HWPOISON: Allow schedule_on_each_cpu() from keventd
[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 #define CREATE_TRACE_POINTS
37 #include <trace/events/workqueue.h>
38
39 /*
40  * The per-CPU workqueue (if single thread, we always use the first
41  * possible cpu).
42  */
43 struct cpu_workqueue_struct {
44
45         spinlock_t lock;
46
47         struct list_head worklist;
48         wait_queue_head_t more_work;
49         struct work_struct *current_work;
50
51         struct workqueue_struct *wq;
52         struct task_struct *thread;
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         int rt;
66 #ifdef CONFIG_LOCKDEP
67         struct lockdep_map lockdep_map;
68 #endif
69 };
70
71 /* Serializes the accesses to the list of workqueues. */
72 static DEFINE_SPINLOCK(workqueue_lock);
73 static LIST_HEAD(workqueues);
74
75 static int singlethread_cpu __read_mostly;
76 static const struct cpumask *cpu_singlethread_map __read_mostly;
77 /*
78  * _cpu_down() first removes CPU from cpu_online_map, then CPU_DEAD
79  * flushes cwq->worklist. This means that flush_workqueue/wait_on_work
80  * which comes in between can't use for_each_online_cpu(). We could
81  * use cpu_possible_map, the cpumask below is more a documentation
82  * than optimization.
83  */
84 static cpumask_var_t cpu_populated_map __read_mostly;
85
86 /* If it's single threaded, it isn't in the list of workqueues. */
87 static inline int is_wq_single_threaded(struct workqueue_struct *wq)
88 {
89         return wq->singlethread;
90 }
91
92 static const struct cpumask *wq_cpu_map(struct workqueue_struct *wq)
93 {
94         return is_wq_single_threaded(wq)
95                 ? cpu_singlethread_map : cpu_populated_map;
96 }
97
98 static
99 struct cpu_workqueue_struct *wq_per_cpu(struct workqueue_struct *wq, int cpu)
100 {
101         if (unlikely(is_wq_single_threaded(wq)))
102                 cpu = singlethread_cpu;
103         return per_cpu_ptr(wq->cpu_wq, cpu);
104 }
105
106 /*
107  * Set the workqueue on which a work item is to be run
108  * - Must *only* be called if the pending flag is set
109  */
110 static inline void set_wq_data(struct work_struct *work,
111                                 struct cpu_workqueue_struct *cwq)
112 {
113         unsigned long new;
114
115         BUG_ON(!work_pending(work));
116
117         new = (unsigned long) cwq | (1UL << WORK_STRUCT_PENDING);
118         new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
119         atomic_long_set(&work->data, new);
120 }
121
122 static inline
123 struct cpu_workqueue_struct *get_wq_data(struct work_struct *work)
124 {
125         return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
126 }
127
128 static void insert_work(struct cpu_workqueue_struct *cwq,
129                         struct work_struct *work, struct list_head *head)
130 {
131         trace_workqueue_insertion(cwq->thread, work);
132
133         set_wq_data(work, cwq);
134         /*
135          * Ensure that we get the right work->data if we see the
136          * result of list_add() below, see try_to_grab_pending().
137          */
138         smp_wmb();
139         list_add_tail(&work->entry, head);
140         wake_up(&cwq->more_work);
141 }
142
143 static void __queue_work(struct cpu_workqueue_struct *cwq,
144                          struct work_struct *work)
145 {
146         unsigned long flags;
147
148         spin_lock_irqsave(&cwq->lock, flags);
149         insert_work(cwq, work, &cwq->worklist);
150         spin_unlock_irqrestore(&cwq->lock, flags);
151 }
152
153 /**
154  * queue_work - queue work on a workqueue
155  * @wq: workqueue to use
156  * @work: work to queue
157  *
158  * Returns 0 if @work was already on a queue, non-zero otherwise.
159  *
160  * We queue the work to the CPU on which it was submitted, but if the CPU dies
161  * it can be processed by another CPU.
162  */
163 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
164 {
165         int ret;
166
167         ret = queue_work_on(get_cpu(), wq, work);
168         put_cpu();
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         while (!list_empty(&cwq->worklist)) {
268                 struct work_struct *work = list_entry(cwq->worklist.next,
269                                                 struct work_struct, entry);
270                 work_func_t f = work->func;
271 #ifdef CONFIG_LOCKDEP
272                 /*
273                  * It is permissible to free the struct work_struct
274                  * from inside the function that is called from it,
275                  * this we need to take into account for lockdep too.
276                  * To avoid bogus "held lock freed" warnings as well
277                  * as problems when looking into work->lockdep_map,
278                  * make a copy and use that here.
279                  */
280                 struct lockdep_map lockdep_map = work->lockdep_map;
281 #endif
282                 trace_workqueue_execution(cwq->thread, work);
283                 cwq->current_work = work;
284                 list_del_init(cwq->worklist.next);
285                 spin_unlock_irq(&cwq->lock);
286
287                 BUG_ON(get_wq_data(work) != cwq);
288                 work_clear_pending(work);
289                 lock_map_acquire(&cwq->wq->lockdep_map);
290                 lock_map_acquire(&lockdep_map);
291                 f(work);
292                 lock_map_release(&lockdep_map);
293                 lock_map_release(&cwq->wq->lockdep_map);
294
295                 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
296                         printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
297                                         "%s/0x%08x/%d\n",
298                                         current->comm, preempt_count(),
299                                         task_pid_nr(current));
300                         printk(KERN_ERR "    last function: ");
301                         print_symbol("%s\n", (unsigned long)f);
302                         debug_show_held_locks(current);
303                         dump_stack();
304                 }
305
306                 spin_lock_irq(&cwq->lock);
307                 cwq->current_work = NULL;
308         }
309         spin_unlock_irq(&cwq->lock);
310 }
311
312 static int worker_thread(void *__cwq)
313 {
314         struct cpu_workqueue_struct *cwq = __cwq;
315         DEFINE_WAIT(wait);
316
317         if (cwq->wq->freezeable)
318                 set_freezable();
319
320         for (;;) {
321                 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
322                 if (!freezing(current) &&
323                     !kthread_should_stop() &&
324                     list_empty(&cwq->worklist))
325                         schedule();
326                 finish_wait(&cwq->more_work, &wait);
327
328                 try_to_freeze();
329
330                 if (kthread_should_stop())
331                         break;
332
333                 run_workqueue(cwq);
334         }
335
336         return 0;
337 }
338
339 struct wq_barrier {
340         struct work_struct      work;
341         struct completion       done;
342 };
343
344 static void wq_barrier_func(struct work_struct *work)
345 {
346         struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
347         complete(&barr->done);
348 }
349
350 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
351                         struct wq_barrier *barr, struct list_head *head)
352 {
353         INIT_WORK(&barr->work, wq_barrier_func);
354         __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
355
356         init_completion(&barr->done);
357
358         insert_work(cwq, &barr->work, head);
359 }
360
361 static int flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
362 {
363         int active = 0;
364         struct wq_barrier barr;
365
366         WARN_ON(cwq->thread == current);
367
368         spin_lock_irq(&cwq->lock);
369         if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
370                 insert_wq_barrier(cwq, &barr, &cwq->worklist);
371                 active = 1;
372         }
373         spin_unlock_irq(&cwq->lock);
374
375         if (active)
376                 wait_for_completion(&barr.done);
377
378         return active;
379 }
380
381 /**
382  * flush_workqueue - ensure that any scheduled work has run to completion.
383  * @wq: workqueue to flush
384  *
385  * Forces execution of the workqueue and blocks until its completion.
386  * This is typically used in driver shutdown handlers.
387  *
388  * We sleep until all works which were queued on entry have been handled,
389  * but we are not livelocked by new incoming ones.
390  *
391  * This function used to run the workqueues itself.  Now we just wait for the
392  * helper threads to do it.
393  */
394 void flush_workqueue(struct workqueue_struct *wq)
395 {
396         const struct cpumask *cpu_map = wq_cpu_map(wq);
397         int cpu;
398
399         might_sleep();
400         lock_map_acquire(&wq->lockdep_map);
401         lock_map_release(&wq->lockdep_map);
402         for_each_cpu(cpu, cpu_map)
403                 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
404 }
405 EXPORT_SYMBOL_GPL(flush_workqueue);
406
407 /**
408  * flush_work - block until a work_struct's callback has terminated
409  * @work: the work which is to be flushed
410  *
411  * Returns false if @work has already terminated.
412  *
413  * It is expected that, prior to calling flush_work(), the caller has
414  * arranged for the work to not be requeued, otherwise it doesn't make
415  * sense to use this function.
416  */
417 int flush_work(struct work_struct *work)
418 {
419         struct cpu_workqueue_struct *cwq;
420         struct list_head *prev;
421         struct wq_barrier barr;
422
423         might_sleep();
424         cwq = get_wq_data(work);
425         if (!cwq)
426                 return 0;
427
428         lock_map_acquire(&cwq->wq->lockdep_map);
429         lock_map_release(&cwq->wq->lockdep_map);
430
431         prev = NULL;
432         spin_lock_irq(&cwq->lock);
433         if (!list_empty(&work->entry)) {
434                 /*
435                  * See the comment near try_to_grab_pending()->smp_rmb().
436                  * If it was re-queued under us we are not going to wait.
437                  */
438                 smp_rmb();
439                 if (unlikely(cwq != get_wq_data(work)))
440                         goto out;
441                 prev = &work->entry;
442         } else {
443                 if (cwq->current_work != work)
444                         goto out;
445                 prev = &cwq->worklist;
446         }
447         insert_wq_barrier(cwq, &barr, prev->next);
448 out:
449         spin_unlock_irq(&cwq->lock);
450         if (!prev)
451                 return 0;
452
453         wait_for_completion(&barr.done);
454         return 1;
455 }
456 EXPORT_SYMBOL_GPL(flush_work);
457
458 /*
459  * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
460  * so this work can't be re-armed in any way.
461  */
462 static int try_to_grab_pending(struct work_struct *work)
463 {
464         struct cpu_workqueue_struct *cwq;
465         int ret = -1;
466
467         if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work)))
468                 return 0;
469
470         /*
471          * The queueing is in progress, or it is already queued. Try to
472          * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
473          */
474
475         cwq = get_wq_data(work);
476         if (!cwq)
477                 return ret;
478
479         spin_lock_irq(&cwq->lock);
480         if (!list_empty(&work->entry)) {
481                 /*
482                  * This work is queued, but perhaps we locked the wrong cwq.
483                  * In that case we must see the new value after rmb(), see
484                  * insert_work()->wmb().
485                  */
486                 smp_rmb();
487                 if (cwq == get_wq_data(work)) {
488                         list_del_init(&work->entry);
489                         ret = 1;
490                 }
491         }
492         spin_unlock_irq(&cwq->lock);
493
494         return ret;
495 }
496
497 static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq,
498                                 struct work_struct *work)
499 {
500         struct wq_barrier barr;
501         int running = 0;
502
503         spin_lock_irq(&cwq->lock);
504         if (unlikely(cwq->current_work == work)) {
505                 insert_wq_barrier(cwq, &barr, cwq->worklist.next);
506                 running = 1;
507         }
508         spin_unlock_irq(&cwq->lock);
509
510         if (unlikely(running))
511                 wait_for_completion(&barr.done);
512 }
513
514 static void wait_on_work(struct work_struct *work)
515 {
516         struct cpu_workqueue_struct *cwq;
517         struct workqueue_struct *wq;
518         const struct cpumask *cpu_map;
519         int cpu;
520
521         might_sleep();
522
523         lock_map_acquire(&work->lockdep_map);
524         lock_map_release(&work->lockdep_map);
525
526         cwq = get_wq_data(work);
527         if (!cwq)
528                 return;
529
530         wq = cwq->wq;
531         cpu_map = wq_cpu_map(wq);
532
533         for_each_cpu(cpu, cpu_map)
534                 wait_on_cpu_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
535 }
536
537 static int __cancel_work_timer(struct work_struct *work,
538                                 struct timer_list* timer)
539 {
540         int ret;
541
542         do {
543                 ret = (timer && likely(del_timer(timer)));
544                 if (!ret)
545                         ret = try_to_grab_pending(work);
546                 wait_on_work(work);
547         } while (unlikely(ret < 0));
548
549         work_clear_pending(work);
550         return ret;
551 }
552
553 /**
554  * cancel_work_sync - block until a work_struct's callback has terminated
555  * @work: the work which is to be flushed
556  *
557  * Returns true if @work was pending.
558  *
559  * cancel_work_sync() will cancel the work if it is queued. If the work's
560  * callback appears to be running, cancel_work_sync() will block until it
561  * has completed.
562  *
563  * It is possible to use this function if the work re-queues itself. It can
564  * cancel the work even if it migrates to another workqueue, however in that
565  * case it only guarantees that work->func() has completed on the last queued
566  * workqueue.
567  *
568  * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
569  * pending, otherwise it goes into a busy-wait loop until the timer expires.
570  *
571  * The caller must ensure that workqueue_struct on which this work was last
572  * queued can't be destroyed before this function returns.
573  */
574 int cancel_work_sync(struct work_struct *work)
575 {
576         return __cancel_work_timer(work, NULL);
577 }
578 EXPORT_SYMBOL_GPL(cancel_work_sync);
579
580 /**
581  * cancel_delayed_work_sync - reliably kill off a delayed work.
582  * @dwork: the delayed work struct
583  *
584  * Returns true if @dwork was pending.
585  *
586  * It is possible to use this function if @dwork rearms itself via queue_work()
587  * or queue_delayed_work(). See also the comment for cancel_work_sync().
588  */
589 int cancel_delayed_work_sync(struct delayed_work *dwork)
590 {
591         return __cancel_work_timer(&dwork->work, &dwork->timer);
592 }
593 EXPORT_SYMBOL(cancel_delayed_work_sync);
594
595 static struct workqueue_struct *keventd_wq __read_mostly;
596
597 /**
598  * schedule_work - put work task in global workqueue
599  * @work: job to be done
600  *
601  * Returns zero if @work was already on the kernel-global workqueue and
602  * non-zero otherwise.
603  *
604  * This puts a job in the kernel-global workqueue if it was not already
605  * queued and leaves it in the same position on the kernel-global
606  * workqueue otherwise.
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         int orig = -1;
671         struct work_struct *works;
672
673         works = alloc_percpu(struct work_struct);
674         if (!works)
675                 return -ENOMEM;
676
677         /*
678          * when running in keventd don't schedule a work item on itself.
679          * Can just call directly because the work queue is already bound.
680          * This also is faster.
681          * Make this a generic parameter for other workqueues?
682          */
683         if (current_is_keventd()) {
684                 orig = raw_smp_processor_id();
685                 INIT_WORK(per_cpu_ptr(works, orig), func);
686                 func(per_cpu_ptr(works, orig));
687         }
688
689         get_online_cpus();
690         for_each_online_cpu(cpu) {
691                 struct work_struct *work = per_cpu_ptr(works, cpu);
692
693                 if (cpu == orig)
694                         continue;
695                 INIT_WORK(work, func);
696                 schedule_work_on(cpu, work);
697         }
698         for_each_online_cpu(cpu) {
699                 if (cpu != orig)
700                         flush_work(per_cpu_ptr(works, cpu));
701         }
702         put_online_cpus();
703         free_percpu(works);
704         return 0;
705 }
706
707 void flush_scheduled_work(void)
708 {
709         flush_workqueue(keventd_wq);
710 }
711 EXPORT_SYMBOL(flush_scheduled_work);
712
713 /**
714  * execute_in_process_context - reliably execute the routine with user context
715  * @fn:         the function to execute
716  * @ew:         guaranteed storage for the execute work structure (must
717  *              be available when the work executes)
718  *
719  * Executes the function immediately if process context is available,
720  * otherwise schedules the function for delayed execution.
721  *
722  * Returns:     0 - function was executed
723  *              1 - function was scheduled for execution
724  */
725 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
726 {
727         if (!in_interrupt()) {
728                 fn(&ew->work);
729                 return 0;
730         }
731
732         INIT_WORK(&ew->work, fn);
733         schedule_work(&ew->work);
734
735         return 1;
736 }
737 EXPORT_SYMBOL_GPL(execute_in_process_context);
738
739 int keventd_up(void)
740 {
741         return keventd_wq != NULL;
742 }
743
744 int current_is_keventd(void)
745 {
746         struct cpu_workqueue_struct *cwq;
747         int cpu = raw_smp_processor_id(); /* preempt-safe: keventd is per-cpu */
748         int ret = 0;
749
750         BUG_ON(!keventd_wq);
751
752         cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
753         if (current == cwq->thread)
754                 ret = 1;
755
756         return ret;
757
758 }
759
760 static struct cpu_workqueue_struct *
761 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
762 {
763         struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
764
765         cwq->wq = wq;
766         spin_lock_init(&cwq->lock);
767         INIT_LIST_HEAD(&cwq->worklist);
768         init_waitqueue_head(&cwq->more_work);
769
770         return cwq;
771 }
772
773 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
774 {
775         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
776         struct workqueue_struct *wq = cwq->wq;
777         const char *fmt = is_wq_single_threaded(wq) ? "%s" : "%s/%d";
778         struct task_struct *p;
779
780         p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
781         /*
782          * Nobody can add the work_struct to this cwq,
783          *      if (caller is __create_workqueue)
784          *              nobody should see this wq
785          *      else // caller is CPU_UP_PREPARE
786          *              cpu is not on cpu_online_map
787          * so we can abort safely.
788          */
789         if (IS_ERR(p))
790                 return PTR_ERR(p);
791         if (cwq->wq->rt)
792                 sched_setscheduler_nocheck(p, SCHED_FIFO, &param);
793         cwq->thread = p;
794
795         trace_workqueue_creation(cwq->thread, cpu);
796
797         return 0;
798 }
799
800 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
801 {
802         struct task_struct *p = cwq->thread;
803
804         if (p != NULL) {
805                 if (cpu >= 0)
806                         kthread_bind(p, cpu);
807                 wake_up_process(p);
808         }
809 }
810
811 struct workqueue_struct *__create_workqueue_key(const char *name,
812                                                 int singlethread,
813                                                 int freezeable,
814                                                 int rt,
815                                                 struct lock_class_key *key,
816                                                 const char *lock_name)
817 {
818         struct workqueue_struct *wq;
819         struct cpu_workqueue_struct *cwq;
820         int err = 0, cpu;
821
822         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
823         if (!wq)
824                 return NULL;
825
826         wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
827         if (!wq->cpu_wq) {
828                 kfree(wq);
829                 return NULL;
830         }
831
832         wq->name = name;
833         lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
834         wq->singlethread = singlethread;
835         wq->freezeable = freezeable;
836         wq->rt = rt;
837         INIT_LIST_HEAD(&wq->list);
838
839         if (singlethread) {
840                 cwq = init_cpu_workqueue(wq, singlethread_cpu);
841                 err = create_workqueue_thread(cwq, singlethread_cpu);
842                 start_workqueue_thread(cwq, -1);
843         } else {
844                 cpu_maps_update_begin();
845                 /*
846                  * We must place this wq on list even if the code below fails.
847                  * cpu_down(cpu) can remove cpu from cpu_populated_map before
848                  * destroy_workqueue() takes the lock, in that case we leak
849                  * cwq[cpu]->thread.
850                  */
851                 spin_lock(&workqueue_lock);
852                 list_add(&wq->list, &workqueues);
853                 spin_unlock(&workqueue_lock);
854                 /*
855                  * We must initialize cwqs for each possible cpu even if we
856                  * are going to call destroy_workqueue() finally. Otherwise
857                  * cpu_up() can hit the uninitialized cwq once we drop the
858                  * lock.
859                  */
860                 for_each_possible_cpu(cpu) {
861                         cwq = init_cpu_workqueue(wq, cpu);
862                         if (err || !cpu_online(cpu))
863                                 continue;
864                         err = create_workqueue_thread(cwq, cpu);
865                         start_workqueue_thread(cwq, cpu);
866                 }
867                 cpu_maps_update_done();
868         }
869
870         if (err) {
871                 destroy_workqueue(wq);
872                 wq = NULL;
873         }
874         return wq;
875 }
876 EXPORT_SYMBOL_GPL(__create_workqueue_key);
877
878 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq)
879 {
880         /*
881          * Our caller is either destroy_workqueue() or CPU_POST_DEAD,
882          * cpu_add_remove_lock protects cwq->thread.
883          */
884         if (cwq->thread == NULL)
885                 return;
886
887         lock_map_acquire(&cwq->wq->lockdep_map);
888         lock_map_release(&cwq->wq->lockdep_map);
889
890         flush_cpu_workqueue(cwq);
891         /*
892          * If the caller is CPU_POST_DEAD and cwq->worklist was not empty,
893          * a concurrent flush_workqueue() can insert a barrier after us.
894          * However, in that case run_workqueue() won't return and check
895          * kthread_should_stop() until it flushes all work_struct's.
896          * When ->worklist becomes empty it is safe to exit because no
897          * more work_structs can be queued on this cwq: flush_workqueue
898          * checks list_empty(), and a "normal" queue_work() can't use
899          * a dead CPU.
900          */
901         trace_workqueue_destruction(cwq->thread);
902         kthread_stop(cwq->thread);
903         cwq->thread = NULL;
904 }
905
906 /**
907  * destroy_workqueue - safely terminate a workqueue
908  * @wq: target workqueue
909  *
910  * Safely destroy a workqueue. All work currently pending will be done first.
911  */
912 void destroy_workqueue(struct workqueue_struct *wq)
913 {
914         const struct cpumask *cpu_map = wq_cpu_map(wq);
915         int cpu;
916
917         cpu_maps_update_begin();
918         spin_lock(&workqueue_lock);
919         list_del(&wq->list);
920         spin_unlock(&workqueue_lock);
921
922         for_each_cpu(cpu, cpu_map)
923                 cleanup_workqueue_thread(per_cpu_ptr(wq->cpu_wq, cpu));
924         cpu_maps_update_done();
925
926         free_percpu(wq->cpu_wq);
927         kfree(wq);
928 }
929 EXPORT_SYMBOL_GPL(destroy_workqueue);
930
931 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
932                                                 unsigned long action,
933                                                 void *hcpu)
934 {
935         unsigned int cpu = (unsigned long)hcpu;
936         struct cpu_workqueue_struct *cwq;
937         struct workqueue_struct *wq;
938         int ret = NOTIFY_OK;
939
940         action &= ~CPU_TASKS_FROZEN;
941
942         switch (action) {
943         case CPU_UP_PREPARE:
944                 cpumask_set_cpu(cpu, cpu_populated_map);
945         }
946 undo:
947         list_for_each_entry(wq, &workqueues, list) {
948                 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
949
950                 switch (action) {
951                 case CPU_UP_PREPARE:
952                         if (!create_workqueue_thread(cwq, cpu))
953                                 break;
954                         printk(KERN_ERR "workqueue [%s] for %i failed\n",
955                                 wq->name, cpu);
956                         action = CPU_UP_CANCELED;
957                         ret = NOTIFY_BAD;
958                         goto undo;
959
960                 case CPU_ONLINE:
961                         start_workqueue_thread(cwq, cpu);
962                         break;
963
964                 case CPU_UP_CANCELED:
965                         start_workqueue_thread(cwq, -1);
966                 case CPU_POST_DEAD:
967                         cleanup_workqueue_thread(cwq);
968                         break;
969                 }
970         }
971
972         switch (action) {
973         case CPU_UP_CANCELED:
974         case CPU_POST_DEAD:
975                 cpumask_clear_cpu(cpu, cpu_populated_map);
976         }
977
978         return ret;
979 }
980
981 #ifdef CONFIG_SMP
982
983 struct work_for_cpu {
984         struct completion completion;
985         long (*fn)(void *);
986         void *arg;
987         long ret;
988 };
989
990 static int do_work_for_cpu(void *_wfc)
991 {
992         struct work_for_cpu *wfc = _wfc;
993         wfc->ret = wfc->fn(wfc->arg);
994         complete(&wfc->completion);
995         return 0;
996 }
997
998 /**
999  * work_on_cpu - run a function in user context on a particular cpu
1000  * @cpu: the cpu to run on
1001  * @fn: the function to run
1002  * @arg: the function arg
1003  *
1004  * This will return the value @fn returns.
1005  * It is up to the caller to ensure that the cpu doesn't go offline.
1006  * The caller must not hold any locks which would prevent @fn from completing.
1007  */
1008 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
1009 {
1010         struct task_struct *sub_thread;
1011         struct work_for_cpu wfc = {
1012                 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
1013                 .fn = fn,
1014                 .arg = arg,
1015         };
1016
1017         sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
1018         if (IS_ERR(sub_thread))
1019                 return PTR_ERR(sub_thread);
1020         kthread_bind(sub_thread, cpu);
1021         wake_up_process(sub_thread);
1022         wait_for_completion(&wfc.completion);
1023         return wfc.ret;
1024 }
1025 EXPORT_SYMBOL_GPL(work_on_cpu);
1026 #endif /* CONFIG_SMP */
1027
1028 void __init init_workqueues(void)
1029 {
1030         alloc_cpumask_var(&cpu_populated_map, GFP_KERNEL);
1031
1032         cpumask_copy(cpu_populated_map, cpu_online_mask);
1033         singlethread_cpu = cpumask_first(cpu_possible_mask);
1034         cpu_singlethread_map = cpumask_of(singlethread_cpu);
1035         hotcpu_notifier(workqueue_cpu_callback, 0);
1036         keventd_wq = create_workqueue("events");
1037         BUG_ON(!keventd_wq);
1038 }