xref: /openbmc/linux/kernel/workqueue.c (revision e868d61272caa648214046a096e5a6bfc068dc8c)
1 /*
2  * linux/kernel/workqueue.c
3  *
4  * Generic mechanism for defining kernel helper threads for running
5  * arbitrary tasks in process context.
6  *
7  * Started by Ingo Molnar, Copyright (C) 2002
8  *
9  * Derived from the taskqueue/keventd code by:
10  *
11  *   David Woodhouse <dwmw2@infradead.org>
12  *   Andrew Morton <andrewm@uow.edu.au>
13  *   Kai Petzke <wpp@marie.physik.tu-berlin.de>
14  *   Theodore Ts'o <tytso@mit.edu>
15  *
16  * Made to use alloc_percpu by Christoph Lameter <clameter@sgi.com>.
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 
36 /*
37  * The per-CPU workqueue (if single thread, we always use the first
38  * possible cpu).
39  */
40 struct cpu_workqueue_struct {
41 
42 	spinlock_t lock;
43 
44 	struct list_head worklist;
45 	wait_queue_head_t more_work;
46 	struct work_struct *current_work;
47 
48 	struct workqueue_struct *wq;
49 	struct task_struct *thread;
50 	int should_stop;
51 
52 	int run_depth;		/* Detect run_workqueue() recursion depth */
53 } ____cacheline_aligned;
54 
55 /*
56  * The externally visible workqueue abstraction is an array of
57  * per-CPU workqueues:
58  */
59 struct workqueue_struct {
60 	struct cpu_workqueue_struct *cpu_wq;
61 	struct list_head list;
62 	const char *name;
63 	int singlethread;
64 	int freezeable;		/* Freeze threads during suspend */
65 };
66 
67 /* All the per-cpu workqueues on the system, for hotplug cpu to add/remove
68    threads to each one as cpus come/go. */
69 static DEFINE_MUTEX(workqueue_mutex);
70 static LIST_HEAD(workqueues);
71 
72 static int singlethread_cpu __read_mostly;
73 static cpumask_t cpu_singlethread_map __read_mostly;
74 /* optimization, we could use cpu_possible_map */
75 static cpumask_t cpu_populated_map __read_mostly;
76 
77 /* If it's single threaded, it isn't in the list of workqueues. */
78 static inline int is_single_threaded(struct workqueue_struct *wq)
79 {
80 	return wq->singlethread;
81 }
82 
83 static const cpumask_t *wq_cpu_map(struct workqueue_struct *wq)
84 {
85 	return is_single_threaded(wq)
86 		? &cpu_singlethread_map : &cpu_populated_map;
87 }
88 
89 static
90 struct cpu_workqueue_struct *wq_per_cpu(struct workqueue_struct *wq, int cpu)
91 {
92 	if (unlikely(is_single_threaded(wq)))
93 		cpu = singlethread_cpu;
94 	return per_cpu_ptr(wq->cpu_wq, cpu);
95 }
96 
97 /*
98  * Set the workqueue on which a work item is to be run
99  * - Must *only* be called if the pending flag is set
100  */
101 static inline void set_wq_data(struct work_struct *work,
102 				struct cpu_workqueue_struct *cwq)
103 {
104 	unsigned long new;
105 
106 	BUG_ON(!work_pending(work));
107 
108 	new = (unsigned long) cwq | (1UL << WORK_STRUCT_PENDING);
109 	new |= WORK_STRUCT_FLAG_MASK & *work_data_bits(work);
110 	atomic_long_set(&work->data, new);
111 }
112 
113 static inline
114 struct cpu_workqueue_struct *get_wq_data(struct work_struct *work)
115 {
116 	return (void *) (atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK);
117 }
118 
119 static void insert_work(struct cpu_workqueue_struct *cwq,
120 				struct work_struct *work, int tail)
121 {
122 	set_wq_data(work, cwq);
123 	/*
124 	 * Ensure that we get the right work->data if we see the
125 	 * result of list_add() below, see try_to_grab_pending().
126 	 */
127 	smp_wmb();
128 	if (tail)
129 		list_add_tail(&work->entry, &cwq->worklist);
130 	else
131 		list_add(&work->entry, &cwq->worklist);
132 	wake_up(&cwq->more_work);
133 }
134 
135 /* Preempt must be disabled. */
136 static void __queue_work(struct cpu_workqueue_struct *cwq,
137 			 struct work_struct *work)
138 {
139 	unsigned long flags;
140 
141 	spin_lock_irqsave(&cwq->lock, flags);
142 	insert_work(cwq, work, 1);
143 	spin_unlock_irqrestore(&cwq->lock, flags);
144 }
145 
146 /**
147  * queue_work - queue work on a workqueue
148  * @wq: workqueue to use
149  * @work: work to queue
150  *
151  * Returns 0 if @work was already on a queue, non-zero otherwise.
152  *
153  * We queue the work to the CPU it was submitted, but there is no
154  * guarantee that it will be processed by that CPU.
155  */
156 int fastcall queue_work(struct workqueue_struct *wq, struct work_struct *work)
157 {
158 	int ret = 0;
159 
160 	if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
161 		BUG_ON(!list_empty(&work->entry));
162 		__queue_work(wq_per_cpu(wq, get_cpu()), work);
163 		put_cpu();
164 		ret = 1;
165 	}
166 	return ret;
167 }
168 EXPORT_SYMBOL_GPL(queue_work);
169 
170 void delayed_work_timer_fn(unsigned long __data)
171 {
172 	struct delayed_work *dwork = (struct delayed_work *)__data;
173 	struct cpu_workqueue_struct *cwq = get_wq_data(&dwork->work);
174 	struct workqueue_struct *wq = cwq->wq;
175 
176 	__queue_work(wq_per_cpu(wq, smp_processor_id()), &dwork->work);
177 }
178 
179 /**
180  * queue_delayed_work - queue work on a workqueue after delay
181  * @wq: workqueue to use
182  * @dwork: delayable work to queue
183  * @delay: number of jiffies to wait before queueing
184  *
185  * Returns 0 if @work was already on a queue, non-zero otherwise.
186  */
187 int fastcall queue_delayed_work(struct workqueue_struct *wq,
188 			struct delayed_work *dwork, unsigned long delay)
189 {
190 	timer_stats_timer_set_start_info(&dwork->timer);
191 	if (delay == 0)
192 		return queue_work(wq, &dwork->work);
193 
194 	return queue_delayed_work_on(-1, wq, dwork, delay);
195 }
196 EXPORT_SYMBOL_GPL(queue_delayed_work);
197 
198 /**
199  * queue_delayed_work_on - queue work on specific CPU after delay
200  * @cpu: CPU number to execute work on
201  * @wq: workqueue to use
202  * @dwork: work to queue
203  * @delay: number of jiffies to wait before queueing
204  *
205  * Returns 0 if @work was already on a queue, non-zero otherwise.
206  */
207 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
208 			struct delayed_work *dwork, unsigned long delay)
209 {
210 	int ret = 0;
211 	struct timer_list *timer = &dwork->timer;
212 	struct work_struct *work = &dwork->work;
213 
214 	if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work))) {
215 		BUG_ON(timer_pending(timer));
216 		BUG_ON(!list_empty(&work->entry));
217 
218 		/* This stores cwq for the moment, for the timer_fn */
219 		set_wq_data(work, wq_per_cpu(wq, raw_smp_processor_id()));
220 		timer->expires = jiffies + delay;
221 		timer->data = (unsigned long)dwork;
222 		timer->function = delayed_work_timer_fn;
223 
224 		if (unlikely(cpu >= 0))
225 			add_timer_on(timer, cpu);
226 		else
227 			add_timer(timer);
228 		ret = 1;
229 	}
230 	return ret;
231 }
232 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
233 
234 static void run_workqueue(struct cpu_workqueue_struct *cwq)
235 {
236 	spin_lock_irq(&cwq->lock);
237 	cwq->run_depth++;
238 	if (cwq->run_depth > 3) {
239 		/* morton gets to eat his hat */
240 		printk("%s: recursion depth exceeded: %d\n",
241 			__FUNCTION__, cwq->run_depth);
242 		dump_stack();
243 	}
244 	while (!list_empty(&cwq->worklist)) {
245 		struct work_struct *work = list_entry(cwq->worklist.next,
246 						struct work_struct, entry);
247 		work_func_t f = work->func;
248 
249 		cwq->current_work = work;
250 		list_del_init(cwq->worklist.next);
251 		spin_unlock_irq(&cwq->lock);
252 
253 		BUG_ON(get_wq_data(work) != cwq);
254 		work_clear_pending(work);
255 		f(work);
256 
257 		if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
258 			printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
259 					"%s/0x%08x/%d\n",
260 					current->comm, preempt_count(),
261 				       	current->pid);
262 			printk(KERN_ERR "    last function: ");
263 			print_symbol("%s\n", (unsigned long)f);
264 			debug_show_held_locks(current);
265 			dump_stack();
266 		}
267 
268 		spin_lock_irq(&cwq->lock);
269 		cwq->current_work = NULL;
270 	}
271 	cwq->run_depth--;
272 	spin_unlock_irq(&cwq->lock);
273 }
274 
275 /*
276  * NOTE: the caller must not touch *cwq if this func returns true
277  */
278 static int cwq_should_stop(struct cpu_workqueue_struct *cwq)
279 {
280 	int should_stop = cwq->should_stop;
281 
282 	if (unlikely(should_stop)) {
283 		spin_lock_irq(&cwq->lock);
284 		should_stop = cwq->should_stop && list_empty(&cwq->worklist);
285 		if (should_stop)
286 			cwq->thread = NULL;
287 		spin_unlock_irq(&cwq->lock);
288 	}
289 
290 	return should_stop;
291 }
292 
293 static int worker_thread(void *__cwq)
294 {
295 	struct cpu_workqueue_struct *cwq = __cwq;
296 	DEFINE_WAIT(wait);
297 
298 	if (!cwq->wq->freezeable)
299 		current->flags |= PF_NOFREEZE;
300 
301 	set_user_nice(current, -5);
302 
303 	for (;;) {
304 		prepare_to_wait(&cwq->more_work, &wait, TASK_INTERRUPTIBLE);
305 		if (!freezing(current) && !cwq->should_stop
306 		    && list_empty(&cwq->worklist))
307 			schedule();
308 		finish_wait(&cwq->more_work, &wait);
309 
310 		try_to_freeze();
311 
312 		if (cwq_should_stop(cwq))
313 			break;
314 
315 		run_workqueue(cwq);
316 	}
317 
318 	return 0;
319 }
320 
321 struct wq_barrier {
322 	struct work_struct	work;
323 	struct completion	done;
324 };
325 
326 static void wq_barrier_func(struct work_struct *work)
327 {
328 	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
329 	complete(&barr->done);
330 }
331 
332 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
333 					struct wq_barrier *barr, int tail)
334 {
335 	INIT_WORK(&barr->work, wq_barrier_func);
336 	__set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
337 
338 	init_completion(&barr->done);
339 
340 	insert_work(cwq, &barr->work, tail);
341 }
342 
343 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
344 {
345 	if (cwq->thread == current) {
346 		/*
347 		 * Probably keventd trying to flush its own queue. So simply run
348 		 * it by hand rather than deadlocking.
349 		 */
350 		run_workqueue(cwq);
351 	} else {
352 		struct wq_barrier barr;
353 		int active = 0;
354 
355 		spin_lock_irq(&cwq->lock);
356 		if (!list_empty(&cwq->worklist) || cwq->current_work != NULL) {
357 			insert_wq_barrier(cwq, &barr, 1);
358 			active = 1;
359 		}
360 		spin_unlock_irq(&cwq->lock);
361 
362 		if (active)
363 			wait_for_completion(&barr.done);
364 	}
365 }
366 
367 /**
368  * flush_workqueue - ensure that any scheduled work has run to completion.
369  * @wq: workqueue to flush
370  *
371  * Forces execution of the workqueue and blocks until its completion.
372  * This is typically used in driver shutdown handlers.
373  *
374  * We sleep until all works which were queued on entry have been handled,
375  * but we are not livelocked by new incoming ones.
376  *
377  * This function used to run the workqueues itself.  Now we just wait for the
378  * helper threads to do it.
379  */
380 void fastcall flush_workqueue(struct workqueue_struct *wq)
381 {
382 	const cpumask_t *cpu_map = wq_cpu_map(wq);
383 	int cpu;
384 
385 	might_sleep();
386 	for_each_cpu_mask(cpu, *cpu_map)
387 		flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
388 }
389 EXPORT_SYMBOL_GPL(flush_workqueue);
390 
391 /*
392  * Upon a successful return, the caller "owns" WORK_STRUCT_PENDING bit,
393  * so this work can't be re-armed in any way.
394  */
395 static int try_to_grab_pending(struct work_struct *work)
396 {
397 	struct cpu_workqueue_struct *cwq;
398 	int ret = 0;
399 
400 	if (!test_and_set_bit(WORK_STRUCT_PENDING, work_data_bits(work)))
401 		return 1;
402 
403 	/*
404 	 * The queueing is in progress, or it is already queued. Try to
405 	 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
406 	 */
407 
408 	cwq = get_wq_data(work);
409 	if (!cwq)
410 		return ret;
411 
412 	spin_lock_irq(&cwq->lock);
413 	if (!list_empty(&work->entry)) {
414 		/*
415 		 * This work is queued, but perhaps we locked the wrong cwq.
416 		 * In that case we must see the new value after rmb(), see
417 		 * insert_work()->wmb().
418 		 */
419 		smp_rmb();
420 		if (cwq == get_wq_data(work)) {
421 			list_del_init(&work->entry);
422 			ret = 1;
423 		}
424 	}
425 	spin_unlock_irq(&cwq->lock);
426 
427 	return ret;
428 }
429 
430 static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq,
431 				struct work_struct *work)
432 {
433 	struct wq_barrier barr;
434 	int running = 0;
435 
436 	spin_lock_irq(&cwq->lock);
437 	if (unlikely(cwq->current_work == work)) {
438 		insert_wq_barrier(cwq, &barr, 0);
439 		running = 1;
440 	}
441 	spin_unlock_irq(&cwq->lock);
442 
443 	if (unlikely(running))
444 		wait_for_completion(&barr.done);
445 }
446 
447 static void wait_on_work(struct work_struct *work)
448 {
449 	struct cpu_workqueue_struct *cwq;
450 	struct workqueue_struct *wq;
451 	const cpumask_t *cpu_map;
452 	int cpu;
453 
454 	might_sleep();
455 
456 	cwq = get_wq_data(work);
457 	if (!cwq)
458 		return;
459 
460 	wq = cwq->wq;
461 	cpu_map = wq_cpu_map(wq);
462 
463 	for_each_cpu_mask(cpu, *cpu_map)
464 		wait_on_cpu_work(per_cpu_ptr(wq->cpu_wq, cpu), work);
465 }
466 
467 /**
468  * cancel_work_sync - block until a work_struct's callback has terminated
469  * @work: the work which is to be flushed
470  *
471  * cancel_work_sync() will cancel the work if it is queued. If the work's
472  * callback appears to be running, cancel_work_sync() will block until it
473  * has completed.
474  *
475  * It is possible to use this function if the work re-queues itself. It can
476  * cancel the work even if it migrates to another workqueue, however in that
477  * case it only guarantees that work->func() has completed on the last queued
478  * workqueue.
479  *
480  * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
481  * pending, otherwise it goes into a busy-wait loop until the timer expires.
482  *
483  * The caller must ensure that workqueue_struct on which this work was last
484  * queued can't be destroyed before this function returns.
485  */
486 void cancel_work_sync(struct work_struct *work)
487 {
488 	while (!try_to_grab_pending(work))
489 		cpu_relax();
490 	wait_on_work(work);
491 	work_clear_pending(work);
492 }
493 EXPORT_SYMBOL_GPL(cancel_work_sync);
494 
495 /**
496  * cancel_rearming_delayed_work - reliably kill off a delayed work.
497  * @dwork: the delayed work struct
498  *
499  * It is possible to use this function if @dwork rearms itself via queue_work()
500  * or queue_delayed_work(). See also the comment for cancel_work_sync().
501  */
502 void cancel_rearming_delayed_work(struct delayed_work *dwork)
503 {
504 	while (!del_timer(&dwork->timer) &&
505 	       !try_to_grab_pending(&dwork->work))
506 		cpu_relax();
507 	wait_on_work(&dwork->work);
508 	work_clear_pending(&dwork->work);
509 }
510 EXPORT_SYMBOL(cancel_rearming_delayed_work);
511 
512 static struct workqueue_struct *keventd_wq __read_mostly;
513 
514 /**
515  * schedule_work - put work task in global workqueue
516  * @work: job to be done
517  *
518  * This puts a job in the kernel-global workqueue.
519  */
520 int fastcall schedule_work(struct work_struct *work)
521 {
522 	return queue_work(keventd_wq, work);
523 }
524 EXPORT_SYMBOL(schedule_work);
525 
526 /**
527  * schedule_delayed_work - put work task in global workqueue after delay
528  * @dwork: job to be done
529  * @delay: number of jiffies to wait or 0 for immediate execution
530  *
531  * After waiting for a given time this puts a job in the kernel-global
532  * workqueue.
533  */
534 int fastcall schedule_delayed_work(struct delayed_work *dwork,
535 					unsigned long delay)
536 {
537 	timer_stats_timer_set_start_info(&dwork->timer);
538 	return queue_delayed_work(keventd_wq, dwork, delay);
539 }
540 EXPORT_SYMBOL(schedule_delayed_work);
541 
542 /**
543  * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
544  * @cpu: cpu to use
545  * @dwork: job to be done
546  * @delay: number of jiffies to wait
547  *
548  * After waiting for a given time this puts a job in the kernel-global
549  * workqueue on the specified CPU.
550  */
551 int schedule_delayed_work_on(int cpu,
552 			struct delayed_work *dwork, unsigned long delay)
553 {
554 	return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
555 }
556 EXPORT_SYMBOL(schedule_delayed_work_on);
557 
558 /**
559  * schedule_on_each_cpu - call a function on each online CPU from keventd
560  * @func: the function to call
561  *
562  * Returns zero on success.
563  * Returns -ve errno on failure.
564  *
565  * Appears to be racy against CPU hotplug.
566  *
567  * schedule_on_each_cpu() is very slow.
568  */
569 int schedule_on_each_cpu(work_func_t func)
570 {
571 	int cpu;
572 	struct work_struct *works;
573 
574 	works = alloc_percpu(struct work_struct);
575 	if (!works)
576 		return -ENOMEM;
577 
578 	preempt_disable();		/* CPU hotplug */
579 	for_each_online_cpu(cpu) {
580 		struct work_struct *work = per_cpu_ptr(works, cpu);
581 
582 		INIT_WORK(work, func);
583 		set_bit(WORK_STRUCT_PENDING, work_data_bits(work));
584 		__queue_work(per_cpu_ptr(keventd_wq->cpu_wq, cpu), work);
585 	}
586 	preempt_enable();
587 	flush_workqueue(keventd_wq);
588 	free_percpu(works);
589 	return 0;
590 }
591 
592 void flush_scheduled_work(void)
593 {
594 	flush_workqueue(keventd_wq);
595 }
596 EXPORT_SYMBOL(flush_scheduled_work);
597 
598 /**
599  * execute_in_process_context - reliably execute the routine with user context
600  * @fn:		the function to execute
601  * @ew:		guaranteed storage for the execute work structure (must
602  *		be available when the work executes)
603  *
604  * Executes the function immediately if process context is available,
605  * otherwise schedules the function for delayed execution.
606  *
607  * Returns:	0 - function was executed
608  *		1 - function was scheduled for execution
609  */
610 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
611 {
612 	if (!in_interrupt()) {
613 		fn(&ew->work);
614 		return 0;
615 	}
616 
617 	INIT_WORK(&ew->work, fn);
618 	schedule_work(&ew->work);
619 
620 	return 1;
621 }
622 EXPORT_SYMBOL_GPL(execute_in_process_context);
623 
624 int keventd_up(void)
625 {
626 	return keventd_wq != NULL;
627 }
628 
629 int current_is_keventd(void)
630 {
631 	struct cpu_workqueue_struct *cwq;
632 	int cpu = smp_processor_id();	/* preempt-safe: keventd is per-cpu */
633 	int ret = 0;
634 
635 	BUG_ON(!keventd_wq);
636 
637 	cwq = per_cpu_ptr(keventd_wq->cpu_wq, cpu);
638 	if (current == cwq->thread)
639 		ret = 1;
640 
641 	return ret;
642 
643 }
644 
645 static struct cpu_workqueue_struct *
646 init_cpu_workqueue(struct workqueue_struct *wq, int cpu)
647 {
648 	struct cpu_workqueue_struct *cwq = per_cpu_ptr(wq->cpu_wq, cpu);
649 
650 	cwq->wq = wq;
651 	spin_lock_init(&cwq->lock);
652 	INIT_LIST_HEAD(&cwq->worklist);
653 	init_waitqueue_head(&cwq->more_work);
654 
655 	return cwq;
656 }
657 
658 static int create_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
659 {
660 	struct workqueue_struct *wq = cwq->wq;
661 	const char *fmt = is_single_threaded(wq) ? "%s" : "%s/%d";
662 	struct task_struct *p;
663 
664 	p = kthread_create(worker_thread, cwq, fmt, wq->name, cpu);
665 	/*
666 	 * Nobody can add the work_struct to this cwq,
667 	 *	if (caller is __create_workqueue)
668 	 *		nobody should see this wq
669 	 *	else // caller is CPU_UP_PREPARE
670 	 *		cpu is not on cpu_online_map
671 	 * so we can abort safely.
672 	 */
673 	if (IS_ERR(p))
674 		return PTR_ERR(p);
675 
676 	cwq->thread = p;
677 	cwq->should_stop = 0;
678 
679 	return 0;
680 }
681 
682 static void start_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
683 {
684 	struct task_struct *p = cwq->thread;
685 
686 	if (p != NULL) {
687 		if (cpu >= 0)
688 			kthread_bind(p, cpu);
689 		wake_up_process(p);
690 	}
691 }
692 
693 struct workqueue_struct *__create_workqueue(const char *name,
694 					    int singlethread, int freezeable)
695 {
696 	struct workqueue_struct *wq;
697 	struct cpu_workqueue_struct *cwq;
698 	int err = 0, cpu;
699 
700 	wq = kzalloc(sizeof(*wq), GFP_KERNEL);
701 	if (!wq)
702 		return NULL;
703 
704 	wq->cpu_wq = alloc_percpu(struct cpu_workqueue_struct);
705 	if (!wq->cpu_wq) {
706 		kfree(wq);
707 		return NULL;
708 	}
709 
710 	wq->name = name;
711 	wq->singlethread = singlethread;
712 	wq->freezeable = freezeable;
713 	INIT_LIST_HEAD(&wq->list);
714 
715 	if (singlethread) {
716 		cwq = init_cpu_workqueue(wq, singlethread_cpu);
717 		err = create_workqueue_thread(cwq, singlethread_cpu);
718 		start_workqueue_thread(cwq, -1);
719 	} else {
720 		mutex_lock(&workqueue_mutex);
721 		list_add(&wq->list, &workqueues);
722 
723 		for_each_possible_cpu(cpu) {
724 			cwq = init_cpu_workqueue(wq, cpu);
725 			if (err || !cpu_online(cpu))
726 				continue;
727 			err = create_workqueue_thread(cwq, cpu);
728 			start_workqueue_thread(cwq, cpu);
729 		}
730 		mutex_unlock(&workqueue_mutex);
731 	}
732 
733 	if (err) {
734 		destroy_workqueue(wq);
735 		wq = NULL;
736 	}
737 	return wq;
738 }
739 EXPORT_SYMBOL_GPL(__create_workqueue);
740 
741 static void cleanup_workqueue_thread(struct cpu_workqueue_struct *cwq, int cpu)
742 {
743 	struct wq_barrier barr;
744 	int alive = 0;
745 
746 	spin_lock_irq(&cwq->lock);
747 	if (cwq->thread != NULL) {
748 		insert_wq_barrier(cwq, &barr, 1);
749 		cwq->should_stop = 1;
750 		alive = 1;
751 	}
752 	spin_unlock_irq(&cwq->lock);
753 
754 	if (alive) {
755 		wait_for_completion(&barr.done);
756 
757 		while (unlikely(cwq->thread != NULL))
758 			cpu_relax();
759 		/*
760 		 * Wait until cwq->thread unlocks cwq->lock,
761 		 * it won't touch *cwq after that.
762 		 */
763 		smp_rmb();
764 		spin_unlock_wait(&cwq->lock);
765 	}
766 }
767 
768 /**
769  * destroy_workqueue - safely terminate a workqueue
770  * @wq: target workqueue
771  *
772  * Safely destroy a workqueue. All work currently pending will be done first.
773  */
774 void destroy_workqueue(struct workqueue_struct *wq)
775 {
776 	const cpumask_t *cpu_map = wq_cpu_map(wq);
777 	struct cpu_workqueue_struct *cwq;
778 	int cpu;
779 
780 	mutex_lock(&workqueue_mutex);
781 	list_del(&wq->list);
782 	mutex_unlock(&workqueue_mutex);
783 
784 	for_each_cpu_mask(cpu, *cpu_map) {
785 		cwq = per_cpu_ptr(wq->cpu_wq, cpu);
786 		cleanup_workqueue_thread(cwq, cpu);
787 	}
788 
789 	free_percpu(wq->cpu_wq);
790 	kfree(wq);
791 }
792 EXPORT_SYMBOL_GPL(destroy_workqueue);
793 
794 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
795 						unsigned long action,
796 						void *hcpu)
797 {
798 	unsigned int cpu = (unsigned long)hcpu;
799 	struct cpu_workqueue_struct *cwq;
800 	struct workqueue_struct *wq;
801 
802 	action &= ~CPU_TASKS_FROZEN;
803 
804 	switch (action) {
805 	case CPU_LOCK_ACQUIRE:
806 		mutex_lock(&workqueue_mutex);
807 		return NOTIFY_OK;
808 
809 	case CPU_LOCK_RELEASE:
810 		mutex_unlock(&workqueue_mutex);
811 		return NOTIFY_OK;
812 
813 	case CPU_UP_PREPARE:
814 		cpu_set(cpu, cpu_populated_map);
815 	}
816 
817 	list_for_each_entry(wq, &workqueues, list) {
818 		cwq = per_cpu_ptr(wq->cpu_wq, cpu);
819 
820 		switch (action) {
821 		case CPU_UP_PREPARE:
822 			if (!create_workqueue_thread(cwq, cpu))
823 				break;
824 			printk(KERN_ERR "workqueue for %i failed\n", cpu);
825 			return NOTIFY_BAD;
826 
827 		case CPU_ONLINE:
828 			start_workqueue_thread(cwq, cpu);
829 			break;
830 
831 		case CPU_UP_CANCELED:
832 			start_workqueue_thread(cwq, -1);
833 		case CPU_DEAD:
834 			cleanup_workqueue_thread(cwq, cpu);
835 			break;
836 		}
837 	}
838 
839 	return NOTIFY_OK;
840 }
841 
842 void __init init_workqueues(void)
843 {
844 	cpu_populated_map = cpu_online_map;
845 	singlethread_cpu = first_cpu(cpu_possible_map);
846 	cpu_singlethread_map = cpumask_of_cpu(singlethread_cpu);
847 	hotcpu_notifier(workqueue_cpu_callback, 0);
848 	keventd_wq = create_workqueue("events");
849 	BUG_ON(!keventd_wq);
850 }
851