workqueue: fix flush_workqueue() vs CPU_DEAD race
[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
47         struct workqueue_struct *wq;
48         struct task_struct *thread;
49         struct work_struct *current_work;
50
51         int run_depth;          /* Detect run_workqueue() recursion depth */
52 } ____cacheline_aligned;
53
54 /*
55  * The externally visible workqueue abstraction is an array of
56  * per-CPU workqueues:
57  */
58 struct workqueue_struct {
59         struct cpu_workqueue_struct *cpu_wq;
60         const char *name;
61         struct list_head list;  /* Empty if single thread */
62         int freezeable;         /* Freeze threads during suspend */
63 };
64
65 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
66    threads to each one as cpus come/go. */
67 static long migrate_sequence __read_mostly;
68 static DEFINE_MUTEX(workqueue_mutex);
69 static LIST_HEAD(workqueues);
70
71 static int singlethread_cpu;
72
73 /* If it's single threaded, it isn't in the list of workqueues. */
74 static inline int is_single_threaded(struct workqueue_struct *wq)
75 {
76         return list_empty(&wq->list);
77 }
78
79 /*
80  * Set the workqueue on which a work item is to be run
81  * - Must *only* be called if the pending flag is set
82  */
83 static inline void set_wq_data(struct work_struct *work, void *wq)
84 {
85         unsigned long new;
86
87         BUG_ON(!work_pending(work));
88
89         new = (unsigned long) wq | (1UL << WORK_STRUCT_PENDING);
90         new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
91         atomic_long_set(&work->data, new);
92 }
93
94 static inline void *get_wq_data(struct work_struct *work)
95 {
96         return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
97 }
98
99 static int __run_work(struct cpu_workqueue_struct *cwq, struct work_struct *work)
100 {
101         int ret = 0;
102         unsigned long flags;
103
104         spin_lock_irqsave(&cwq->lock, flags);
105         /*
106          * We need to re-validate the work info after we've gotten
107          * the cpu_workqueue lock. We can run the work now iff:
108          *
109          *  - the wq_data still matches the cpu_workqueue_struct
110          *  - AND the work is still marked pending
111          *  - AND the work is still on a list (which will be this
112          *    workqueue_struct list)
113          *
114          * All these conditions are important, because we
115          * need to protect against the work being run right
116          * now on another CPU (all but the last one might be
117          * true if it's currently running and has not been
118          * released yet, for example).
119          */
120         if (get_wq_data(work) == cwq
121             && work_pending(work)
122             && !list_empty(&work->entry)) {
123                 work_func_t f = work->func;
124                 cwq->current_work = work;
125                 list_del_init(&work->entry);
126                 spin_unlock_irqrestore(&cwq->lock, flags);
127
128                 if (!test_bit(WORK_STRUCT_NOAUTOREL, work_data_bits(work)))
129                         work_release(work);
130                 f(work);
131
132                 spin_lock_irqsave(&cwq->lock, flags);
133                 cwq->current_work = NULL;
134                 ret = 1;
135         }
136         spin_unlock_irqrestore(&cwq->lock, flags);
137         return ret;
138 }
139
140 /**
141  * run_scheduled_work - run scheduled work synchronously
142  * @work: work to run
143  *
144  * This checks if the work was pending, and runs it
145  * synchronously if so. It returns a boolean to indicate
146  * whether it had any scheduled work to run or not.
147  *
148  * NOTE! This _only_ works for normal work_structs. You
149  * CANNOT use this for delayed work, because the wq data
150  * for delayed work will not point properly to the per-
151  * CPU workqueue struct, but will change!
152  */
153 int fastcall run_scheduled_work(struct work_struct *work)
154 {
155         for (;;) {
156                 struct cpu_workqueue_struct *cwq;
157
158                 if (!work_pending(work))
159                         return 0;
160                 if (list_empty(&work->entry))
161                         return 0;
162                 /* NOTE! This depends intimately on __queue_work! */
163                 cwq = get_wq_data(work);
164                 if (!cwq)
165                         return 0;
166                 if (__run_work(cwq, work))
167                         return 1;
168         }
169 }
170 EXPORT_SYMBOL(run_scheduled_work);
171
172 static void insert_work(struct cpu_workqueue_struct *cwq,
173                                 struct work_struct *work, int tail)
174 {
175         set_wq_data(work, cwq);
176         if (tail)
177                 list_add_tail(&work->entry, &cwq->worklist);
178         else
179                 list_add(&work->entry, &cwq->worklist);
180         wake_up(&cwq->more_work);
181 }
182
183 /* Preempt must be disabled. */
184 static void __queue_work(struct cpu_workqueue_struct *cwq,
185                          struct work_struct *work)
186 {
187         unsigned long flags;
188
189         spin_lock_irqsave(&cwq->lock, flags);
190         insert_work(cwq, work, 1);
191         spin_unlock_irqrestore(&cwq->lock, flags);
192 }
193
194 /**
195  * queue_work - queue work on a workqueue
196  * @wq: workqueue to use
197  * @work: work to queue
198  *
199  * Returns 0 if @work was already on a queue, non-zero otherwise.
200  *
201  * We queue the work to the CPU it was submitted, but there is no
202  * guarantee that it will be processed by that CPU.
203  */
204 int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
205 {
206         int ret = 0, cpu = get_cpu();
207
208         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
209                 if (unlikely(is_single_threaded(wq)))
210                         cpu = singlethread_cpu;
211                 BUG_ON(!list_empty(&work->entry));
212                 __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
213                 ret = 1;
214         }
215         put_cpu();
216         return ret;
217 }
218 EXPORT_SYMBOL_GPL(queue_work);
219
220 void delayed_work_timer_fn(unsigned long __data)
221 {
222         struct delayed_work *dwork = (struct delayed_work *)__data;
223         struct workqueue_struct *wq = get_wq_data(&dwork->work);
224         int cpu = smp_processor_id();
225
226         if (unlikely(is_single_threaded(wq)))
227                 cpu = singlethread_cpu;
228
229         __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), &dwork->work);
230 }
231
232 /**
233  * queue_delayed_work - queue work on a workqueue after delay
234  * @wq: workqueue to use
235  * @dwork: delayable work to queue
236  * @delay: number of jiffies to wait before queueing
237  *
238  * Returns 0 if @work was already on a queue, non-zero otherwise.
239  */
240 int fastcall queue_delayed_work(struct workqueue_struct *wq,
241                         struct delayed_work *dwork, unsigned long delay)
242 {
243         int ret = 0;
244         struct timer_list *timer = &dwork->timer;
245         struct work_struct *work = &dwork->work;
246
247         timer_stats_timer_set_start_info(timer);
248         if (delay == 0)
249                 return queue_work(wq, work);
250
251         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
252                 BUG_ON(timer_pending(timer));
253                 BUG_ON(!list_empty(&work->entry));
254
255                 /* This stores wq for the moment, for the timer_fn */
256                 set_wq_data(work, wq);
257                 timer->expires = jiffies + delay;
258                 timer->data = (unsigned long)dwork;
259                 timer->function = delayed_work_timer_fn;
260                 add_timer(timer);
261                 ret = 1;
262         }
263         return ret;
264 }
265 EXPORT_SYMBOL_GPL(queue_delayed_work);
266
267 /**
268  * queue_delayed_work_on - queue work on specific CPU after delay
269  * @cpu: CPU number to execute work on
270  * @wq: workqueue to use
271  * @dwork: work to queue
272  * @delay: number of jiffies to wait before queueing
273  *
274  * Returns 0 if @work was already on a queue, non-zero otherwise.
275  */
276 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
277                         struct delayed_work *dwork, unsigned long delay)
278 {
279         int ret = 0;
280         struct timer_list *timer = &dwork->timer;
281         struct work_struct *work = &dwork->work;
282
283         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
284                 BUG_ON(timer_pending(timer));
285                 BUG_ON(!list_empty(&work->entry));
286
287                 /* This stores wq for the moment, for the timer_fn */
288                 set_wq_data(work, wq);
289                 timer->expires = jiffies + delay;
290                 timer->data = (unsigned long)dwork;
291                 timer->function = delayed_work_timer_fn;
292                 add_timer_on(timer, cpu);
293                 ret = 1;
294         }
295         return ret;
296 }
297 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
298
299 static void run_workqueue(struct cpu_workqueue_struct *cwq)
300 {
301         unsigned long flags;
302
303         /*
304          * Keep taking off work from the queue until
305          * done.
306          */
307         spin_lock_irqsave(&cwq->lock, flags);
308         cwq->run_depth++;
309         if (cwq->run_depth > 3) {
310                 /* morton gets to eat his hat */
311                 printk("%s: recursion depth exceeded: %d\n",
312                         __FUNCTION__, cwq->run_depth);
313                 dump_stack();
314         }
315         while (!list_empty(&cwq->worklist)) {
316                 struct work_struct *work = list_entry(cwq->worklist.next,
317                                                 struct work_struct, entry);
318                 work_func_t f = work->func;
319
320                 cwq->current_work = work;
321                 list_del_init(cwq->worklist.next);
322                 spin_unlock_irqrestore(&cwq->lock, flags);
323
324                 BUG_ON(get_wq_data(work) != cwq);
325                 if (!test_bit(WORK_STRUCT_NOAUTOREL, work_data_bits(work)))
326                         work_release(work);
327                 f(work);
328
329                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
330                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
331                                         "%s/0x%08x/%d\n",
332                                         current->comm, preempt_count(),
333                                         current->pid);
334                         printk(KERN_ERR "    last function: ");
335                         print_symbol("%s\n", (unsigned long)f);
336                         debug_show_held_locks(current);
337                         dump_stack();
338                 }
339
340                 spin_lock_irqsave(&cwq->lock, flags);
341                 cwq->current_work = NULL;
342         }
343         cwq->run_depth--;
344         spin_unlock_irqrestore(&cwq->lock, flags);
345 }
346
347 static int worker_thread(void *__cwq)
348 {
349         struct cpu_workqueue_struct *cwq = __cwq;
350         DECLARE_WAITQUEUE(wait, current);
351         struct k_sigaction sa;
352         sigset_t blocked;
353
354         if (!cwq->wq->freezeable)
355                 current->flags |= PF_NOFREEZE;
356
357         set_user_nice(current, -5);
358
359         /* Block and flush all signals */
360         sigfillset(&blocked);
361         sigprocmask(SIG_BLOCK, &blocked, NULL);
362         flush_signals(current);
363
364         /*
365          * We inherited MPOL_INTERLEAVE from the booting kernel.
366          * Set MPOL_DEFAULT to insure node local allocations.
367          */
368         numa_default_policy();
369
370         /* SIG_IGN makes children autoreap: see do_notify_parent(). */
371         sa.sa.sa_handler = SIG_IGN;
372         sa.sa.sa_flags = 0;
373         siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
374         do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
375
376         set_current_state(TASK_INTERRUPTIBLE);
377         while (!kthread_should_stop()) {
378                 if (cwq->wq->freezeable)
379                         try_to_freeze();
380
381                 add_wait_queue(&cwq->more_work, &wait);
382                 if (list_empty(&cwq->worklist))
383                         schedule();
384                 else
385                         __set_current_state(TASK_RUNNING);
386                 remove_wait_queue(&cwq->more_work, &wait);
387
388                 if (!list_empty(&cwq->worklist))
389                         run_workqueue(cwq);
390                 set_current_state(TASK_INTERRUPTIBLE);
391         }
392         __set_current_state(TASK_RUNNING);
393         return 0;
394 }
395
396 struct wq_barrier {
397         struct work_struct      work;
398         struct completion       done;
399 };
400
401 static void wq_barrier_func(struct work_struct *work)
402 {
403         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
404         complete(&barr->done);
405 }
406
407 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
408                                         struct wq_barrier *barr, int tail)
409 {
410         INIT_WORK(&barr->work, wq_barrier_func);
411         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
412
413         init_completion(&barr->done);
414
415         insert_work(cwq, &barr->work, tail);
416 }
417
418 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
419 {
420         if (cwq->thread == current) {
421                 /*
422                  * Probably keventd trying to flush its own queue. So simply run
423                  * it by hand rather than deadlocking.
424                  */
425                 run_workqueue(cwq);
426         } else {
427                 struct wq_barrier barr;
428                 int active = 0;
429
430                 spin_lock_irq(&cwq->lock);
431                 if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
432                         insert_wq_barrier(cwq, &barr, 1);
433                         active = 1;
434                 }
435                 spin_unlock_irq(&cwq->lock);
436
437                 if (active)
438                         wait_for_completion(&barr.done);
439         }
440 }
441
442 /**
443  * flush_workqueue - ensure that any scheduled work has run to completion.
444  * @wq: workqueue to flush
445  *
446  * Forces execution of the workqueue and blocks until its completion.
447  * This is typically used in driver shutdown handlers.
448  *
449  * We sleep until all works which were queued on entry have been handled,
450  * but we are not livelocked by new incoming ones.
451  *
452  * This function used to run the workqueues itself.  Now we just wait for the
453  * helper threads to do it.
454  */
455 void fastcall flush_workqueue(struct workqueue_struct *wq)
456 {
457         if (is_single_threaded(wq)) {
458                 /* Always use first cpu's area. */
459                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, singlethread_cpu));
460         } else {
461                 long sequence;
462                 int cpu;
463 again:
464                 sequence = migrate_sequence;
465
466                 for_each_possible_cpu(cpu)
467                         flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
468
469                 if (unlikely(sequence != migrate_sequence))
470                         goto again;
471         }
472 }
473 EXPORT_SYMBOL_GPL(flush_workqueue);
474
475 static void wait_on_work(struct cpu_workqueue_struct *cwq,
476                                 struct work_struct *work)
477 {
478         struct wq_barrier barr;
479         int running = 0;
480
481         spin_lock_irq(&cwq->lock);
482         if (unlikely(cwq->current_work == work)) {
483                 insert_wq_barrier(cwq, &barr, 0);
484                 running = 1;
485         }
486         spin_unlock_irq(&cwq->lock);
487
488         if (unlikely(running)) {
489                 mutex_unlock(&workqueue_mutex);
490                 wait_for_completion(&barr.done);
491                 mutex_lock(&workqueue_mutex);
492         }
493 }
494
495 /**
496  * flush_work - block until a work_struct's callback has terminated
497  * @wq: the workqueue on which the work is queued
498  * @work: the work which is to be flushed
499  *
500  * flush_work() will attempt to cancel the work if it is queued.  If the work's
501  * callback appears to be running, flush_work() will block until it has
502  * completed.
503  *
504  * flush_work() is designed to be used when the caller is tearing down data
505  * structures which the callback function operates upon.  It is expected that,
506  * prior to calling flush_work(), the caller has arranged for the work to not
507  * be requeued.
508  */
509 void flush_work(struct workqueue_struct *wq, struct work_struct *work)
510 {
511         struct cpu_workqueue_struct *cwq;
512
513         mutex_lock(&workqueue_mutex);
514         cwq = get_wq_data(work);
515         /* Was it ever queued ? */
516         if (!cwq)
517                 goto out;
518
519         /*
520          * This work can't be re-queued, and the lock above protects us
521          * from take_over_work(), no need to re-check that get_wq_data()
522          * is still the same when we take cwq->lock.
523          */
524         spin_lock_irq(&cwq->lock);
525         list_del_init(&work->entry);
526         work_release(work);
527         spin_unlock_irq(&cwq->lock);
528
529         if (is_single_threaded(wq)) {
530                 /* Always use first cpu's area. */
531                 wait_on_work(per_cpu_ptr(wq->cpu_wq, singlethread_cpu), work);
532         } else {
533                 int cpu;
534
535                 for_each_online_cpu(cpu)
536                         wait_on_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
537         }
538 out:
539         mutex_unlock(&workqueue_mutex);
540 }
541 EXPORT_SYMBOL_GPL(flush_work);
542
543 static void init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
544 {
545         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
546
547         cwq->wq = wq;
548         spin_lock_init(&cwq->lock);
549         INIT_LIST_HEAD(&cwq->worklist);
550         init_waitqueue_head(&cwq->more_work);
551 }
552
553 static struct task_struct *create_workqueue_thread(struct workqueue_struct *wq,
554                                                         int cpu)
555 {
556         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
557         struct task_struct *p;
558
559         if (is_single_threaded(wq))
560                 p = kthread_create(worker_thread, cwq, "%s", wq->name);
561         else
562                 p = kthread_create(worker_thread, cwq, "%s/%d", wq->name, cpu);
563         if (IS_ERR(p))
564                 return NULL;
565         cwq->thread = p;
566         return p;
567 }
568
569 struct workqueue_struct *__create_workqueue(const char *name,
570                                             int singlethread, int freezeable)
571 {
572         int cpu, destroy = 0;
573         struct workqueue_struct *wq;
574         struct task_struct *p;
575
576         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
577         if (!wq)
578                 return NULL;
579
580         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
581         if (!wq->cpu_wq) {
582                 kfree(wq);
583                 return NULL;
584         }
585
586         wq->name = name;
587         wq->freezeable = freezeable;
588
589         mutex_lock(&workqueue_mutex);
590         if (singlethread) {
591                 INIT_LIST_HEAD(&wq->list);
592                 init_cpu_workqueue(wq, singlethread_cpu);
593                 p = create_workqueue_thread(wq, singlethread_cpu);
594                 if (!p)
595                         destroy = 1;
596                 else
597                         wake_up_process(p);
598         } else {
599                 list_add(&wq->list, &workqueues);
600                 for_each_possible_cpu(cpu) {
601                         init_cpu_workqueue(wq, cpu);
602                         if (!cpu_online(cpu))
603                                 continue;
604
605                         p = create_workqueue_thread(wq, cpu);
606                         if (p) {
607                                 kthread_bind(p, cpu);
608                                 wake_up_process(p);
609                         } else
610                                 destroy = 1;
611                 }
612         }
613         mutex_unlock(&workqueue_mutex);
614
615         /*
616          * Was there any error during startup? If yes then clean up:
617          */
618         if (destroy) {
619                 destroy_workqueue(wq);
620                 wq = NULL;
621         }
622         return wq;
623 }
624 EXPORT_SYMBOL_GPL(__create_workqueue);
625
626 static void cleanup_workqueue_thread(struct workqueue_struct *wq, int cpu)
627 {
628         struct cpu_workqueue_struct *cwq;
629         unsigned long flags;
630         struct task_struct *p;
631
632         cwq = per_cpu_ptr(wq->cpu_wq, cpu);
633         spin_lock_irqsave(&cwq->lock, flags);
634         p = cwq->thread;
635         cwq->thread = NULL;
636         spin_unlock_irqrestore(&cwq->lock, flags);
637         if (p)
638                 kthread_stop(p);
639 }
640
641 /**
642  * destroy_workqueue - safely terminate a workqueue
643  * @wq: target workqueue
644  *
645  * Safely destroy a workqueue. All work currently pending will be done first.
646  */
647 void destroy_workqueue(struct workqueue_struct *wq)
648 {
649         int cpu;
650
651         flush_workqueue(wq);
652
653         /* We don't need the distraction of CPUs appearing and vanishing. */
654         mutex_lock(&workqueue_mutex);
655         if (is_single_threaded(wq))
656                 cleanup_workqueue_thread(wq, singlethread_cpu);
657         else {
658                 for_each_online_cpu(cpu)
659                         cleanup_workqueue_thread(wq, cpu);
660                 list_del(&wq->list);
661         }
662         mutex_unlock(&workqueue_mutex);
663         free_percpu(wq->cpu_wq);
664         kfree(wq);
665 }
666 EXPORT_SYMBOL_GPL(destroy_workqueue);
667
668 static struct workqueue_struct *keventd_wq;
669
670 /**
671  * schedule_work - put work task in global workqueue
672  * @work: job to be done
673  *
674  * This puts a job in the kernel-global workqueue.
675  */
676 int fastcall schedule_work(struct work_struct *work)
677 {
678         return queue_work(keventd_wq, work);
679 }
680 EXPORT_SYMBOL(schedule_work);
681
682 /**
683  * schedule_delayed_work - put work task in global workqueue after delay
684  * @dwork: job to be done
685  * @delay: number of jiffies to wait or 0 for immediate execution
686  *
687  * After waiting for a given time this puts a job in the kernel-global
688  * workqueue.
689  */
690 int fastcall schedule_delayed_work(struct delayed_work *dwork,
691                                         unsigned long delay)
692 {
693         timer_stats_timer_set_start_info(&dwork->timer);
694         return queue_delayed_work(keventd_wq, dwork, delay);
695 }
696 EXPORT_SYMBOL(schedule_delayed_work);
697
698 /**
699  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
700  * @cpu: cpu to use
701  * @dwork: job to be done
702  * @delay: number of jiffies to wait
703  *
704  * After waiting for a given time this puts a job in the kernel-global
705  * workqueue on the specified CPU.
706  */
707 int schedule_delayed_work_on(int cpu,
708                         struct delayed_work *dwork, unsigned long delay)
709 {
710         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
711 }
712 EXPORT_SYMBOL(schedule_delayed_work_on);
713
714 /**
715  * schedule_on_each_cpu - call a function on each online CPU from keventd
716  * @func: the function to call
717  *
718  * Returns zero on success.
719  * Returns -ve errno on failure.
720  *
721  * Appears to be racy against CPU hotplug.
722  *
723  * schedule_on_each_cpu() is very slow.
724  */
725 int schedule_on_each_cpu(work_func_t func)
726 {
727         int cpu;
728         struct work_struct *works;
729
730         works = alloc_percpu(struct work_struct);
731         if (!works)
732                 return -ENOMEM;
733
734         preempt_disable();              /* CPU hotplug */
735         for_each_online_cpu(cpu) {
736                 struct work_struct *work = per_cpu_ptr(works, cpu);
737
738                 INIT_WORK(work, func);
739                 set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
740                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
741         }
742         preempt_enable();
743         flush_workqueue(keventd_wq);
744         free_percpu(works);
745         return 0;
746 }
747
748 void flush_scheduled_work(void)
749 {
750         flush_workqueue(keventd_wq);
751 }
752 EXPORT_SYMBOL(flush_scheduled_work);
753
754 void flush_work_keventd(struct work_struct *work)
755 {
756         flush_work(keventd_wq, work);
757 }
758 EXPORT_SYMBOL(flush_work_keventd);
759
760 /**
761  * cancel_rearming_delayed_workqueue - reliably kill off a delayed work whose handler rearms the delayed work.
762  * @wq:   the controlling workqueue structure
763  * @dwork: the delayed work struct
764  */
765 void cancel_rearming_delayed_workqueue(struct workqueue_struct *wq,
766                                        struct delayed_work *dwork)
767 {
768         while (!cancel_delayed_work(dwork))
769                 flush_workqueue(wq);
770 }
771 EXPORT_SYMBOL(cancel_rearming_delayed_workqueue);
772
773 /**
774  * cancel_rearming_delayed_work - reliably kill off a delayed keventd work whose handler rearms the delayed work.
775  * @dwork: the delayed work struct
776  */
777 void cancel_rearming_delayed_work(struct delayed_work *dwork)
778 {
779         cancel_rearming_delayed_workqueue(keventd_wq, dwork);
780 }
781 EXPORT_SYMBOL(cancel_rearming_delayed_work);
782
783 /**
784  * execute_in_process_context - reliably execute the routine with user context
785  * @fn:         the function to execute
786  * @ew:         guaranteed storage for the execute work structure (must
787  *              be available when the work executes)
788  *
789  * Executes the function immediately if process context is available,
790  * otherwise schedules the function for delayed execution.
791  *
792  * Returns:     0 - function was executed
793  *              1 - function was scheduled for execution
794  */
795 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
796 {
797         if (!in_interrupt()) {
798                 fn(&ew->work);
799                 return 0;
800         }
801
802         INIT_WORK(&ew->work, fn);
803         schedule_work(&ew->work);
804
805         return 1;
806 }
807 EXPORT_SYMBOL_GPL(execute_in_process_context);
808
809 int keventd_up(void)
810 {
811         return keventd_wq != NULL;
812 }
813
814 int current_is_keventd(void)
815 {
816         struct cpu_workqueue_struct *cwq;
817         int cpu = smp_processor_id();   /* preempt-safe: keventd is per-cpu */
818         int ret = 0;
819
820         BUG_ON(!keventd_wq);
821
822         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
823         if (current == cwq->thread)
824                 ret = 1;
825
826         return ret;
827
828 }
829
830 /* Take the work from this (downed) CPU. */
831 static void take_over_work(struct workqueue_struct *wq, unsigned int cpu)
832 {
833         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
834         struct list_head list;
835         struct work_struct *work;
836
837         spin_lock_irq(&cwq->lock);
838         list_replace_init(&cwq->worklist, &list);
839         migrate_sequence++;
840
841         while (!list_empty(&list)) {
842                 printk("Taking work for %s\n", wq->name);
843                 work = list_entry(list.next,struct work_struct,entry);
844                 list_del(&work->entry);
845                 __queue_work(per_cpu_ptr(wq->cpu_wq, smp_processor_id()), work);
846         }
847         spin_unlock_irq(&cwq->lock);
848 }
849
850 /* We're holding the cpucontrol mutex here */
851 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
852                                   unsigned long action,
853                                   void *hcpu)
854 {
855         unsigned int hotcpu = (unsigned long)hcpu;
856         struct workqueue_struct *wq;
857
858         switch (action) {
859         case CPU_UP_PREPARE:
860                 mutex_lock(&workqueue_mutex);
861                 /* Create a new workqueue thread for it. */
862                 list_for_each_entry(wq, &workqueues, list) {
863                         if (!create_workqueue_thread(wq, hotcpu)) {
864                                 printk("workqueue for %i failed\n", hotcpu);
865                                 return NOTIFY_BAD;
866                         }
867                 }
868                 break;
869
870         case CPU_ONLINE:
871                 /* Kick off worker threads. */
872                 list_for_each_entry(wq, &workqueues, list) {
873                         struct cpu_workqueue_struct *cwq;
874
875                         cwq = per_cpu_ptr(wq->cpu_wq, hotcpu);
876                         kthread_bind(cwq->thread, hotcpu);
877                         wake_up_process(cwq->thread);
878                 }
879                 mutex_unlock(&workqueue_mutex);
880                 break;
881
882         case CPU_UP_CANCELED:
883                 list_for_each_entry(wq, &workqueues, list) {
884                         if (!per_cpu_ptr(wq->cpu_wq, hotcpu)->thread)
885                                 continue;
886                         /* Unbind so it can run. */
887                         kthread_bind(per_cpu_ptr(wq->cpu_wq, hotcpu)->thread,
888                                      any_online_cpu(cpu_online_map));
889                         cleanup_workqueue_thread(wq, hotcpu);
890                 }
891                 mutex_unlock(&workqueue_mutex);
892                 break;
893
894         case CPU_DOWN_PREPARE:
895                 mutex_lock(&workqueue_mutex);
896                 break;
897
898         case CPU_DOWN_FAILED:
899                 mutex_unlock(&workqueue_mutex);
900                 break;
901
902         case CPU_DEAD:
903                 list_for_each_entry(wq, &workqueues, list)
904                         cleanup_workqueue_thread(wq, hotcpu);
905                 list_for_each_entry(wq, &workqueues, list)
906                         take_over_work(wq, hotcpu);
907                 mutex_unlock(&workqueue_mutex);
908                 break;
909         }
910
911         return NOTIFY_OK;
912 }
913
914 void init_workqueues(void)
915 {
916         singlethread_cpu = first_cpu(cpu_possible_map);
917         hotcpu_notifier(workqueue_cpu_callback, 0);
918         keventd_wq = create_workqueue("events");
919         BUG_ON(!keventd_wq);
920 }
921