workqueues: lockdep annotations for flush_work()
[linux-2.6.git] / kernel / workqueue.c
1 /*
2  * linux/kernel/workqueue.c
3  *
4  * Generic mechanism for defining kernel helper threads for running
5  * arbitrary tasks in process context.
6  *
7  * Started by Ingo Molnar, Copyright (C) 2002
8  *
9  * Derived from the taskqueue/keventd code by:
10  *
11  *   David Woodhouse <dwmw2@infradead.org>
12  *   Andrew Morton <andrewm@uow.edu.au>
13  *   Kai Petzke <wpp@marie.physik.tu-berlin.de>
14  *   Theodore Ts'o <tytso@mit.edu>
15  *
16  * Made to use alloc_percpu by Christoph Lameter.
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
37 /*
38  * The per-CPU workqueue (if single thread, we always use the first
39  * possible cpu).
40  */
41 struct cpu_workqueue_struct {
42
43         spinlock_t lock;
44
45         struct list_head worklist;
46         wait_queue_head_t more_work;
47         struct work_struct *current_work;
48
49         struct workqueue_struct *wq;
50         struct task_struct *thread;
51
52         int run_depth;          /* Detect run_workqueue() recursion depth */
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 /* Serializes the accesses to the list of workqueues. */
71 static DEFINE_SPINLOCK(workqueue_lock);
72 static LIST_HEAD(workqueues);
73
74 static int singlethread_cpu __read_mostly;
75 static cpumask_t cpu_singlethread_map __read_mostly;
76 /*
77  * _cpu_down() first removes CPU from cpu_online_map, then CPU_DEAD
78  * flushes cwq->worklist. This means that flush_workqueue/wait_on_work
79  * which comes in between can't use for_each_online_cpu(). We could
80  * use cpu_possible_map, the cpumask below is more a documentation
81  * than optimization.
82  */
83 static cpumask_t cpu_populated_map __read_mostly;
84
85 /* If it's single threaded, it isn't in the list of workqueues. */
86 static inline int is_single_threaded(struct workqueue_struct *wq)
87 {
88         return wq->singlethread;
89 }
90
91 static const cpumask_t *wq_cpu_map(struct workqueue_struct *wq)
92 {
93         return is_single_threaded(wq)
94                 ? &cpu_singlethread_map : &cpu_populated_map;
95 }
96
97 static
98 struct cpu_workqueue_struct *wq_per_cpu(struct workqueue_struct *wq, int cpu)
99 {
100         if (unlikely(is_single_threaded(wq)))
101                 cpu = singlethread_cpu;
102         return per_cpu_ptr(wq->cpu_wq, cpu);
103 }
104
105 /*
106  * Set the workqueue on which a work item is to be run
107  * - Must *only* be called if the pending flag is set
108  */
109 static inline void set_wq_data(struct work_struct *work,
110                                 struct cpu_workqueue_struct *cwq)
111 {
112         unsigned long new;
113
114         BUG_ON(!work_pending(work));
115
116         new = (unsigned long) cwq | (1UL << WORK_STRUCT_PENDING);
117         new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
118         atomic_long_set(&work->data, new);
119 }
120
121 static inline
122 struct cpu_workqueue_struct *get_wq_data(struct work_struct *work)
123 {
124         return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
125 }
126
127 static void insert_work(struct cpu_workqueue_struct *cwq,
128                         struct work_struct *work, struct list_head *head)
129 {
130         set_wq_data(work, cwq);
131         /*
132          * Ensure that we get the right work->data if we see the
133          * result of list_add() below, see try_to_grab_pending().
134          */
135         smp_wmb();
136         list_add_tail(&work->entry, head);
137         wake_up(&cwq->more_work);
138 }
139
140 static void __queue_work(struct cpu_workqueue_struct *cwq,
141                          struct work_struct *work)
142 {
143         unsigned long flags;
144
145         spin_lock_irqsave(&cwq->lock, flags);
146         insert_work(cwq, work, &cwq->worklist);
147         spin_unlock_irqrestore(&cwq->lock, flags);
148 }
149
150 /**
151  * queue_work - queue work on a workqueue
152  * @wq: workqueue to use
153  * @work: work to queue
154  *
155  * Returns 0 if @work was already on a queue, non-zero otherwise.
156  *
157  * We queue the work to the CPU on which it was submitted, but if the CPU dies
158  * it can be processed by another CPU.
159  */
160 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
161 {
162         int ret = 0;
163
164         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
165                 BUG_ON(!list_empty(&work->entry));
166                 __queue_work(wq_per_cpu(wq, get_cpu()), work);
167                 put_cpu();
168                 ret = 1;
169         }
170         return ret;
171 }
172 EXPORT_SYMBOL_GPL(queue_work);
173
174 /**
175  * queue_work_on - queue work on specific cpu
176  * @cpu: CPU number to execute work on
177  * @wq: workqueue to use
178  * @work: work to queue
179  *
180  * Returns 0 if @work was already on a queue, non-zero otherwise.
181  *
182  * We queue the work to a specific CPU, the caller must ensure it
183  * can't go away.
184  */
185 int
186 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
187 {
188         int ret = 0;
189
190         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
191                 BUG_ON(!list_empty(&work->entry));
192                 __queue_work(wq_per_cpu(wq, cpu), work);
193                 ret = 1;
194         }
195         return ret;
196 }
197 EXPORT_SYMBOL_GPL(queue_work_on);
198
199 static void delayed_work_timer_fn(unsigned long __data)
200 {
201         struct delayed_work *dwork = (struct delayed_work *)__data;
202         struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
203         struct workqueue_struct *wq = cwq->wq;
204
205         __queue_work(wq_per_cpu(wq, smp_processor_id()), &dwork->work);
206 }
207
208 /**
209  * queue_delayed_work - queue work on a workqueue after delay
210  * @wq: workqueue to use
211  * @dwork: delayable work to queue
212  * @delay: number of jiffies to wait before queueing
213  *
214  * Returns 0 if @work was already on a queue, non-zero otherwise.
215  */
216 int queue_delayed_work(struct workqueue_struct *wq,
217                         struct delayed_work *dwork, unsigned long delay)
218 {
219         if (delay == 0)
220                 return queue_work(wq, &dwork->work);
221
222         return queue_delayed_work_on(-1, wq, dwork, delay);
223 }
224 EXPORT_SYMBOL_GPL(queue_delayed_work);
225
226 /**
227  * queue_delayed_work_on - queue work on specific CPU after delay
228  * @cpu: CPU number to execute work on
229  * @wq: workqueue to use
230  * @dwork: work to queue
231  * @delay: number of jiffies to wait before queueing
232  *
233  * Returns 0 if @work was already on a queue, non-zero otherwise.
234  */
235 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
236                         struct delayed_work *dwork, unsigned long delay)
237 {
238         int ret = 0;
239         struct timer_list *timer = &dwork->timer;
240         struct work_struct *work = &dwork->work;
241
242         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
243                 BUG_ON(timer_pending(timer));
244                 BUG_ON(!list_empty(&work->entry));
245
246                 timer_stats_timer_set_start_info(&dwork->timer);
247
248                 /* This stores cwq for the moment, for the timer_fn */
249                 set_wq_data(work, wq_per_cpu(wq, raw_smp_processor_id()));
250                 timer->expires = jiffies + delay;
251                 timer->data = (unsigned long)dwork;
252                 timer->function = delayed_work_timer_fn;
253
254                 if (unlikely(cpu >= 0))
255                         add_timer_on(timer, cpu);
256                 else
257                         add_timer(timer);
258                 ret = 1;
259         }
260         return ret;
261 }
262 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
263
264 static void run_workqueue(struct cpu_workqueue_struct *cwq)
265 {
266         spin_lock_irq(&cwq->lock);
267         cwq->run_depth++;
268         if (cwq->run_depth > 3) {
269                 /* morton gets to eat his hat */
270                 printk("%s: recursion depth exceeded: %d\n",
271                         __func__, cwq->run_depth);
272                 dump_stack();
273         }
274         while (!list_empty(&cwq->worklist)) {
275                 struct work_struct *work = list_entry(cwq->worklist.next,
276                                                 struct work_struct, entry);
277                 work_func_t f = work->func;
278 #ifdef CONFIG_LOCKDEP
279                 /*
280                  * It is permissible to free the struct work_struct
281                  * from inside the function that is called from it,
282                  * this we need to take into account for lockdep too.
283                  * To avoid bogus "held lock freed" warnings as well
284                  * as problems when looking into work->lockdep_map,
285                  * make a copy and use that here.
286                  */
287                 struct lockdep_map lockdep_map = work->lockdep_map;
288 #endif
289
290                 cwq->current_work = work;
291                 list_del_init(cwq->worklist.next);
292                 spin_unlock_irq(&cwq->lock);
293
294                 BUG_ON(get_wq_data(work) != cwq);
295                 work_clear_pending(work);
296                 lock_acquire(&cwq->wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
297                 lock_acquire(&lockdep_map, 0, 0, 0, 2, _THIS_IP_);
298                 f(work);
299                 lock_release(&lockdep_map, 1, _THIS_IP_);
300                 lock_release(&cwq->wq->lockdep_map, 1, _THIS_IP_);
301
302                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
303                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
304                                         "%s/0x%08x/%d\n",
305                                         current->comm, preempt_count(),
306                                         task_pid_nr(current));
307                         printk(KERN_ERR "    last function: ");
308                         print_symbol("%s\n", (unsigned long)f);
309                         debug_show_held_locks(current);
310                         dump_stack();
311                 }
312
313                 spin_lock_irq(&cwq->lock);
314                 cwq->current_work = NULL;
315         }
316         cwq->run_depth--;
317         spin_unlock_irq(&cwq->lock);
318 }
319
320 static int worker_thread(void *__cwq)
321 {
322         struct cpu_workqueue_struct *cwq = __cwq;
323         DEFINE_WAIT(wait);
324
325         if (cwq->wq->freezeable)
326                 set_freezable();
327
328         set_user_nice(current, -5);
329
330         for (;;) {
331                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
332                 if (!freezing(current) &&
333                     !kthread_should_stop() &&
334                     list_empty(&cwq->worklist))
335                         schedule();
336                 finish_wait(&cwq->more_work, &wait);
337
338                 try_to_freeze();
339
340                 if (kthread_should_stop())
341                         break;
342
343                 run_workqueue(cwq);
344         }
345
346         return 0;
347 }
348
349 struct wq_barrier {
350         struct work_struct      work;
351         struct completion       done;
352 };
353
354 static void wq_barrier_func(struct work_struct *work)
355 {
356         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
357         complete(&barr->done);
358 }
359
360 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
361                         struct wq_barrier *barr, struct list_head *head)
362 {
363         INIT_WORK(&barr->work, wq_barrier_func);
364         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
365
366         init_completion(&barr->done);
367
368         insert_work(cwq, &barr->work, head);
369 }
370
371 static int flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
372 {
373         int active;
374
375         if (cwq->thread == current) {
376                 /*
377                  * Probably keventd trying to flush its own queue. So simply run
378                  * it by hand rather than deadlocking.
379                  */
380                 run_workqueue(cwq);
381                 active = 1;
382         } else {
383                 struct wq_barrier barr;
384
385                 active = 0;
386                 spin_lock_irq(&cwq->lock);
387                 if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
388                         insert_wq_barrier(cwq, &barr, &cwq->worklist);
389                         active = 1;
390                 }
391                 spin_unlock_irq(&cwq->lock);
392
393                 if (active)
394                         wait_for_completion(&barr.done);
395         }
396
397         return active;
398 }
399
400 /**
401  * flush_workqueue - ensure that any scheduled work has run to completion.
402  * @wq: workqueue to flush
403  *
404  * Forces execution of the workqueue and blocks until its completion.
405  * This is typically used in driver shutdown handlers.
406  *
407  * We sleep until all works which were queued on entry have been handled,
408  * but we are not livelocked by new incoming ones.
409  *
410  * This function used to run the workqueues itself.  Now we just wait for the
411  * helper threads to do it.
412  */
413 void flush_workqueue(struct workqueue_struct *wq)
414 {
415         const cpumask_t *cpu_map = wq_cpu_map(wq);
416         int cpu;
417
418         might_sleep();
419         lock_acquire(&wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
420         lock_release(&wq->lockdep_map, 1, _THIS_IP_);
421         for_each_cpu_mask_nr(cpu, *cpu_map)
422                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
423 }
424 EXPORT_SYMBOL_GPL(flush_workqueue);
425
426 /**
427  * flush_work - block until a work_struct's callback has terminated
428  * @work: the work which is to be flushed
429  *
430  * Returns false if @work has already terminated.
431  *
432  * It is expected that, prior to calling flush_work(), the caller has
433  * arranged for the work to not be requeued, otherwise it doesn't make
434  * sense to use this function.
435  */
436 int flush_work(struct work_struct *work)
437 {
438         struct cpu_workqueue_struct *cwq;
439         struct list_head *prev;
440         struct wq_barrier barr;
441
442         might_sleep();
443         cwq = get_wq_data(work);
444         if (!cwq)
445                 return 0;
446
447         lock_acquire(&cwq->wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
448         lock_release(&cwq->wq->lockdep_map, 1, _THIS_IP_);
449
450         prev = NULL;
451         spin_lock_irq(&cwq->lock);
452         if (!list_empty(&work->entry)) {
453                 /*
454                  * See the comment near try_to_grab_pending()->smp_rmb().
455                  * If it was re-queued under us we are not going to wait.
456                  */
457                 smp_rmb();
458                 if (unlikely(cwq != get_wq_data(work)))
459                         goto out;
460                 prev = &work->entry;
461         } else {
462                 if (cwq->current_work != work)
463                         goto out;
464                 prev = &cwq->worklist;
465         }
466         insert_wq_barrier(cwq, &barr, prev->next);
467 out:
468         spin_unlock_irq(&cwq->lock);
469         if (!prev)
470                 return 0;
471
472         wait_for_completion(&barr.done);
473         return 1;
474 }
475 EXPORT_SYMBOL_GPL(flush_work);
476
477 /*
478  * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
479  * so this work can't be re-armed in any way.
480  */
481 static int try_to_grab_pending(struct work_struct *work)
482 {
483         struct cpu_workqueue_struct *cwq;
484         int ret = -1;
485
486         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work)))
487                 return 0;
488
489         /*
490          * The queueing is in progress, or it is already queued. Try to
491          * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
492          */
493
494         cwq = get_wq_data(work);
495         if (!cwq)
496                 return ret;
497
498         spin_lock_irq(&cwq->lock);
499         if (!list_empty(&work->entry)) {
500                 /*
501                  * This work is queued, but perhaps we locked the wrong cwq.
502                  * In that case we must see the new value after rmb(), see
503                  * insert_work()->wmb().
504                  */
505                 smp_rmb();
506                 if (cwq == get_wq_data(work)) {
507                         list_del_init(&work->entry);
508                         ret = 1;
509                 }
510         }
511         spin_unlock_irq(&cwq->lock);
512
513         return ret;
514 }
515
516 static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq,
517                                 struct work_struct *work)
518 {
519         struct wq_barrier barr;
520         int running = 0;
521
522         spin_lock_irq(&cwq->lock);
523         if (unlikely(cwq->current_work == work)) {
524                 insert_wq_barrier(cwq, &barr, cwq->worklist.next);
525                 running = 1;
526         }
527         spin_unlock_irq(&cwq->lock);
528
529         if (unlikely(running))
530                 wait_for_completion(&barr.done);
531 }
532
533 static void wait_on_work(struct work_struct *work)
534 {
535         struct cpu_workqueue_struct *cwq;
536         struct workqueue_struct *wq;
537         const cpumask_t *cpu_map;
538         int cpu;
539
540         might_sleep();
541
542         lock_acquire(&work->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
543         lock_release(&work->lockdep_map, 1, _THIS_IP_);
544
545         cwq = get_wq_data(work);
546         if (!cwq)
547                 return;
548
549         wq = cwq->wq;
550         cpu_map = wq_cpu_map(wq);
551
552         for_each_cpu_mask_nr(cpu, *cpu_map)
553                 wait_on_cpu_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
554 }
555
556 static int __cancel_work_timer(struct work_struct *work,
557                                 struct timer_list* timer)
558 {
559         int ret;
560
561         do {
562                 ret = (timer && likely(del_timer(timer)));
563                 if (!ret)
564                         ret = try_to_grab_pending(work);
565                 wait_on_work(work);
566         } while (unlikely(ret < 0));
567
568         work_clear_pending(work);
569         return ret;
570 }
571
572 /**
573  * cancel_work_sync - block until a work_struct's callback has terminated
574  * @work: the work which is to be flushed
575  *
576  * Returns true if @work was pending.
577  *
578  * cancel_work_sync() will cancel the work if it is queued. If the work's
579  * callback appears to be running, cancel_work_sync() will block until it
580  * has completed.
581  *
582  * It is possible to use this function if the work re-queues itself. It can
583  * cancel the work even if it migrates to another workqueue, however in that
584  * case it only guarantees that work->func() has completed on the last queued
585  * workqueue.
586  *
587  * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
588  * pending, otherwise it goes into a busy-wait loop until the timer expires.
589  *
590  * The caller must ensure that workqueue_struct on which this work was last
591  * queued can't be destroyed before this function returns.
592  */
593 int cancel_work_sync(struct work_struct *work)
594 {
595         return __cancel_work_timer(work, NULL);
596 }
597 EXPORT_SYMBOL_GPL(cancel_work_sync);
598
599 /**
600  * cancel_delayed_work_sync - reliably kill off a delayed work.
601  * @dwork: the delayed work struct
602  *
603  * Returns true if @dwork was pending.
604  *
605  * It is possible to use this function if @dwork rearms itself via queue_work()
606  * or queue_delayed_work(). See also the comment for cancel_work_sync().
607  */
608 int cancel_delayed_work_sync(struct delayed_work *dwork)
609 {
610         return __cancel_work_timer(&dwork->work, &dwork->timer);
611 }
612 EXPORT_SYMBOL(cancel_delayed_work_sync);
613
614 static struct workqueue_struct *keventd_wq __read_mostly;
615
616 /**
617  * schedule_work - put work task in global workqueue
618  * @work: job to be done
619  *
620  * This puts a job in the kernel-global workqueue.
621  */
622 int schedule_work(struct work_struct *work)
623 {
624         return queue_work(keventd_wq, work);
625 }
626 EXPORT_SYMBOL(schedule_work);
627
628 /*
629  * schedule_work_on - put work task on a specific cpu
630  * @cpu: cpu to put the work task on
631  * @work: job to be done
632  *
633  * This puts a job on a specific cpu
634  */
635 int schedule_work_on(int cpu, struct work_struct *work)
636 {
637         return queue_work_on(cpu, keventd_wq, work);
638 }
639 EXPORT_SYMBOL(schedule_work_on);
640
641 /**
642  * schedule_delayed_work - put work task in global workqueue after delay
643  * @dwork: job to be done
644  * @delay: number of jiffies to wait or 0 for immediate execution
645  *
646  * After waiting for a given time this puts a job in the kernel-global
647  * workqueue.
648  */
649 int schedule_delayed_work(struct delayed_work *dwork,
650                                         unsigned long delay)
651 {
652         return queue_delayed_work(keventd_wq, dwork, delay);
653 }
654 EXPORT_SYMBOL(schedule_delayed_work);
655
656 /**
657  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
658  * @cpu: cpu to use
659  * @dwork: job to be done
660  * @delay: number of jiffies to wait
661  *
662  * After waiting for a given time this puts a job in the kernel-global
663  * workqueue on the specified CPU.
664  */
665 int schedule_delayed_work_on(int cpu,
666                         struct delayed_work *dwork, unsigned long delay)
667 {
668         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
669 }
670 EXPORT_SYMBOL(schedule_delayed_work_on);
671
672 /**
673  * schedule_on_each_cpu - call a function on each online CPU from keventd
674  * @func: the function to call
675  *
676  * Returns zero on success.
677  * Returns -ve errno on failure.
678  *
679  * schedule_on_each_cpu() is very slow.
680  */
681 int schedule_on_each_cpu(work_func_t func)
682 {
683         int cpu;
684         struct work_struct *works;
685
686         works = alloc_percpu(struct work_struct);
687         if (!works)
688                 return -ENOMEM;
689
690         get_online_cpus();
691         for_each_online_cpu(cpu) {
692                 struct work_struct *work = per_cpu_ptr(works, cpu);
693
694                 INIT_WORK(work, func);
695                 set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
696                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
697         }
698         for_each_online_cpu(cpu)
699                 flush_work(per_cpu_ptr(works, cpu));
700         put_online_cpus();
701         free_percpu(works);
702         return 0;
703 }
704
705 void flush_scheduled_work(void)
706 {
707         flush_workqueue(keventd_wq);
708 }
709 EXPORT_SYMBOL(flush_scheduled_work);
710
711 /**
712  * execute_in_process_context - reliably execute the routine with user context
713  * @fn:         the function to execute
714  * @ew:         guaranteed storage for the execute work structure (must
715  *              be available when the work executes)
716  *
717  * Executes the function immediately if process context is available,
718  * otherwise schedules the function for delayed execution.
719  *
720  * Returns:     0 - function was executed
721  *              1 - function was scheduled for execution
722  */
723 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
724 {
725         if (!in_interrupt()) {
726                 fn(&ew->work);
727                 return 0;
728         }
729
730         INIT_WORK(&ew->work, fn);
731         schedule_work(&ew->work);
732
733         return 1;
734 }
735 EXPORT_SYMBOL_GPL(execute_in_process_context);
736
737 int keventd_up(void)
738 {
739         return keventd_wq != NULL;
740 }
741
742 int current_is_keventd(void)
743 {
744         struct cpu_workqueue_struct *cwq;
745         int cpu = raw_smp_processor_id(); /* preempt-safe: keventd is per-cpu */
746         int ret = 0;
747
748         BUG_ON(!keventd_wq);
749
750         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
751         if (current == cwq->thread)
752                 ret = 1;
753
754         return ret;
755
756 }
757
758 static struct cpu_workqueue_struct *
759 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
760 {
761         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
762
763         cwq->wq = wq;
764         spin_lock_init(&cwq->lock);
765         INIT_LIST_HEAD(&cwq->worklist);
766         init_waitqueue_head(&cwq->more_work);
767
768         return cwq;
769 }
770
771 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
772 {
773         struct workqueue_struct *wq = cwq->wq;
774         const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
775         struct task_struct *p;
776
777         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
778         /*
779          * Nobody can add the work_struct to this cwq,
780          *      if (caller is __create_workqueue)
781          *              nobody should see this wq
782          *      else // caller is CPU_UP_PREPARE
783          *              cpu is not on cpu_online_map
784          * so we can abort safely.
785          */
786         if (IS_ERR(p))
787                 return PTR_ERR(p);
788
789         cwq->thread = p;
790
791         return 0;
792 }
793
794 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
795 {
796         struct task_struct *p = cwq->thread;
797
798         if (p != NULL) {
799                 if (cpu >= 0)
800                         kthread_bind(p, cpu);
801                 wake_up_process(p);
802         }
803 }
804
805 struct workqueue_struct *__create_workqueue_key(const char *name,
806                                                 int singlethread,
807                                                 int freezeable,
808                                                 struct lock_class_key *key,
809                                                 const char *lock_name)
810 {
811         struct workqueue_struct *wq;
812         struct cpu_workqueue_struct *cwq;
813         int err = 0, cpu;
814
815         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
816         if (!wq)
817                 return NULL;
818
819         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
820         if (!wq->cpu_wq) {
821                 kfree(wq);
822                 return NULL;
823         }
824
825         wq->name = name;
826         lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
827         wq->singlethread = singlethread;
828         wq->freezeable = freezeable;
829         INIT_LIST_HEAD(&wq->list);
830
831         if (singlethread) {
832                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
833                 err = create_workqueue_thread(cwq, singlethread_cpu);
834                 start_workqueue_thread(cwq, -1);
835         } else {
836                 cpu_maps_update_begin();
837                 spin_lock(&workqueue_lock);
838                 list_add(&wq->list, &workqueues);
839                 spin_unlock(&workqueue_lock);
840
841                 for_each_possible_cpu(cpu) {
842                         cwq = init_cpu_workqueue(wq, cpu);
843                         if (err || !cpu_online(cpu))
844                                 continue;
845                         err = create_workqueue_thread(cwq, cpu);
846                         start_workqueue_thread(cwq, cpu);
847                 }
848                 cpu_maps_update_done();
849         }
850
851         if (err) {
852                 destroy_workqueue(wq);
853                 wq = NULL;
854         }
855         return wq;
856 }
857 EXPORT_SYMBOL_GPL(__create_workqueue_key);
858
859 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq)
860 {
861         /*
862          * Our caller is either destroy_workqueue() or CPU_POST_DEAD,
863          * cpu_add_remove_lock protects cwq->thread.
864          */
865         if (cwq->thread == NULL)
866                 return;
867
868         lock_acquire(&cwq->wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
869         lock_release(&cwq->wq->lockdep_map, 1, _THIS_IP_);
870
871         flush_cpu_workqueue(cwq);
872         /*
873          * If the caller is CPU_POST_DEAD and cwq->worklist was not empty,
874          * a concurrent flush_workqueue() can insert a barrier after us.
875          * However, in that case run_workqueue() won't return and check
876          * kthread_should_stop() until it flushes all work_struct's.
877          * When ->worklist becomes empty it is safe to exit because no
878          * more work_structs can be queued on this cwq: flush_workqueue
879          * checks list_empty(), and a "normal" queue_work() can't use
880          * a dead CPU.
881          */
882         kthread_stop(cwq->thread);
883         cwq->thread = NULL;
884 }
885
886 /**
887  * destroy_workqueue - safely terminate a workqueue
888  * @wq: target workqueue
889  *
890  * Safely destroy a workqueue. All work currently pending will be done first.
891  */
892 void destroy_workqueue(struct workqueue_struct *wq)
893 {
894         const cpumask_t *cpu_map = wq_cpu_map(wq);
895         int cpu;
896
897         cpu_maps_update_begin();
898         spin_lock(&workqueue_lock);
899         list_del(&wq->list);
900         spin_unlock(&workqueue_lock);
901
902         for_each_cpu_mask_nr(cpu, *cpu_map)
903                 cleanup_workqueue_thread(per_cpu_ptr(wq->cpu_wq, cpu));
904         cpu_maps_update_done();
905
906         free_percpu(wq->cpu_wq);
907         kfree(wq);
908 }
909 EXPORT_SYMBOL_GPL(destroy_workqueue);
910
911 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
912                                                 unsigned long action,
913                                                 void *hcpu)
914 {
915         unsigned int cpu = (unsigned long)hcpu;
916         struct cpu_workqueue_struct *cwq;
917         struct workqueue_struct *wq;
918
919         action &= ~CPU_TASKS_FROZEN;
920
921         switch (action) {
922         case CPU_UP_PREPARE:
923                 cpu_set(cpu, cpu_populated_map);
924         }
925
926         list_for_each_entry(wq, &workqueues, list) {
927                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
928
929                 switch (action) {
930                 case CPU_UP_PREPARE:
931                         if (!create_workqueue_thread(cwq, cpu))
932                                 break;
933                         printk(KERN_ERR "workqueue [%s] for %i failed\n",
934                                 wq->name, cpu);
935                         return NOTIFY_BAD;
936
937                 case CPU_ONLINE:
938                         start_workqueue_thread(cwq, cpu);
939                         break;
940
941                 case CPU_UP_CANCELED:
942                         start_workqueue_thread(cwq, -1);
943                 case CPU_POST_DEAD:
944                         cleanup_workqueue_thread(cwq);
945                         break;
946                 }
947         }
948
949         switch (action) {
950         case CPU_UP_CANCELED:
951         case CPU_POST_DEAD:
952                 cpu_clear(cpu, cpu_populated_map);
953         }
954
955         return NOTIFY_OK;
956 }
957
958 void __init init_workqueues(void)
959 {
960         cpu_populated_map = cpu_online_map;
961         singlethread_cpu = first_cpu(cpu_possible_map);
962         cpu_singlethread_map = cpumask_of_cpu(singlethread_cpu);
963         hotcpu_notifier(workqueue_cpu_callback, 0);
964         keventd_wq = create_workqueue("events");
965         BUG_ON(!keventd_wq);
966 }