KVM: Timer event should not unconditionally unhalt vcpu.
[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
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 #include <trace/workqueue.h>
37
38 /*
39  * The per-CPU workqueue (if single thread, we always use the first
40  * possible cpu).
41  */
42 struct cpu_workqueue_struct {
43
44         spinlock_t lock;
45
46         struct list_head worklist;
47         wait_queue_head_t more_work;
48         struct work_struct *current_work;
49
50         struct workqueue_struct *wq;
51         struct task_struct *thread;
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         struct list_head list;
61         const char *name;
62         int singlethread;
63         int freezeable;         /* Freeze threads during suspend */
64         int rt;
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 const struct cpumask *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_var_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_wq_single_threaded(struct workqueue_struct *wq)
87 {
88         return wq->singlethread;
89 }
90
91 static const struct cpumask *wq_cpu_map(struct workqueue_struct *wq)
92 {
93         return is_wq_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_wq_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 DEFINE_TRACE(workqueue_insertion);
128
129 static void insert_work(struct cpu_workqueue_struct *cwq,
130                         struct work_struct *work, struct list_head *head)
131 {
132         trace_workqueue_insertion(cwq->thread, work);
133
134         set_wq_data(work, cwq);
135         /*
136          * Ensure that we get the right work->data if we see the
137          * result of list_add() below, see try_to_grab_pending().
138          */
139         smp_wmb();
140         list_add_tail(&work->entry, head);
141         wake_up(&cwq->more_work);
142 }
143
144 static void __queue_work(struct cpu_workqueue_struct *cwq,
145                          struct work_struct *work)
146 {
147         unsigned long flags;
148
149         spin_lock_irqsave(&cwq->lock, flags);
150         insert_work(cwq, work, &cwq->worklist);
151         spin_unlock_irqrestore(&cwq->lock, flags);
152 }
153
154 /**
155  * queue_work - queue work on a workqueue
156  * @wq: workqueue to use
157  * @work: work to queue
158  *
159  * Returns 0 if @work was already on a queue, non-zero otherwise.
160  *
161  * We queue the work to the CPU on which it was submitted, but if the CPU dies
162  * it can be processed by another CPU.
163  */
164 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
165 {
166         int ret;
167
168         ret = queue_work_on(get_cpu(), wq, work);
169         put_cpu();
170
171         return ret;
172 }
173 EXPORT_SYMBOL_GPL(queue_work);
174
175 /**
176  * queue_work_on - queue work on specific cpu
177  * @cpu: CPU number to execute work on
178  * @wq: workqueue to use
179  * @work: work to queue
180  *
181  * Returns 0 if @work was already on a queue, non-zero otherwise.
182  *
183  * We queue the work to a specific CPU, the caller must ensure it
184  * can't go away.
185  */
186 int
187 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
188 {
189         int ret = 0;
190
191         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
192                 BUG_ON(!list_empty(&work->entry));
193                 __queue_work(wq_per_cpu(wq, cpu), work);
194                 ret = 1;
195         }
196         return ret;
197 }
198 EXPORT_SYMBOL_GPL(queue_work_on);
199
200 static void delayed_work_timer_fn(unsigned long __data)
201 {
202         struct delayed_work *dwork = (struct delayed_work *)__data;
203         struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
204         struct workqueue_struct *wq = cwq->wq;
205
206         __queue_work(wq_per_cpu(wq, smp_processor_id()), &dwork->work);
207 }
208
209 /**
210  * queue_delayed_work - queue work on a workqueue after delay
211  * @wq: workqueue to use
212  * @dwork: delayable work to queue
213  * @delay: number of jiffies to wait before queueing
214  *
215  * Returns 0 if @work was already on a queue, non-zero otherwise.
216  */
217 int queue_delayed_work(struct workqueue_struct *wq,
218                         struct delayed_work *dwork, unsigned long delay)
219 {
220         if (delay == 0)
221                 return queue_work(wq, &dwork->work);
222
223         return queue_delayed_work_on(-1, wq, dwork, delay);
224 }
225 EXPORT_SYMBOL_GPL(queue_delayed_work);
226
227 /**
228  * queue_delayed_work_on - queue work on specific CPU after delay
229  * @cpu: CPU number to execute work on
230  * @wq: workqueue to use
231  * @dwork: work to queue
232  * @delay: number of jiffies to wait before queueing
233  *
234  * Returns 0 if @work was already on a queue, non-zero otherwise.
235  */
236 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
237                         struct delayed_work *dwork, unsigned long delay)
238 {
239         int ret = 0;
240         struct timer_list *timer = &dwork->timer;
241         struct work_struct *work = &dwork->work;
242
243         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
244                 BUG_ON(timer_pending(timer));
245                 BUG_ON(!list_empty(&work->entry));
246
247                 timer_stats_timer_set_start_info(&dwork->timer);
248
249                 /* This stores cwq for the moment, for the timer_fn */
250                 set_wq_data(work, wq_per_cpu(wq, raw_smp_processor_id()));
251                 timer->expires = jiffies + delay;
252                 timer->data = (unsigned long)dwork;
253                 timer->function = delayed_work_timer_fn;
254
255                 if (unlikely(cpu >= 0))
256                         add_timer_on(timer, cpu);
257                 else
258                         add_timer(timer);
259                 ret = 1;
260         }
261         return ret;
262 }
263 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
264
265 DEFINE_TRACE(workqueue_execution);
266
267 static void run_workqueue(struct cpu_workqueue_struct *cwq)
268 {
269         spin_lock_irq(&cwq->lock);
270         while (!list_empty(&cwq->worklist)) {
271                 struct work_struct *work = list_entry(cwq->worklist.next,
272                                                 struct work_struct, entry);
273                 work_func_t f = work->func;
274 #ifdef CONFIG_LOCKDEP
275                 /*
276                  * It is permissible to free the struct work_struct
277                  * from inside the function that is called from it,
278                  * this we need to take into account for lockdep too.
279                  * To avoid bogus "held lock freed" warnings as well
280                  * as problems when looking into work->lockdep_map,
281                  * make a copy and use that here.
282                  */
283                 struct lockdep_map lockdep_map = work->lockdep_map;
284 #endif
285                 trace_workqueue_execution(cwq->thread, work);
286                 cwq->current_work = work;
287                 list_del_init(cwq->worklist.next);
288                 spin_unlock_irq(&cwq->lock);
289
290                 BUG_ON(get_wq_data(work) != cwq);
291                 work_clear_pending(work);
292                 lock_map_acquire(&cwq->wq->lockdep_map);
293                 lock_map_acquire(&lockdep_map);
294                 f(work);
295                 lock_map_release(&lockdep_map);
296                 lock_map_release(&cwq->wq->lockdep_map);
297
298                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
299                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
300                                         "%s/0x%08x/%d\n",
301                                         current->comm, preempt_count(),
302                                         task_pid_nr(current));
303                         printk(KERN_ERR "    last function: ");
304                         print_symbol("%s\n", (unsigned long)f);
305                         debug_show_held_locks(current);
306                         dump_stack();
307                 }
308
309                 spin_lock_irq(&cwq->lock);
310                 cwq->current_work = NULL;
311         }
312         spin_unlock_irq(&cwq->lock);
313 }
314
315 static int worker_thread(void *__cwq)
316 {
317         struct cpu_workqueue_struct *cwq = __cwq;
318         DEFINE_WAIT(wait);
319
320         if (cwq->wq->freezeable)
321                 set_freezable();
322
323         set_user_nice(current, -5);
324
325         for (;;) {
326                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
327                 if (!freezing(current) &&
328                     !kthread_should_stop() &&
329                     list_empty(&cwq->worklist))
330                         schedule();
331                 finish_wait(&cwq->more_work, &wait);
332
333                 try_to_freeze();
334
335                 if (kthread_should_stop())
336                         break;
337
338                 run_workqueue(cwq);
339         }
340
341         return 0;
342 }
343
344 struct wq_barrier {
345         struct work_struct      work;
346         struct completion       done;
347 };
348
349 static void wq_barrier_func(struct work_struct *work)
350 {
351         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
352         complete(&barr->done);
353 }
354
355 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
356                         struct wq_barrier *barr, struct list_head *head)
357 {
358         INIT_WORK(&barr->work, wq_barrier_func);
359         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
360
361         init_completion(&barr->done);
362
363         insert_work(cwq, &barr->work, head);
364 }
365
366 static int flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
367 {
368         int active = 0;
369         struct wq_barrier barr;
370
371         WARN_ON(cwq->thread == current);
372
373         spin_lock_irq(&cwq->lock);
374         if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
375                 insert_wq_barrier(cwq, &barr, &cwq->worklist);
376                 active = 1;
377         }
378         spin_unlock_irq(&cwq->lock);
379
380         if (active)
381                 wait_for_completion(&barr.done);
382
383         return active;
384 }
385
386 /**
387  * flush_workqueue - ensure that any scheduled work has run to completion.
388  * @wq: workqueue to flush
389  *
390  * Forces execution of the workqueue and blocks until its completion.
391  * This is typically used in driver shutdown handlers.
392  *
393  * We sleep until all works which were queued on entry have been handled,
394  * but we are not livelocked by new incoming ones.
395  *
396  * This function used to run the workqueues itself.  Now we just wait for the
397  * helper threads to do it.
398  */
399 void flush_workqueue(struct workqueue_struct *wq)
400 {
401         const struct cpumask *cpu_map = wq_cpu_map(wq);
402         int cpu;
403
404         might_sleep();
405         lock_map_acquire(&wq->lockdep_map);
406         lock_map_release(&wq->lockdep_map);
407         for_each_cpu(cpu, cpu_map)
408                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
409 }
410 EXPORT_SYMBOL_GPL(flush_workqueue);
411
412 /**
413  * flush_work - block until a work_struct's callback has terminated
414  * @work: the work which is to be flushed
415  *
416  * Returns false if @work has already terminated.
417  *
418  * It is expected that, prior to calling flush_work(), the caller has
419  * arranged for the work to not be requeued, otherwise it doesn't make
420  * sense to use this function.
421  */
422 int flush_work(struct work_struct *work)
423 {
424         struct cpu_workqueue_struct *cwq;
425         struct list_head *prev;
426         struct wq_barrier barr;
427
428         might_sleep();
429         cwq = get_wq_data(work);
430         if (!cwq)
431                 return 0;
432
433         lock_map_acquire(&cwq->wq->lockdep_map);
434         lock_map_release(&cwq->wq->lockdep_map);
435
436         prev = NULL;
437         spin_lock_irq(&cwq->lock);
438         if (!list_empty(&work->entry)) {
439                 /*
440                  * See the comment near try_to_grab_pending()->smp_rmb().
441                  * If it was re-queued under us we are not going to wait.
442                  */
443                 smp_rmb();
444                 if (unlikely(cwq != get_wq_data(work)))
445                         goto out;
446                 prev = &work->entry;
447         } else {
448                 if (cwq->current_work != work)
449                         goto out;
450                 prev = &cwq->worklist;
451         }
452         insert_wq_barrier(cwq, &barr, prev->next);
453 out:
454         spin_unlock_irq(&cwq->lock);
455         if (!prev)
456                 return 0;
457
458         wait_for_completion(&barr.done);
459         return 1;
460 }
461 EXPORT_SYMBOL_GPL(flush_work);
462
463 /*
464  * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
465  * so this work can't be re-armed in any way.
466  */
467 static int try_to_grab_pending(struct work_struct *work)
468 {
469         struct cpu_workqueue_struct *cwq;
470         int ret = -1;
471
472         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work)))
473                 return 0;
474
475         /*
476          * The queueing is in progress, or it is already queued. Try to
477          * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
478          */
479
480         cwq = get_wq_data(work);
481         if (!cwq)
482                 return ret;
483
484         spin_lock_irq(&cwq->lock);
485         if (!list_empty(&work->entry)) {
486                 /*
487                  * This work is queued, but perhaps we locked the wrong cwq.
488                  * In that case we must see the new value after rmb(), see
489                  * insert_work()->wmb().
490                  */
491                 smp_rmb();
492                 if (cwq == get_wq_data(work)) {
493                         list_del_init(&work->entry);
494                         ret = 1;
495                 }
496         }
497         spin_unlock_irq(&cwq->lock);
498
499         return ret;
500 }
501
502 static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq,
503                                 struct work_struct *work)
504 {
505         struct wq_barrier barr;
506         int running = 0;
507
508         spin_lock_irq(&cwq->lock);
509         if (unlikely(cwq->current_work == work)) {
510                 insert_wq_barrier(cwq, &barr, cwq->worklist.next);
511                 running = 1;
512         }
513         spin_unlock_irq(&cwq->lock);
514
515         if (unlikely(running))
516                 wait_for_completion(&barr.done);
517 }
518
519 static void wait_on_work(struct work_struct *work)
520 {
521         struct cpu_workqueue_struct *cwq;
522         struct workqueue_struct *wq;
523         const struct cpumask *cpu_map;
524         int cpu;
525
526         might_sleep();
527
528         lock_map_acquire(&work->lockdep_map);
529         lock_map_release(&work->lockdep_map);
530
531         cwq = get_wq_data(work);
532         if (!cwq)
533                 return;
534
535         wq = cwq->wq;
536         cpu_map = wq_cpu_map(wq);
537
538         for_each_cpu(cpu, cpu_map)
539                 wait_on_cpu_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
540 }
541
542 static int __cancel_work_timer(struct work_struct *work,
543                                 struct timer_list* timer)
544 {
545         int ret;
546
547         do {
548                 ret = (timer && likely(del_timer(timer)));
549                 if (!ret)
550                         ret = try_to_grab_pending(work);
551                 wait_on_work(work);
552         } while (unlikely(ret < 0));
553
554         work_clear_pending(work);
555         return ret;
556 }
557
558 /**
559  * cancel_work_sync - block until a work_struct's callback has terminated
560  * @work: the work which is to be flushed
561  *
562  * Returns true if @work was pending.
563  *
564  * cancel_work_sync() will cancel the work if it is queued. If the work's
565  * callback appears to be running, cancel_work_sync() will block until it
566  * has completed.
567  *
568  * It is possible to use this function if the work re-queues itself. It can
569  * cancel the work even if it migrates to another workqueue, however in that
570  * case it only guarantees that work->func() has completed on the last queued
571  * workqueue.
572  *
573  * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
574  * pending, otherwise it goes into a busy-wait loop until the timer expires.
575  *
576  * The caller must ensure that workqueue_struct on which this work was last
577  * queued can't be destroyed before this function returns.
578  */
579 int cancel_work_sync(struct work_struct *work)
580 {
581         return __cancel_work_timer(work, NULL);
582 }
583 EXPORT_SYMBOL_GPL(cancel_work_sync);
584
585 /**
586  * cancel_delayed_work_sync - reliably kill off a delayed work.
587  * @dwork: the delayed work struct
588  *
589  * Returns true if @dwork was pending.
590  *
591  * It is possible to use this function if @dwork rearms itself via queue_work()
592  * or queue_delayed_work(). See also the comment for cancel_work_sync().
593  */
594 int cancel_delayed_work_sync(struct delayed_work *dwork)
595 {
596         return __cancel_work_timer(&dwork->work, &dwork->timer);
597 }
598 EXPORT_SYMBOL(cancel_delayed_work_sync);
599
600 static struct workqueue_struct *keventd_wq __read_mostly;
601
602 /**
603  * schedule_work - put work task in global workqueue
604  * @work: job to be done
605  *
606  * This puts a job in the kernel-global workqueue.
607  */
608 int schedule_work(struct work_struct *work)
609 {
610         return queue_work(keventd_wq, work);
611 }
612 EXPORT_SYMBOL(schedule_work);
613
614 /*
615  * schedule_work_on - put work task on a specific cpu
616  * @cpu: cpu to put the work task on
617  * @work: job to be done
618  *
619  * This puts a job on a specific cpu
620  */
621 int schedule_work_on(int cpu, struct work_struct *work)
622 {
623         return queue_work_on(cpu, keventd_wq, work);
624 }
625 EXPORT_SYMBOL(schedule_work_on);
626
627 /**
628  * schedule_delayed_work - put work task in global workqueue after delay
629  * @dwork: job to be done
630  * @delay: number of jiffies to wait or 0 for immediate execution
631  *
632  * After waiting for a given time this puts a job in the kernel-global
633  * workqueue.
634  */
635 int schedule_delayed_work(struct delayed_work *dwork,
636                                         unsigned long delay)
637 {
638         return queue_delayed_work(keventd_wq, dwork, delay);
639 }
640 EXPORT_SYMBOL(schedule_delayed_work);
641
642 /**
643  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
644  * @cpu: cpu to use
645  * @dwork: job to be done
646  * @delay: number of jiffies to wait
647  *
648  * After waiting for a given time this puts a job in the kernel-global
649  * workqueue on the specified CPU.
650  */
651 int schedule_delayed_work_on(int cpu,
652                         struct delayed_work *dwork, unsigned long delay)
653 {
654         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
655 }
656 EXPORT_SYMBOL(schedule_delayed_work_on);
657
658 /**
659  * schedule_on_each_cpu - call a function on each online CPU from keventd
660  * @func: the function to call
661  *
662  * Returns zero on success.
663  * Returns -ve errno on failure.
664  *
665  * schedule_on_each_cpu() is very slow.
666  */
667 int schedule_on_each_cpu(work_func_t func)
668 {
669         int cpu;
670         struct work_struct *works;
671
672         works = alloc_percpu(struct work_struct);
673         if (!works)
674                 return -ENOMEM;
675
676         get_online_cpus();
677         for_each_online_cpu(cpu) {
678                 struct work_struct *work = per_cpu_ptr(works, cpu);
679
680                 INIT_WORK(work, func);
681                 schedule_work_on(cpu, work);
682         }
683         for_each_online_cpu(cpu)
684                 flush_work(per_cpu_ptr(works, cpu));
685         put_online_cpus();
686         free_percpu(works);
687         return 0;
688 }
689
690 void flush_scheduled_work(void)
691 {
692         flush_workqueue(keventd_wq);
693 }
694 EXPORT_SYMBOL(flush_scheduled_work);
695
696 /**
697  * execute_in_process_context - reliably execute the routine with user context
698  * @fn:         the function to execute
699  * @ew:         guaranteed storage for the execute work structure (must
700  *              be available when the work executes)
701  *
702  * Executes the function immediately if process context is available,
703  * otherwise schedules the function for delayed execution.
704  *
705  * Returns:     0 - function was executed
706  *              1 - function was scheduled for execution
707  */
708 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
709 {
710         if (!in_interrupt()) {
711                 fn(&ew->work);
712                 return 0;
713         }
714
715         INIT_WORK(&ew->work, fn);
716         schedule_work(&ew->work);
717
718         return 1;
719 }
720 EXPORT_SYMBOL_GPL(execute_in_process_context);
721
722 int keventd_up(void)
723 {
724         return keventd_wq != NULL;
725 }
726
727 int current_is_keventd(void)
728 {
729         struct cpu_workqueue_struct *cwq;
730         int cpu = raw_smp_processor_id(); /* preempt-safe: keventd is per-cpu */
731         int ret = 0;
732
733         BUG_ON(!keventd_wq);
734
735         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
736         if (current == cwq->thread)
737                 ret = 1;
738
739         return ret;
740
741 }
742
743 static struct cpu_workqueue_struct *
744 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
745 {
746         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
747
748         cwq->wq = wq;
749         spin_lock_init(&cwq->lock);
750         INIT_LIST_HEAD(&cwq->worklist);
751         init_waitqueue_head(&cwq->more_work);
752
753         return cwq;
754 }
755
756 DEFINE_TRACE(workqueue_creation);
757
758 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
759 {
760         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
761         struct workqueue_struct *wq = cwq->wq;
762         const char *fmt = is_wq_single_threaded(wq) ? "%s" : "%s/%d";
763         struct task_struct *p;
764
765         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
766         /*
767          * Nobody can add the work_struct to this cwq,
768          *      if (caller is __create_workqueue)
769          *              nobody should see this wq
770          *      else // caller is CPU_UP_PREPARE
771          *              cpu is not on cpu_online_map
772          * so we can abort safely.
773          */
774         if (IS_ERR(p))
775                 return PTR_ERR(p);
776         if (cwq->wq->rt)
777                 sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
778         cwq->thread = p;
779
780         trace_workqueue_creation(cwq->thread, cpu);
781
782         return 0;
783 }
784
785 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
786 {
787         struct task_struct *p = cwq->thread;
788
789         if (p != NULL) {
790                 if (cpu >= 0)
791                         kthread_bind(p, cpu);
792                 wake_up_process(p);
793         }
794 }
795
796 struct workqueue_struct *__create_workqueue_key(const char *name,
797                                                 int singlethread,
798                                                 int freezeable,
799                                                 int rt,
800                                                 struct lock_class_key *key,
801                                                 const char *lock_name)
802 {
803         struct workqueue_struct *wq;
804         struct cpu_workqueue_struct *cwq;
805         int err = 0, cpu;
806
807         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
808         if (!wq)
809                 return NULL;
810
811         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
812         if (!wq->cpu_wq) {
813                 kfree(wq);
814                 return NULL;
815         }
816
817         wq->name = name;
818         lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
819         wq->singlethread = singlethread;
820         wq->freezeable = freezeable;
821         wq->rt = rt;
822         INIT_LIST_HEAD(&wq->list);
823
824         if (singlethread) {
825                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
826                 err = create_workqueue_thread(cwq, singlethread_cpu);
827                 start_workqueue_thread(cwq, -1);
828         } else {
829                 cpu_maps_update_begin();
830                 /*
831                  * We must place this wq on list even if the code below fails.
832                  * cpu_down(cpu) can remove cpu from cpu_populated_map before
833                  * destroy_workqueue() takes the lock, in that case we leak
834                  * cwq[cpu]->thread.
835                  */
836                 spin_lock(&workqueue_lock);
837                 list_add(&wq->list, &workqueues);
838                 spin_unlock(&workqueue_lock);
839                 /*
840                  * We must initialize cwqs for each possible cpu even if we
841                  * are going to call destroy_workqueue() finally. Otherwise
842                  * cpu_up() can hit the uninitialized cwq once we drop the
843                  * lock.
844                  */
845                 for_each_possible_cpu(cpu) {
846                         cwq = init_cpu_workqueue(wq, cpu);
847                         if (err || !cpu_online(cpu))
848                                 continue;
849                         err = create_workqueue_thread(cwq, cpu);
850                         start_workqueue_thread(cwq, cpu);
851                 }
852                 cpu_maps_update_done();
853         }
854
855         if (err) {
856                 destroy_workqueue(wq);
857                 wq = NULL;
858         }
859         return wq;
860 }
861 EXPORT_SYMBOL_GPL(__create_workqueue_key);
862
863 DEFINE_TRACE(workqueue_destruction);
864
865 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq)
866 {
867         /*
868          * Our caller is either destroy_workqueue() or CPU_POST_DEAD,
869          * cpu_add_remove_lock protects cwq->thread.
870          */
871         if (cwq->thread == NULL)
872                 return;
873
874         lock_map_acquire(&cwq->wq->lockdep_map);
875         lock_map_release(&cwq->wq->lockdep_map);
876
877         flush_cpu_workqueue(cwq);
878         /*
879          * If the caller is CPU_POST_DEAD and cwq->worklist was not empty,
880          * a concurrent flush_workqueue() can insert a barrier after us.
881          * However, in that case run_workqueue() won't return and check
882          * kthread_should_stop() until it flushes all work_struct's.
883          * When ->worklist becomes empty it is safe to exit because no
884          * more work_structs can be queued on this cwq: flush_workqueue
885          * checks list_empty(), and a "normal" queue_work() can't use
886          * a dead CPU.
887          */
888         trace_workqueue_destruction(cwq->thread);
889         kthread_stop(cwq->thread);
890         cwq->thread = NULL;
891 }
892
893 /**
894  * destroy_workqueue - safely terminate a workqueue
895  * @wq: target workqueue
896  *
897  * Safely destroy a workqueue. All work currently pending will be done first.
898  */
899 void destroy_workqueue(struct workqueue_struct *wq)
900 {
901         const struct cpumask *cpu_map = wq_cpu_map(wq);
902         int cpu;
903
904         cpu_maps_update_begin();
905         spin_lock(&workqueue_lock);
906         list_del(&wq->list);
907         spin_unlock(&workqueue_lock);
908
909         for_each_cpu(cpu, cpu_map)
910                 cleanup_workqueue_thread(per_cpu_ptr(wq->cpu_wq, cpu));
911         cpu_maps_update_done();
912
913         free_percpu(wq->cpu_wq);
914         kfree(wq);
915 }
916 EXPORT_SYMBOL_GPL(destroy_workqueue);
917
918 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
919                                                 unsigned long action,
920                                                 void *hcpu)
921 {
922         unsigned int cpu = (unsigned long)hcpu;
923         struct cpu_workqueue_struct *cwq;
924         struct workqueue_struct *wq;
925         int ret = NOTIFY_OK;
926
927         action &= ~CPU_TASKS_FROZEN;
928
929         switch (action) {
930         case CPU_UP_PREPARE:
931                 cpumask_set_cpu(cpu, cpu_populated_map);
932         }
933 undo:
934         list_for_each_entry(wq, &workqueues, list) {
935                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
936
937                 switch (action) {
938                 case CPU_UP_PREPARE:
939                         if (!create_workqueue_thread(cwq, cpu))
940                                 break;
941                         printk(KERN_ERR "workqueue [%s] for %i failed\n",
942                                 wq->name, cpu);
943                         action = CPU_UP_CANCELED;
944                         ret = NOTIFY_BAD;
945                         goto undo;
946
947                 case CPU_ONLINE:
948                         start_workqueue_thread(cwq, cpu);
949                         break;
950
951                 case CPU_UP_CANCELED:
952                         start_workqueue_thread(cwq, -1);
953                 case CPU_POST_DEAD:
954                         cleanup_workqueue_thread(cwq);
955                         break;
956                 }
957         }
958
959         switch (action) {
960         case CPU_UP_CANCELED:
961         case CPU_POST_DEAD:
962                 cpumask_clear_cpu(cpu, cpu_populated_map);
963         }
964
965         return ret;
966 }
967
968 #ifdef CONFIG_SMP
969
970 struct work_for_cpu {
971         struct completion completion;
972         long (*fn)(void *);
973         void *arg;
974         long ret;
975 };
976
977 static int do_work_for_cpu(void *_wfc)
978 {
979         struct work_for_cpu *wfc = _wfc;
980         wfc->ret = wfc->fn(wfc->arg);
981         complete(&wfc->completion);
982         return 0;
983 }
984
985 /**
986  * work_on_cpu - run a function in user context on a particular cpu
987  * @cpu: the cpu to run on
988  * @fn: the function to run
989  * @arg: the function arg
990  *
991  * This will return the value @fn returns.
992  * It is up to the caller to ensure that the cpu doesn't go offline.
993  * The caller must not hold any locks which would prevent @fn from completing.
994  */
995 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
996 {
997         struct task_struct *sub_thread;
998         struct work_for_cpu wfc = {
999                 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
1000                 .fn = fn,
1001                 .arg = arg,
1002         };
1003
1004         sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
1005         if (IS_ERR(sub_thread))
1006                 return PTR_ERR(sub_thread);
1007         kthread_bind(sub_thread, cpu);
1008         wake_up_process(sub_thread);
1009         wait_for_completion(&wfc.completion);
1010         return wfc.ret;
1011 }
1012 EXPORT_SYMBOL_GPL(work_on_cpu);
1013 #endif /* CONFIG_SMP */
1014
1015 void __init init_workqueues(void)
1016 {
1017         alloc_cpumask_var(&cpu_populated_map, GFP_KERNEL);
1018
1019         cpumask_copy(cpu_populated_map, cpu_online_mask);
1020         singlethread_cpu = cpumask_first(cpu_possible_mask);
1021         cpu_singlethread_map = cpumask_of(singlethread_cpu);
1022         hotcpu_notifier(workqueue_cpu_callback, 0);
1023         keventd_wq = create_workqueue("events");
1024         BUG_ON(!keventd_wq);
1025 }