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