2 * linux/kernel/workqueue.c
4 * Generic mechanism for defining kernel helper threads for running
5 * arbitrary tasks in process context.
7 * Started by Ingo Molnar, Copyright (C) 2002
9 * Derived from the taskqueue/keventd code by:
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>
16 * Made to use alloc_percpu by Christoph Lameter <clameter@sgi.com>.
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>
37 * The per-CPU workqueue (if single thread, we always use the first
40 * The sequence counters are for flush_scheduled_work(). It wants to wait
41 * until all currently-scheduled works are completed, but it doesn't
42 * want to be livelocked by new, incoming ones. So it waits until
43 * remove_sequence is >= the insert_sequence which pertained when
44 * flush_scheduled_work() was called.
46 struct cpu_workqueue_struct {
50 long remove_sequence; /* Least-recently added (next to run) */
51 long insert_sequence; /* Next to add */
53 struct list_head worklist;
54 wait_queue_head_t more_work;
55 wait_queue_head_t work_done;
57 struct workqueue_struct *wq;
58 struct task_struct *thread;
60 int run_depth; /* Detect run_workqueue() recursion depth */
62 int freezeable; /* Freeze the thread during suspend */
63 } ____cacheline_aligned;
66 * The externally visible workqueue abstraction is an array of
69 struct workqueue_struct {
70 struct cpu_workqueue_struct *cpu_wq;
72 struct list_head list; /* Empty if single thread */
75 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
76 threads to each one as cpus come/go. */
77 static DEFINE_MUTEX(workqueue_mutex);
78 static LIST_HEAD(workqueues);
80 static int singlethread_cpu;
82 /* If it's single threaded, it isn't in the list of workqueues. */
83 static inline int is_single_threaded(struct workqueue_struct *wq)
85 return list_empty(&wq->list);
88 static inline void set_wq_data(struct work_struct *work, void *wq)
90 unsigned long new, old, res;
92 /* assume the pending flag is already set and that the task has already
93 * been queued on this workqueue */
94 new = (unsigned long) wq | (1UL << WORK_STRUCT_PENDING);
95 res = work->management;
99 new = (unsigned long) wq;
100 new |= (old & WORK_STRUCT_FLAG_MASK);
101 res = cmpxchg(&work->management, old, new);
102 } while (res != old);
106 static inline void *get_wq_data(struct work_struct *work)
108 return (void *) (work->management & WORK_STRUCT_WQ_DATA_MASK);
111 static int __run_work(struct cpu_workqueue_struct *cwq, struct work_struct *work)
116 spin_lock_irqsave(&cwq->lock, flags);
118 * We need to re-validate the work info after we've gotten
119 * the cpu_workqueue lock. We can run the work now iff:
121 * - the wq_data still matches the cpu_workqueue_struct
122 * - AND the work is still marked pending
123 * - AND the work is still on a list (which will be this
124 * workqueue_struct list)
126 * All these conditions are important, because we
127 * need to protect against the work being run right
128 * now on another CPU (all but the last one might be
129 * true if it's currently running and has not been
130 * released yet, for example).
132 if (get_wq_data(work) == cwq
133 && work_pending(work)
134 && !list_empty(&work->entry)) {
135 work_func_t f = work->func;
136 list_del_init(&work->entry);
137 spin_unlock_irqrestore(&cwq->lock, flags);
139 if (!test_bit(WORK_STRUCT_NOAUTOREL, &work->management))
143 spin_lock_irqsave(&cwq->lock, flags);
144 cwq->remove_sequence++;
145 wake_up(&cwq->work_done);
148 spin_unlock_irqrestore(&cwq->lock, flags);
153 * run_scheduled_work - run scheduled work synchronously
156 * This checks if the work was pending, and runs it
157 * synchronously if so. It returns a boolean to indicate
158 * whether it had any scheduled work to run or not.
160 * NOTE! This _only_ works for normal work_structs. You
161 * CANNOT use this for delayed work, because the wq data
162 * for delayed work will not point properly to the per-
163 * CPU workqueue struct, but will change!
165 int fastcall run_scheduled_work(struct work_struct *work)
168 struct cpu_workqueue_struct *cwq;
170 if (!work_pending(work))
172 if (list_empty(&work->entry))
174 /* NOTE! This depends intimately on __queue_work! */
175 cwq = get_wq_data(work);
178 if (__run_work(cwq, work))
182 EXPORT_SYMBOL(run_scheduled_work);
184 /* Preempt must be disabled. */
185 static void __queue_work(struct cpu_workqueue_struct *cwq,
186 struct work_struct *work)
190 spin_lock_irqsave(&cwq->lock, flags);
191 set_wq_data(work, cwq);
192 list_add_tail(&work->entry, &cwq->worklist);
193 cwq->insert_sequence++;
194 wake_up(&cwq->more_work);
195 spin_unlock_irqrestore(&cwq->lock, flags);
199 * queue_work - queue work on a workqueue
200 * @wq: workqueue to use
201 * @work: work to queue
203 * Returns 0 if @work was already on a queue, non-zero otherwise.
205 * We queue the work to the CPU it was submitted, but there is no
206 * guarantee that it will be processed by that CPU.
208 int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
210 int ret = 0, cpu = get_cpu();
212 if (!test_and_set_bit(WORK_STRUCT_PENDING, &work->management)) {
213 if (unlikely(is_single_threaded(wq)))
214 cpu = singlethread_cpu;
215 BUG_ON(!list_empty(&work->entry));
216 __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
222 EXPORT_SYMBOL_GPL(queue_work);
224 static void delayed_work_timer_fn(unsigned long __data)
226 struct delayed_work *dwork = (struct delayed_work *)__data;
227 struct workqueue_struct *wq = get_wq_data(&dwork->work);
228 int cpu = smp_processor_id();
230 if (unlikely(is_single_threaded(wq)))
231 cpu = singlethread_cpu;
233 __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), &dwork->work);
237 * queue_delayed_work - queue work on a workqueue after delay
238 * @wq: workqueue to use
239 * @work: delayable work to queue
240 * @delay: number of jiffies to wait before queueing
242 * Returns 0 if @work was already on a queue, non-zero otherwise.
244 int fastcall queue_delayed_work(struct workqueue_struct *wq,
245 struct delayed_work *dwork, unsigned long delay)
248 struct timer_list *timer = &dwork->timer;
249 struct work_struct *work = &dwork->work;
252 return queue_work(wq, work);
254 if (!test_and_set_bit(WORK_STRUCT_PENDING, &work->management)) {
255 BUG_ON(timer_pending(timer));
256 BUG_ON(!list_empty(&work->entry));
258 /* This stores wq for the moment, for the timer_fn */
259 set_wq_data(work, wq);
260 timer->expires = jiffies + delay;
261 timer->data = (unsigned long)dwork;
262 timer->function = delayed_work_timer_fn;
268 EXPORT_SYMBOL_GPL(queue_delayed_work);
271 * queue_delayed_work_on - queue work on specific CPU after delay
272 * @cpu: CPU number to execute work on
273 * @wq: workqueue to use
274 * @work: work to queue
275 * @delay: number of jiffies to wait before queueing
277 * Returns 0 if @work was already on a queue, non-zero otherwise.
279 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
280 struct delayed_work *dwork, unsigned long delay)
283 struct timer_list *timer = &dwork->timer;
284 struct work_struct *work = &dwork->work;
286 if (!test_and_set_bit(WORK_STRUCT_PENDING, &work->management)) {
287 BUG_ON(timer_pending(timer));
288 BUG_ON(!list_empty(&work->entry));
290 /* This stores wq for the moment, for the timer_fn */
291 set_wq_data(work, wq);
292 timer->expires = jiffies + delay;
293 timer->data = (unsigned long)dwork;
294 timer->function = delayed_work_timer_fn;
295 add_timer_on(timer, cpu);
300 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
302 static void run_workqueue(struct cpu_workqueue_struct *cwq)
307 * Keep taking off work from the queue until
310 spin_lock_irqsave(&cwq->lock, flags);
312 if (cwq->run_depth > 3) {
313 /* morton gets to eat his hat */
314 printk("%s: recursion depth exceeded: %d\n",
315 __FUNCTION__, cwq->run_depth);
318 while (!list_empty(&cwq->worklist)) {
319 struct work_struct *work = list_entry(cwq->worklist.next,
320 struct work_struct, entry);
321 work_func_t f = work->func;
323 list_del_init(cwq->worklist.next);
324 spin_unlock_irqrestore(&cwq->lock, flags);
326 BUG_ON(get_wq_data(work) != cwq);
327 if (!test_bit(WORK_STRUCT_NOAUTOREL, &work->management))
331 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
332 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
334 current->comm, preempt_count(),
336 printk(KERN_ERR " last function: ");
337 print_symbol("%s\n", (unsigned long)f);
338 debug_show_held_locks(current);
342 spin_lock_irqsave(&cwq->lock, flags);
343 cwq->remove_sequence++;
344 wake_up(&cwq->work_done);
347 spin_unlock_irqrestore(&cwq->lock, flags);
350 static int worker_thread(void *__cwq)
352 struct cpu_workqueue_struct *cwq = __cwq;
353 DECLARE_WAITQUEUE(wait, current);
354 struct k_sigaction sa;
357 if (!cwq->freezeable)
358 current->flags |= PF_NOFREEZE;
360 set_user_nice(current, -5);
362 /* Block and flush all signals */
363 sigfillset(&blocked);
364 sigprocmask(SIG_BLOCK, &blocked, NULL);
365 flush_signals(current);
368 * We inherited MPOL_INTERLEAVE from the booting kernel.
369 * Set MPOL_DEFAULT to insure node local allocations.
371 numa_default_policy();
373 /* SIG_IGN makes children autoreap: see do_notify_parent(). */
374 sa.sa.sa_handler = SIG_IGN;
376 siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
377 do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
379 set_current_state(TASK_INTERRUPTIBLE);
380 while (!kthread_should_stop()) {
384 add_wait_queue(&cwq->more_work, &wait);
385 if (list_empty(&cwq->worklist))
388 __set_current_state(TASK_RUNNING);
389 remove_wait_queue(&cwq->more_work, &wait);
391 if (!list_empty(&cwq->worklist))
393 set_current_state(TASK_INTERRUPTIBLE);
395 __set_current_state(TASK_RUNNING);
399 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
401 if (cwq->thread == current) {
403 * Probably keventd trying to flush its own queue. So simply run
404 * it by hand rather than deadlocking.
409 long sequence_needed;
411 spin_lock_irq(&cwq->lock);
412 sequence_needed = cwq->insert_sequence;
414 while (sequence_needed - cwq->remove_sequence > 0) {
415 prepare_to_wait(&cwq->work_done, &wait,
416 TASK_UNINTERRUPTIBLE);
417 spin_unlock_irq(&cwq->lock);
419 spin_lock_irq(&cwq->lock);
421 finish_wait(&cwq->work_done, &wait);
422 spin_unlock_irq(&cwq->lock);
427 * flush_workqueue - ensure that any scheduled work has run to completion.
428 * @wq: workqueue to flush
430 * Forces execution of the workqueue and blocks until its completion.
431 * This is typically used in driver shutdown handlers.
433 * This function will sample each workqueue's current insert_sequence number and
434 * will sleep until the head sequence is greater than or equal to that. This
435 * means that we sleep until all works which were queued on entry have been
436 * handled, but we are not livelocked by new incoming ones.
438 * This function used to run the workqueues itself. Now we just wait for the
439 * helper threads to do it.
441 void fastcall flush_workqueue(struct workqueue_struct *wq)
445 if (is_single_threaded(wq)) {
446 /* Always use first cpu's area. */
447 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, singlethread_cpu));
451 mutex_lock(&workqueue_mutex);
452 for_each_online_cpu(cpu)
453 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
454 mutex_unlock(&workqueue_mutex);
457 EXPORT_SYMBOL_GPL(flush_workqueue);
459 static struct task_struct *create_workqueue_thread(struct workqueue_struct *wq,
460 int cpu, int freezeable)
462 struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
463 struct task_struct *p;
465 spin_lock_init(&cwq->lock);
468 cwq->insert_sequence = 0;
469 cwq->remove_sequence = 0;
470 cwq->freezeable = freezeable;
471 INIT_LIST_HEAD(&cwq->worklist);
472 init_waitqueue_head(&cwq->more_work);
473 init_waitqueue_head(&cwq->work_done);
475 if (is_single_threaded(wq))
476 p = kthread_create(worker_thread, cwq, "%s", wq->name);
478 p = kthread_create(worker_thread, cwq, "%s/%d", wq->name, cpu);
485 struct workqueue_struct *__create_workqueue(const char *name,
486 int singlethread, int freezeable)
488 int cpu, destroy = 0;
489 struct workqueue_struct *wq;
490 struct task_struct *p;
492 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
496 wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
503 mutex_lock(&workqueue_mutex);
505 INIT_LIST_HEAD(&wq->list);
506 p = create_workqueue_thread(wq, singlethread_cpu, freezeable);
512 list_add(&wq->list, &workqueues);
513 for_each_online_cpu(cpu) {
514 p = create_workqueue_thread(wq, cpu, freezeable);
516 kthread_bind(p, cpu);
522 mutex_unlock(&workqueue_mutex);
525 * Was there any error during startup? If yes then clean up:
528 destroy_workqueue(wq);
533 EXPORT_SYMBOL_GPL(__create_workqueue);
535 static void cleanup_workqueue_thread(struct workqueue_struct *wq, int cpu)
537 struct cpu_workqueue_struct *cwq;
539 struct task_struct *p;
541 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
542 spin_lock_irqsave(&cwq->lock, flags);
545 spin_unlock_irqrestore(&cwq->lock, flags);
551 * destroy_workqueue - safely terminate a workqueue
552 * @wq: target workqueue
554 * Safely destroy a workqueue. All work currently pending will be done first.
556 void destroy_workqueue(struct workqueue_struct *wq)
562 /* We don't need the distraction of CPUs appearing and vanishing. */
563 mutex_lock(&workqueue_mutex);
564 if (is_single_threaded(wq))
565 cleanup_workqueue_thread(wq, singlethread_cpu);
567 for_each_online_cpu(cpu)
568 cleanup_workqueue_thread(wq, cpu);
571 mutex_unlock(&workqueue_mutex);
572 free_percpu(wq->cpu_wq);
575 EXPORT_SYMBOL_GPL(destroy_workqueue);
577 static struct workqueue_struct *keventd_wq;
580 * schedule_work - put work task in global workqueue
581 * @work: job to be done
583 * This puts a job in the kernel-global workqueue.
585 int fastcall schedule_work(struct work_struct *work)
587 return queue_work(keventd_wq, work);
589 EXPORT_SYMBOL(schedule_work);
592 * schedule_delayed_work - put work task in global workqueue after delay
593 * @dwork: job to be done
594 * @delay: number of jiffies to wait or 0 for immediate execution
596 * After waiting for a given time this puts a job in the kernel-global
599 int fastcall schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
601 return queue_delayed_work(keventd_wq, dwork, delay);
603 EXPORT_SYMBOL(schedule_delayed_work);
606 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
608 * @dwork: job to be done
609 * @delay: number of jiffies to wait
611 * After waiting for a given time this puts a job in the kernel-global
612 * workqueue on the specified CPU.
614 int schedule_delayed_work_on(int cpu,
615 struct delayed_work *dwork, unsigned long delay)
617 return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
619 EXPORT_SYMBOL(schedule_delayed_work_on);
622 * schedule_on_each_cpu - call a function on each online CPU from keventd
623 * @func: the function to call
625 * Returns zero on success.
626 * Returns -ve errno on failure.
628 * Appears to be racy against CPU hotplug.
630 * schedule_on_each_cpu() is very slow.
632 int schedule_on_each_cpu(work_func_t func)
635 struct work_struct *works;
637 works = alloc_percpu(struct work_struct);
641 mutex_lock(&workqueue_mutex);
642 for_each_online_cpu(cpu) {
643 INIT_WORK(per_cpu_ptr(works, cpu), func);
644 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu),
645 per_cpu_ptr(works, cpu));
647 mutex_unlock(&workqueue_mutex);
648 flush_workqueue(keventd_wq);
653 void flush_scheduled_work(void)
655 flush_workqueue(keventd_wq);
657 EXPORT_SYMBOL(flush_scheduled_work);
660 * cancel_rearming_delayed_workqueue - reliably kill off a delayed
661 * work whose handler rearms the delayed work.
662 * @wq: the controlling workqueue structure
663 * @dwork: the delayed work struct
665 void cancel_rearming_delayed_workqueue(struct workqueue_struct *wq,
666 struct delayed_work *dwork)
668 while (!cancel_delayed_work(dwork))
671 EXPORT_SYMBOL(cancel_rearming_delayed_workqueue);
674 * cancel_rearming_delayed_work - reliably kill off a delayed keventd
675 * work whose handler rearms the delayed work.
676 * @dwork: the delayed work struct
678 void cancel_rearming_delayed_work(struct delayed_work *dwork)
680 cancel_rearming_delayed_workqueue(keventd_wq, dwork);
682 EXPORT_SYMBOL(cancel_rearming_delayed_work);
685 * execute_in_process_context - reliably execute the routine with user context
686 * @fn: the function to execute
687 * @ew: guaranteed storage for the execute work structure (must
688 * be available when the work executes)
690 * Executes the function immediately if process context is available,
691 * otherwise schedules the function for delayed execution.
693 * Returns: 0 - function was executed
694 * 1 - function was scheduled for execution
696 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
698 if (!in_interrupt()) {
703 INIT_WORK(&ew->work, fn);
704 schedule_work(&ew->work);
708 EXPORT_SYMBOL_GPL(execute_in_process_context);
712 return keventd_wq != NULL;
715 int current_is_keventd(void)
717 struct cpu_workqueue_struct *cwq;
718 int cpu = smp_processor_id(); /* preempt-safe: keventd is per-cpu */
723 cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
724 if (current == cwq->thread)
731 /* Take the work from this (downed) CPU. */
732 static void take_over_work(struct workqueue_struct *wq, unsigned int cpu)
734 struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
735 struct list_head list;
736 struct work_struct *work;
738 spin_lock_irq(&cwq->lock);
739 list_replace_init(&cwq->worklist, &list);
741 while (!list_empty(&list)) {
742 printk("Taking work for %s\n", wq->name);
743 work = list_entry(list.next,struct work_struct,entry);
744 list_del(&work->entry);
745 __queue_work(per_cpu_ptr(wq->cpu_wq, smp_processor_id()), work);
747 spin_unlock_irq(&cwq->lock);
750 /* We're holding the cpucontrol mutex here */
751 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
752 unsigned long action,
755 unsigned int hotcpu = (unsigned long)hcpu;
756 struct workqueue_struct *wq;
760 mutex_lock(&workqueue_mutex);
761 /* Create a new workqueue thread for it. */
762 list_for_each_entry(wq, &workqueues, list) {
763 if (!create_workqueue_thread(wq, hotcpu, 0)) {
764 printk("workqueue for %i failed\n", hotcpu);
771 /* Kick off worker threads. */
772 list_for_each_entry(wq, &workqueues, list) {
773 struct cpu_workqueue_struct *cwq;
775 cwq = per_cpu_ptr(wq->cpu_wq, hotcpu);
776 kthread_bind(cwq->thread, hotcpu);
777 wake_up_process(cwq->thread);
779 mutex_unlock(&workqueue_mutex);
782 case CPU_UP_CANCELED:
783 list_for_each_entry(wq, &workqueues, list) {
784 if (!per_cpu_ptr(wq->cpu_wq, hotcpu)->thread)
786 /* Unbind so it can run. */
787 kthread_bind(per_cpu_ptr(wq->cpu_wq, hotcpu)->thread,
788 any_online_cpu(cpu_online_map));
789 cleanup_workqueue_thread(wq, hotcpu);
791 mutex_unlock(&workqueue_mutex);
794 case CPU_DOWN_PREPARE:
795 mutex_lock(&workqueue_mutex);
798 case CPU_DOWN_FAILED:
799 mutex_unlock(&workqueue_mutex);
803 list_for_each_entry(wq, &workqueues, list)
804 cleanup_workqueue_thread(wq, hotcpu);
805 list_for_each_entry(wq, &workqueues, list)
806 take_over_work(wq, hotcpu);
807 mutex_unlock(&workqueue_mutex);
814 void init_workqueues(void)
816 singlethread_cpu = first_cpu(cpu_possible_map);
817 hotcpu_notifier(workqueue_cpu_callback, 0);
818 keventd_wq = create_workqueue("events");