = per_cpu_ptr(works, cpu);
873
874                 INIT_WORK(work, func);
875                 if (cpu != orig)
876                         schedule_work_on(cpu, work);
877         }
878         if (orig >= 0)
879                 func(per_cpu_ptr(works, orig));
880
881         for_each_online_cpu(cpu)
882                 flush_work(per_cpu_ptr(works, cpu));
883
884         put_online_cpus();
885         free_percpu(works);
886         return 0;
887 }
888
889 /**
890  * flush_scheduled_work - ensure that any scheduled work has run to completion.
891  *
892  * Forces execution of the kernel-global workqueue and blocks until its
893  * completion.
894  *
895  * Think twice before calling this function!  It's very easy to get into
896  * trouble if you don't take great care.  Either of the following situations
897  * will lead to deadlock:
898  *
899  *      One of the work items currently on the workqueue needs to acquire
900  *      a lock held by your code or its caller.
901  *
902  *      Your code is running in the context of a work routine.
903  *
904  * They will be detected by lockdep when they occur, but the first might not
905  * occur very often.  It depends on what work items are on the workqueue and
906  * what locks they need, which you have no control over.
907  *
908  * In most situations flushing the entire workqueue is overkill; you merely
909  * need to know that a particular work item isn't queued and isn't running.
910  * In such cases you should use cancel_delayed_work_sync() or
911  * cancel_work_sync() instead.
912  */
913 void flush_scheduled_work(void)
914 {
915         flush_workqueue(keventd_wq);
916 }
917 EXPORT_SYMBOL(flush_scheduled_work);
918
919 /**
920  * execute_in_process_context - reliably execute the routine with user context
921  * @fn:         the function to execute
922  * @ew:         guaranteed storage for the execute work structure (must
923  *              be available when the work executes)
924  *
925  * Executes the function immediately if process context is available,
926  * otherwise schedules the function for delayed execution.
927  *
928  * Returns:     0 - function was executed
929  *              1 - function was scheduled for execution
930  */
931 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
932 {
933         if (!in_interrupt()) {
934                 fn(&ew->work);
935                 return 0;
936         }
937
938         INIT_WORK(&ew->work, fn);
939         schedule_work(&ew->work);
940
941         return 1;
942 }
943 EXPORT_SYMBOL_GPL(execute_in_process_context);
944
945 int keventd_up(void)
946 {
947         return keventd_wq != NULL;
948 }
949
950 int current_is_keventd(void)
951 {
952         struct cpu_workqueue_struct *cwq;
953         int cpu = raw_smp_processor_id(); /* preempt-safe: keventd is per-cpu */
954         int ret = 0;
955
956         BUG_ON(!keventd_wq);
957
958         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
959         if (current == cwq->thread)
960                 ret = 1;
961
962         return ret;
963
964 }
965
966 static struct cpu_workqueue_struct *
967 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
968 {
969         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
970
971         cwq->wq = wq;
972         spin_lock_init(&cwq->lock);
973         INIT_LIST_HEAD(&cwq->worklist);
974         init_waitqueue_head(&cwq->more_work);
975
976         return cwq;
977 }
978
979 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
980 {
981         struct workqueue_struct *wq = cwq->wq;
982         const char *fmt = is_wq_single_threaded(wq) ? "%s" : "%s/%d";
983         struct task_struct *p;
984
985         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
986         /*
987          * Nobody can add the work_struct to this cwq,
988          *      if (caller is __create_workqueue)
989          *              nobody should see this wq
990          *      else // caller is CPU_UP_PREPARE
991          *              cpu is not on cpu_online_map
992          * so we can abort safely.
993          */
994         if (IS_ERR(p))
995                 return PTR_ERR(p);
996         cwq->thread = p;
997
998         trace_workqueue_creation(cwq->thread, cpu);
999
1000         return 0;
1001 }
1002
1003 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
1004 {
1005         struct task_struct *p = cwq->thread;
1006
1007         if (p != NULL) {
1008                 if (cpu >= 0)
1009                         kthread_bind(p, cpu);
1010                 wake_up_process(p);
1011         }
1012 }
1013
1014 struct workqueue_struct *__create_workqueue_key(const char *name,
1015                                                 unsigned int flags,
1016                                                 struct lock_class_key *key,
1017                                                 const char *lock_name)
1018 {
1019         struct workqueue_struct *wq;
1020         struct cpu_workqueue_struct *cwq;
1021         int err = 0, cpu;
1022
1023         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
1024         if (!wq)
1025                 goto err;
1026
1027         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
1028         if (!wq->cpu_wq)
1029                 goto err;
1030
1031         wq->flags = flags;
1032         wq->name = name;
1033         lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
1034         INIT_LIST_HEAD(&wq->list);
1035
1036         if (flags & WQ_SINGLE_THREAD) {
1037                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
1038                 err = create_workqueue_thread(cwq, singlethread_cpu);
1039                 start_workqueue_thread(cwq, -1);
1040         } else {
1041                 cpu_maps_update_begin();
1042                 /*
1043                  * We must place this wq on list even if the code below fails.
1044                  * cpu_down(cpu) can remove cpu from cpu_populated_map before
1045                  * destroy_workqueue() takes the lock, in that case we leak
1046                  * cwq[cpu]->thread.
1047                  */
1048                 spin_lock(&workqueue_lock);
1049                 list_add(&wq->list, &workqueues);
1050                 spin_unlock(&workqueue_lock);
1051                 /*
1052                  * We must initialize cwqs for each possible cpu even if we
1053                  * are going to call destroy_workqueue() finally. Otherwise
1054                  * cpu_up() can hit the uninitialized cwq once we drop the
1055                  * lock.
1056                  */
1057                 for_each_possible_cpu(cpu) {
1058                         cwq = init_cpu_workqueue(wq, cpu);
1059                         if (err || !cpu_online(cpu))
1060                                 continue;
1061                         err = create_workqueue_thread(cwq, cpu);
1062                         start_workqueue_thread(cwq, cpu);
1063                 }
1064                 cpu_maps_update_done();
1065         }
1066
1067         if (err) {
1068                 destroy_workqueue(wq);
1069                 wq = NULL;
1070         }
1071         return wq;
1072 err:
1073         if (wq) {
1074                 free_percpu(wq->cpu_wq);
1075                 kfree(wq);
1076         }
1077         return NULL;
1078 }
1079 EXPORT_SYMBOL_GPL(__create_workqueue_key);
1080
1081 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq)
1082 {
1083         /*
1084          * Our caller is either destroy_workqueue() or CPU_POST_DEAD,
1085          * cpu_add_remove_lock protects cwq->thread.
1086          */
1087         if (cwq->thread == NULL)
1088                 return;
1089
1090         lock_map_acquire(&cwq->wq->lockdep_map);
1091         lock_map_release(&cwq->wq->lockdep_map);
1092
1093         flush_cpu_workqueue(cwq);
1094         /*
1095          * If the caller is CPU_POST_DEAD and cwq->worklist was not empty,
1096          * a concurrent flush_workqueue() can insert a barrier after us.
1097          * However, in that case run_workqueue() won't return and check
1098          * kthread_should_stop() until it flushes all work_struct's.
1099          * When ->worklist becomes empty it is safe to exit because no
1100          * more work_structs can be queued on this cwq: flush_workqueue
1101          * checks list_empty(), and a "normal" queue_work() can't use
1102          * a dead CPU.
1103          */
1104         trace_workqueue_destruction(cwq->thread);
1105         kthread_stop(cwq->thread);
1106         cwq->thread = NULL;
1107 }
1108
1109 /**
1110  * destroy_workqueue - safely terminate a workqueue
1111  * @wq: target workqueue
1112  *
1113  * Safely destroy a workqueue. All work currently pending will be done first.
1114  */
1115 void destroy_workqueue(struct workqueue_struct *wq)
1116 {
1117         const struct cpumask *cpu_map = wq_cpu_map(wq);
1118         int cpu;
1119
1120         cpu_maps_update_begin();
1121         spin_lock(&workqueue_lock);
1122         list_del(&wq->list);
1123         spin_unlock(&workqueue_lock);
1124
1125         for_each_cpu(cpu, cpu_map)
1126                 cleanup_workqueue_thread(per_cpu_ptr(wq->cpu_wq, cpu));
1127         cpu_maps_update_done();
1128
1129         free_percpu(wq->cpu_wq);
1130         kfree(wq);
1131 }
1132 EXPORT_SYMBOL_GPL(destroy_workqueue);
1133
1134 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
1135                                                 unsigned long action,
1136                                                 void *hcpu)
1137 {
1138         unsigned int cpu = (unsigned long)hcpu;
1139         struct cpu_workqueue_struct *cwq;
1140         struct workqueue_struct *wq;
1141         int err = 0;
1142
1143         action &= ~CPU_TASKS_FROZEN;
1144
1145         switch (action) {
1146         case CPU_UP_PREPARE:
1147                 cpumask_set_cpu(cpu, cpu_populated_map);
1148         }
1149 undo:
1150         list_for_each_entry(wq, &workqueues, list) {
1151                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
1152
1153                 switch (action) {
1154                 case CPU_UP_PREPARE:
1155                         err = create_workqueue_thread(cwq, cpu);
1156                         if (!err)
1157                                 break;
1158                         printk(KERN_ERR "workqueue [%s] for %i failed\n",
1159                                 wq->name, cpu);
1160                         action = CPU_UP_CANCELED;
1161                         err = -ENOMEM;
1162                         goto undo;
1163
1164                 case CPU_ONLINE:
1165                         start_workqueue_thread(cwq, cpu);
1166                         break;
1167
1168                 case CPU_UP_CANCELED:
1169                         start_workqueue_thread(cwq, -1);
1170                 case CPU_POST_DEAD:
1171                         cleanup_workqueue_thread(cwq);
1172                         break;
1173                 }
1174         }
1175
1176         switch (action) {
1177         case CPU_UP_CANCELED:
1178         case CPU_POST_DEAD:
1179                 cpumask_clear_cpu(cpu, cpu_populated_map);
1180         }
1181
1182         return notifier_from_errno(err);
1183 }
1184
1185 #ifdef CONFIG_SMP
1186
1187 struct work_for_cpu {
1188         struct completion completion;
1189         long (*fn)(void *);
1190         void *arg;
1191         long ret;
1192 };
1193
1194 static int do_work_for_cpu(void *_wfc)
1195 {
1196         struct work_for_cpu *wfc = _wfc;
1197         wfc->ret = wfc->fn(wfc->arg);
1198         complete(&wfc->completion);
1199         return 0;
1200 }
1201
1202 /**
1203  * work_on_cpu - run a function in user context on a particular cpu
1204  * @cpu: the cpu to run on
1205  * @fn: the function to run
1206  * @arg: the function arg
1207  *
1208  * This will return the value @fn returns.
1209  * It is up to the caller to ensure that the cpu doesn't go offline.
1210  * The caller must not hold any locks which would prevent @fn from completing.
1211  */
1212 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
1213 {
1214         struct task_struct *sub_thread;
1215         struct work_for_cpu wfc = {
1216                 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
1217                 .fn = fn,
1218                 .arg = arg,
1219         };
1220
1221         sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
1222         if (IS_ERR(sub_thread))
1223                 return PTR_ERR(sub_thread);
1224         kthread_bind(sub_thread, cpu);
1225         wake_up_process(sub_thread);
1226         wait_for_completion(&wfc.completion);
1227         return wfc.ret;
1228 }
1229 EXPORT_SYMBOL_GPL(work_on_cpu);
1230 #endif /* CONFIG_SMP */
1231
1232 void __init init_workqueues(void)
1233 {
1234         alloc_cpumask_var(&cpu_populated_map, GFP_KERNEL);
1235
1236         cpumask_copy(cpu_populated_map, cpu_online_mask);
1237         singlethread_cpu = cpumask_first(cpu_possible_mask);
1238         cpu_singlethread_map = cpumask_of(singlethread_cpu);
1239         hotcpu_notifier(workqueue_cpu_callback, 0);
1240         keventd_wq = create_workqueue("events");
1241         BUG_ON(!keventd_wq);
1242 }