lockdep: map_acquire
[linux-2.6.git] / kernel / workqueue.c
1 /*
2  * linux/kernel/workqueue.c
3  *
4  * Generic mechanism for defining kernel helper threads for running
5  * arbitrary tasks in process context.
6  *
7  * Started by Ingo Molnar, Copyright (C) 2002
8  *
9  * Derived from the taskqueue/keventd code by:
10  *
11  *   David Woodhouse <dwmw2@infradead.org>
12  *   Andrew Morton <andrewm@uow.edu.au>
13  *   Kai Petzke <wpp@marie.physik.tu-berlin.de>
14  *   Theodore Ts'o <tytso@mit.edu>
15  *
16  * Made to use alloc_percpu by Christoph Lameter.
17  */
18
19 #include <linux/module.h>
20 #include <linux/kernel.h>
21 #include <linux/sched.h>
22 #include <linux/init.h>
23 #include <linux/signal.h>
24 #include <linux/completion.h>
25 #include <linux/workqueue.h>
26 #include <linux/slab.h>
27 #include <linux/cpu.h>
28 #include <linux/notifier.h>
29 #include <linux/kthread.h>
30 #include <linux/hardirq.h>
31 #include <linux/mempolicy.h>
32 #include <linux/freezer.h>
33 #include <linux/kallsyms.h>
34 #include <linux/debug_locks.h>
35 #include <linux/lockdep.h>
36
37 /*
38  * The per-CPU workqueue (if single thread, we always use the first
39  * possible cpu).
40  */
41 struct cpu_workqueue_struct {
42
43         spinlock_t lock;
44
45         struct list_head worklist;
46         wait_queue_head_t more_work;
47         struct work_struct *current_work;
48
49         struct workqueue_struct *wq;
50         struct task_struct *thread;
51
52         int run_depth;          /* Detect run_workqueue() recursion depth */
53 } ____cacheline_aligned;
54
55 /*
56  * The externally visible workqueue abstraction is an array of
57  * per-CPU workqueues:
58  */
59 struct workqueue_struct {
60         struct cpu_workqueue_struct *cpu_wq;
61         struct list_head list;
62         const char *name;
63         int singlethread;
64         int freezeable;         /* Freeze threads during suspend */
65 #ifdef CONFIG_LOCKDEP
66         struct lockdep_map lockdep_map;
67 #endif
68 };
69
70 /* Serializes the accesses to the list of workqueues. */
71 static DEFINE_SPINLOCK(workqueue_lock);
72 static LIST_HEAD(workqueues);
73
74 static int singlethread_cpu __read_mostly;
75 static cpumask_t cpu_singlethread_map __read_mostly;
76 /*
77  * _cpu_down() first removes CPU from cpu_online_map, then CPU_DEAD
78  * flushes cwq->worklist. This means that flush_workqueue/wait_on_work
79  * which comes in between can't use for_each_online_cpu(). We could
80  * use cpu_possible_map, the cpumask below is more a documentation
81  * than optimization.
82  */
83 static cpumask_t cpu_populated_map __read_mostly;
84
85 /* If it's single threaded, it isn't in the list of workqueues. */
86 static inline int is_single_threaded(struct workqueue_struct *wq)
87 {
88         return wq->singlethread;
89 }
90
91 static const cpumask_t *wq_cpu_map(struct workqueue_struct *wq)
92 {
93         return is_single_threaded(wq)
94                 ? &cpu_singlethread_map : &cpu_populated_map;
95 }
96
97 static
98 struct cpu_workqueue_struct *wq_per_cpu(struct workqueue_struct *wq, int cpu)
99 {
100         if (unlikely(is_single_threaded(wq)))
101                 cpu = singlethread_cpu;
102         return per_cpu_ptr(wq->cpu_wq, cpu);
103 }
104
105 /*
106  * Set the workqueue on which a work item is to be run
107  * - Must *only* be called if the pending flag is set
108  */
109 static inline void set_wq_data(struct work_struct *work,
110                                 struct cpu_workqueue_struct *cwq)
111 {
112         unsigned long new;
113
114         BUG_ON(!work_pending(work));
115
116         new = (unsigned long) cwq | (1UL << WORK_STRUCT_PENDING);
117         new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
118         atomic_long_set(&work->data, new);
119 }
120
121 static inline
122 struct cpu_workqueue_struct *get_wq_data(struct work_struct *work)
123 {
124         return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
125 }
126
127 static void insert_work(struct cpu_workqueue_struct *cwq,
128                         struct work_struct *work, struct list_head *head)
129 {
130         set_wq_data(work, cwq);
131         /*
132          * Ensure that we get the right work->data if we see the
133          * result of list_add() below, see try_to_grab_pending().
134          */
135         smp_wmb();
136         list_add_tail(&work->entry, head);
137         wake_up(&cwq->more_work);
138 }
139
140 static void __queue_work(struct cpu_workqueue_struct *cwq,
141                          struct work_struct *work)
142 {
143         unsigned long flags;
144
145         spin_lock_irqsave(&cwq->lock, flags);
146         insert_work(cwq, work, &cwq->worklist);
147         spin_unlock_irqrestore(&cwq->lock, flags);
148 }
149
150 /**
151  * queue_work - queue work on a workqueue
152  * @wq: workqueue to use
153  * @work: work to queue
154  *
155  * Returns 0 if @work was already on a queue, non-zero otherwise.
156  *
157  * We queue the work to the CPU on which it was submitted, but if the CPU dies
158  * it can be processed by another CPU.
159  */
160 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
161 {
162         int ret;
163
164         ret = queue_work_on(get_cpu(), wq, work);
165         put_cpu();
166
167         return ret;
168 }
169 EXPORT_SYMBOL_GPL(queue_work);
170
171 /**
172  * queue_work_on - queue work on specific cpu
173  * @cpu: CPU number to execute work on
174  * @wq: workqueue to use
175  * @work: work to queue
176  *
177  * Returns 0 if @work was already on a queue, non-zero otherwise.
178  *
179  * We queue the work to a specific CPU, the caller must ensure it
180  * can't go away.
181  */
182 int
183 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
184 {
185         int ret = 0;
186
187         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
188                 BUG_ON(!list_empty(&work->entry));
189                 __queue_work(wq_per_cpu(wq, cpu), work);
190                 ret = 1;
191         }
192         return ret;
193 }
194 EXPORT_SYMBOL_GPL(queue_work_on);
195
196 static void delayed_work_timer_fn(unsigned long __data)
197 {
198         struct delayed_work *dwork = (struct delayed_work *)__data;
199         struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
200         struct workqueue_struct *wq = cwq->wq;
201
202         __queue_work(wq_per_cpu(wq, smp_processor_id()), &dwork->work);
203 }
204
205 /**
206  * queue_delayed_work - queue work on a workqueue after delay
207  * @wq: workqueue to use
208  * @dwork: delayable 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(struct workqueue_struct *wq,
214                         struct delayed_work *dwork, unsigned long delay)
215 {
216         if (delay == 0)
217                 return queue_work(wq, &dwork->work);
218
219         return queue_delayed_work_on(-1, wq, dwork, delay);
220 }
221 EXPORT_SYMBOL_GPL(queue_delayed_work);
222
223 /**
224  * queue_delayed_work_on - queue work on specific CPU after delay
225  * @cpu: CPU number to execute work on
226  * @wq: workqueue to use
227  * @dwork: work to queue
228  * @delay: number of jiffies to wait before queueing
229  *
230  * Returns 0 if @work was already on a queue, non-zero otherwise.
231  */
232 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
233                         struct delayed_work *dwork, unsigned long delay)
234 {
235         int ret = 0;
236         struct timer_list *timer = &dwork->timer;
237         struct work_struct *work = &dwork->work;
238
239         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
240                 BUG_ON(timer_pending(timer));
241                 BUG_ON(!list_empty(&work->entry));
242
243                 timer_stats_timer_set_start_info(&dwork->timer);
244
245                 /* This stores cwq for the moment, for the timer_fn */
246                 set_wq_data(work, wq_per_cpu(wq, raw_smp_processor_id()));
247                 timer->expires = jiffies + delay;
248                 timer->data = (unsigned long)dwork;
249                 timer->function = delayed_work_timer_fn;
250
251                 if (unlikely(cpu >= 0))
252                         add_timer_on(timer, cpu);
253                 else
254                         add_timer(timer);
255                 ret = 1;
256         }
257         return ret;
258 }
259 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
260
261 static void run_workqueue(struct cpu_workqueue_struct *cwq)
262 {
263         spin_lock_irq(&cwq->lock);
264         cwq->run_depth++;
265         if (cwq->run_depth > 3) {
266                 /* morton gets to eat his hat */
267                 printk("%s: recursion depth exceeded: %d\n",
268                         __func__, cwq->run_depth);
269                 dump_stack();
270         }
271         while (!list_empty(&cwq->worklist)) {
272                 struct work_struct *work = list_entry(cwq->worklist.next,
273                                                 struct work_struct, entry);
274                 work_func_t f = work->func;
275 #ifdef CONFIG_LOCKDEP
276                 /*
277                  * It is permissible to free the struct work_struct
278                  * from inside the function that is called from it,
279                  * this we need to take into account for lockdep too.
280                  * To avoid bogus "held lock freed" warnings as well
281                  * as problems when looking into work->lockdep_map,
282                  * make a copy and use that here.
283                  */
284                 struct lockdep_map lockdep_map = work->lockdep_map;
285 #endif
286
287                 cwq->current_work = work;
288                 list_del_init(cwq->worklist.next);
289                 spin_unlock_irq(&cwq->lock);
290
291                 BUG_ON(get_wq_data(work) != cwq);
292                 work_clear_pending(work);
293                 map_acquire(&cwq->wq->lockdep_map);
294                 map_acquire(&lockdep_map);
295                 f(work);
296                 map_release(&lockdep_map);
297                 map_release(&cwq->wq->lockdep_map);
298
299                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
300                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
301                                         "%s/0x%08x/%d\n",
302                                         current->comm, preempt_count(),
303                                         task_pid_nr(current));
304                         printk(KERN_ERR "    last function: ");
305                         print_symbol("%s\n", (unsigned long)f);
306                         debug_show_held_locks(current);
307                         dump_stack();
308                 }
309
310                 spin_lock_irq(&cwq->lock);
311                 cwq->current_work = NULL;
312         }
313         cwq->run_depth--;
314         spin_unlock_irq(&cwq->lock);
315 }
316
317 static int worker_thread(void *__cwq)
318 {
319         struct cpu_workqueue_struct *cwq = __cwq;
320         DEFINE_WAIT(wait);
321
322         if (cwq->wq->freezeable)
323                 set_freezable();
324
325         set_user_nice(current, -5);
326
327         for (;;) {
328                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
329                 if (!freezing(current) &&
330                     !kthread_should_stop() &&
331                     list_empty(&cwq->worklist))
332                         schedule();
333                 finish_wait(&cwq->more_work, &wait);
334
335                 try_to_freeze();
336
337                 if (kthread_should_stop())
338                         break;
339
340                 run_workqueue(cwq);
341         }
342
343         return 0;
344 }
345
346 struct wq_barrier {
347         struct work_struct      work;
348         struct completion       done;
349 };
350
351 static void wq_barrier_func(struct work_struct *work)
352 {
353         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
354         complete(&barr->done);
355 }
356
357 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
358                         struct wq_barrier *barr, struct list_head *head)
359 {
360         INIT_WORK(&barr->work, wq_barrier_func);
361         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
362
363         init_completion(&barr->done);
364
365         insert_work(cwq, &barr->work, head);
366 }
367
368 static int flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
369 {
370         int active;
371
372         if (cwq->thread == current) {
373                 /*
374                  * Probably keventd trying to flush its own queue. So simply run
375                  * it by hand rather than deadlocking.
376                  */
377                 run_workqueue(cwq);
378                 active = 1;
379         } else {
380                 struct wq_barrier barr;
381
382                 active = 0;
383                 spin_lock_irq(&cwq->lock);
384                 if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
385                         insert_wq_barrier(cwq, &barr, &cwq->worklist);
386                         active = 1;
387                 }
388                 spin_unlock_irq(&cwq->lock);
389
390                 if (active)
391                         wait_for_completion(&barr.done);
392         }
393
394         return active;
395 }
396
397 /**
398  * flush_workqueue - ensure that any scheduled work has run to completion.
399  * @wq: workqueue to flush
400  *
401  * Forces execution of the workqueue and blocks until its completion.
402  * This is typically used in driver shutdown handlers.
403  *
404  * We sleep until all works which were queued on entry have been handled,
405  * but we are not livelocked by new incoming ones.
406  *
407  * This function used to run the workqueues itself.  Now we just wait for the
408  * helper threads to do it.
409  */
410 void flush_workqueue(struct workqueue_struct *wq)
411 {
412         const cpumask_t *cpu_map = wq_cpu_map(wq);
413         int cpu;
414
415         might_sleep();
416         map_acquire(&wq->lockdep_map);
417         map_release(&wq->lockdep_map);
418         for_each_cpu_mask_nr(cpu, *cpu_map)
419                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
420 }
421 EXPORT_SYMBOL_GPL(flush_workqueue);
422
423 /**
424  * flush_work - block until a work_struct's callback has terminated
425  * @work: the work which is to be flushed
426  *
427  * Returns false if @work has already terminated.
428  *
429  * It is expected that, prior to calling flush_work(), the caller has
430  * arranged for the work to not be requeued, otherwise it doesn't make
431  * sense to use this function.
432  */
433 int flush_work(struct work_struct *work)
434 {
435         struct cpu_workqueue_struct *cwq;
436         struct list_head *prev;
437         struct wq_barrier barr;
438
439         might_sleep();
440         cwq = get_wq_data(work);
441         if (!cwq)
442                 return 0;
443
444         map_acquire(&cwq->wq->lockdep_map);
445         map_release(&cwq->wq->lockdep_map);
446
447         prev = NULL;
448         spin_lock_irq(&cwq->lock);
449         if (!list_empty(&work->entry)) {
450                 /*
451                  * See the comment near try_to_grab_pending()->smp_rmb().
452                  * If it was re-queued under us we are not going to wait.
453                  */
454                 smp_rmb();
455                 if (unlikely(cwq != get_wq_data(work)))
456                         goto out;
457                 prev = &work->entry;
458         } else {
459                 if (cwq->current_work != work)
460                         goto out;
461                 prev = &cwq->worklist;
462         }
463         insert_wq_barrier(cwq, &barr, prev->next);
464 out:
465         spin_unlock_irq(&cwq->lock);
466         if (!prev)
467                 return 0;
468
469         wait_for_completion(&barr.done);
470         return 1;
471 }
472 EXPORT_SYMBOL_GPL(flush_work);
473
474 /*
475  * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
476  * so this work can't be re-armed in any way.
477  */
478 static int try_to_grab_pending(struct work_struct *work)
479 {
480         struct cpu_workqueue_struct *cwq;
481         int ret = -1;
482
483         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work)))
484                 return 0;
485
486         /*
487          * The queueing is in progress, or it is already queued. Try to
488          * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
489          */
490
491         cwq = get_wq_data(work);
492         if (!cwq)
493                 return ret;
494
495         spin_lock_irq(&cwq->lock);
496         if (!list_empty(&work->entry)) {
497                 /*
498                  * This work is queued, but perhaps we locked the wrong cwq.
499                  * In that case we must see the new value after rmb(), see
500                  * insert_work()->wmb().
501                  */
502                 smp_rmb();
503                 if (cwq == get_wq_data(work)) {
504                         list_del_init(&work->entry);
505                         ret = 1;
506                 }
507         }
508         spin_unlock_irq(&cwq->lock);
509
510         return ret;
511 }
512
513 static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq,
514                                 struct work_struct *work)
515 {
516         struct wq_barrier barr;
517         int running = 0;
518
519         spin_lock_irq(&cwq->lock);
520         if (unlikely(cwq->current_work == work)) {
521                 insert_wq_barrier(cwq, &barr, cwq->worklist.next);
522                 running = 1;
523         }
524         spin_unlock_irq(&cwq->lock);
525
526         if (unlikely(running))
527                 wait_for_completion(&barr.done);
528 }
529
530 static void wait_on_work(struct work_struct *work)
531 {
532         struct cpu_workqueue_struct *cwq;
533         struct workqueue_struct *wq;
534         const cpumask_t *cpu_map;
535         int cpu;
536
537         might_sleep();
538
539         map_acquire(&work->lockdep_map);
540         map_release(&work->lockdep_map);
541
542         cwq = get_wq_data(work);
543         if (!cwq)
544                 return;
545
546         wq = cwq->wq;
547         cpu_map = wq_cpu_map(wq);
548
549         for_each_cpu_mask_nr(cpu, *cpu_map)
550                 wait_on_cpu_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
551 }
552
553 static int __cancel_work_timer(struct work_struct *work,
554                                 struct timer_list* timer)
555 {
556         int ret;
557
558         do {
559                 ret = (timer && likely(del_timer(timer)));
560                 if (!ret)
561                         ret = try_to_grab_pending(work);
562                 wait_on_work(work);
563         } while (unlikely(ret < 0));
564
565         work_clear_pending(work);
566         return ret;
567 }
568
569 /**
570  * cancel_work_sync - block until a work_struct's callback has terminated
571  * @work: the work which is to be flushed
572  *
573  * Returns true if @work was pending.
574  *
575  * cancel_work_sync() will cancel the work if it is queued. If the work's
576  * callback appears to be running, cancel_work_sync() will block until it
577  * has completed.
578  *
579  * It is possible to use this function if the work re-queues itself. It can
580  * cancel the work even if it migrates to another workqueue, however in that
581  * case it only guarantees that work->func() has completed on the last queued
582  * workqueue.
583  *
584  * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
585  * pending, otherwise it goes into a busy-wait loop until the timer expires.
586  *
587  * The caller must ensure that workqueue_struct on which this work was last
588  * queued can't be destroyed before this function returns.
589  */
590 int cancel_work_sync(struct work_struct *work)
591 {
592         return __cancel_work_timer(work, NULL);
593 }
594 EXPORT_SYMBOL_GPL(cancel_work_sync);
595
596 /**
597  * cancel_delayed_work_sync - reliably kill off a delayed work.
598  * @dwork: the delayed work struct
599  *
600  * Returns true if @dwork was pending.
601  *
602  * It is possible to use this function if @dwork rearms itself via queue_work()
603  * or queue_delayed_work(). See also the comment for cancel_work_sync().
604  */
605 int cancel_delayed_work_sync(struct delayed_work *dwork)
606 {
607         return __cancel_work_timer(&dwork->work, &dwork->timer);
608 }
609 EXPORT_SYMBOL(cancel_delayed_work_sync);
610
611 static struct workqueue_struct *keventd_wq __read_mostly;
612
613 /**
614  * schedule_work - put work task in global workqueue
615  * @work: job to be done
616  *
617  * This puts a job in the kernel-global workqueue.
618  */
619 int schedule_work(struct work_struct *work)
620 {
621         return queue_work(keventd_wq, work);
622 }
623 EXPORT_SYMBOL(schedule_work);
624
625 /*
626  * schedule_work_on - put work task on a specific cpu
627  * @cpu: cpu to put the work task on
628  * @work: job to be done
629  *
630  * This puts a job on a specific cpu
631  */
632 int schedule_work_on(int cpu, struct work_struct *work)
633 {
634         return queue_work_on(cpu, keventd_wq, work);
635 }
636 EXPORT_SYMBOL(schedule_work_on);
637
638 /**
639  * schedule_delayed_work - put work task in global workqueue after delay
640  * @dwork: job to be done
641  * @delay: number of jiffies to wait or 0 for immediate execution
642  *
643  * After waiting for a given time this puts a job in the kernel-global
644  * workqueue.
645  */
646 int schedule_delayed_work(struct delayed_work *dwork,
647                                         unsigned long delay)
648 {
649         return queue_delayed_work(keventd_wq, dwork, delay);
650 }
651 EXPORT_SYMBOL(schedule_delayed_work);
652
653 /**
654  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
655  * @cpu: cpu to use
656  * @dwork: job to be done
657  * @delay: number of jiffies to wait
658  *
659  * After waiting for a given time this puts a job in the kernel-global
660  * workqueue on the specified CPU.
661  */
662 int schedule_delayed_work_on(int cpu,
663                         struct delayed_work *dwork, unsigned long delay)
664 {
665         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
666 }
667 EXPORT_SYMBOL(schedule_delayed_work_on);
668
669 /**
670  * schedule_on_each_cpu - call a function on each online CPU from keventd
671  * @func: the function to call
672  *
673  * Returns zero on success.
674  * Returns -ve errno on failure.
675  *
676  * schedule_on_each_cpu() is very slow.
677  */
678 int schedule_on_each_cpu(work_func_t func)
679 {
680         int cpu;
681         struct work_struct *works;
682
683         works = alloc_percpu(struct work_struct);
684         if (!works)
685                 return -ENOMEM;
686
687         get_online_cpus();
688         for_each_online_cpu(cpu) {
689                 struct work_struct *work = per_cpu_ptr(works, cpu);
690
691                 INIT_WORK(work, func);
692                 schedule_work_on(cpu, work);
693         }
694         for_each_online_cpu(cpu)
695                 flush_work(per_cpu_ptr(works, cpu));
696         put_online_cpus();
697         free_percpu(works);
698         return 0;
699 }
700
701 void flush_scheduled_work(void)
702 {
703         flush_workqueue(keventd_wq);
704 }
705 EXPORT_SYMBOL(flush_scheduled_work);
706
707 /**
708  * execute_in_process_context - reliably execute the routine with user context
709  * @fn:         the function to execute
710  * @ew:         guaranteed storage for the execute work structure (must
711  *              be available when the work executes)
712  *
713  * Executes the function immediately if process context is available,
714  * otherwise schedules the function for delayed execution.
715  *
716  * Returns:     0 - function was executed
717  *              1 - function was scheduled for execution
718  */
719 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
720 {
721         if (!in_interrupt()) {
722                 fn(&ew->work);
723                 return 0;
724         }
725
726         INIT_WORK(&ew->work, fn);
727         schedule_work(&ew->work);
728
729         return 1;
730 }
731 EXPORT_SYMBOL_GPL(execute_in_process_context);
732
733 int keventd_up(void)
734 {
735         return keventd_wq != NULL;
736 }
737
738 int current_is_keventd(void)
739 {
740         struct cpu_workqueue_struct *cwq;
741         int cpu = raw_smp_processor_id(); /* preempt-safe: keventd is per-cpu */
742         int ret = 0;
743
744         BUG_ON(!keventd_wq);
745
746         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
747         if (current == cwq->thread)
748                 ret = 1;
749
750         return ret;
751
752 }
753
754 static struct cpu_workqueue_struct *
755 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
756 {
757         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
758
759         cwq->wq = wq;
760         spin_lock_init(&cwq->lock);
761         INIT_LIST_HEAD(&cwq->worklist);
762         init_waitqueue_head(&cwq->more_work);
763
764         return cwq;
765 }
766
767 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
768 {
769         struct workqueue_struct *wq = cwq->wq;
770         const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
771         struct task_struct *p;
772
773         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
774         /*
775          * Nobody can add the work_struct to this cwq,
776          *      if (caller is __create_workqueue)
777          *              nobody should see this wq
778          *      else // caller is CPU_UP_PREPARE
779          *              cpu is not on cpu_online_map
780          * so we can abort safely.
781          */
782         if (IS_ERR(p))
783                 return PTR_ERR(p);
784
785         cwq->thread = p;
786
787         return 0;
788 }
789
790 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
791 {
792         struct task_struct *p = cwq->thread;
793
794         if (p != NULL) {
795                 if (cpu >= 0)
796                         kthread_bind(p, cpu);
797                 wake_up_process(p);
798         }
799 }
800
801 struct workqueue_struct *__create_workqueue_key(const char *name,
802                                                 int singlethread,
803                                                 int freezeable,
804                                                 struct lock_class_key *key,
805                                                 const char *lock_name)
806 {
807         struct workqueue_struct *wq;
808         struct cpu_workqueue_struct *cwq;
809         int err = 0, cpu;
810
811         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
812         if (!wq)
813                 return NULL;
814
815         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
816         if (!wq->cpu_wq) {
817                 kfree(wq);
818                 return NULL;
819         }
820
821         wq->name = name;
822         lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
823         wq->singlethread = singlethread;
824         wq->freezeable = freezeable;
825         INIT_LIST_HEAD(&wq->list);
826
827         if (singlethread) {
828                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
829                 err = create_workqueue_thread(cwq, singlethread_cpu);
830                 start_workqueue_thread(cwq, -1);
831         } else {
832                 cpu_maps_update_begin();
833                 spin_lock(&workqueue_lock);
834                 list_add(&wq->list, &workqueues);
835                 spin_unlock(&workqueue_lock);
836
837                 for_each_possible_cpu(cpu) {
838                         cwq = init_cpu_workqueue(wq, cpu);
839                         if (err || !cpu_online(cpu))
840                                 continue;
841                         err = create_workqueue_thread(cwq, cpu);
842                         start_workqueue_thread(cwq, cpu);
843                 }
844                 cpu_maps_update_done();
845         }
846
847         if (err) {
848                 destroy_workqueue(wq);
849                 wq = NULL;
850         }
851         return wq;
852 }
853 EXPORT_SYMBOL_GPL(__create_workqueue_key);
854
855 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq)
856 {
857         /*
858          * Our caller is either destroy_workqueue() or CPU_POST_DEAD,
859          * cpu_add_remove_lock protects cwq->thread.
860          */
861         if (cwq->thread == NULL)
862                 return;
863
864         map_acquire(&cwq->wq->lockdep_map);
865         map_release(&cwq->wq->lockdep_map);
866
867         flush_cpu_workqueue(cwq);
868         /*
869          * If the caller is CPU_POST_DEAD and cwq->worklist was not empty,
870          * a concurrent flush_workqueue() can insert a barrier after us.
871          * However, in that case run_workqueue() won't return and check
872          * kthread_should_stop() until it flushes all work_struct's.
873          * When ->worklist becomes empty it is safe to exit because no
874          * more work_structs can be queued on this cwq: flush_workqueue
875          * checks list_empty(), and a "normal" queue_work() can't use
876          * a dead CPU.
877          */
878         kthread_stop(cwq->thread);
879         cwq->thread = NULL;
880 }
881
882 /**
883  * destroy_workqueue - safely terminate a workqueue
884  * @wq: target workqueue
885  *
886  * Safely destroy a workqueue. All work currently pending will be done first.
887  */
888 void destroy_workqueue(struct workqueue_struct *wq)
889 {
890         const cpumask_t *cpu_map = wq_cpu_map(wq);
891         int cpu;
892
893         cpu_maps_update_begin();
894         spin_lock(&workqueue_lock);
895         list_del(&wq->list);
896         spin_unlock(&workqueue_lock);
897
898         for_each_cpu_mask_nr(cpu, *cpu_map)
899                 cleanup_workqueue_thread(per_cpu_ptr(wq->cpu_wq, cpu));
900         cpu_maps_update_done();
901
902         free_percpu(wq->cpu_wq);
903         kfree(wq);
904 }
905 EXPORT_SYMBOL_GPL(destroy_workqueue);
906
907 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
908                                                 unsigned long action,
909                                                 void *hcpu)
910 {
911         unsigned int cpu = (unsigned long)hcpu;
912         struct cpu_workqueue_struct *cwq;
913         struct workqueue_struct *wq;
914         int ret = NOTIFY_OK;
915
916         action &= ~CPU_TASKS_FROZEN;
917
918         switch (action) {
919         case CPU_UP_PREPARE:
920                 cpu_set(cpu, cpu_populated_map);
921         }
922 undo:
923         list_for_each_entry(wq, &workqueues, list) {
924                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
925
926                 switch (action) {
927                 case CPU_UP_PREPARE:
928                         if (!create_workqueue_thread(cwq, cpu))
929                                 break;
930                         printk(KERN_ERR "workqueue [%s] for %i failed\n",
931                                 wq->name, cpu);
932                         action = CPU_UP_CANCELED;
933                         ret = NOTIFY_BAD;
934                         goto undo;
935
936                 case CPU_ONLINE:
937                         start_workqueue_thread(cwq, cpu);
938                         break;
939
940                 case CPU_UP_CANCELED:
941                         start_workqueue_thread(cwq, -1);
942                 case CPU_POST_DEAD:
943                         cleanup_workqueue_thread(cwq);
944                         break;
945                 }
946         }
947
948         switch (action) {
949         case CPU_UP_CANCELED:
950         case CPU_POST_DEAD:
951                 cpu_clear(cpu, cpu_populated_map);
952         }
953
954         return ret;
955 }
956
957 void __init init_workqueues(void)
958 {
959         cpu_populated_map = cpu_online_map;
960         singlethread_cpu = first_cpu(cpu_possible_map);
961         cpu_singlethread_map = cpumask_of_cpu(singlethread_cpu);
962         hotcpu_notifier(workqueue_cpu_callback, 0);
963         keventd_wq = create_workqueue("events");
964         BUG_ON(!keventd_wq);
965 }