workqueues: schedule_on_each_cpu: use flush_work()
[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 = 0;
163
164         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
165                 BUG_ON(!list_empty(&work->entry));
166                 __queue_work(wq_per_cpu(wq, get_cpu()), work);
167                 put_cpu();
168                 ret = 1;
169         }
170         return ret;
171 }
172 EXPORT_SYMBOL_GPL(queue_work);
173
174 /**
175  * queue_work_on - queue work on specific cpu
176  * @cpu: CPU number to execute work on
177  * @wq: workqueue to use
178  * @work: work to queue
179  *
180  * Returns 0 if @work was already on a queue, non-zero otherwise.
181  *
182  * We queue the work to a specific CPU, the caller must ensure it
183  * can't go away.
184  */
185 int
186 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
187 {
188         int ret = 0;
189
190         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
191                 BUG_ON(!list_empty(&work->entry));
192                 __queue_work(wq_per_cpu(wq, cpu), work);
193                 ret = 1;
194         }
195         return ret;
196 }
197 EXPORT_SYMBOL_GPL(queue_work_on);
198
199 static void delayed_work_timer_fn(unsigned long __data)
200 {
201         struct delayed_work *dwork = (struct delayed_work *)__data;
202         struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
203         struct workqueue_struct *wq = cwq->wq;
204
205         __queue_work(wq_per_cpu(wq, smp_processor_id()), &dwork->work);
206 }
207
208 /**
209  * queue_delayed_work - queue work on a workqueue after delay
210  * @wq: workqueue to use
211  * @dwork: delayable work to queue
212  * @delay: number of jiffies to wait before queueing
213  *
214  * Returns 0 if @work was already on a queue, non-zero otherwise.
215  */
216 int queue_delayed_work(struct workqueue_struct *wq,
217                         struct delayed_work *dwork, unsigned long delay)
218 {
219         if (delay == 0)
220                 return queue_work(wq, &dwork->work);
221
222         return queue_delayed_work_on(-1, wq, dwork, delay);
223 }
224 EXPORT_SYMBOL_GPL(queue_delayed_work);
225
226 /**
227  * queue_delayed_work_on - queue work on specific CPU after delay
228  * @cpu: CPU number to execute work on
229  * @wq: workqueue to use
230  * @dwork: work to queue
231  * @delay: number of jiffies to wait before queueing
232  *
233  * Returns 0 if @work was already on a queue, non-zero otherwise.
234  */
235 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
236                         struct delayed_work *dwork, unsigned long delay)
237 {
238         int ret = 0;
239         struct timer_list *timer = &dwork->timer;
240         struct work_struct *work = &dwork->work;
241
242         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
243                 BUG_ON(timer_pending(timer));
244                 BUG_ON(!list_empty(&work->entry));
245
246                 timer_stats_timer_set_start_info(&dwork->timer);
247
248                 /* This stores cwq for the moment, for the timer_fn */
249                 set_wq_data(work, wq_per_cpu(wq, raw_smp_processor_id()));
250                 timer->expires = jiffies + delay;
251                 timer->data = (unsigned long)dwork;
252                 timer->function = delayed_work_timer_fn;
253
254                 if (unlikely(cpu >= 0))
255                         add_timer_on(timer, cpu);
256                 else
257                         add_timer(timer);
258                 ret = 1;
259         }
260         return ret;
261 }
262 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
263
264 static void run_workqueue(struct cpu_workqueue_struct *cwq)
265 {
266         spin_lock_irq(&cwq->lock);
267         cwq->run_depth++;
268         if (cwq->run_depth > 3) {
269                 /* morton gets to eat his hat */
270                 printk("%s: recursion depth exceeded: %d\n",
271                         __func__, cwq->run_depth);
272                 dump_stack();
273         }
274         while (!list_empty(&cwq->worklist)) {
275                 struct work_struct *work = list_entry(cwq->worklist.next,
276                                                 struct work_struct, entry);
277                 work_func_t f = work->func;
278 #ifdef CONFIG_LOCKDEP
279                 /*
280                  * It is permissible to free the struct work_struct
281                  * from inside the function that is called from it,
282                  * this we need to take into account for lockdep too.
283                  * To avoid bogus "held lock freed" warnings as well
284                  * as problems when looking into work->lockdep_map,
285                  * make a copy and use that here.
286                  */
287                 struct lockdep_map lockdep_map = work->lockdep_map;
288 #endif
289
290                 cwq->current_work = work;
291                 list_del_init(cwq->worklist.next);
292                 spin_unlock_irq(&cwq->lock);
293
294                 BUG_ON(get_wq_data(work) != cwq);
295                 work_clear_pending(work);
296                 lock_acquire(&cwq->wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
297                 lock_acquire(&lockdep_map, 0, 0, 0, 2, _THIS_IP_);
298                 f(work);
299                 lock_release(&lockdep_map, 1, _THIS_IP_);
300                 lock_release(&cwq->wq->lockdep_map, 1, _THIS_IP_);
301
302                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
303                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
304                                         "%s/0x%08x/%d\n",
305                                         current->comm, preempt_count(),
306                                         task_pid_nr(current));
307                         printk(KERN_ERR "    last function: ");
308                         print_symbol("%s\n", (unsigned long)f);
309                         debug_show_held_locks(current);
310                         dump_stack();
311                 }
312
313                 spin_lock_irq(&cwq->lock);
314                 cwq->current_work = NULL;
315         }
316         cwq->run_depth--;
317         spin_unlock_irq(&cwq->lock);
318 }
319
320 static int worker_thread(void *__cwq)
321 {
322         struct cpu_workqueue_struct *cwq = __cwq;
323         DEFINE_WAIT(wait);
324
325         if (cwq->wq->freezeable)
326                 set_freezable();
327
328         set_user_nice(current, -5);
329
330         for (;;) {
331                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
332                 if (!freezing(current) &&
333                     !kthread_should_stop() &&
334                     list_empty(&cwq->worklist))
335                         schedule();
336                 finish_wait(&cwq->more_work, &wait);
337
338                 try_to_freeze();
339
340                 if (kthread_should_stop())
341                         break;
342
343                 run_workqueue(cwq);
344         }
345
346         return 0;
347 }
348
349 struct wq_barrier {
350         struct work_struct      work;
351         struct completion       done;
352 };
353
354 static void wq_barrier_func(struct work_struct *work)
355 {
356         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
357         complete(&barr->done);
358 }
359
360 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
361                         struct wq_barrier *barr, struct list_head *head)
362 {
363         INIT_WORK(&barr->work, wq_barrier_func);
364         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
365
366         init_completion(&barr->done);
367
368         insert_work(cwq, &barr->work, head);
369 }
370
371 static int flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
372 {
373         int active;
374
375         if (cwq->thread == current) {
376                 /*
377                  * Probably keventd trying to flush its own queue. So simply run
378                  * it by hand rather than deadlocking.
379                  */
380                 run_workqueue(cwq);
381                 active = 1;
382         } else {
383                 struct wq_barrier barr;
384
385                 active = 0;
386                 spin_lock_irq(&cwq->lock);
387                 if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
388                         insert_wq_barrier(cwq, &barr, &cwq->worklist);
389                         active = 1;
390                 }
391                 spin_unlock_irq(&cwq->lock);
392
393                 if (active)
394                         wait_for_completion(&barr.done);
395         }
396
397         return active;
398 }
399
400 /**
401  * flush_workqueue - ensure that any scheduled work has run to completion.
402  * @wq: workqueue to flush
403  *
404  * Forces execution of the workqueue and blocks until its completion.
405  * This is typically used in driver shutdown handlers.
406  *
407  * We sleep until all works which were queued on entry have been handled,
408  * but we are not livelocked by new incoming ones.
409  *
410  * This function used to run the workqueues itself.  Now we just wait for the
411  * helper threads to do it.
412  */
413 void flush_workqueue(struct workqueue_struct *wq)
414 {
415         const cpumask_t *cpu_map = wq_cpu_map(wq);
416         int cpu;
417
418         might_sleep();
419         lock_acquire(&wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
420         lock_release(&wq->lockdep_map, 1, _THIS_IP_);
421         for_each_cpu_mask_nr(cpu, *cpu_map)
422                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
423 }
424 EXPORT_SYMBOL_GPL(flush_workqueue);
425
426 /**
427  * flush_work - block until a work_struct's callback has terminated
428  * @work: the work which is to be flushed
429  *
430  * It is expected that, prior to calling flush_work(), the caller has
431  * arranged for the work to not be requeued, otherwise it doesn't make
432  * sense to use this function.
433  */
434 int flush_work(struct work_struct *work)
435 {
436         struct cpu_workqueue_struct *cwq;
437         struct list_head *prev;
438         struct wq_barrier barr;
439
440         might_sleep();
441         cwq = get_wq_data(work);
442         if (!cwq)
443                 return 0;
444
445         prev = NULL;
446         spin_lock_irq(&cwq->lock);
447         if (!list_empty(&work->entry)) {
448                 /*
449                  * See the comment near try_to_grab_pending()->smp_rmb().
450                  * If it was re-queued under us we are not going to wait.
451                  */
452                 smp_rmb();
453                 if (unlikely(cwq != get_wq_data(work)))
454                         goto out;
455                 prev = &work->entry;
456         } else {
457                 if (cwq->current_work != work)
458                         goto out;
459                 prev = &cwq->worklist;
460         }
461         insert_wq_barrier(cwq, &barr, prev->next);
462 out:
463         spin_unlock_irq(&cwq->lock);
464         if (!prev)
465                 return 0;
466
467         wait_for_completion(&barr.done);
468         return 1;
469 }
470 EXPORT_SYMBOL_GPL(flush_work);
471
472 /*
473  * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
474  * so this work can't be re-armed in any way.
475  */
476 static int try_to_grab_pending(struct work_struct *work)
477 {
478         struct cpu_workqueue_struct *cwq;
479         int ret = -1;
480
481         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work)))
482                 return 0;
483
484         /*
485          * The queueing is in progress, or it is already queued. Try to
486          * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
487          */
488
489         cwq = get_wq_data(work);
490         if (!cwq)
491                 return ret;
492
493         spin_lock_irq(&cwq->lock);
494         if (!list_empty(&work->entry)) {
495                 /*
496                  * This work is queued, but perhaps we locked the wrong cwq.
497                  * In that case we must see the new value after rmb(), see
498                  * insert_work()->wmb().
499                  */
500                 smp_rmb();
501                 if (cwq == get_wq_data(work)) {
502                         list_del_init(&work->entry);
503                         ret = 1;
504                 }
505         }
506         spin_unlock_irq(&cwq->lock);
507
508         return ret;
509 }
510
511 static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq,
512                                 struct work_struct *work)
513 {
514         struct wq_barrier barr;
515         int running = 0;
516
517         spin_lock_irq(&cwq->lock);
518         if (unlikely(cwq->current_work == work)) {
519                 insert_wq_barrier(cwq, &barr, cwq->worklist.next);
520                 running = 1;
521         }
522         spin_unlock_irq(&cwq->lock);
523
524         if (unlikely(running))
525                 wait_for_completion(&barr.done);
526 }
527
528 static void wait_on_work(struct work_struct *work)
529 {
530         struct cpu_workqueue_struct *cwq;
531         struct workqueue_struct *wq;
532         const cpumask_t *cpu_map;
533         int cpu;
534
535         might_sleep();
536
537         lock_acquire(&work->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
538         lock_release(&work->lockdep_map, 1, _THIS_IP_);
539
540         cwq = get_wq_data(work);
541         if (!cwq)
542                 return;
543
544         wq = cwq->wq;
545         cpu_map = wq_cpu_map(wq);
546
547         for_each_cpu_mask_nr(cpu, *cpu_map)
548                 wait_on_cpu_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
549 }
550
551 static int __cancel_work_timer(struct work_struct *work,
552                                 struct timer_list* timer)
553 {
554         int ret;
555
556         do {
557                 ret = (timer && likely(del_timer(timer)));
558                 if (!ret)
559                         ret = try_to_grab_pending(work);
560                 wait_on_work(work);
561         } while (unlikely(ret < 0));
562
563         work_clear_pending(work);
564         return ret;
565 }
566
567 /**
568  * cancel_work_sync - block until a work_struct's callback has terminated
569  * @work: the work which is to be flushed
570  *
571  * Returns true if @work was pending.
572  *
573  * cancel_work_sync() will cancel the work if it is queued. If the work's
574  * callback appears to be running, cancel_work_sync() will block until it
575  * has completed.
576  *
577  * It is possible to use this function if the work re-queues itself. It can
578  * cancel the work even if it migrates to another workqueue, however in that
579  * case it only guarantees that work->func() has completed on the last queued
580  * workqueue.
581  *
582  * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
583  * pending, otherwise it goes into a busy-wait loop until the timer expires.
584  *
585  * The caller must ensure that workqueue_struct on which this work was last
586  * queued can't be destroyed before this function returns.
587  */
588 int cancel_work_sync(struct work_struct *work)
589 {
590         return __cancel_work_timer(work, NULL);
591 }
592 EXPORT_SYMBOL_GPL(cancel_work_sync);
593
594 /**
595  * cancel_delayed_work_sync - reliably kill off a delayed work.
596  * @dwork: the delayed work struct
597  *
598  * Returns true if @dwork was pending.
599  *
600  * It is possible to use this function if @dwork rearms itself via queue_work()
601  * or queue_delayed_work(). See also the comment for cancel_work_sync().
602  */
603 int cancel_delayed_work_sync(struct delayed_work *dwork)
604 {
605         return __cancel_work_timer(&dwork->work, &dwork->timer);
606 }
607 EXPORT_SYMBOL(cancel_delayed_work_sync);
608
609 static struct workqueue_struct *keventd_wq __read_mostly;
610
611 /**
612  * schedule_work - put work task in global workqueue
613  * @work: job to be done
614  *
615  * This puts a job in the kernel-global workqueue.
616  */
617 int schedule_work(struct work_struct *work)
618 {
619         return queue_work(keventd_wq, work);
620 }
621 EXPORT_SYMBOL(schedule_work);
622
623 /*
624  * schedule_work_on - put work task on a specific cpu
625  * @cpu: cpu to put the work task on
626  * @work: job to be done
627  *
628  * This puts a job on a specific cpu
629  */
630 int schedule_work_on(int cpu, struct work_struct *work)
631 {
632         return queue_work_on(cpu, keventd_wq, work);
633 }
634 EXPORT_SYMBOL(schedule_work_on);
635
636 /**
637  * schedule_delayed_work - put work task in global workqueue after delay
638  * @dwork: job to be done
639  * @delay: number of jiffies to wait or 0 for immediate execution
640  *
641  * After waiting for a given time this puts a job in the kernel-global
642  * workqueue.
643  */
644 int schedule_delayed_work(struct delayed_work *dwork,
645                                         unsigned long delay)
646 {
647         return queue_delayed_work(keventd_wq, dwork, delay);
648 }
649 EXPORT_SYMBOL(schedule_delayed_work);
650
651 /**
652  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
653  * @cpu: cpu to use
654  * @dwork: job to be done
655  * @delay: number of jiffies to wait
656  *
657  * After waiting for a given time this puts a job in the kernel-global
658  * workqueue on the specified CPU.
659  */
660 int schedule_delayed_work_on(int cpu,
661                         struct delayed_work *dwork, unsigned long delay)
662 {
663         return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
664 }
665 EXPORT_SYMBOL(schedule_delayed_work_on);
666
667 /**
668  * schedule_on_each_cpu - call a function on each online CPU from keventd
669  * @func: the function to call
670  *
671  * Returns zero on success.
672  * Returns -ve errno on failure.
673  *
674  * schedule_on_each_cpu() is very slow.
675  */
676 int schedule_on_each_cpu(work_func_t func)
677 {
678         int cpu;
679         struct work_struct *works;
680
681         works = alloc_percpu(struct work_struct);
682         if (!works)
683                 return -ENOMEM;
684
685         get_online_cpus();
686         for_each_online_cpu(cpu) {
687                 struct work_struct *work = per_cpu_ptr(works, cpu);
688
689                 INIT_WORK(work, func);
690                 set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
691                 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
692         }
693         for_each_online_cpu(cpu)
694                 flush_work(per_cpu_ptr(works, cpu));
695         put_online_cpus();
696         free_percpu(works);
697         return 0;
698 }
699
700 void flush_scheduled_work(void)
701 {
702         flush_workqueue(keventd_wq);
703 }
704 EXPORT_SYMBOL(flush_scheduled_work);
705
706 /**
707  * execute_in_process_context - reliably execute the routine with user context
708  * @fn:         the function to execute
709  * @ew:         guaranteed storage for the execute work structure (must
710  *              be available when the work executes)
711  *
712  * Executes the function immediately if process context is available,
713  * otherwise schedules the function for delayed execution.
714  *
715  * Returns:     0 - function was executed
716  *              1 - function was scheduled for execution
717  */
718 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
719 {
720         if (!in_interrupt()) {
721                 fn(&ew->work);
722                 return 0;
723         }
724
725         INIT_WORK(&ew->work, fn);
726         schedule_work(&ew->work);
727
728         return 1;
729 }
730 EXPORT_SYMBOL_GPL(execute_in_process_context);
731
732 int keventd_up(void)
733 {
734         return keventd_wq != NULL;
735 }
736
737 int current_is_keventd(void)
738 {
739         struct cpu_workqueue_struct *cwq;
740         int cpu = raw_smp_processor_id(); /* preempt-safe: keventd is per-cpu */
741         int ret = 0;
742
743         BUG_ON(!keventd_wq);
744
745         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
746         if (current == cwq->thread)
747                 ret = 1;
748
749         return ret;
750
751 }
752
753 static struct cpu_workqueue_struct *
754 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
755 {
756         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
757
758         cwq->wq = wq;
759         spin_lock_init(&cwq->lock);
760         INIT_LIST_HEAD(&cwq->worklist);
761         init_waitqueue_head(&cwq->more_work);
762
763         return cwq;
764 }
765
766 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
767 {
768         struct workqueue_struct *wq = cwq->wq;
769         const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
770         struct task_struct *p;
771
772         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
773         /*
774          * Nobody can add the work_struct to this cwq,
775          *      if (caller is __create_workqueue)
776          *              nobody should see this wq
777          *      else // caller is CPU_UP_PREPARE
778          *              cpu is not on cpu_online_map
779          * so we can abort safely.
780          */
781         if (IS_ERR(p))
782                 return PTR_ERR(p);
783
784         cwq->thread = p;
785
786         return 0;
787 }
788
789 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
790 {
791         struct task_struct *p = cwq->thread;
792
793         if (p != NULL) {
794                 if (cpu >= 0)
795                         kthread_bind(p, cpu);
796                 wake_up_process(p);
797         }
798 }
799
800 struct workqueue_struct *__create_workqueue_key(const char *name,
801                                                 int singlethread,
802                                                 int freezeable,
803                                                 struct lock_class_key *key,
804                                                 const char *lock_name)
805 {
806         struct workqueue_struct *wq;
807         struct cpu_workqueue_struct *cwq;
808         int err = 0, cpu;
809
810         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
811         if (!wq)
812                 return NULL;
813
814         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
815         if (!wq->cpu_wq) {
816                 kfree(wq);
817                 return NULL;
818         }
819
820         wq->name = name;
821         lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
822         wq->singlethread = singlethread;
823         wq->freezeable = freezeable;
824         INIT_LIST_HEAD(&wq->list);
825
826         if (singlethread) {
827                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
828                 err = create_workqueue_thread(cwq, singlethread_cpu);
829                 start_workqueue_thread(cwq, -1);
830         } else {
831                 get_online_cpus();
832                 spin_lock(&workqueue_lock);
833                 list_add(&wq->list, &workqueues);
834                 spin_unlock(&workqueue_lock);
835
836                 for_each_possible_cpu(cpu) {
837                         cwq = init_cpu_workqueue(wq, cpu);
838                         if (err || !cpu_online(cpu))
839                                 continue;
840                         err = create_workqueue_thread(cwq, cpu);
841                         start_workqueue_thread(cwq, cpu);
842                 }
843                 put_online_cpus();
844         }
845
846         if (err) {
847                 destroy_workqueue(wq);
848                 wq = NULL;
849         }
850         return wq;
851 }
852 EXPORT_SYMBOL_GPL(__create_workqueue_key);
853
854 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq)
855 {
856         /*
857          * Our caller is either destroy_workqueue() or CPU_DEAD,
858          * get_online_cpus() protects cwq->thread.
859          */
860         if (cwq->thread == NULL)
861                 return;
862
863         lock_acquire(&cwq->wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
864         lock_release(&cwq->wq->lockdep_map, 1, _THIS_IP_);
865
866         flush_cpu_workqueue(cwq);
867         /*
868          * If the caller is CPU_DEAD and cwq->worklist was not empty,
869          * a concurrent flush_workqueue() can insert a barrier after us.
870          * However, in that case run_workqueue() won't return and check
871          * kthread_should_stop() until it flushes all work_struct's.
872          * When ->worklist becomes empty it is safe to exit because no
873          * more work_structs can be queued on this cwq: flush_workqueue
874          * checks list_empty(), and a "normal" queue_work() can't use
875          * a dead CPU.
876          */
877         kthread_stop(cwq->thread);
878         cwq->thread = NULL;
879 }
880
881 /**
882  * destroy_workqueue - safely terminate a workqueue
883  * @wq: target workqueue
884  *
885  * Safely destroy a workqueue. All work currently pending will be done first.
886  */
887 void destroy_workqueue(struct workqueue_struct *wq)
888 {
889         const cpumask_t *cpu_map = wq_cpu_map(wq);
890         int cpu;
891
892         get_online_cpus();
893         spin_lock(&workqueue_lock);
894         list_del(&wq->list);
895         spin_unlock(&workqueue_lock);
896
897         for_each_cpu_mask_nr(cpu, *cpu_map)
898                 cleanup_workqueue_thread(per_cpu_ptr(wq->cpu_wq, cpu));
899         put_online_cpus();
900
901         free_percpu(wq->cpu_wq);
902         kfree(wq);
903 }
904 EXPORT_SYMBOL_GPL(destroy_workqueue);
905
906 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
907                                                 unsigned long action,
908                                                 void *hcpu)
909 {
910         unsigned int cpu = (unsigned long)hcpu;
911         struct cpu_workqueue_struct *cwq;
912         struct workqueue_struct *wq;
913
914         action &= ~CPU_TASKS_FROZEN;
915
916         switch (action) {
917         case CPU_UP_PREPARE:
918                 cpu_set(cpu, cpu_populated_map);
919         }
920
921         list_for_each_entry(wq, &workqueues, list) {
922                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
923
924                 switch (action) {
925                 case CPU_UP_PREPARE:
926                         if (!create_workqueue_thread(cwq, cpu))
927                                 break;
928                         printk(KERN_ERR "workqueue [%s] for %i failed\n",
929                                 wq->name, cpu);
930                         return NOTIFY_BAD;
931
932                 case CPU_ONLINE:
933                         start_workqueue_thread(cwq, cpu);
934                         break;
935
936                 case CPU_UP_CANCELED:
937                         start_workqueue_thread(cwq, -1);
938                 case CPU_DEAD:
939                         cleanup_workqueue_thread(cwq);
940                         break;
941                 }
942         }
943
944         switch (action) {
945         case CPU_UP_CANCELED:
946         case CPU_DEAD:
947                 cpu_clear(cpu, cpu_populated_map);
948         }
949
950         return NOTIFY_OK;
951 }
952
953 void __init init_workqueues(void)
954 {
955         cpu_populated_map = cpu_online_map;
956         singlethread_cpu = first_cpu(cpu_possible_map);
957         cpu_singlethread_map = cpumask_of_cpu(singlethread_cpu);
958         hotcpu_notifier(workqueue_cpu_callback, 0);
959         keventd_wq = create_workqueue("events");
960         BUG_ON(!keventd_wq);
961 }