workqueues: shift kthread_bind() from CPU_UP_PREPARE to CPU_ONLINE
[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
36 /*
37  * The per-CPU workqueue (if single thread, we always use the first
38  * possible cpu).
39  */
40 struct cpu_workqueue_struct {
41
42         spinlock_t lock;
43
44         struct list_head worklist;
45         wait_queue_head_t more_work;
46         struct work_struct *current_work;
47
48         struct workqueue_struct *wq;
49         struct task_struct *thread;
50         int should_stop;
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 };
66
67 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
68    threads to each one as cpus come/go. */
69 static DEFINE_MUTEX(workqueue_mutex);
70 static LIST_HEAD(workqueues);
71
72 static int singlethread_cpu __read_mostly;
73 static cpumask_t cpu_singlethread_map __read_mostly;
74 /* optimization, we could use cpu_possible_map */
75 static cpumask_t cpu_populated_map __read_mostly;
76
77 /* If it's single threaded, it isn't in the list of workqueues. */
78 static inline int is_single_threaded(struct workqueue_struct *wq)
79 {
80         return wq->singlethread;
81 }
82
83 static const cpumask_t *wq_cpu_map(struct workqueue_struct *wq)
84 {
85         return is_single_threaded(wq)
86                 ? &cpu_singlethread_map : &cpu_populated_map;
87 }
88
89 /*
90  * Set the workqueue on which a work item is to be run
91  * - Must *only* be called if the pending flag is set
92  */
93 static inline void set_wq_data(struct work_struct *work, void *wq)
94 {
95         unsigned long new;
96
97         BUG_ON(!work_pending(work));
98
99         new = (unsigned long) wq | (1UL << WORK_STRUCT_PENDING);
100         new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
101         atomic_long_set(&work->data, new);
102 }
103
104 static inline void *get_wq_data(struct work_struct *work)
105 {
106         return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
107 }
108
109 static void insert_work(struct cpu_workqueue_struct *cwq,
110                                 struct work_struct *work, int tail)
111 {
112         set_wq_data(work, cwq);
113         if (tail)
114                 list_add_tail(&work->entry, &cwq->worklist);
115         else
116                 list_add(&work->entry, &cwq->worklist);
117         wake_up(&cwq->more_work);
118 }
119
120 /* Preempt must be disabled. */
121 static void __queue_work(struct cpu_workqueue_struct *cwq,
122                          struct work_struct *work)
123 {
124         unsigned long flags;
125
126         spin_lock_irqsave(&cwq->lock, flags);
127         insert_work(cwq, work, 1);
128         spin_unlock_irqrestore(&cwq->lock, flags);
129 }
130
131 /**
132  * queue_work - queue work on a workqueue
133  * @wq: workqueue to use
134  * @work: work to queue
135  *
136  * Returns 0 if @work was already on a queue, non-zero otherwise.
137  *
138  * We queue the work to the CPU it was submitted, but there is no
139  * guarantee that it will be processed by that CPU.
140  */
141 int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
142 {
143         int ret = 0, cpu = get_cpu();
144
145         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
146                 if (unlikely(is_single_threaded(wq)))
147                         cpu = singlethread_cpu;
148                 BUG_ON(!list_empty(&work->entry));
149                 __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
150                 ret = 1;
151         }
152         put_cpu();
153         return ret;
154 }
155 EXPORT_SYMBOL_GPL(queue_work);
156
157 void delayed_work_timer_fn(unsigned long __data)
158 {
159         struct delayed_work *dwork = (struct delayed_work *)__data;
160         struct workqueue_struct *wq = get_wq_data(&dwork->work);
161         int cpu = smp_processor_id();
162
163         if (unlikely(is_single_threaded(wq)))
164                 cpu = singlethread_cpu;
165
166         __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), &dwork->work);
167 }
168
169 /**
170  * queue_delayed_work - queue work on a workqueue after delay
171  * @wq: workqueue to use
172  * @dwork: delayable work to queue
173  * @delay: number of jiffies to wait before queueing
174  *
175  * Returns 0 if @work was already on a queue, non-zero otherwise.
176  */
177 int fastcall queue_delayed_work(struct workqueue_struct *wq,
178                         struct delayed_work *dwork, unsigned long delay)
179 {
180         int ret = 0;
181         struct timer_list *timer = &dwork->timer;
182         struct work_struct *work = &dwork->work;
183
184         timer_stats_timer_set_start_info(timer);
185         if (delay == 0)
186                 return queue_work(wq, work);
187
188         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
189                 BUG_ON(timer_pending(timer));
190                 BUG_ON(!list_empty(&work->entry));
191
192                 /* This stores wq for the moment, for the timer_fn */
193                 set_wq_data(work, wq);
194                 timer->expires = jiffies + delay;
195                 timer->data = (unsigned long)dwork;
196                 timer->function = delayed_work_timer_fn;
197                 add_timer(timer);
198                 ret = 1;
199         }
200         return ret;
201 }
202 EXPORT_SYMBOL_GPL(queue_delayed_work);
203
204 /**
205  * queue_delayed_work_on - queue work on specific CPU after delay
206  * @cpu: CPU number to execute work on
207  * @wq: workqueue to use
208  * @dwork: work to queue
209  * @delay: number of jiffies to wait before queueing
210  *
211  * Returns 0 if @work was already on a queue, non-zero otherwise.
212  */
213 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
214                         struct delayed_work *dwork, unsigned long delay)
215 {
216         int ret = 0;
217         struct timer_list *timer = &dwork->timer;
218         struct work_struct *work = &dwork->work;
219
220         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
221                 BUG_ON(timer_pending(timer));
222                 BUG_ON(!list_empty(&work->entry));
223
224                 /* This stores wq for the moment, for the timer_fn */
225                 set_wq_data(work, wq);
226                 timer->expires = jiffies + delay;
227                 timer->data = (unsigned long)dwork;
228                 timer->function = delayed_work_timer_fn;
229                 add_timer_on(timer, cpu);
230                 ret = 1;
231         }
232         return ret;
233 }
234 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
235
236 static void run_workqueue(struct cpu_workqueue_struct *cwq)
237 {
238         spin_lock_irq(&cwq->lock);
239         cwq->run_depth++;
240         if (cwq->run_depth > 3) {
241                 /* morton gets to eat his hat */
242                 printk("%s: recursion depth exceeded: %d\n",
243                         __FUNCTION__, cwq->run_depth);
244                 dump_stack();
245         }
246         while (!list_empty(&cwq->worklist)) {
247                 struct work_struct *work = list_entry(cwq->worklist.next,
248                                                 struct work_struct, entry);
249                 work_func_t f = work->func;
250
251                 cwq->current_work = work;
252                 list_del_init(cwq->worklist.next);
253                 spin_unlock_irq(&cwq->lock);
254
255                 BUG_ON(get_wq_data(work) != cwq);
256                 if (!test_bit(WORK_STRUCT_NOAUTOREL, work_data_bits(work)))
257                         work_release(work);
258                 f(work);
259
260                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
261                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
262                                         "%s/0x%08x/%d\n",
263                                         current->comm, preempt_count(),
264                                         current->pid);
265                         printk(KERN_ERR "    last function: ");
266                         print_symbol("%s\n", (unsigned long)f);
267                         debug_show_held_locks(current);
268                         dump_stack();
269                 }
270
271                 spin_lock_irq(&cwq->lock);
272                 cwq->current_work = NULL;
273         }
274         cwq->run_depth--;
275         spin_unlock_irq(&cwq->lock);
276 }
277
278 /*
279  * NOTE: the caller must not touch *cwq if this func returns true
280  */
281 static int cwq_should_stop(struct cpu_workqueue_struct *cwq)
282 {
283         int should_stop = cwq->should_stop;
284
285         if (unlikely(should_stop)) {
286                 spin_lock_irq(&cwq->lock);
287                 should_stop = cwq->should_stop && list_empty(&cwq->worklist);
288                 if (should_stop)
289                         cwq->thread = NULL;
290                 spin_unlock_irq(&cwq->lock);
291         }
292
293         return should_stop;
294 }
295
296 static int worker_thread(void *__cwq)
297 {
298         struct cpu_workqueue_struct *cwq = __cwq;
299         DEFINE_WAIT(wait);
300         struct k_sigaction sa;
301         sigset_t blocked;
302
303         if (!cwq->wq->freezeable)
304                 current->flags |= PF_NOFREEZE;
305
306         set_user_nice(current, -5);
307
308         /* Block and flush all signals */
309         sigfillset(&blocked);
310         sigprocmask(SIG_BLOCK, &blocked, NULL);
311         flush_signals(current);
312
313         /*
314          * We inherited MPOL_INTERLEAVE from the booting kernel.
315          * Set MPOL_DEFAULT to insure node local allocations.
316          */
317         numa_default_policy();
318
319         /* SIG_IGN makes children autoreap: see do_notify_parent(). */
320         sa.sa.sa_handler = SIG_IGN;
321         sa.sa.sa_flags = 0;
322         siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
323         do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
324
325         for (;;) {
326                 if (cwq->wq->freezeable)
327                         try_to_freeze();
328
329                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
330                 if (!cwq->should_stop && list_empty(&cwq->worklist))
331                         schedule();
332                 finish_wait(&cwq->more_work, &wait);
333
334                 if (cwq_should_stop(cwq))
335                         break;
336
337                 run_workqueue(cwq);
338         }
339
340         return 0;
341 }
342
343 struct wq_barrier {
344         struct work_struct      work;
345         struct completion       done;
346 };
347
348 static void wq_barrier_func(struct work_struct *work)
349 {
350         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
351         complete(&barr->done);
352 }
353
354 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
355                                         struct wq_barrier *barr, int tail)
356 {
357         INIT_WORK(&barr->work, wq_barrier_func);
358         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
359
360         init_completion(&barr->done);
361
362         insert_work(cwq, &barr->work, tail);
363 }
364
365 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
366 {
367         if (cwq->thread == current) {
368                 /*
369                  * Probably keventd trying to flush its own queue. So simply run
370                  * it by hand rather than deadlocking.
371                  */
372                 run_workqueue(cwq);
373         } else {
374                 struct wq_barrier barr;
375                 int active = 0;
376
377                 spin_lock_irq(&cwq->lock);
378                 if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
379                         insert_wq_barrier(cwq, &barr, 1);
380                         active = 1;
381                 }
382                 spin_unlock_irq(&cwq->lock);
383
384                 if (active)
385                         wait_for_completion(&barr.done);
386         }
387 }
388
389 /**
390  * flush_workqueue - ensure that any scheduled work has run to completion.
391  * @wq: workqueue to flush
392  *
393  * Forces execution of the workqueue and blocks until its completion.
394  * This is typically used in driver shutdown handlers.
395  *
396  * We sleep until all works which were queued on entry have been handled,
397  * but we are not livelocked by new incoming ones.
398  *
399  * This function used to run the workqueues itself.  Now we just wait for the
400  * helper threads to do it.
401  */
402 void fastcall flush_workqueue(struct workqueue_struct *wq)
403 {
404         const cpumask_t *cpu_map = wq_cpu_map(wq);
405         int cpu;
406
407         might_sleep();
408         for_each_cpu_mask(cpu, *cpu_map)
409                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
410 }
411 EXPORT_SYMBOL_GPL(flush_workqueue);
412
413 static void wait_on_work(struct cpu_workqueue_struct *cwq,
414                                 struct work_struct *work)
415 {
416         struct wq_barrier barr;
417         int running = 0;
418
419         spin_lock_irq(&cwq->lock);
420         if (unlikely(cwq->current_work == work)) {
421                 insert_wq_barrier(cwq, &barr, 0);
422                 running = 1;
423         }
424         spin_unlock_irq(&cwq->lock);
425
426         if (unlikely(running))
427                 wait_for_completion(&barr.done);
428 }
429
430 /**
431  * flush_work - block until a work_struct's callback has terminated
432  * @wq: the workqueue on which the work is queued
433  * @work: the work which is to be flushed
434  *
435  * flush_work() will attempt to cancel the work if it is queued.  If the work's
436  * callback appears to be running, flush_work() will block until it has
437  * completed.
438  *
439  * flush_work() is designed to be used when the caller is tearing down data
440  * structures which the callback function operates upon.  It is expected that,
441  * prior to calling flush_work(), the caller has arranged for the work to not
442  * be requeued.
443  */
444 void flush_work(struct workqueue_struct *wq, struct work_struct *work)
445 {
446         const cpumask_t *cpu_map = wq_cpu_map(wq);
447         struct cpu_workqueue_struct *cwq;
448         int cpu;
449
450         might_sleep();
451
452         cwq = get_wq_data(work);
453         /* Was it ever queued ? */
454         if (!cwq)
455                 return;
456
457         /*
458          * This work can't be re-queued, no need to re-check that
459          * get_wq_data() is still the same when we take cwq->lock.
460          */
461         spin_lock_irq(&cwq->lock);
462         list_del_init(&work->entry);
463         work_release(work);
464         spin_unlock_irq(&cwq->lock);
465
466         for_each_cpu_mask(cpu, *cpu_map)
467                 wait_on_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
468 }
469 EXPORT_SYMBOL_GPL(flush_work);
470
471
472 static struct workqueue_struct *keventd_wq;
473
474 /**
475  * schedule_work - put work task in global workqueue
476  * @work: job to be done
477  *
478  * This puts a job in the kernel-global workqueue.
479  */
480 int fastcall schedule_work(struct work_struct *work)
481 {
482         return queue_work(keventd_wq, work);
483 }
484 EXPORT_SYMBOL(schedule_work);
485
486 /**
487  * schedule_delayed_work - put work task in global workqueue after delay
488  * @dwork: job to be done
489  * @delay: number of jiffies to wait or 0 for immediate execution
490  *
491  * After waiting for a given time this puts a job in the kernel-global
492  * workqueue.
493  */
494 int fastcall schedule_delayed_work(struct delayed_work *dwork,
495                                         unsigned long delay)
496 {
497         timer_stats_timer_set_start_info(&dwork->timer);
498         return queue_delayed_work(keventd_wq, dwork, delay);
499 }
500 EXPORT_SYMBOL(schedule_delayed_work);
501
502 /**
503  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
504  * @cpu: cpu to use
505  * @dwork: job to be done
506  * @delay: number of jiffies to wait
507  *
508  * After waiting for a given time this puts a job in the kernel-global
509  * workqueue on the specified CPU.
510  */
511 int schedule_delayed_work_on(int cpu,
512                         struct delayed_work *dwork, unsigned long delay)
513 {
514         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
515 }
516 EXPORT_SYMBOL(schedule_delayed_work_on);
517
518 /**
519  * schedule_on_each_cpu - call a function on each online CPU from keventd
520  * @func: the function to call
521  *
522  * Returns zero on success.
523  * Returns -ve errno on failure.
524  *
525  * Appears to be racy against CPU hotplug.
526  *
527  * schedule_on_each_cpu() is very slow.
528  */
529 int schedule_on_each_cpu(work_func_t func)
530 {
531         int cpu;
532         struct work_struct *works;
533
534         works = alloc_percpu(struct work_struct);
535         if (!works)
536                 return -ENOMEM;
537
538         preempt_disable();              /* CPU hotplug */
539         for_each_online_cpu(cpu) {
540                 struct work_struct *work = per_cpu_ptr(works, cpu);
541
542                 INIT_WORK(work, func);
543                 set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
544                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
545         }
546         preempt_enable();
547         flush_workqueue(keventd_wq);
548         free_percpu(works);
549         return 0;
550 }
551
552 void flush_scheduled_work(void)
553 {
554         flush_workqueue(keventd_wq);
555 }
556 EXPORT_SYMBOL(flush_scheduled_work);
557
558 void flush_work_keventd(struct work_struct *work)
559 {
560         flush_work(keventd_wq, work);
561 }
562 EXPORT_SYMBOL(flush_work_keventd);
563
564 /**
565  * cancel_rearming_delayed_workqueue - reliably kill off a delayed work whose handler rearms the delayed work.
566  * @wq:   the controlling workqueue structure
567  * @dwork: the delayed work struct
568  */
569 void cancel_rearming_delayed_workqueue(struct workqueue_struct *wq,
570                                        struct delayed_work *dwork)
571 {
572         /* Was it ever queued ? */
573         if (!get_wq_data(&dwork->work))
574                 return;
575
576         while (!cancel_delayed_work(dwork))
577                 flush_workqueue(wq);
578 }
579 EXPORT_SYMBOL(cancel_rearming_delayed_workqueue);
580
581 /**
582  * cancel_rearming_delayed_work - reliably kill off a delayed keventd work whose handler rearms the delayed work.
583  * @dwork: the delayed work struct
584  */
585 void cancel_rearming_delayed_work(struct delayed_work *dwork)
586 {
587         cancel_rearming_delayed_workqueue(keventd_wq, dwork);
588 }
589 EXPORT_SYMBOL(cancel_rearming_delayed_work);
590
591 /**
592  * execute_in_process_context - reliably execute the routine with user context
593  * @fn:         the function to execute
594  * @ew:         guaranteed storage for the execute work structure (must
595  *              be available when the work executes)
596  *
597  * Executes the function immediately if process context is available,
598  * otherwise schedules the function for delayed execution.
599  *
600  * Returns:     0 - function was executed
601  *              1 - function was scheduled for execution
602  */
603 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
604 {
605         if (!in_interrupt()) {
606                 fn(&ew->work);
607                 return 0;
608         }
609
610         INIT_WORK(&ew->work, fn);
611         schedule_work(&ew->work);
612
613         return 1;
614 }
615 EXPORT_SYMBOL_GPL(execute_in_process_context);
616
617 int keventd_up(void)
618 {
619         return keventd_wq != NULL;
620 }
621
622 int current_is_keventd(void)
623 {
624         struct cpu_workqueue_struct *cwq;
625         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
626         int ret = 0;
627
628         BUG_ON(!keventd_wq);
629
630         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
631         if (current == cwq->thread)
632                 ret = 1;
633
634         return ret;
635
636 }
637
638 static struct cpu_workqueue_struct *
639 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
640 {
641         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
642
643         cwq->wq = wq;
644         spin_lock_init(&cwq->lock);
645         INIT_LIST_HEAD(&cwq->worklist);
646         init_waitqueue_head(&cwq->more_work);
647
648         return cwq;
649 }
650
651 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
652 {
653         struct workqueue_struct *wq = cwq->wq;
654         const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
655         struct task_struct *p;
656
657         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
658         /*
659          * Nobody can add the work_struct to this cwq,
660          *      if (caller is __create_workqueue)
661          *              nobody should see this wq
662          *      else // caller is CPU_UP_PREPARE
663          *              cpu is not on cpu_online_map
664          * so we can abort safely.
665          */
666         if (IS_ERR(p))
667                 return PTR_ERR(p);
668
669         cwq->thread = p;
670         cwq->should_stop = 0;
671
672         return 0;
673 }
674
675 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
676 {
677         struct task_struct *p = cwq->thread;
678
679         if (p != NULL) {
680                 if (cpu >= 0)
681                         kthread_bind(p, cpu);
682                 wake_up_process(p);
683         }
684 }
685
686 struct workqueue_struct *__create_workqueue(const char *name,
687                                             int singlethread, int freezeable)
688 {
689         struct workqueue_struct *wq;
690         struct cpu_workqueue_struct *cwq;
691         int err = 0, cpu;
692
693         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
694         if (!wq)
695                 return NULL;
696
697         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
698         if (!wq->cpu_wq) {
699                 kfree(wq);
700                 return NULL;
701         }
702
703         wq->name = name;
704         wq->singlethread = singlethread;
705         wq->freezeable = freezeable;
706         INIT_LIST_HEAD(&wq->list);
707
708         if (singlethread) {
709                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
710                 err = create_workqueue_thread(cwq, singlethread_cpu);
711                 start_workqueue_thread(cwq, -1);
712         } else {
713                 mutex_lock(&workqueue_mutex);
714                 list_add(&wq->list, &workqueues);
715
716                 for_each_possible_cpu(cpu) {
717                         cwq = init_cpu_workqueue(wq, cpu);
718                         if (err || !cpu_online(cpu))
719                                 continue;
720                         err = create_workqueue_thread(cwq, cpu);
721                         start_workqueue_thread(cwq, cpu);
722                 }
723                 mutex_unlock(&workqueue_mutex);
724         }
725
726         if (err) {
727                 destroy_workqueue(wq);
728                 wq = NULL;
729         }
730         return wq;
731 }
732 EXPORT_SYMBOL_GPL(__create_workqueue);
733
734 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
735 {
736         struct wq_barrier barr;
737         int alive = 0;
738
739         spin_lock_irq(&cwq->lock);
740         if (cwq->thread != NULL) {
741                 insert_wq_barrier(cwq, &barr, 1);
742                 cwq->should_stop = 1;
743                 alive = 1;
744         }
745         spin_unlock_irq(&cwq->lock);
746
747         if (alive) {
748                 wait_for_completion(&barr.done);
749
750                 while (unlikely(cwq->thread != NULL))
751                         cpu_relax();
752                 /*
753                  * Wait until cwq->thread unlocks cwq->lock,
754                  * it won't touch *cwq after that.
755                  */
756                 smp_rmb();
757                 spin_unlock_wait(&cwq->lock);
758         }
759 }
760
761 /**
762  * destroy_workqueue - safely terminate a workqueue
763  * @wq: target workqueue
764  *
765  * Safely destroy a workqueue. All work currently pending will be done first.
766  */
767 void destroy_workqueue(struct workqueue_struct *wq)
768 {
769         const cpumask_t *cpu_map = wq_cpu_map(wq);
770         struct cpu_workqueue_struct *cwq;
771         int cpu;
772
773         mutex_lock(&workqueue_mutex);
774         list_del(&wq->list);
775         mutex_unlock(&workqueue_mutex);
776
777         for_each_cpu_mask(cpu, *cpu_map) {
778                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
779                 cleanup_workqueue_thread(cwq, cpu);
780         }
781
782         free_percpu(wq->cpu_wq);
783         kfree(wq);
784 }
785 EXPORT_SYMBOL_GPL(destroy_workqueue);
786
787 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
788                                                 unsigned long action,
789                                                 void *hcpu)
790 {
791         unsigned int cpu = (unsigned long)hcpu;
792         struct cpu_workqueue_struct *cwq;
793         struct workqueue_struct *wq;
794
795         switch (action) {
796         case CPU_LOCK_ACQUIRE:
797                 mutex_lock(&workqueue_mutex);
798                 return NOTIFY_OK;
799
800         case CPU_LOCK_RELEASE:
801                 mutex_unlock(&workqueue_mutex);
802                 return NOTIFY_OK;
803
804         case CPU_UP_PREPARE:
805                 cpu_set(cpu, cpu_populated_map);
806         }
807
808         list_for_each_entry(wq, &workqueues, list) {
809                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
810
811                 switch (action) {
812                 case CPU_UP_PREPARE:
813                         if (!create_workqueue_thread(cwq, cpu))
814                                 break;
815                         printk(KERN_ERR "workqueue for %i failed\n", cpu);
816                         return NOTIFY_BAD;
817
818                 case CPU_ONLINE:
819                         start_workqueue_thread(cwq, cpu);
820                         break;
821
822                 case CPU_UP_CANCELED:
823                         start_workqueue_thread(cwq, -1);
824                 case CPU_DEAD:
825                         cleanup_workqueue_thread(cwq, cpu);
826                         break;
827                 }
828         }
829
830         return NOTIFY_OK;
831 }
832
833 void __init init_workqueues(void)
834 {
835         cpu_populated_map = cpu_online_map;
836         singlethread_cpu = first_cpu(cpu_possible_map);
837         cpu_singlethread_map = cpumask_of_cpu(singlethread_cpu);
838         hotcpu_notifier(workqueue_cpu_callback, 0);
839         keventd_wq = create_workqueue("events");
840         BUG_ON(!keventd_wq);
841 }