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>
33 * The per-CPU workqueue (if single thread, we always use the first
36 * The sequence counters are for flush_scheduled_work(). It wants to wait
37 * until all currently-scheduled works are completed, but it doesn't
38 * want to be livelocked by new, incoming ones. So it waits until
39 * remove_sequence is >= the insert_sequence which pertained when
40 * flush_scheduled_work() was called.
42 struct cpu_workqueue_struct {
46 long remove_sequence; /* Least-recently added (next to run) */
47 long insert_sequence; /* Next to add */
49 struct list_head worklist;
50 wait_queue_head_t more_work;
51 wait_queue_head_t work_done;
53 struct workqueue_struct *wq;
54 struct task_struct *thread;
56 int run_depth; /* Detect run_workqueue() recursion depth */
57 } ____cacheline_aligned;
60 * The externally visible workqueue abstraction is an array of
63 struct workqueue_struct {
64 struct cpu_workqueue_struct *cpu_wq;
66 struct list_head list; /* Empty if single thread */
69 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
70 threads to each one as cpus come/go. */
71 static DEFINE_MUTEX(workqueue_mutex);
72 static LIST_HEAD(workqueues);
74 static int singlethread_cpu;
76 /* If it's single threaded, it isn't in the list of workqueues. */
77 static inline int is_single_threaded(struct workqueue_struct *wq)
79 return list_empty(&wq->list);
82 /* Preempt must be disabled. */
83 static void __queue_work(struct cpu_workqueue_struct *cwq,
84 struct work_struct *work)
88 spin_lock_irqsave(&cwq->lock, flags);
90 list_add_tail(&work->entry, &cwq->worklist);
91 cwq->insert_sequence++;
92 wake_up(&cwq->more_work);
93 spin_unlock_irqrestore(&cwq->lock, flags);
97 * queue_work - queue work on a workqueue
98 * @wq: workqueue to use
99 * @work: work to queue
101 * Returns non-zero if it was successfully added.
103 * We queue the work to the CPU it was submitted, but there is no
104 * guarantee that it will be processed by that CPU.
106 int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
108 int ret = 0, cpu = get_cpu();
110 if (!test_and_set_bit(0, &work->pending)) {
111 if (unlikely(is_single_threaded(wq)))
112 cpu = singlethread_cpu;
113 BUG_ON(!list_empty(&work->entry));
114 __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
120 EXPORT_SYMBOL_GPL(queue_work);
122 static void delayed_work_timer_fn(unsigned long __data)
124 struct work_struct *work = (struct work_struct *)__data;
125 struct workqueue_struct *wq = work->wq_data;
126 int cpu = smp_processor_id();
128 if (unlikely(is_single_threaded(wq)))
129 cpu = singlethread_cpu;
131 __queue_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
135 * queue_delayed_work - queue work on a workqueue after delay
136 * @wq: workqueue to use
137 * @work: work to queue
138 * @delay: number of jiffies to wait before queueing
140 * Returns non-zero if it was successfully added.
142 int fastcall queue_delayed_work(struct workqueue_struct *wq,
143 struct work_struct *work, unsigned long delay)
146 struct timer_list *timer = &work->timer;
148 if (!test_and_set_bit(0, &work->pending)) {
149 BUG_ON(timer_pending(timer));
150 BUG_ON(!list_empty(&work->entry));
152 /* This stores wq for the moment, for the timer_fn */
154 timer->expires = jiffies + delay;
155 timer->data = (unsigned long)work;
156 timer->function = delayed_work_timer_fn;
162 EXPORT_SYMBOL_GPL(queue_delayed_work);
165 * queue_delayed_work_on - queue work on specific CPU after delay
166 * @cpu: CPU number to execute work on
167 * @wq: workqueue to use
168 * @work: work to queue
169 * @delay: number of jiffies to wait before queueing
171 * Returns non-zero if it was successfully added.
173 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
174 struct work_struct *work, unsigned long delay)
177 struct timer_list *timer = &work->timer;
179 if (!test_and_set_bit(0, &work->pending)) {
180 BUG_ON(timer_pending(timer));
181 BUG_ON(!list_empty(&work->entry));
183 /* This stores wq for the moment, for the timer_fn */
185 timer->expires = jiffies + delay;
186 timer->data = (unsigned long)work;
187 timer->function = delayed_work_timer_fn;
188 add_timer_on(timer, cpu);
193 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
195 static void run_workqueue(struct cpu_workqueue_struct *cwq)
200 * Keep taking off work from the queue until
203 spin_lock_irqsave(&cwq->lock, flags);
205 if (cwq->run_depth > 3) {
206 /* morton gets to eat his hat */
207 printk("%s: recursion depth exceeded: %d\n",
208 __FUNCTION__, cwq->run_depth);
211 while (!list_empty(&cwq->worklist)) {
212 struct work_struct *work = list_entry(cwq->worklist.next,
213 struct work_struct, entry);
214 void (*f) (void *) = work->func;
215 void *data = work->data;
217 list_del_init(cwq->worklist.next);
218 spin_unlock_irqrestore(&cwq->lock, flags);
220 BUG_ON(work->wq_data != cwq);
221 clear_bit(0, &work->pending);
224 spin_lock_irqsave(&cwq->lock, flags);
225 cwq->remove_sequence++;
226 wake_up(&cwq->work_done);
229 spin_unlock_irqrestore(&cwq->lock, flags);
232 static int worker_thread(void *__cwq)
234 struct cpu_workqueue_struct *cwq = __cwq;
235 DECLARE_WAITQUEUE(wait, current);
236 struct k_sigaction sa;
239 current->flags |= PF_NOFREEZE;
241 set_user_nice(current, -5);
243 /* Block and flush all signals */
244 sigfillset(&blocked);
245 sigprocmask(SIG_BLOCK, &blocked, NULL);
246 flush_signals(current);
248 /* SIG_IGN makes children autoreap: see do_notify_parent(). */
249 sa.sa.sa_handler = SIG_IGN;
251 siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
252 do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
254 set_current_state(TASK_INTERRUPTIBLE);
255 while (!kthread_should_stop()) {
256 add_wait_queue(&cwq->more_work, &wait);
257 if (list_empty(&cwq->worklist))
260 __set_current_state(TASK_RUNNING);
261 remove_wait_queue(&cwq->more_work, &wait);
263 if (!list_empty(&cwq->worklist))
265 set_current_state(TASK_INTERRUPTIBLE);
267 __set_current_state(TASK_RUNNING);
271 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
273 if (cwq->thread == current) {
275 * Probably keventd trying to flush its own queue. So simply run
276 * it by hand rather than deadlocking.
281 long sequence_needed;
283 spin_lock_irq(&cwq->lock);
284 sequence_needed = cwq->insert_sequence;
286 while (sequence_needed - cwq->remove_sequence > 0) {
287 prepare_to_wait(&cwq->work_done, &wait,
288 TASK_UNINTERRUPTIBLE);
289 spin_unlock_irq(&cwq->lock);
291 spin_lock_irq(&cwq->lock);
293 finish_wait(&cwq->work_done, &wait);
294 spin_unlock_irq(&cwq->lock);
299 * flush_workqueue - ensure that any scheduled work has run to completion.
300 * @wq: workqueue to flush
302 * Forces execution of the workqueue and blocks until its completion.
303 * This is typically used in driver shutdown handlers.
305 * This function will sample each workqueue's current insert_sequence number and
306 * will sleep until the head sequence is greater than or equal to that. This
307 * means that we sleep until all works which were queued on entry have been
308 * handled, but we are not livelocked by new incoming ones.
310 * This function used to run the workqueues itself. Now we just wait for the
311 * helper threads to do it.
313 void fastcall flush_workqueue(struct workqueue_struct *wq)
317 if (is_single_threaded(wq)) {
318 /* Always use first cpu's area. */
319 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, singlethread_cpu));
323 mutex_lock(&workqueue_mutex);
324 for_each_online_cpu(cpu)
325 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
326 mutex_unlock(&workqueue_mutex);
329 EXPORT_SYMBOL_GPL(flush_workqueue);
331 static struct task_struct *create_workqueue_thread(struct workqueue_struct *wq,
334 struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
335 struct task_struct *p;
337 spin_lock_init(&cwq->lock);
340 cwq->insert_sequence = 0;
341 cwq->remove_sequence = 0;
342 INIT_LIST_HEAD(&cwq->worklist);
343 init_waitqueue_head(&cwq->more_work);
344 init_waitqueue_head(&cwq->work_done);
346 if (is_single_threaded(wq))
347 p = kthread_create(worker_thread, cwq, "%s", wq->name);
349 p = kthread_create(worker_thread, cwq, "%s/%d", wq->name, cpu);
356 struct workqueue_struct *__create_workqueue(const char *name,
359 int cpu, destroy = 0;
360 struct workqueue_struct *wq;
361 struct task_struct *p;
363 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
367 wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
374 mutex_lock(&workqueue_mutex);
376 INIT_LIST_HEAD(&wq->list);
377 p = create_workqueue_thread(wq, singlethread_cpu);
383 list_add(&wq->list, &workqueues);
384 for_each_online_cpu(cpu) {
385 p = create_workqueue_thread(wq, cpu);
387 kthread_bind(p, cpu);
393 mutex_unlock(&workqueue_mutex);
396 * Was there any error during startup? If yes then clean up:
399 destroy_workqueue(wq);
404 EXPORT_SYMBOL_GPL(__create_workqueue);
406 static void cleanup_workqueue_thread(struct workqueue_struct *wq, int cpu)
408 struct cpu_workqueue_struct *cwq;
410 struct task_struct *p;
412 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
413 spin_lock_irqsave(&cwq->lock, flags);
416 spin_unlock_irqrestore(&cwq->lock, flags);
422 * destroy_workqueue - safely terminate a workqueue
423 * @wq: target workqueue
425 * Safely destroy a workqueue. All work currently pending will be done first.
427 void destroy_workqueue(struct workqueue_struct *wq)
433 /* We don't need the distraction of CPUs appearing and vanishing. */
434 mutex_lock(&workqueue_mutex);
435 if (is_single_threaded(wq))
436 cleanup_workqueue_thread(wq, singlethread_cpu);
438 for_each_online_cpu(cpu)
439 cleanup_workqueue_thread(wq, cpu);
442 mutex_unlock(&workqueue_mutex);
443 free_percpu(wq->cpu_wq);
446 EXPORT_SYMBOL_GPL(destroy_workqueue);
448 static struct workqueue_struct *keventd_wq;
451 * schedule_work - put work task in global workqueue
452 * @work: job to be done
454 * This puts a job in the kernel-global workqueue.
456 int fastcall schedule_work(struct work_struct *work)
458 return queue_work(keventd_wq, work);
460 EXPORT_SYMBOL(schedule_work);
463 * schedule_delayed_work - put work task in global workqueue after delay
464 * @work: job to be done
465 * @delay: number of jiffies to wait
467 * After waiting for a given time this puts a job in the kernel-global
470 int fastcall schedule_delayed_work(struct work_struct *work, unsigned long delay)
472 return queue_delayed_work(keventd_wq, work, delay);
474 EXPORT_SYMBOL(schedule_delayed_work);
477 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
479 * @work: job to be done
480 * @delay: number of jiffies to wait
482 * After waiting for a given time this puts a job in the kernel-global
483 * workqueue on the specified CPU.
485 int schedule_delayed_work_on(int cpu,
486 struct work_struct *work, unsigned long delay)
488 return queue_delayed_work_on(cpu, keventd_wq, work, delay);
490 EXPORT_SYMBOL(schedule_delayed_work_on);
493 * schedule_on_each_cpu - call a function on each online CPU from keventd
494 * @func: the function to call
495 * @info: a pointer to pass to func()
497 * Returns zero on success.
498 * Returns -ve errno on failure.
500 * Appears to be racy against CPU hotplug.
502 * schedule_on_each_cpu() is very slow.
504 int schedule_on_each_cpu(void (*func)(void *info), void *info)
507 struct work_struct *works;
509 works = alloc_percpu(struct work_struct);
513 mutex_lock(&workqueue_mutex);
514 for_each_online_cpu(cpu) {
515 INIT_WORK(per_cpu_ptr(works, cpu), func, info);
516 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu),
517 per_cpu_ptr(works, cpu));
519 mutex_unlock(&workqueue_mutex);
520 flush_workqueue(keventd_wq);
525 void flush_scheduled_work(void)
527 flush_workqueue(keventd_wq);
529 EXPORT_SYMBOL(flush_scheduled_work);
532 * cancel_rearming_delayed_workqueue - reliably kill off a delayed
533 * work whose handler rearms the delayed work.
534 * @wq: the controlling workqueue structure
535 * @work: the delayed work struct
537 void cancel_rearming_delayed_workqueue(struct workqueue_struct *wq,
538 struct work_struct *work)
540 while (!cancel_delayed_work(work))
543 EXPORT_SYMBOL(cancel_rearming_delayed_workqueue);
546 * cancel_rearming_delayed_work - reliably kill off a delayed keventd
547 * work whose handler rearms the delayed work.
548 * @work: the delayed work struct
550 void cancel_rearming_delayed_work(struct work_struct *work)
552 cancel_rearming_delayed_workqueue(keventd_wq, work);
554 EXPORT_SYMBOL(cancel_rearming_delayed_work);
557 * execute_in_process_context - reliably execute the routine with user context
558 * @fn: the function to execute
559 * @data: data to pass to the function
560 * @ew: guaranteed storage for the execute work structure (must
561 * be available when the work executes)
563 * Executes the function immediately if process context is available,
564 * otherwise schedules the function for delayed execution.
566 * Returns: 0 - function was executed
567 * 1 - function was scheduled for execution
569 int execute_in_process_context(void (*fn)(void *data), void *data,
570 struct execute_work *ew)
572 if (!in_interrupt()) {
577 INIT_WORK(&ew->work, fn, data);
578 schedule_work(&ew->work);
582 EXPORT_SYMBOL_GPL(execute_in_process_context);
586 return keventd_wq != NULL;
589 int current_is_keventd(void)
591 struct cpu_workqueue_struct *cwq;
592 int cpu = smp_processor_id(); /* preempt-safe: keventd is per-cpu */
597 cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
598 if (current == cwq->thread)
605 #ifdef CONFIG_HOTPLUG_CPU
606 /* Take the work from this (downed) CPU. */
607 static void take_over_work(struct workqueue_struct *wq, unsigned int cpu)
609 struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
610 struct list_head list;
611 struct work_struct *work;
613 spin_lock_irq(&cwq->lock);
614 list_replace_init(&cwq->worklist, &list);
616 while (!list_empty(&list)) {
617 printk("Taking work for %s\n", wq->name);
618 work = list_entry(list.next,struct work_struct,entry);
619 list_del(&work->entry);
620 __queue_work(per_cpu_ptr(wq->cpu_wq, smp_processor_id()), work);
622 spin_unlock_irq(&cwq->lock);
625 /* We're holding the cpucontrol mutex here */
626 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
627 unsigned long action,
630 unsigned int hotcpu = (unsigned long)hcpu;
631 struct workqueue_struct *wq;
635 mutex_lock(&workqueue_mutex);
636 /* Create a new workqueue thread for it. */
637 list_for_each_entry(wq, &workqueues, list) {
638 if (!create_workqueue_thread(wq, hotcpu)) {
639 printk("workqueue for %i failed\n", hotcpu);
646 /* Kick off worker threads. */
647 list_for_each_entry(wq, &workqueues, list) {
648 struct cpu_workqueue_struct *cwq;
650 cwq = per_cpu_ptr(wq->cpu_wq, hotcpu);
651 kthread_bind(cwq->thread, hotcpu);
652 wake_up_process(cwq->thread);
654 mutex_unlock(&workqueue_mutex);
657 case CPU_UP_CANCELED:
658 list_for_each_entry(wq, &workqueues, list) {
659 if (!per_cpu_ptr(wq->cpu_wq, hotcpu)->thread)
661 /* Unbind so it can run. */
662 kthread_bind(per_cpu_ptr(wq->cpu_wq, hotcpu)->thread,
663 any_online_cpu(cpu_online_map));
664 cleanup_workqueue_thread(wq, hotcpu);
666 mutex_unlock(&workqueue_mutex);
669 case CPU_DOWN_PREPARE:
670 mutex_lock(&workqueue_mutex);
673 case CPU_DOWN_FAILED:
674 mutex_unlock(&workqueue_mutex);
678 list_for_each_entry(wq, &workqueues, list)
679 cleanup_workqueue_thread(wq, hotcpu);
680 list_for_each_entry(wq, &workqueues, list)
681 take_over_work(wq, hotcpu);
682 mutex_unlock(&workqueue_mutex);
690 void init_workqueues(void)
692 singlethread_cpu = first_cpu(cpu_possible_map);
693 hotcpu_notifier(workqueue_cpu_callback, 0);
694 keventd_wq = create_workqueue("events");