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