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.
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>
38 * The per-CPU workqueue (if single thread, we always use the first
41 struct cpu_workqueue_struct {
45 struct list_head worklist;
46 wait_queue_head_t more_work;
47 struct work_struct *current_work;
49 struct workqueue_struct *wq;
50 struct task_struct *thread;
52 int run_depth; /* Detect run_workqueue() recursion depth */
53 } ____cacheline_aligned;
56 * The externally visible workqueue abstraction is an array of
59 struct workqueue_struct {
60 struct cpu_workqueue_struct *cpu_wq;
61 struct list_head list;
64 int freezeable; /* Freeze threads during suspend */
66 struct lockdep_map lockdep_map;
70 /* Serializes the accesses to the list of workqueues. */
71 static DEFINE_SPINLOCK(workqueue_lock);
72 static LIST_HEAD(workqueues);
74 static int singlethread_cpu __read_mostly;
75 static cpumask_t cpu_singlethread_map __read_mostly;
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
83 static cpumask_t cpu_populated_map __read_mostly;
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)
88 return wq->singlethread;
91 static const cpumask_t *wq_cpu_map(struct workqueue_struct *wq)
93 return is_single_threaded(wq)
94 ? &cpu_singlethread_map : &cpu_populated_map;
98 struct cpu_workqueue_struct *wq_per_cpu(struct workqueue_struct *wq, int cpu)
100 if (unlikely(is_single_threaded(wq)))
101 cpu = singlethread_cpu;
102 return per_cpu_ptr(wq->cpu_wq, cpu);
106 * Set the workqueue on which a work item is to be run
107 * - Must *only* be called if the pending flag is set
109 static inline void set_wq_data(struct work_struct *work,
110 struct cpu_workqueue_struct *cwq)
114 BUG_ON(!work_pending(work));
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);
122 struct cpu_workqueue_struct *get_wq_data(struct work_struct *work)
124 return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
127 static void insert_work(struct cpu_workqueue_struct *cwq,
128 struct work_struct *work, int tail)
130 set_wq_data(work, cwq);
132 * Ensure that we get the right work->data if we see the
133 * result of list_add() below, see try_to_grab_pending().
137 list_add_tail(&work->entry, &cwq->worklist);
139 list_add(&work->entry, &cwq->worklist);
140 wake_up(&cwq->more_work);
143 static void __queue_work(struct cpu_workqueue_struct *cwq,
144 struct work_struct *work)
148 spin_lock_irqsave(&cwq->lock, flags);
149 insert_work(cwq, work, 1);
150 spin_unlock_irqrestore(&cwq->lock, flags);
154 * queue_work - queue work on a workqueue
155 * @wq: workqueue to use
156 * @work: work to queue
158 * Returns 0 if @work was already on a queue, non-zero otherwise.
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.
163 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
167 if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
168 BUG_ON(!list_empty(&work->entry));
169 __queue_work(wq_per_cpu(wq, get_cpu()), work);
175 EXPORT_SYMBOL_GPL(queue_work);
178 * queue_work_on - queue work on specific cpu
179 * @cpu: CPU number to execute work on
180 * @wq: workqueue to use
181 * @work: work to queue
183 * Returns 0 if @work was already on a queue, non-zero otherwise.
185 * We queue the work to a specific CPU, the caller must ensure it
189 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
193 if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
194 BUG_ON(!list_empty(&work->entry));
195 __queue_work(wq_per_cpu(wq, cpu), work);
200 EXPORT_SYMBOL_GPL(queue_work_on);
202 static void delayed_work_timer_fn(unsigned long __data)
204 struct delayed_work *dwork = (struct delayed_work *)__data;
205 struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
206 struct workqueue_struct *wq = cwq->wq;
208 __queue_work(wq_per_cpu(wq, smp_processor_id()), &dwork->work);
212 * queue_delayed_work - queue work on a workqueue after delay
213 * @wq: workqueue to use
214 * @dwork: delayable work to queue
215 * @delay: number of jiffies to wait before queueing
217 * Returns 0 if @work was already on a queue, non-zero otherwise.
219 int queue_delayed_work(struct workqueue_struct *wq,
220 struct delayed_work *dwork, unsigned long delay)
223 return queue_work(wq, &dwork->work);
225 return queue_delayed_work_on(-1, wq, dwork, delay);
227 EXPORT_SYMBOL_GPL(queue_delayed_work);
230 * queue_delayed_work_on - queue work on specific CPU after delay
231 * @cpu: CPU number to execute work on
232 * @wq: workqueue to use
233 * @dwork: work to queue
234 * @delay: number of jiffies to wait before queueing
236 * Returns 0 if @work was already on a queue, non-zero otherwise.
238 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
239 struct delayed_work *dwork, unsigned long delay)
242 struct timer_list *timer = &dwork->timer;
243 struct work_struct *work = &dwork->work;
245 if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
246 BUG_ON(timer_pending(timer));
247 BUG_ON(!list_empty(&work->entry));
249 timer_stats_timer_set_start_info(&dwork->timer);
251 /* This stores cwq for the moment, for the timer_fn */
252 set_wq_data(work, wq_per_cpu(wq, raw_smp_processor_id()));
253 timer->expires = jiffies + delay;
254 timer->data = (unsigned long)dwork;
255 timer->function = delayed_work_timer_fn;
257 if (unlikely(cpu >= 0))
258 add_timer_on(timer, cpu);
265 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
267 static void run_workqueue(struct cpu_workqueue_struct *cwq)
269 spin_lock_irq(&cwq->lock);
271 if (cwq->run_depth > 3) {
272 /* morton gets to eat his hat */
273 printk("%s: recursion depth exceeded: %d\n",
274 __func__, cwq->run_depth);
277 while (!list_empty(&cwq->worklist)) {
278 struct work_struct *work = list_entry(cwq->worklist.next,
279 struct work_struct, entry);
280 work_func_t f = work->func;
281 #ifdef CONFIG_LOCKDEP
283 * It is permissible to free the struct work_struct
284 * from inside the function that is called from it,
285 * this we need to take into account for lockdep too.
286 * To avoid bogus "held lock freed" warnings as well
287 * as problems when looking into work->lockdep_map,
288 * make a copy and use that here.
290 struct lockdep_map lockdep_map = work->lockdep_map;
293 cwq->current_work = work;
294 list_del_init(cwq->worklist.next);
295 spin_unlock_irq(&cwq->lock);
297 BUG_ON(get_wq_data(work) != cwq);
298 work_clear_pending(work);
299 lock_acquire(&cwq->wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
300 lock_acquire(&lockdep_map, 0, 0, 0, 2, _THIS_IP_);
302 lock_release(&lockdep_map, 1, _THIS_IP_);
303 lock_release(&cwq->wq->lockdep_map, 1, _THIS_IP_);
305 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
306 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
308 current->comm, preempt_count(),
309 task_pid_nr(current));
310 printk(KERN_ERR " last function: ");
311 print_symbol("%s\n", (unsigned long)f);
312 debug_show_held_locks(current);
316 spin_lock_irq(&cwq->lock);
317 cwq->current_work = NULL;
320 spin_unlock_irq(&cwq->lock);
323 static int worker_thread(void *__cwq)
325 struct cpu_workqueue_struct *cwq = __cwq;
328 if (cwq->wq->freezeable)
331 set_user_nice(current, -5);
334 prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
335 if (!freezing(current) &&
336 !kthread_should_stop() &&
337 list_empty(&cwq->worklist))
339 finish_wait(&cwq->more_work, &wait);
343 if (kthread_should_stop())
353 struct work_struct work;
354 struct completion done;
357 static void wq_barrier_func(struct work_struct *work)
359 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
360 complete(&barr->done);
363 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
364 struct wq_barrier *barr, int tail)
366 INIT_WORK(&barr->work, wq_barrier_func);
367 __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
369 init_completion(&barr->done);
371 insert_work(cwq, &barr->work, tail);
374 static int flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
378 if (cwq->thread == current) {
380 * Probably keventd trying to flush its own queue. So simply run
381 * it by hand rather than deadlocking.
386 struct wq_barrier barr;
389 spin_lock_irq(&cwq->lock);
390 if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
391 insert_wq_barrier(cwq, &barr, 1);
394 spin_unlock_irq(&cwq->lock);
397 wait_for_completion(&barr.done);
404 * flush_workqueue - ensure that any scheduled work has run to completion.
405 * @wq: workqueue to flush
407 * Forces execution of the workqueue and blocks until its completion.
408 * This is typically used in driver shutdown handlers.
410 * We sleep until all works which were queued on entry have been handled,
411 * but we are not livelocked by new incoming ones.
413 * This function used to run the workqueues itself. Now we just wait for the
414 * helper threads to do it.
416 void flush_workqueue(struct workqueue_struct *wq)
418 const cpumask_t *cpu_map = wq_cpu_map(wq);
422 lock_acquire(&wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
423 lock_release(&wq->lockdep_map, 1, _THIS_IP_);
424 for_each_cpu_mask_nr(cpu, *cpu_map)
425 flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
427 EXPORT_SYMBOL_GPL(flush_workqueue);
430 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
431 * so this work can't be re-armed in any way.
433 static int try_to_grab_pending(struct work_struct *work)
435 struct cpu_workqueue_struct *cwq;
438 if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work)))
442 * The queueing is in progress, or it is already queued. Try to
443 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
446 cwq = get_wq_data(work);
450 spin_lock_irq(&cwq->lock);
451 if (!list_empty(&work->entry)) {
453 * This work is queued, but perhaps we locked the wrong cwq.
454 * In that case we must see the new value after rmb(), see
455 * insert_work()->wmb().
458 if (cwq == get_wq_data(work)) {
459 list_del_init(&work->entry);
463 spin_unlock_irq(&cwq->lock);
468 static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq,
469 struct work_struct *work)
471 struct wq_barrier barr;
474 spin_lock_irq(&cwq->lock);
475 if (unlikely(cwq->current_work == work)) {
476 insert_wq_barrier(cwq, &barr, 0);
479 spin_unlock_irq(&cwq->lock);
481 if (unlikely(running))
482 wait_for_completion(&barr.done);
485 static void wait_on_work(struct work_struct *work)
487 struct cpu_workqueue_struct *cwq;
488 struct workqueue_struct *wq;
489 const cpumask_t *cpu_map;
494 lock_acquire(&work->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
495 lock_release(&work->lockdep_map, 1, _THIS_IP_);
497 cwq = get_wq_data(work);
502 cpu_map = wq_cpu_map(wq);
504 for_each_cpu_mask_nr(cpu, *cpu_map)
505 wait_on_cpu_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
508 static int __cancel_work_timer(struct work_struct *work,
509 struct timer_list* timer)
514 ret = (timer && likely(del_timer(timer)));
516 ret = try_to_grab_pending(work);
518 } while (unlikely(ret < 0));
520 work_clear_pending(work);
525 * cancel_work_sync - block until a work_struct's callback has terminated
526 * @work: the work which is to be flushed
528 * Returns true if @work was pending.
530 * cancel_work_sync() will cancel the work if it is queued. If the work's
531 * callback appears to be running, cancel_work_sync() will block until it
534 * It is possible to use this function if the work re-queues itself. It can
535 * cancel the work even if it migrates to another workqueue, however in that
536 * case it only guarantees that work->func() has completed on the last queued
539 * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
540 * pending, otherwise it goes into a busy-wait loop until the timer expires.
542 * The caller must ensure that workqueue_struct on which this work was last
543 * queued can't be destroyed before this function returns.
545 int cancel_work_sync(struct work_struct *work)
547 return __cancel_work_timer(work, NULL);
549 EXPORT_SYMBOL_GPL(cancel_work_sync);
552 * cancel_delayed_work_sync - reliably kill off a delayed work.
553 * @dwork: the delayed work struct
555 * Returns true if @dwork was pending.
557 * It is possible to use this function if @dwork rearms itself via queue_work()
558 * or queue_delayed_work(). See also the comment for cancel_work_sync().
560 int cancel_delayed_work_sync(struct delayed_work *dwork)
562 return __cancel_work_timer(&dwork->work, &dwork->timer);
564 EXPORT_SYMBOL(cancel_delayed_work_sync);
566 static struct workqueue_struct *keventd_wq __read_mostly;
569 * schedule_work - put work task in global workqueue
570 * @work: job to be done
572 * This puts a job in the kernel-global workqueue.
574 int schedule_work(struct work_struct *work)
576 return queue_work(keventd_wq, work);
578 EXPORT_SYMBOL(schedule_work);
581 * schedule_work_on - put work task on a specific cpu
582 * @cpu: cpu to put the work task on
583 * @work: job to be done
585 * This puts a job on a specific cpu
587 int schedule_work_on(int cpu, struct work_struct *work)
589 return queue_work_on(cpu, keventd_wq, work);
591 EXPORT_SYMBOL(schedule_work_on);
594 * schedule_delayed_work - put work task in global workqueue after delay
595 * @dwork: job to be done
596 * @delay: number of jiffies to wait or 0 for immediate execution
598 * After waiting for a given time this puts a job in the kernel-global
601 int schedule_delayed_work(struct delayed_work *dwork,
604 return queue_delayed_work(keventd_wq, dwork, delay);
606 EXPORT_SYMBOL(schedule_delayed_work);
609 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
611 * @dwork: job to be done
612 * @delay: number of jiffies to wait
614 * After waiting for a given time this puts a job in the kernel-global
615 * workqueue on the specified CPU.
617 int schedule_delayed_work_on(int cpu,
618 struct delayed_work *dwork, unsigned long delay)
620 return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
622 EXPORT_SYMBOL(schedule_delayed_work_on);
625 * schedule_on_each_cpu - call a function on each online CPU from keventd
626 * @func: the function to call
628 * Returns zero on success.
629 * Returns -ve errno on failure.
631 * schedule_on_each_cpu() is very slow.
633 int schedule_on_each_cpu(work_func_t func)
636 struct work_struct *works;
638 works = alloc_percpu(struct work_struct);
643 for_each_online_cpu(cpu) {
644 struct work_struct *work = per_cpu_ptr(works, cpu);
646 INIT_WORK(work, func);
647 set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
648 __queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
650 flush_workqueue(keventd_wq);
656 void flush_scheduled_work(void)
658 flush_workqueue(keventd_wq);
660 EXPORT_SYMBOL(flush_scheduled_work);
663 * execute_in_process_context - reliably execute the routine with user context
664 * @fn: the function to execute
665 * @ew: guaranteed storage for the execute work structure (must
666 * be available when the work executes)
668 * Executes the function immediately if process context is available,
669 * otherwise schedules the function for delayed execution.
671 * Returns: 0 - function was executed
672 * 1 - function was scheduled for execution
674 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
676 if (!in_interrupt()) {
681 INIT_WORK(&ew->work, fn);
682 schedule_work(&ew->work);
686 EXPORT_SYMBOL_GPL(execute_in_process_context);
690 return keventd_wq != NULL;
693 int current_is_keventd(void)
695 struct cpu_workqueue_struct *cwq;
696 int cpu = raw_smp_processor_id(); /* preempt-safe: keventd is per-cpu */
701 cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
702 if (current == cwq->thread)
709 static struct cpu_workqueue_struct *
710 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
712 struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
715 spin_lock_init(&cwq->lock);
716 INIT_LIST_HEAD(&cwq->worklist);
717 init_waitqueue_head(&cwq->more_work);
722 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
724 struct workqueue_struct *wq = cwq->wq;
725 const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
726 struct task_struct *p;
728 p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
730 * Nobody can add the work_struct to this cwq,
731 * if (caller is __create_workqueue)
732 * nobody should see this wq
733 * else // caller is CPU_UP_PREPARE
734 * cpu is not on cpu_online_map
735 * so we can abort safely.
745 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
747 struct task_struct *p = cwq->thread;
751 kthread_bind(p, cpu);
756 struct workqueue_struct *__create_workqueue_key(const char *name,
759 struct lock_class_key *key,
760 const char *lock_name)
762 struct workqueue_struct *wq;
763 struct cpu_workqueue_struct *cwq;
766 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
770 wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
777 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
778 wq->singlethread = singlethread;
779 wq->freezeable = freezeable;
780 INIT_LIST_HEAD(&wq->list);
783 cwq = init_cpu_workqueue(wq, singlethread_cpu);
784 err = create_workqueue_thread(cwq, singlethread_cpu);
785 start_workqueue_thread(cwq, -1);
788 spin_lock(&workqueue_lock);
789 list_add(&wq->list, &workqueues);
790 spin_unlock(&workqueue_lock);
792 for_each_possible_cpu(cpu) {
793 cwq = init_cpu_workqueue(wq, cpu);
794 if (err || !cpu_online(cpu))
796 err = create_workqueue_thread(cwq, cpu);
797 start_workqueue_thread(cwq, cpu);
803 destroy_workqueue(wq);
808 EXPORT_SYMBOL_GPL(__create_workqueue_key);
810 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq)
813 * Our caller is either destroy_workqueue() or CPU_DEAD,
814 * get_online_cpus() protects cwq->thread.
816 if (cwq->thread == NULL)
819 lock_acquire(&cwq->wq->lockdep_map, 0, 0, 0, 2, _THIS_IP_);
820 lock_release(&cwq->wq->lockdep_map, 1, _THIS_IP_);
822 flush_cpu_workqueue(cwq);
824 * If the caller is CPU_DEAD and cwq->worklist was not empty,
825 * a concurrent flush_workqueue() can insert a barrier after us.
826 * However, in that case run_workqueue() won't return and check
827 * kthread_should_stop() until it flushes all work_struct's.
828 * When ->worklist becomes empty it is safe to exit because no
829 * more work_structs can be queued on this cwq: flush_workqueue
830 * checks list_empty(), and a "normal" queue_work() can't use
833 kthread_stop(cwq->thread);
838 * destroy_workqueue - safely terminate a workqueue
839 * @wq: target workqueue
841 * Safely destroy a workqueue. All work currently pending will be done first.
843 void destroy_workqueue(struct workqueue_struct *wq)
845 const cpumask_t *cpu_map = wq_cpu_map(wq);
849 spin_lock(&workqueue_lock);
851 spin_unlock(&workqueue_lock);
853 for_each_cpu_mask_nr(cpu, *cpu_map)
854 cleanup_workqueue_thread(per_cpu_ptr(wq->cpu_wq, cpu));
857 free_percpu(wq->cpu_wq);
860 EXPORT_SYMBOL_GPL(destroy_workqueue);
862 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
863 unsigned long action,
866 unsigned int cpu = (unsigned long)hcpu;
867 struct cpu_workqueue_struct *cwq;
868 struct workqueue_struct *wq;
870 action &= ~CPU_TASKS_FROZEN;
874 cpu_set(cpu, cpu_populated_map);
877 list_for_each_entry(wq, &workqueues, list) {
878 cwq = per_cpu_ptr(wq->cpu_wq, cpu);
882 if (!create_workqueue_thread(cwq, cpu))
884 printk(KERN_ERR "workqueue [%s] for %i failed\n",
889 start_workqueue_thread(cwq, cpu);
892 case CPU_UP_CANCELED:
893 start_workqueue_thread(cwq, -1);
895 cleanup_workqueue_thread(cwq);
901 case CPU_UP_CANCELED:
903 cpu_clear(cpu, cpu_populated_map);
909 void __init init_workqueues(void)
911 cpu_populated_map = cpu_online_map;
912 singlethread_cpu = first_cpu(cpu_possible_map);
913 cpu_singlethread_map = cpumask_of_cpu(singlethread_cpu);
914 hotcpu_notifier(workqueue_cpu_callback, 0);
915 keventd_wq = create_workqueue("events");