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