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>
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>
36 #include <linux/idr.h>
38 #include "workqueue_sched.h"
41 /* global_cwq flags */
42 GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
43 GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */
44 GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
45 GCWQ_FREEZING = 1 << 3, /* freeze in progress */
48 WORKER_STARTED = 1 << 0, /* started */
49 WORKER_DIE = 1 << 1, /* die die die */
50 WORKER_IDLE = 1 << 2, /* is idle */
51 WORKER_PREP = 1 << 3, /* preparing to run works */
52 WORKER_ROGUE = 1 << 4, /* not bound to any cpu */
53 WORKER_REBIND = 1 << 5, /* mom is home, come back */
55 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND,
57 /* gcwq->trustee_state */
58 TRUSTEE_START = 0, /* start */
59 TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */
60 TRUSTEE_BUTCHER = 2, /* butcher workers */
61 TRUSTEE_RELEASE = 3, /* release workers */
62 TRUSTEE_DONE = 4, /* trustee is done */
64 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
65 BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
66 BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
68 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
69 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
71 MAYDAY_INITIAL_TIMEOUT = HZ / 100, /* call for help after 10ms */
72 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
73 CREATE_COOLDOWN = HZ, /* time to breath after fail */
74 TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */
77 * Rescue workers are used only on emergencies and shared by
80 RESCUER_NICE_LEVEL = -20,
84 * Structure fields follow one of the following exclusion rules.
86 * I: Set during initialization and read-only afterwards.
88 * P: Preemption protected. Disabling preemption is enough and should
89 * only be modified and accessed from the local cpu.
91 * L: gcwq->lock protected. Access with gcwq->lock held.
93 * X: During normal operation, modification requires gcwq->lock and
94 * should be done only from local cpu. Either disabling preemption
95 * on local cpu or grabbing gcwq->lock is enough for read access.
96 * While trustee is in charge, it's identical to L.
98 * F: wq->flush_mutex protected.
100 * W: workqueue_lock protected.
106 * The poor guys doing the actual heavy lifting. All on-duty workers
107 * are either serving the manager role, on idle list or on busy hash.
110 /* on idle list while idle, on busy hash table while busy */
112 struct list_head entry; /* L: while idle */
113 struct hlist_node hentry; /* L: while busy */
116 struct work_struct *current_work; /* L: work being processed */
117 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
118 struct list_head scheduled; /* L: scheduled works */
119 struct task_struct *task; /* I: worker task */
120 struct global_cwq *gcwq; /* I: the associated gcwq */
121 /* 64 bytes boundary on 64bit, 32 on 32bit */
122 unsigned long last_active; /* L: last active timestamp */
123 unsigned int flags; /* X: flags */
124 int id; /* I: worker id */
125 struct work_struct rebind_work; /* L: rebind worker to cpu */
129 * Global per-cpu workqueue. There's one and only one for each cpu
130 * and all works are queued and processed here regardless of their
134 spinlock_t lock; /* the gcwq lock */
135 struct list_head worklist; /* L: list of pending works */
136 unsigned int cpu; /* I: the associated cpu */
137 unsigned int flags; /* L: GCWQ_* flags */
139 int nr_workers; /* L: total number of workers */
140 int nr_idle; /* L: currently idle ones */
142 /* workers are chained either in the idle_list or busy_hash */
143 struct list_head idle_list; /* X: list of idle workers */
144 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
145 /* L: hash of busy workers */
147 struct timer_list idle_timer; /* L: worker idle timeout */
148 struct timer_list mayday_timer; /* L: SOS timer for dworkers */
150 struct ida worker_ida; /* L: for worker IDs */
152 struct task_struct *trustee; /* L: for gcwq shutdown */
153 unsigned int trustee_state; /* L: trustee state */
154 wait_queue_head_t trustee_wait; /* trustee wait */
155 struct worker *first_idle; /* L: first idle worker */
156 } ____cacheline_aligned_in_smp;
159 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
160 * work_struct->data are used for flags and thus cwqs need to be
161 * aligned at two's power of the number of flag bits.
163 struct cpu_workqueue_struct {
164 struct global_cwq *gcwq; /* I: the associated gcwq */
165 struct workqueue_struct *wq; /* I: the owning workqueue */
166 int work_color; /* L: current color */
167 int flush_color; /* L: flushing color */
168 int nr_in_flight[WORK_NR_COLORS];
169 /* L: nr of in_flight works */
170 int nr_active; /* L: nr of active works */
171 int max_active; /* L: max active works */
172 struct list_head delayed_works; /* L: delayed works */
176 * Structure used to wait for workqueue flush.
179 struct list_head list; /* F: list of flushers */
180 int flush_color; /* F: flush color waiting for */
181 struct completion done; /* flush completion */
185 * The externally visible workqueue abstraction is an array of
186 * per-CPU workqueues:
188 struct workqueue_struct {
189 unsigned int flags; /* I: WQ_* flags */
190 struct cpu_workqueue_struct *cpu_wq; /* I: cwq's */
191 struct list_head list; /* W: list of all workqueues */
193 struct mutex flush_mutex; /* protects wq flushing */
194 int work_color; /* F: current work color */
195 int flush_color; /* F: current flush color */
196 atomic_t nr_cwqs_to_flush; /* flush in progress */
197 struct wq_flusher *first_flusher; /* F: first flusher */
198 struct list_head flusher_queue; /* F: flush waiters */
199 struct list_head flusher_overflow; /* F: flush overflow list */
201 unsigned long single_cpu; /* cpu for single cpu wq */
203 cpumask_var_t mayday_mask; /* cpus requesting rescue */
204 struct worker *rescuer; /* I: rescue worker */
206 int saved_max_active; /* I: saved cwq max_active */
207 const char *name; /* I: workqueue name */
208 #ifdef CONFIG_LOCKDEP
209 struct lockdep_map lockdep_map;
213 #define for_each_busy_worker(worker, i, pos, gcwq) \
214 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
215 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
217 #ifdef CONFIG_DEBUG_OBJECTS_WORK
219 static struct debug_obj_descr work_debug_descr;
222 * fixup_init is called when:
223 * - an active object is initialized
225 static int work_fixup_init(void *addr, enum debug_obj_state state)
227 struct work_struct *work = addr;
230 case ODEBUG_STATE_ACTIVE:
231 cancel_work_sync(work);
232 debug_object_init(work, &work_debug_descr);
240 * fixup_activate is called when:
241 * - an active object is activated
242 * - an unknown object is activated (might be a statically initialized object)
244 static int work_fixup_activate(void *addr, enum debug_obj_state state)
246 struct work_struct *work = addr;
250 case ODEBUG_STATE_NOTAVAILABLE:
252 * This is not really a fixup. The work struct was
253 * statically initialized. We just make sure that it
254 * is tracked in the object tracker.
256 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
257 debug_object_init(work, &work_debug_descr);
258 debug_object_activate(work, &work_debug_descr);
264 case ODEBUG_STATE_ACTIVE:
273 * fixup_free is called when:
274 * - an active object is freed
276 static int work_fixup_free(void *addr, enum debug_obj_state state)
278 struct work_struct *work = addr;
281 case ODEBUG_STATE_ACTIVE:
282 cancel_work_sync(work);
283 debug_object_free(work, &work_debug_descr);
290 static struct debug_obj_descr work_debug_descr = {
291 .name = "work_struct",
292 .fixup_init = work_fixup_init,
293 .fixup_activate = work_fixup_activate,
294 .fixup_free = work_fixup_free,
297 static inline void debug_work_activate(struct work_struct *work)
299 debug_object_activate(work, &work_debug_descr);
302 static inline void debug_work_deactivate(struct work_struct *work)
304 debug_object_deactivate(work, &work_debug_descr);
307 void __init_work(struct work_struct *work, int onstack)
310 debug_object_init_on_stack(work, &work_debug_descr);
312 debug_object_init(work, &work_debug_descr);
314 EXPORT_SYMBOL_GPL(__init_work);
316 void destroy_work_on_stack(struct work_struct *work)
318 debug_object_free(work, &work_debug_descr);
320 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
323 static inline void debug_work_activate(struct work_struct *work) { }
324 static inline void debug_work_deactivate(struct work_struct *work) { }
327 /* Serializes the accesses to the list of workqueues. */
328 static DEFINE_SPINLOCK(workqueue_lock);
329 static LIST_HEAD(workqueues);
330 static bool workqueue_freezing; /* W: have wqs started freezing? */
333 * The almighty global cpu workqueues. nr_running is the only field
334 * which is expected to be used frequently by other cpus via
335 * try_to_wake_up(). Put it in a separate cacheline.
337 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
338 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
340 static int worker_thread(void *__worker);
342 static struct global_cwq *get_gcwq(unsigned int cpu)
344 return &per_cpu(global_cwq, cpu);
347 static atomic_t *get_gcwq_nr_running(unsigned int cpu)
349 return &per_cpu(gcwq_nr_running, cpu);
352 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
353 struct workqueue_struct *wq)
355 return per_cpu_ptr(wq->cpu_wq, cpu);
358 static unsigned int work_color_to_flags(int color)
360 return color << WORK_STRUCT_COLOR_SHIFT;
363 static int get_work_color(struct work_struct *work)
365 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
366 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
369 static int work_next_color(int color)
371 return (color + 1) % WORK_NR_COLORS;
375 * Work data points to the cwq while a work is on queue. Once
376 * execution starts, it points to the cpu the work was last on. This
377 * can be distinguished by comparing the data value against
380 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
381 * cwq, cpu or clear work->data. These functions should only be
382 * called while the work is owned - ie. while the PENDING bit is set.
384 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
385 * corresponding to a work. gcwq is available once the work has been
386 * queued anywhere after initialization. cwq is available only from
387 * queueing until execution starts.
389 static inline void set_work_data(struct work_struct *work, unsigned long data,
392 BUG_ON(!work_pending(work));
393 atomic_long_set(&work->data, data | flags | work_static(work));
396 static void set_work_cwq(struct work_struct *work,
397 struct cpu_workqueue_struct *cwq,
398 unsigned long extra_flags)
400 set_work_data(work, (unsigned long)cwq,
401 WORK_STRUCT_PENDING | extra_flags);
404 static void set_work_cpu(struct work_struct *work, unsigned int cpu)
406 set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
409 static void clear_work_data(struct work_struct *work)
411 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
414 static inline unsigned long get_work_data(struct work_struct *work)
416 return atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK;
419 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
421 unsigned long data = get_work_data(work);
423 return data >= PAGE_OFFSET ? (void *)data : NULL;
426 static struct global_cwq *get_work_gcwq(struct work_struct *work)
428 unsigned long data = get_work_data(work);
431 if (data >= PAGE_OFFSET)
432 return ((struct cpu_workqueue_struct *)data)->gcwq;
434 cpu = data >> WORK_STRUCT_FLAG_BITS;
438 BUG_ON(cpu >= num_possible_cpus());
439 return get_gcwq(cpu);
443 * Policy functions. These define the policies on how the global
444 * worker pool is managed. Unless noted otherwise, these functions
445 * assume that they're being called with gcwq->lock held.
449 * Need to wake up a worker? Called from anything but currently
452 static bool need_more_worker(struct global_cwq *gcwq)
454 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
456 return !list_empty(&gcwq->worklist) && !atomic_read(nr_running);
459 /* Can I start working? Called from busy but !running workers. */
460 static bool may_start_working(struct global_cwq *gcwq)
462 return gcwq->nr_idle;
465 /* Do I need to keep working? Called from currently running workers. */
466 static bool keep_working(struct global_cwq *gcwq)
468 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
470 return !list_empty(&gcwq->worklist) && atomic_read(nr_running) <= 1;
473 /* Do we need a new worker? Called from manager. */
474 static bool need_to_create_worker(struct global_cwq *gcwq)
476 return need_more_worker(gcwq) && !may_start_working(gcwq);
479 /* Do I need to be the manager? */
480 static bool need_to_manage_workers(struct global_cwq *gcwq)
482 return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS;
485 /* Do we have too many workers and should some go away? */
486 static bool too_many_workers(struct global_cwq *gcwq)
488 bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
489 int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
490 int nr_busy = gcwq->nr_workers - nr_idle;
492 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
499 /* Return the first worker. Safe with preemption disabled */
500 static struct worker *first_worker(struct global_cwq *gcwq)
502 if (unlikely(list_empty(&gcwq->idle_list)))
505 return list_first_entry(&gcwq->idle_list, struct worker, entry);
509 * wake_up_worker - wake up an idle worker
510 * @gcwq: gcwq to wake worker for
512 * Wake up the first idle worker of @gcwq.
515 * spin_lock_irq(gcwq->lock).
517 static void wake_up_worker(struct global_cwq *gcwq)
519 struct worker *worker = first_worker(gcwq);
522 wake_up_process(worker->task);
526 * wq_worker_waking_up - a worker is waking up
527 * @task: task waking up
528 * @cpu: CPU @task is waking up to
530 * This function is called during try_to_wake_up() when a worker is
534 * spin_lock_irq(rq->lock)
536 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
538 struct worker *worker = kthread_data(task);
540 if (likely(!(worker->flags & WORKER_NOT_RUNNING)))
541 atomic_inc(get_gcwq_nr_running(cpu));
545 * wq_worker_sleeping - a worker is going to sleep
546 * @task: task going to sleep
547 * @cpu: CPU in question, must be the current CPU number
549 * This function is called during schedule() when a busy worker is
550 * going to sleep. Worker on the same cpu can be woken up by
551 * returning pointer to its task.
554 * spin_lock_irq(rq->lock)
557 * Worker task on @cpu to wake up, %NULL if none.
559 struct task_struct *wq_worker_sleeping(struct task_struct *task,
562 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
563 struct global_cwq *gcwq = get_gcwq(cpu);
564 atomic_t *nr_running = get_gcwq_nr_running(cpu);
566 if (unlikely(worker->flags & WORKER_NOT_RUNNING))
569 /* this can only happen on the local cpu */
570 BUG_ON(cpu != raw_smp_processor_id());
573 * The counterpart of the following dec_and_test, implied mb,
574 * worklist not empty test sequence is in insert_work().
575 * Please read comment there.
577 * NOT_RUNNING is clear. This means that trustee is not in
578 * charge and we're running on the local cpu w/ rq lock held
579 * and preemption disabled, which in turn means that none else
580 * could be manipulating idle_list, so dereferencing idle_list
581 * without gcwq lock is safe.
583 if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
584 to_wakeup = first_worker(gcwq);
585 return to_wakeup ? to_wakeup->task : NULL;
589 * worker_set_flags - set worker flags and adjust nr_running accordingly
590 * @worker: worker to set flags for
591 * @flags: flags to set
592 * @wakeup: wakeup an idle worker if necessary
594 * Set @flags in @worker->flags and adjust nr_running accordingly. If
595 * nr_running becomes zero and @wakeup is %true, an idle worker is
599 * spin_lock_irq(gcwq->lock).
601 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
604 struct global_cwq *gcwq = worker->gcwq;
607 * If transitioning into NOT_RUNNING, adjust nr_running and
608 * wake up an idle worker as necessary if requested by
611 if ((flags & WORKER_NOT_RUNNING) &&
612 !(worker->flags & WORKER_NOT_RUNNING)) {
613 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
616 if (atomic_dec_and_test(nr_running) &&
617 !list_empty(&gcwq->worklist))
618 wake_up_worker(gcwq);
620 atomic_dec(nr_running);
623 worker->flags |= flags;
627 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
628 * @worker: worker to set flags for
629 * @flags: flags to clear
631 * Clear @flags in @worker->flags and adjust nr_running accordingly.
634 * spin_lock_irq(gcwq->lock).
636 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
638 struct global_cwq *gcwq = worker->gcwq;
639 unsigned int oflags = worker->flags;
641 worker->flags &= ~flags;
643 /* if transitioning out of NOT_RUNNING, increment nr_running */
644 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
645 if (!(worker->flags & WORKER_NOT_RUNNING))
646 atomic_inc(get_gcwq_nr_running(gcwq->cpu));
650 * busy_worker_head - return the busy hash head for a work
651 * @gcwq: gcwq of interest
652 * @work: work to be hashed
654 * Return hash head of @gcwq for @work.
657 * spin_lock_irq(gcwq->lock).
660 * Pointer to the hash head.
662 static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
663 struct work_struct *work)
665 const int base_shift = ilog2(sizeof(struct work_struct));
666 unsigned long v = (unsigned long)work;
668 /* simple shift and fold hash, do we need something better? */
670 v += v >> BUSY_WORKER_HASH_ORDER;
671 v &= BUSY_WORKER_HASH_MASK;
673 return &gcwq->busy_hash[v];
677 * __find_worker_executing_work - find worker which is executing a work
678 * @gcwq: gcwq of interest
679 * @bwh: hash head as returned by busy_worker_head()
680 * @work: work to find worker for
682 * Find a worker which is executing @work on @gcwq. @bwh should be
683 * the hash head obtained by calling busy_worker_head() with the same
687 * spin_lock_irq(gcwq->lock).
690 * Pointer to worker which is executing @work if found, NULL
693 static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
694 struct hlist_head *bwh,
695 struct work_struct *work)
697 struct worker *worker;
698 struct hlist_node *tmp;
700 hlist_for_each_entry(worker, tmp, bwh, hentry)
701 if (worker->current_work == work)
707 * find_worker_executing_work - find worker which is executing a work
708 * @gcwq: gcwq of interest
709 * @work: work to find worker for
711 * Find a worker which is executing @work on @gcwq. This function is
712 * identical to __find_worker_executing_work() except that this
713 * function calculates @bwh itself.
716 * spin_lock_irq(gcwq->lock).
719 * Pointer to worker which is executing @work if found, NULL
722 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
723 struct work_struct *work)
725 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
730 * insert_work - insert a work into gcwq
731 * @cwq: cwq @work belongs to
732 * @work: work to insert
733 * @head: insertion point
734 * @extra_flags: extra WORK_STRUCT_* flags to set
736 * Insert @work which belongs to @cwq into @gcwq after @head.
737 * @extra_flags is or'd to work_struct flags.
740 * spin_lock_irq(gcwq->lock).
742 static void insert_work(struct cpu_workqueue_struct *cwq,
743 struct work_struct *work, struct list_head *head,
744 unsigned int extra_flags)
746 struct global_cwq *gcwq = cwq->gcwq;
748 /* we own @work, set data and link */
749 set_work_cwq(work, cwq, extra_flags);
752 * Ensure that we get the right work->data if we see the
753 * result of list_add() below, see try_to_grab_pending().
757 list_add_tail(&work->entry, head);
760 * Ensure either worker_sched_deactivated() sees the above
761 * list_add_tail() or we see zero nr_running to avoid workers
762 * lying around lazily while there are works to be processed.
766 if (!atomic_read(get_gcwq_nr_running(gcwq->cpu)))
767 wake_up_worker(gcwq);
771 * cwq_unbind_single_cpu - unbind cwq from single cpu workqueue processing
772 * @cwq: cwq to unbind
774 * Try to unbind @cwq from single cpu workqueue processing. If
775 * @cwq->wq is frozen, unbind is delayed till the workqueue is thawed.
778 * spin_lock_irq(gcwq->lock).
780 static void cwq_unbind_single_cpu(struct cpu_workqueue_struct *cwq)
782 struct workqueue_struct *wq = cwq->wq;
783 struct global_cwq *gcwq = cwq->gcwq;
785 BUG_ON(wq->single_cpu != gcwq->cpu);
787 * Unbind from workqueue if @cwq is not frozen. If frozen,
788 * thaw_workqueues() will either restart processing on this
789 * cpu or unbind if empty. This keeps works queued while
790 * frozen fully ordered and flushable.
792 if (likely(!(gcwq->flags & GCWQ_FREEZING))) {
793 smp_wmb(); /* paired with cmpxchg() in __queue_work() */
794 wq->single_cpu = NR_CPUS;
798 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
799 struct work_struct *work)
801 struct global_cwq *gcwq;
802 struct cpu_workqueue_struct *cwq;
803 struct list_head *worklist;
807 debug_work_activate(work);
810 * Determine gcwq to use. SINGLE_CPU is inherently
811 * NON_REENTRANT, so test it first.
813 if (!(wq->flags & WQ_SINGLE_CPU)) {
814 struct global_cwq *last_gcwq;
817 * It's multi cpu. If @wq is non-reentrant and @work
818 * was previously on a different cpu, it might still
819 * be running there, in which case the work needs to
820 * be queued on that cpu to guarantee non-reentrance.
822 gcwq = get_gcwq(cpu);
823 if (wq->flags & WQ_NON_REENTRANT &&
824 (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
825 struct worker *worker;
827 spin_lock_irqsave(&last_gcwq->lock, flags);
829 worker = find_worker_executing_work(last_gcwq, work);
831 if (worker && worker->current_cwq->wq == wq)
834 /* meh... not running there, queue here */
835 spin_unlock_irqrestore(&last_gcwq->lock, flags);
836 spin_lock_irqsave(&gcwq->lock, flags);
839 spin_lock_irqsave(&gcwq->lock, flags);
841 unsigned int req_cpu = cpu;
844 * It's a bit more complex for single cpu workqueues.
845 * We first need to determine which cpu is going to be
846 * used. If no cpu is currently serving this
847 * workqueue, arbitrate using atomic accesses to
848 * wq->single_cpu; otherwise, use the current one.
851 cpu = wq->single_cpu;
852 arbitrate = cpu == NR_CPUS;
856 gcwq = get_gcwq(cpu);
857 spin_lock_irqsave(&gcwq->lock, flags);
860 * The following cmpxchg() is a full barrier paired
861 * with smp_wmb() in cwq_unbind_single_cpu() and
862 * guarantees that all changes to wq->st_* fields are
863 * visible on the new cpu after this point.
866 cmpxchg(&wq->single_cpu, NR_CPUS, cpu);
868 if (unlikely(wq->single_cpu != cpu)) {
869 spin_unlock_irqrestore(&gcwq->lock, flags);
874 /* gcwq determined, get cwq and queue */
875 cwq = get_cwq(gcwq->cpu, wq);
877 BUG_ON(!list_empty(&work->entry));
879 cwq->nr_in_flight[cwq->work_color]++;
881 if (likely(cwq->nr_active < cwq->max_active)) {
883 worklist = &gcwq->worklist;
885 worklist = &cwq->delayed_works;
887 insert_work(cwq, work, worklist, work_color_to_flags(cwq->work_color));
889 spin_unlock_irqrestore(&gcwq->lock, flags);
893 * queue_work - queue work on a workqueue
894 * @wq: workqueue to use
895 * @work: work to queue
897 * Returns 0 if @work was already on a queue, non-zero otherwise.
899 * We queue the work to the CPU on which it was submitted, but if the CPU dies
900 * it can be processed by another CPU.
902 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
906 ret = queue_work_on(get_cpu(), wq, work);
911 EXPORT_SYMBOL_GPL(queue_work);
914 * queue_work_on - queue work on specific cpu
915 * @cpu: CPU number to execute work on
916 * @wq: workqueue to use
917 * @work: work to queue
919 * Returns 0 if @work was already on a queue, non-zero otherwise.
921 * We queue the work to a specific CPU, the caller must ensure it
925 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
929 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
930 __queue_work(cpu, wq, work);
935 EXPORT_SYMBOL_GPL(queue_work_on);
937 static void delayed_work_timer_fn(unsigned long __data)
939 struct delayed_work *dwork = (struct delayed_work *)__data;
940 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
942 __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
946 * queue_delayed_work - queue work on a workqueue after delay
947 * @wq: workqueue to use
948 * @dwork: delayable work to queue
949 * @delay: number of jiffies to wait before queueing
951 * Returns 0 if @work was already on a queue, non-zero otherwise.
953 int queue_delayed_work(struct workqueue_struct *wq,
954 struct delayed_work *dwork, unsigned long delay)
957 return queue_work(wq, &dwork->work);
959 return queue_delayed_work_on(-1, wq, dwork, delay);
961 EXPORT_SYMBOL_GPL(queue_delayed_work);
964 * queue_delayed_work_on - queue work on specific CPU after delay
965 * @cpu: CPU number to execute work on
966 * @wq: workqueue to use
967 * @dwork: work to queue
968 * @delay: number of jiffies to wait before queueing
970 * Returns 0 if @work was already on a queue, non-zero otherwise.
972 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
973 struct delayed_work *dwork, unsigned long delay)
976 struct timer_list *timer = &dwork->timer;
977 struct work_struct *work = &dwork->work;
979 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
980 struct global_cwq *gcwq = get_work_gcwq(work);
981 unsigned int lcpu = gcwq ? gcwq->cpu : raw_smp_processor_id();
983 BUG_ON(timer_pending(timer));
984 BUG_ON(!list_empty(&work->entry));
986 timer_stats_timer_set_start_info(&dwork->timer);
988 * This stores cwq for the moment, for the timer_fn.
989 * Note that the work's gcwq is preserved to allow
990 * reentrance detection for delayed works.
992 set_work_cwq(work, get_cwq(lcpu, wq), 0);
993 timer->expires = jiffies + delay;
994 timer->data = (unsigned long)dwork;
995 timer->function = delayed_work_timer_fn;
997 if (unlikely(cpu >= 0))
998 add_timer_on(timer, cpu);
1005 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1008 * worker_enter_idle - enter idle state
1009 * @worker: worker which is entering idle state
1011 * @worker is entering idle state. Update stats and idle timer if
1015 * spin_lock_irq(gcwq->lock).
1017 static void worker_enter_idle(struct worker *worker)
1019 struct global_cwq *gcwq = worker->gcwq;
1021 BUG_ON(worker->flags & WORKER_IDLE);
1022 BUG_ON(!list_empty(&worker->entry) &&
1023 (worker->hentry.next || worker->hentry.pprev));
1025 worker_set_flags(worker, WORKER_IDLE, false);
1027 worker->last_active = jiffies;
1029 /* idle_list is LIFO */
1030 list_add(&worker->entry, &gcwq->idle_list);
1032 if (likely(!(worker->flags & WORKER_ROGUE))) {
1033 if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1034 mod_timer(&gcwq->idle_timer,
1035 jiffies + IDLE_WORKER_TIMEOUT);
1037 wake_up_all(&gcwq->trustee_wait);
1041 * worker_leave_idle - leave idle state
1042 * @worker: worker which is leaving idle state
1044 * @worker is leaving idle state. Update stats.
1047 * spin_lock_irq(gcwq->lock).
1049 static void worker_leave_idle(struct worker *worker)
1051 struct global_cwq *gcwq = worker->gcwq;
1053 BUG_ON(!(worker->flags & WORKER_IDLE));
1054 worker_clr_flags(worker, WORKER_IDLE);
1056 list_del_init(&worker->entry);
1060 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1063 * Works which are scheduled while the cpu is online must at least be
1064 * scheduled to a worker which is bound to the cpu so that if they are
1065 * flushed from cpu callbacks while cpu is going down, they are
1066 * guaranteed to execute on the cpu.
1068 * This function is to be used by rogue workers and rescuers to bind
1069 * themselves to the target cpu and may race with cpu going down or
1070 * coming online. kthread_bind() can't be used because it may put the
1071 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1072 * verbatim as it's best effort and blocking and gcwq may be
1073 * [dis]associated in the meantime.
1075 * This function tries set_cpus_allowed() and locks gcwq and verifies
1076 * the binding against GCWQ_DISASSOCIATED which is set during
1077 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1078 * idle state or fetches works without dropping lock, it can guarantee
1079 * the scheduling requirement described in the first paragraph.
1082 * Might sleep. Called without any lock but returns with gcwq->lock
1086 * %true if the associated gcwq is online (@worker is successfully
1087 * bound), %false if offline.
1089 static bool worker_maybe_bind_and_lock(struct worker *worker)
1091 struct global_cwq *gcwq = worker->gcwq;
1092 struct task_struct *task = worker->task;
1096 * The following call may fail, succeed or succeed
1097 * without actually migrating the task to the cpu if
1098 * it races with cpu hotunplug operation. Verify
1099 * against GCWQ_DISASSOCIATED.
1101 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1103 spin_lock_irq(&gcwq->lock);
1104 if (gcwq->flags & GCWQ_DISASSOCIATED)
1106 if (task_cpu(task) == gcwq->cpu &&
1107 cpumask_equal(¤t->cpus_allowed,
1108 get_cpu_mask(gcwq->cpu)))
1110 spin_unlock_irq(&gcwq->lock);
1112 /* CPU has come up inbetween, retry migration */
1118 * Function for worker->rebind_work used to rebind rogue busy workers
1119 * to the associated cpu which is coming back online. This is
1120 * scheduled by cpu up but can race with other cpu hotplug operations
1121 * and may be executed twice without intervening cpu down.
1123 static void worker_rebind_fn(struct work_struct *work)
1125 struct worker *worker = container_of(work, struct worker, rebind_work);
1126 struct global_cwq *gcwq = worker->gcwq;
1128 if (worker_maybe_bind_and_lock(worker))
1129 worker_clr_flags(worker, WORKER_REBIND);
1131 spin_unlock_irq(&gcwq->lock);
1134 static struct worker *alloc_worker(void)
1136 struct worker *worker;
1138 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1140 INIT_LIST_HEAD(&worker->entry);
1141 INIT_LIST_HEAD(&worker->scheduled);
1142 INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1143 /* on creation a worker is in !idle && prep state */
1144 worker->flags = WORKER_PREP;
1150 * create_worker - create a new workqueue worker
1151 * @gcwq: gcwq the new worker will belong to
1152 * @bind: whether to set affinity to @cpu or not
1154 * Create a new worker which is bound to @gcwq. The returned worker
1155 * can be started by calling start_worker() or destroyed using
1159 * Might sleep. Does GFP_KERNEL allocations.
1162 * Pointer to the newly created worker.
1164 static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
1167 struct worker *worker = NULL;
1169 spin_lock_irq(&gcwq->lock);
1170 while (ida_get_new(&gcwq->worker_ida, &id)) {
1171 spin_unlock_irq(&gcwq->lock);
1172 if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1174 spin_lock_irq(&gcwq->lock);
1176 spin_unlock_irq(&gcwq->lock);
1178 worker = alloc_worker();
1182 worker->gcwq = gcwq;
1185 worker->task = kthread_create(worker_thread, worker, "kworker/%u:%d",
1187 if (IS_ERR(worker->task))
1191 * A rogue worker will become a regular one if CPU comes
1192 * online later on. Make sure every worker has
1193 * PF_THREAD_BOUND set.
1196 kthread_bind(worker->task, gcwq->cpu);
1198 worker->task->flags |= PF_THREAD_BOUND;
1203 spin_lock_irq(&gcwq->lock);
1204 ida_remove(&gcwq->worker_ida, id);
1205 spin_unlock_irq(&gcwq->lock);
1212 * start_worker - start a newly created worker
1213 * @worker: worker to start
1215 * Make the gcwq aware of @worker and start it.
1218 * spin_lock_irq(gcwq->lock).
1220 static void start_worker(struct worker *worker)
1222 worker_set_flags(worker, WORKER_STARTED, false);
1223 worker->gcwq->nr_workers++;
1224 worker_enter_idle(worker);
1225 wake_up_process(worker->task);
1229 * destroy_worker - destroy a workqueue worker
1230 * @worker: worker to be destroyed
1232 * Destroy @worker and adjust @gcwq stats accordingly.
1235 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1237 static void destroy_worker(struct worker *worker)
1239 struct global_cwq *gcwq = worker->gcwq;
1240 int id = worker->id;
1242 /* sanity check frenzy */
1243 BUG_ON(worker->current_work);
1244 BUG_ON(!list_empty(&worker->scheduled));
1246 if (worker->flags & WORKER_STARTED)
1248 if (worker->flags & WORKER_IDLE)
1251 list_del_init(&worker->entry);
1252 worker_set_flags(worker, WORKER_DIE, false);
1254 spin_unlock_irq(&gcwq->lock);
1256 kthread_stop(worker->task);
1259 spin_lock_irq(&gcwq->lock);
1260 ida_remove(&gcwq->worker_ida, id);
1263 static void idle_worker_timeout(unsigned long __gcwq)
1265 struct global_cwq *gcwq = (void *)__gcwq;
1267 spin_lock_irq(&gcwq->lock);
1269 if (too_many_workers(gcwq)) {
1270 struct worker *worker;
1271 unsigned long expires;
1273 /* idle_list is kept in LIFO order, check the last one */
1274 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1275 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1277 if (time_before(jiffies, expires))
1278 mod_timer(&gcwq->idle_timer, expires);
1280 /* it's been idle for too long, wake up manager */
1281 gcwq->flags |= GCWQ_MANAGE_WORKERS;
1282 wake_up_worker(gcwq);
1286 spin_unlock_irq(&gcwq->lock);
1289 static bool send_mayday(struct work_struct *work)
1291 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1292 struct workqueue_struct *wq = cwq->wq;
1294 if (!(wq->flags & WQ_RESCUER))
1297 /* mayday mayday mayday */
1298 if (!cpumask_test_and_set_cpu(cwq->gcwq->cpu, wq->mayday_mask))
1299 wake_up_process(wq->rescuer->task);
1303 static void gcwq_mayday_timeout(unsigned long __gcwq)
1305 struct global_cwq *gcwq = (void *)__gcwq;
1306 struct work_struct *work;
1308 spin_lock_irq(&gcwq->lock);
1310 if (need_to_create_worker(gcwq)) {
1312 * We've been trying to create a new worker but
1313 * haven't been successful. We might be hitting an
1314 * allocation deadlock. Send distress signals to
1317 list_for_each_entry(work, &gcwq->worklist, entry)
1321 spin_unlock_irq(&gcwq->lock);
1323 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1327 * maybe_create_worker - create a new worker if necessary
1328 * @gcwq: gcwq to create a new worker for
1330 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1331 * have at least one idle worker on return from this function. If
1332 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1333 * sent to all rescuers with works scheduled on @gcwq to resolve
1334 * possible allocation deadlock.
1336 * On return, need_to_create_worker() is guaranteed to be false and
1337 * may_start_working() true.
1340 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1341 * multiple times. Does GFP_KERNEL allocations. Called only from
1345 * false if no action was taken and gcwq->lock stayed locked, true
1348 static bool maybe_create_worker(struct global_cwq *gcwq)
1350 if (!need_to_create_worker(gcwq))
1353 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1354 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1357 struct worker *worker;
1359 spin_unlock_irq(&gcwq->lock);
1361 worker = create_worker(gcwq, true);
1363 del_timer_sync(&gcwq->mayday_timer);
1364 spin_lock_irq(&gcwq->lock);
1365 start_worker(worker);
1366 BUG_ON(need_to_create_worker(gcwq));
1370 if (!need_to_create_worker(gcwq))
1373 spin_unlock_irq(&gcwq->lock);
1374 __set_current_state(TASK_INTERRUPTIBLE);
1375 schedule_timeout(CREATE_COOLDOWN);
1376 spin_lock_irq(&gcwq->lock);
1377 if (!need_to_create_worker(gcwq))
1381 spin_unlock_irq(&gcwq->lock);
1382 del_timer_sync(&gcwq->mayday_timer);
1383 spin_lock_irq(&gcwq->lock);
1384 if (need_to_create_worker(gcwq))
1390 * maybe_destroy_worker - destroy workers which have been idle for a while
1391 * @gcwq: gcwq to destroy workers for
1393 * Destroy @gcwq workers which have been idle for longer than
1394 * IDLE_WORKER_TIMEOUT.
1397 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1398 * multiple times. Called only from manager.
1401 * false if no action was taken and gcwq->lock stayed locked, true
1404 static bool maybe_destroy_workers(struct global_cwq *gcwq)
1408 while (too_many_workers(gcwq)) {
1409 struct worker *worker;
1410 unsigned long expires;
1412 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1413 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1415 if (time_before(jiffies, expires)) {
1416 mod_timer(&gcwq->idle_timer, expires);
1420 destroy_worker(worker);
1428 * manage_workers - manage worker pool
1431 * Assume the manager role and manage gcwq worker pool @worker belongs
1432 * to. At any given time, there can be only zero or one manager per
1433 * gcwq. The exclusion is handled automatically by this function.
1435 * The caller can safely start processing works on false return. On
1436 * true return, it's guaranteed that need_to_create_worker() is false
1437 * and may_start_working() is true.
1440 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1441 * multiple times. Does GFP_KERNEL allocations.
1444 * false if no action was taken and gcwq->lock stayed locked, true if
1445 * some action was taken.
1447 static bool manage_workers(struct worker *worker)
1449 struct global_cwq *gcwq = worker->gcwq;
1452 if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1455 gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1456 gcwq->flags |= GCWQ_MANAGING_WORKERS;
1459 * Destroy and then create so that may_start_working() is true
1462 ret |= maybe_destroy_workers(gcwq);
1463 ret |= maybe_create_worker(gcwq);
1465 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1468 * The trustee might be waiting to take over the manager
1469 * position, tell it we're done.
1471 if (unlikely(gcwq->trustee))
1472 wake_up_all(&gcwq->trustee_wait);
1478 * move_linked_works - move linked works to a list
1479 * @work: start of series of works to be scheduled
1480 * @head: target list to append @work to
1481 * @nextp: out paramter for nested worklist walking
1483 * Schedule linked works starting from @work to @head. Work series to
1484 * be scheduled starts at @work and includes any consecutive work with
1485 * WORK_STRUCT_LINKED set in its predecessor.
1487 * If @nextp is not NULL, it's updated to point to the next work of
1488 * the last scheduled work. This allows move_linked_works() to be
1489 * nested inside outer list_for_each_entry_safe().
1492 * spin_lock_irq(gcwq->lock).
1494 static void move_linked_works(struct work_struct *work, struct list_head *head,
1495 struct work_struct **nextp)
1497 struct work_struct *n;
1500 * Linked worklist will always end before the end of the list,
1501 * use NULL for list head.
1503 list_for_each_entry_safe_from(work, n, NULL, entry) {
1504 list_move_tail(&work->entry, head);
1505 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1510 * If we're already inside safe list traversal and have moved
1511 * multiple works to the scheduled queue, the next position
1512 * needs to be updated.
1518 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1520 struct work_struct *work = list_first_entry(&cwq->delayed_works,
1521 struct work_struct, entry);
1523 move_linked_works(work, &cwq->gcwq->worklist, NULL);
1528 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1529 * @cwq: cwq of interest
1530 * @color: color of work which left the queue
1532 * A work either has completed or is removed from pending queue,
1533 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1536 * spin_lock_irq(gcwq->lock).
1538 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
1540 /* ignore uncolored works */
1541 if (color == WORK_NO_COLOR)
1544 cwq->nr_in_flight[color]--;
1547 if (!list_empty(&cwq->delayed_works)) {
1548 /* one down, submit a delayed one */
1549 if (cwq->nr_active < cwq->max_active)
1550 cwq_activate_first_delayed(cwq);
1551 } else if (!cwq->nr_active && cwq->wq->flags & WQ_SINGLE_CPU) {
1552 /* this was the last work, unbind from single cpu */
1553 cwq_unbind_single_cpu(cwq);
1556 /* is flush in progress and are we at the flushing tip? */
1557 if (likely(cwq->flush_color != color))
1560 /* are there still in-flight works? */
1561 if (cwq->nr_in_flight[color])
1564 /* this cwq is done, clear flush_color */
1565 cwq->flush_color = -1;
1568 * If this was the last cwq, wake up the first flusher. It
1569 * will handle the rest.
1571 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1572 complete(&cwq->wq->first_flusher->done);
1576 * process_one_work - process single work
1578 * @work: work to process
1580 * Process @work. This function contains all the logics necessary to
1581 * process a single work including synchronization against and
1582 * interaction with other workers on the same cpu, queueing and
1583 * flushing. As long as context requirement is met, any worker can
1584 * call this function to process a work.
1587 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1589 static void process_one_work(struct worker *worker, struct work_struct *work)
1591 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1592 struct global_cwq *gcwq = cwq->gcwq;
1593 struct hlist_head *bwh = busy_worker_head(gcwq, work);
1594 work_func_t f = work->func;
1596 struct worker *collision;
1597 #ifdef CONFIG_LOCKDEP
1599 * It is permissible to free the struct work_struct from
1600 * inside the function that is called from it, this we need to
1601 * take into account for lockdep too. To avoid bogus "held
1602 * lock freed" warnings as well as problems when looking into
1603 * work->lockdep_map, make a copy and use that here.
1605 struct lockdep_map lockdep_map = work->lockdep_map;
1608 * A single work shouldn't be executed concurrently by
1609 * multiple workers on a single cpu. Check whether anyone is
1610 * already processing the work. If so, defer the work to the
1611 * currently executing one.
1613 collision = __find_worker_executing_work(gcwq, bwh, work);
1614 if (unlikely(collision)) {
1615 move_linked_works(work, &collision->scheduled, NULL);
1619 /* claim and process */
1620 debug_work_deactivate(work);
1621 hlist_add_head(&worker->hentry, bwh);
1622 worker->current_work = work;
1623 worker->current_cwq = cwq;
1624 work_color = get_work_color(work);
1626 /* record the current cpu number in the work data and dequeue */
1627 set_work_cpu(work, gcwq->cpu);
1628 list_del_init(&work->entry);
1630 spin_unlock_irq(&gcwq->lock);
1632 work_clear_pending(work);
1633 lock_map_acquire(&cwq->wq->lockdep_map);
1634 lock_map_acquire(&lockdep_map);
1636 lock_map_release(&lockdep_map);
1637 lock_map_release(&cwq->wq->lockdep_map);
1639 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1640 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1642 current->comm, preempt_count(), task_pid_nr(current));
1643 printk(KERN_ERR " last function: ");
1644 print_symbol("%s\n", (unsigned long)f);
1645 debug_show_held_locks(current);
1649 spin_lock_irq(&gcwq->lock);
1651 /* we're done with it, release */
1652 hlist_del_init(&worker->hentry);
1653 worker->current_work = NULL;
1654 worker->current_cwq = NULL;
1655 cwq_dec_nr_in_flight(cwq, work_color);
1659 * process_scheduled_works - process scheduled works
1662 * Process all scheduled works. Please note that the scheduled list
1663 * may change while processing a work, so this function repeatedly
1664 * fetches a work from the top and executes it.
1667 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1670 static void process_scheduled_works(struct worker *worker)
1672 while (!list_empty(&worker->scheduled)) {
1673 struct work_struct *work = list_first_entry(&worker->scheduled,
1674 struct work_struct, entry);
1675 process_one_work(worker, work);
1680 * worker_thread - the worker thread function
1683 * The gcwq worker thread function. There's a single dynamic pool of
1684 * these per each cpu. These workers process all works regardless of
1685 * their specific target workqueue. The only exception is works which
1686 * belong to workqueues with a rescuer which will be explained in
1689 static int worker_thread(void *__worker)
1691 struct worker *worker = __worker;
1692 struct global_cwq *gcwq = worker->gcwq;
1694 /* tell the scheduler that this is a workqueue worker */
1695 worker->task->flags |= PF_WQ_WORKER;
1697 spin_lock_irq(&gcwq->lock);
1699 /* DIE can be set only while we're idle, checking here is enough */
1700 if (worker->flags & WORKER_DIE) {
1701 spin_unlock_irq(&gcwq->lock);
1702 worker->task->flags &= ~PF_WQ_WORKER;
1706 worker_leave_idle(worker);
1708 /* no more worker necessary? */
1709 if (!need_more_worker(gcwq))
1712 /* do we need to manage? */
1713 if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1717 * ->scheduled list can only be filled while a worker is
1718 * preparing to process a work or actually processing it.
1719 * Make sure nobody diddled with it while I was sleeping.
1721 BUG_ON(!list_empty(&worker->scheduled));
1724 * When control reaches this point, we're guaranteed to have
1725 * at least one idle worker or that someone else has already
1726 * assumed the manager role.
1728 worker_clr_flags(worker, WORKER_PREP);
1731 struct work_struct *work =
1732 list_first_entry(&gcwq->worklist,
1733 struct work_struct, entry);
1735 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1736 /* optimization path, not strictly necessary */
1737 process_one_work(worker, work);
1738 if (unlikely(!list_empty(&worker->scheduled)))
1739 process_scheduled_works(worker);
1741 move_linked_works(work, &worker->scheduled, NULL);
1742 process_scheduled_works(worker);
1744 } while (keep_working(gcwq));
1746 worker_set_flags(worker, WORKER_PREP, false);
1748 if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1752 * gcwq->lock is held and there's no work to process and no
1753 * need to manage, sleep. Workers are woken up only while
1754 * holding gcwq->lock or from local cpu, so setting the
1755 * current state before releasing gcwq->lock is enough to
1756 * prevent losing any event.
1758 worker_enter_idle(worker);
1759 __set_current_state(TASK_INTERRUPTIBLE);
1760 spin_unlock_irq(&gcwq->lock);
1766 * rescuer_thread - the rescuer thread function
1767 * @__wq: the associated workqueue
1769 * Workqueue rescuer thread function. There's one rescuer for each
1770 * workqueue which has WQ_RESCUER set.
1772 * Regular work processing on a gcwq may block trying to create a new
1773 * worker which uses GFP_KERNEL allocation which has slight chance of
1774 * developing into deadlock if some works currently on the same queue
1775 * need to be processed to satisfy the GFP_KERNEL allocation. This is
1776 * the problem rescuer solves.
1778 * When such condition is possible, the gcwq summons rescuers of all
1779 * workqueues which have works queued on the gcwq and let them process
1780 * those works so that forward progress can be guaranteed.
1782 * This should happen rarely.
1784 static int rescuer_thread(void *__wq)
1786 struct workqueue_struct *wq = __wq;
1787 struct worker *rescuer = wq->rescuer;
1788 struct list_head *scheduled = &rescuer->scheduled;
1791 set_user_nice(current, RESCUER_NICE_LEVEL);
1793 set_current_state(TASK_INTERRUPTIBLE);
1795 if (kthread_should_stop())
1798 for_each_cpu(cpu, wq->mayday_mask) {
1799 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
1800 struct global_cwq *gcwq = cwq->gcwq;
1801 struct work_struct *work, *n;
1803 __set_current_state(TASK_RUNNING);
1804 cpumask_clear_cpu(cpu, wq->mayday_mask);
1806 /* migrate to the target cpu if possible */
1807 rescuer->gcwq = gcwq;
1808 worker_maybe_bind_and_lock(rescuer);
1811 * Slurp in all works issued via this workqueue and
1814 BUG_ON(!list_empty(&rescuer->scheduled));
1815 list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
1816 if (get_work_cwq(work) == cwq)
1817 move_linked_works(work, scheduled, &n);
1819 process_scheduled_works(rescuer);
1820 spin_unlock_irq(&gcwq->lock);
1828 struct work_struct work;
1829 struct completion done;
1832 static void wq_barrier_func(struct work_struct *work)
1834 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
1835 complete(&barr->done);
1839 * insert_wq_barrier - insert a barrier work
1840 * @cwq: cwq to insert barrier into
1841 * @barr: wq_barrier to insert
1842 * @target: target work to attach @barr to
1843 * @worker: worker currently executing @target, NULL if @target is not executing
1845 * @barr is linked to @target such that @barr is completed only after
1846 * @target finishes execution. Please note that the ordering
1847 * guarantee is observed only with respect to @target and on the local
1850 * Currently, a queued barrier can't be canceled. This is because
1851 * try_to_grab_pending() can't determine whether the work to be
1852 * grabbed is at the head of the queue and thus can't clear LINKED
1853 * flag of the previous work while there must be a valid next work
1854 * after a work with LINKED flag set.
1856 * Note that when @worker is non-NULL, @target may be modified
1857 * underneath us, so we can't reliably determine cwq from @target.
1860 * spin_lock_irq(gcwq->lock).
1862 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
1863 struct wq_barrier *barr,
1864 struct work_struct *target, struct worker *worker)
1866 struct list_head *head;
1867 unsigned int linked = 0;
1870 * debugobject calls are safe here even with gcwq->lock locked
1871 * as we know for sure that this will not trigger any of the
1872 * checks and call back into the fixup functions where we
1875 INIT_WORK_ON_STACK(&barr->work, wq_barrier_func);
1876 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
1877 init_completion(&barr->done);
1880 * If @target is currently being executed, schedule the
1881 * barrier to the worker; otherwise, put it after @target.
1884 head = worker->scheduled.next;
1886 unsigned long *bits = work_data_bits(target);
1888 head = target->entry.next;
1889 /* there can already be other linked works, inherit and set */
1890 linked = *bits & WORK_STRUCT_LINKED;
1891 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
1894 debug_work_activate(&barr->work);
1895 insert_work(cwq, &barr->work, head,
1896 work_color_to_flags(WORK_NO_COLOR) | linked);
1900 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
1901 * @wq: workqueue being flushed
1902 * @flush_color: new flush color, < 0 for no-op
1903 * @work_color: new work color, < 0 for no-op
1905 * Prepare cwqs for workqueue flushing.
1907 * If @flush_color is non-negative, flush_color on all cwqs should be
1908 * -1. If no cwq has in-flight commands at the specified color, all
1909 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
1910 * has in flight commands, its cwq->flush_color is set to
1911 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
1912 * wakeup logic is armed and %true is returned.
1914 * The caller should have initialized @wq->first_flusher prior to
1915 * calling this function with non-negative @flush_color. If
1916 * @flush_color is negative, no flush color update is done and %false
1919 * If @work_color is non-negative, all cwqs should have the same
1920 * work_color which is previous to @work_color and all will be
1921 * advanced to @work_color.
1924 * mutex_lock(wq->flush_mutex).
1927 * %true if @flush_color >= 0 and there's something to flush. %false
1930 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
1931 int flush_color, int work_color)
1936 if (flush_color >= 0) {
1937 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
1938 atomic_set(&wq->nr_cwqs_to_flush, 1);
1941 for_each_possible_cpu(cpu) {
1942 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
1943 struct global_cwq *gcwq = cwq->gcwq;
1945 spin_lock_irq(&gcwq->lock);
1947 if (flush_color >= 0) {
1948 BUG_ON(cwq->flush_color != -1);
1950 if (cwq->nr_in_flight[flush_color]) {
1951 cwq->flush_color = flush_color;
1952 atomic_inc(&wq->nr_cwqs_to_flush);
1957 if (work_color >= 0) {
1958 BUG_ON(work_color != work_next_color(cwq->work_color));
1959 cwq->work_color = work_color;
1962 spin_unlock_irq(&gcwq->lock);
1965 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
1966 complete(&wq->first_flusher->done);
1972 * flush_workqueue - ensure that any scheduled work has run to completion.
1973 * @wq: workqueue to flush
1975 * Forces execution of the workqueue and blocks until its completion.
1976 * This is typically used in driver shutdown handlers.
1978 * We sleep until all works which were queued on entry have been handled,
1979 * but we are not livelocked by new incoming ones.
1981 void flush_workqueue(struct workqueue_struct *wq)
1983 struct wq_flusher this_flusher = {
1984 .list = LIST_HEAD_INIT(this_flusher.list),
1986 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
1990 lock_map_acquire(&wq->lockdep_map);
1991 lock_map_release(&wq->lockdep_map);
1993 mutex_lock(&wq->flush_mutex);
1996 * Start-to-wait phase
1998 next_color = work_next_color(wq->work_color);
2000 if (next_color != wq->flush_color) {
2002 * Color space is not full. The current work_color
2003 * becomes our flush_color and work_color is advanced
2006 BUG_ON(!list_empty(&wq->flusher_overflow));
2007 this_flusher.flush_color = wq->work_color;
2008 wq->work_color = next_color;
2010 if (!wq->first_flusher) {
2011 /* no flush in progress, become the first flusher */
2012 BUG_ON(wq->flush_color != this_flusher.flush_color);
2014 wq->first_flusher = &this_flusher;
2016 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2018 /* nothing to flush, done */
2019 wq->flush_color = next_color;
2020 wq->first_flusher = NULL;
2025 BUG_ON(wq->flush_color == this_flusher.flush_color);
2026 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2027 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2031 * Oops, color space is full, wait on overflow queue.
2032 * The next flush completion will assign us
2033 * flush_color and transfer to flusher_queue.
2035 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2038 mutex_unlock(&wq->flush_mutex);
2040 wait_for_completion(&this_flusher.done);
2043 * Wake-up-and-cascade phase
2045 * First flushers are responsible for cascading flushes and
2046 * handling overflow. Non-first flushers can simply return.
2048 if (wq->first_flusher != &this_flusher)
2051 mutex_lock(&wq->flush_mutex);
2053 wq->first_flusher = NULL;
2055 BUG_ON(!list_empty(&this_flusher.list));
2056 BUG_ON(wq->flush_color != this_flusher.flush_color);
2059 struct wq_flusher *next, *tmp;
2061 /* complete all the flushers sharing the current flush color */
2062 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2063 if (next->flush_color != wq->flush_color)
2065 list_del_init(&next->list);
2066 complete(&next->done);
2069 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2070 wq->flush_color != work_next_color(wq->work_color));
2072 /* this flush_color is finished, advance by one */
2073 wq->flush_color = work_next_color(wq->flush_color);
2075 /* one color has been freed, handle overflow queue */
2076 if (!list_empty(&wq->flusher_overflow)) {
2078 * Assign the same color to all overflowed
2079 * flushers, advance work_color and append to
2080 * flusher_queue. This is the start-to-wait
2081 * phase for these overflowed flushers.
2083 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2084 tmp->flush_color = wq->work_color;
2086 wq->work_color = work_next_color(wq->work_color);
2088 list_splice_tail_init(&wq->flusher_overflow,
2089 &wq->flusher_queue);
2090 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2093 if (list_empty(&wq->flusher_queue)) {
2094 BUG_ON(wq->flush_color != wq->work_color);
2099 * Need to flush more colors. Make the next flusher
2100 * the new first flusher and arm cwqs.
2102 BUG_ON(wq->flush_color == wq->work_color);
2103 BUG_ON(wq->flush_color != next->flush_color);
2105 list_del_init(&next->list);
2106 wq->first_flusher = next;
2108 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2112 * Meh... this color is already done, clear first
2113 * flusher and repeat cascading.
2115 wq->first_flusher = NULL;
2119 mutex_unlock(&wq->flush_mutex);
2121 EXPORT_SYMBOL_GPL(flush_workqueue);
2124 * flush_work - block until a work_struct's callback has terminated
2125 * @work: the work which is to be flushed
2127 * Returns false if @work has already terminated.
2129 * It is expected that, prior to calling flush_work(), the caller has
2130 * arranged for the work to not be requeued, otherwise it doesn't make
2131 * sense to use this function.
2133 int flush_work(struct work_struct *work)
2135 struct worker *worker = NULL;
2136 struct global_cwq *gcwq;
2137 struct cpu_workqueue_struct *cwq;
2138 struct wq_barrier barr;
2141 gcwq = get_work_gcwq(work);
2145 spin_lock_irq(&gcwq->lock);
2146 if (!list_empty(&work->entry)) {
2148 * See the comment near try_to_grab_pending()->smp_rmb().
2149 * If it was re-queued to a different gcwq under us, we
2150 * are not going to wait.
2153 cwq = get_work_cwq(work);
2154 if (unlikely(!cwq || gcwq != cwq->gcwq))
2157 worker = find_worker_executing_work(gcwq, work);
2160 cwq = worker->current_cwq;
2163 insert_wq_barrier(cwq, &barr, work, worker);
2164 spin_unlock_irq(&gcwq->lock);
2166 lock_map_acquire(&cwq->wq->lockdep_map);
2167 lock_map_release(&cwq->wq->lockdep_map);
2169 wait_for_completion(&barr.done);
2170 destroy_work_on_stack(&barr.work);
2173 spin_unlock_irq(&gcwq->lock);
2176 EXPORT_SYMBOL_GPL(flush_work);
2179 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2180 * so this work can't be re-armed in any way.
2182 static int try_to_grab_pending(struct work_struct *work)
2184 struct global_cwq *gcwq;
2187 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2191 * The queueing is in progress, or it is already queued. Try to
2192 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2194 gcwq = get_work_gcwq(work);
2198 spin_lock_irq(&gcwq->lock);
2199 if (!list_empty(&work->entry)) {
2201 * This work is queued, but perhaps we locked the wrong gcwq.
2202 * In that case we must see the new value after rmb(), see
2203 * insert_work()->wmb().
2206 if (gcwq == get_work_gcwq(work)) {
2207 debug_work_deactivate(work);
2208 list_del_init(&work->entry);
2209 cwq_dec_nr_in_flight(get_work_cwq(work),
2210 get_work_color(work));
2214 spin_unlock_irq(&gcwq->lock);
2219 static void wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2221 struct wq_barrier barr;
2222 struct worker *worker;
2224 spin_lock_irq(&gcwq->lock);
2226 worker = find_worker_executing_work(gcwq, work);
2227 if (unlikely(worker))
2228 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2230 spin_unlock_irq(&gcwq->lock);
2232 if (unlikely(worker)) {
2233 wait_for_completion(&barr.done);
2234 destroy_work_on_stack(&barr.work);
2238 static void wait_on_work(struct work_struct *work)
2244 lock_map_acquire(&work->lockdep_map);
2245 lock_map_release(&work->lockdep_map);
2247 for_each_possible_cpu(cpu)
2248 wait_on_cpu_work(get_gcwq(cpu), work);
2251 static int __cancel_work_timer(struct work_struct *work,
2252 struct timer_list* timer)
2257 ret = (timer && likely(del_timer(timer)));
2259 ret = try_to_grab_pending(work);
2261 } while (unlikely(ret < 0));
2263 clear_work_data(work);
2268 * cancel_work_sync - block until a work_struct's callback has terminated
2269 * @work: the work which is to be flushed
2271 * Returns true if @work was pending.
2273 * cancel_work_sync() will cancel the work if it is queued. If the work's
2274 * callback appears to be running, cancel_work_sync() will block until it
2277 * It is possible to use this function if the work re-queues itself. It can
2278 * cancel the work even if it migrates to another workqueue, however in that
2279 * case it only guarantees that work->func() has completed on the last queued
2282 * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
2283 * pending, otherwise it goes into a busy-wait loop until the timer expires.
2285 * The caller must ensure that workqueue_struct on which this work was last
2286 * queued can't be destroyed before this function returns.
2288 int cancel_work_sync(struct work_struct *work)
2290 return __cancel_work_timer(work, NULL);
2292 EXPORT_SYMBOL_GPL(cancel_work_sync);
2295 * cancel_delayed_work_sync - reliably kill off a delayed work.
2296 * @dwork: the delayed work struct
2298 * Returns true if @dwork was pending.
2300 * It is possible to use this function if @dwork rearms itself via queue_work()
2301 * or queue_delayed_work(). See also the comment for cancel_work_sync().
2303 int cancel_delayed_work_sync(struct delayed_work *dwork)
2305 return __cancel_work_timer(&dwork->work, &dwork->timer);
2307 EXPORT_SYMBOL(cancel_delayed_work_sync);
2309 static struct workqueue_struct *keventd_wq __read_mostly;
2312 * schedule_work - put work task in global workqueue
2313 * @work: job to be done
2315 * Returns zero if @work was already on the kernel-global workqueue and
2316 * non-zero otherwise.
2318 * This puts a job in the kernel-global workqueue if it was not already
2319 * queued and leaves it in the same position on the kernel-global
2320 * workqueue otherwise.
2322 int schedule_work(struct work_struct *work)
2324 return queue_work(keventd_wq, work);
2326 EXPORT_SYMBOL(schedule_work);
2329 * schedule_work_on - put work task on a specific cpu
2330 * @cpu: cpu to put the work task on
2331 * @work: job to be done
2333 * This puts a job on a specific cpu
2335 int schedule_work_on(int cpu, struct work_struct *work)
2337 return queue_work_on(cpu, keventd_wq, work);
2339 EXPORT_SYMBOL(schedule_work_on);
2342 * schedule_delayed_work - put work task in global workqueue after delay
2343 * @dwork: job to be done
2344 * @delay: number of jiffies to wait or 0 for immediate execution
2346 * After waiting for a given time this puts a job in the kernel-global
2349 int schedule_delayed_work(struct delayed_work *dwork,
2350 unsigned long delay)
2352 return queue_delayed_work(keventd_wq, dwork, delay);
2354 EXPORT_SYMBOL(schedule_delayed_work);
2357 * flush_delayed_work - block until a dwork_struct's callback has terminated
2358 * @dwork: the delayed work which is to be flushed
2360 * Any timeout is cancelled, and any pending work is run immediately.
2362 void flush_delayed_work(struct delayed_work *dwork)
2364 if (del_timer_sync(&dwork->timer)) {
2365 __queue_work(get_cpu(), get_work_cwq(&dwork->work)->wq,
2369 flush_work(&dwork->work);
2371 EXPORT_SYMBOL(flush_delayed_work);
2374 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2376 * @dwork: job to be done
2377 * @delay: number of jiffies to wait
2379 * After waiting for a given time this puts a job in the kernel-global
2380 * workqueue on the specified CPU.
2382 int schedule_delayed_work_on(int cpu,
2383 struct delayed_work *dwork, unsigned long delay)
2385 return queue_delayed_work_on(cpu, keventd_wq, dwork, delay);
2387 EXPORT_SYMBOL(schedule_delayed_work_on);
2390 * schedule_on_each_cpu - call a function on each online CPU from keventd
2391 * @func: the function to call
2393 * Returns zero on success.
2394 * Returns -ve errno on failure.
2396 * schedule_on_each_cpu() is very slow.
2398 int schedule_on_each_cpu(work_func_t func)
2401 struct work_struct *works;
2403 works = alloc_percpu(struct work_struct);
2409 for_each_online_cpu(cpu) {
2410 struct work_struct *work = per_cpu_ptr(works, cpu);
2412 INIT_WORK(work, func);
2413 schedule_work_on(cpu, work);
2416 for_each_online_cpu(cpu)
2417 flush_work(per_cpu_ptr(works, cpu));
2425 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2427 * Forces execution of the kernel-global workqueue and blocks until its
2430 * Think twice before calling this function! It's very easy to get into
2431 * trouble if you don't take great care. Either of the following situations
2432 * will lead to deadlock:
2434 * One of the work items currently on the workqueue needs to acquire
2435 * a lock held by your code or its caller.
2437 * Your code is running in the context of a work routine.
2439 * They will be detected by lockdep when they occur, but the first might not
2440 * occur very often. It depends on what work items are on the workqueue and
2441 * what locks they need, which you have no control over.
2443 * In most situations flushing the entire workqueue is overkill; you merely
2444 * need to know that a particular work item isn't queued and isn't running.
2445 * In such cases you should use cancel_delayed_work_sync() or
2446 * cancel_work_sync() instead.
2448 void flush_scheduled_work(void)
2450 flush_workqueue(keventd_wq);
2452 EXPORT_SYMBOL(flush_scheduled_work);
2455 * execute_in_process_context - reliably execute the routine with user context
2456 * @fn: the function to execute
2457 * @ew: guaranteed storage for the execute work structure (must
2458 * be available when the work executes)
2460 * Executes the function immediately if process context is available,
2461 * otherwise schedules the function for delayed execution.
2463 * Returns: 0 - function was executed
2464 * 1 - function was scheduled for execution
2466 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2468 if (!in_interrupt()) {
2473 INIT_WORK(&ew->work, fn);
2474 schedule_work(&ew->work);
2478 EXPORT_SYMBOL_GPL(execute_in_process_context);
2480 int keventd_up(void)
2482 return keventd_wq != NULL;
2485 static struct cpu_workqueue_struct *alloc_cwqs(void)
2488 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2489 * Make sure that the alignment isn't lower than that of
2490 * unsigned long long.
2492 const size_t size = sizeof(struct cpu_workqueue_struct);
2493 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2494 __alignof__(unsigned long long));
2495 struct cpu_workqueue_struct *cwqs;
2500 * On UP, percpu allocator doesn't honor alignment parameter
2501 * and simply uses arch-dependent default. Allocate enough
2502 * room to align cwq and put an extra pointer at the end
2503 * pointing back to the originally allocated pointer which
2504 * will be used for free.
2506 * FIXME: This really belongs to UP percpu code. Update UP
2507 * percpu code to honor alignment and remove this ugliness.
2509 ptr = __alloc_percpu(size + align + sizeof(void *), 1);
2510 cwqs = PTR_ALIGN(ptr, align);
2511 *(void **)per_cpu_ptr(cwqs + 1, 0) = ptr;
2513 /* On SMP, percpu allocator can do it itself */
2514 cwqs = __alloc_percpu(size, align);
2516 /* just in case, make sure it's actually aligned */
2517 BUG_ON(!IS_ALIGNED((unsigned long)cwqs, align));
2521 static void free_cwqs(struct cpu_workqueue_struct *cwqs)
2524 /* on UP, the pointer to free is stored right after the cwq */
2526 free_percpu(*(void **)per_cpu_ptr(cwqs + 1, 0));
2532 static int wq_clamp_max_active(int max_active, const char *name)
2534 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
2535 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2536 "is out of range, clamping between %d and %d\n",
2537 max_active, name, 1, WQ_MAX_ACTIVE);
2539 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
2542 struct workqueue_struct *__create_workqueue_key(const char *name,
2545 struct lock_class_key *key,
2546 const char *lock_name)
2548 struct workqueue_struct *wq;
2551 max_active = wq_clamp_max_active(max_active, name);
2553 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
2557 wq->cpu_wq = alloc_cwqs();
2562 wq->saved_max_active = max_active;
2563 mutex_init(&wq->flush_mutex);
2564 atomic_set(&wq->nr_cwqs_to_flush, 0);
2565 INIT_LIST_HEAD(&wq->flusher_queue);
2566 INIT_LIST_HEAD(&wq->flusher_overflow);
2567 wq->single_cpu = NR_CPUS;
2570 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
2571 INIT_LIST_HEAD(&wq->list);
2573 for_each_possible_cpu(cpu) {
2574 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2575 struct global_cwq *gcwq = get_gcwq(cpu);
2577 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
2580 cwq->flush_color = -1;
2581 cwq->max_active = max_active;
2582 INIT_LIST_HEAD(&cwq->delayed_works);
2585 if (flags & WQ_RESCUER) {
2586 struct worker *rescuer;
2588 if (!alloc_cpumask_var(&wq->mayday_mask, GFP_KERNEL))
2591 wq->rescuer = rescuer = alloc_worker();
2595 rescuer->task = kthread_create(rescuer_thread, wq, "%s", name);
2596 if (IS_ERR(rescuer->task))
2599 wq->rescuer = rescuer;
2600 rescuer->task->flags |= PF_THREAD_BOUND;
2601 wake_up_process(rescuer->task);
2605 * workqueue_lock protects global freeze state and workqueues
2606 * list. Grab it, set max_active accordingly and add the new
2607 * workqueue to workqueues list.
2609 spin_lock(&workqueue_lock);
2611 if (workqueue_freezing && wq->flags & WQ_FREEZEABLE)
2612 for_each_possible_cpu(cpu)
2613 get_cwq(cpu, wq)->max_active = 0;
2615 list_add(&wq->list, &workqueues);
2617 spin_unlock(&workqueue_lock);
2622 free_cwqs(wq->cpu_wq);
2623 free_cpumask_var(wq->mayday_mask);
2629 EXPORT_SYMBOL_GPL(__create_workqueue_key);
2632 * destroy_workqueue - safely terminate a workqueue
2633 * @wq: target workqueue
2635 * Safely destroy a workqueue. All work currently pending will be done first.
2637 void destroy_workqueue(struct workqueue_struct *wq)
2641 flush_workqueue(wq);
2644 * wq list is used to freeze wq, remove from list after
2645 * flushing is complete in case freeze races us.
2647 spin_lock(&workqueue_lock);
2648 list_del(&wq->list);
2649 spin_unlock(&workqueue_lock);
2652 for_each_possible_cpu(cpu) {
2653 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2656 for (i = 0; i < WORK_NR_COLORS; i++)
2657 BUG_ON(cwq->nr_in_flight[i]);
2658 BUG_ON(cwq->nr_active);
2659 BUG_ON(!list_empty(&cwq->delayed_works));
2662 if (wq->flags & WQ_RESCUER) {
2663 kthread_stop(wq->rescuer->task);
2664 free_cpumask_var(wq->mayday_mask);
2667 free_cwqs(wq->cpu_wq);
2670 EXPORT_SYMBOL_GPL(destroy_workqueue);
2675 * There are two challenges in supporting CPU hotplug. Firstly, there
2676 * are a lot of assumptions on strong associations among work, cwq and
2677 * gcwq which make migrating pending and scheduled works very
2678 * difficult to implement without impacting hot paths. Secondly,
2679 * gcwqs serve mix of short, long and very long running works making
2680 * blocked draining impractical.
2682 * This is solved by allowing a gcwq to be detached from CPU, running
2683 * it with unbound (rogue) workers and allowing it to be reattached
2684 * later if the cpu comes back online. A separate thread is created
2685 * to govern a gcwq in such state and is called the trustee of the
2688 * Trustee states and their descriptions.
2690 * START Command state used on startup. On CPU_DOWN_PREPARE, a
2691 * new trustee is started with this state.
2693 * IN_CHARGE Once started, trustee will enter this state after
2694 * assuming the manager role and making all existing
2695 * workers rogue. DOWN_PREPARE waits for trustee to
2696 * enter this state. After reaching IN_CHARGE, trustee
2697 * tries to execute the pending worklist until it's empty
2698 * and the state is set to BUTCHER, or the state is set
2701 * BUTCHER Command state which is set by the cpu callback after
2702 * the cpu has went down. Once this state is set trustee
2703 * knows that there will be no new works on the worklist
2704 * and once the worklist is empty it can proceed to
2705 * killing idle workers.
2707 * RELEASE Command state which is set by the cpu callback if the
2708 * cpu down has been canceled or it has come online
2709 * again. After recognizing this state, trustee stops
2710 * trying to drain or butcher and clears ROGUE, rebinds
2711 * all remaining workers back to the cpu and releases
2714 * DONE Trustee will enter this state after BUTCHER or RELEASE
2717 * trustee CPU draining
2718 * took over down complete
2719 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
2721 * | CPU is back online v return workers |
2722 * ----------------> RELEASE --------------
2726 * trustee_wait_event_timeout - timed event wait for trustee
2727 * @cond: condition to wait for
2728 * @timeout: timeout in jiffies
2730 * wait_event_timeout() for trustee to use. Handles locking and
2731 * checks for RELEASE request.
2734 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2735 * multiple times. To be used by trustee.
2738 * Positive indicating left time if @cond is satisfied, 0 if timed
2739 * out, -1 if canceled.
2741 #define trustee_wait_event_timeout(cond, timeout) ({ \
2742 long __ret = (timeout); \
2743 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
2745 spin_unlock_irq(&gcwq->lock); \
2746 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
2747 (gcwq->trustee_state == TRUSTEE_RELEASE), \
2749 spin_lock_irq(&gcwq->lock); \
2751 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
2755 * trustee_wait_event - event wait for trustee
2756 * @cond: condition to wait for
2758 * wait_event() for trustee to use. Automatically handles locking and
2759 * checks for CANCEL request.
2762 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2763 * multiple times. To be used by trustee.
2766 * 0 if @cond is satisfied, -1 if canceled.
2768 #define trustee_wait_event(cond) ({ \
2770 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
2771 __ret1 < 0 ? -1 : 0; \
2774 static int __cpuinit trustee_thread(void *__gcwq)
2776 struct global_cwq *gcwq = __gcwq;
2777 struct worker *worker;
2778 struct work_struct *work;
2779 struct hlist_node *pos;
2783 BUG_ON(gcwq->cpu != smp_processor_id());
2785 spin_lock_irq(&gcwq->lock);
2787 * Claim the manager position and make all workers rogue.
2788 * Trustee must be bound to the target cpu and can't be
2791 BUG_ON(gcwq->cpu != smp_processor_id());
2792 rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
2795 gcwq->flags |= GCWQ_MANAGING_WORKERS;
2797 list_for_each_entry(worker, &gcwq->idle_list, entry)
2798 worker_set_flags(worker, WORKER_ROGUE, false);
2800 for_each_busy_worker(worker, i, pos, gcwq)
2801 worker_set_flags(worker, WORKER_ROGUE, false);
2804 * Call schedule() so that we cross rq->lock and thus can
2805 * guarantee sched callbacks see the rogue flag. This is
2806 * necessary as scheduler callbacks may be invoked from other
2809 spin_unlock_irq(&gcwq->lock);
2811 spin_lock_irq(&gcwq->lock);
2814 * Sched callbacks are disabled now. gcwq->nr_running should
2815 * be zero and will stay that way, making need_more_worker()
2816 * and keep_working() always return true as long as the
2817 * worklist is not empty.
2819 WARN_ON_ONCE(atomic_read(get_gcwq_nr_running(gcwq->cpu)) != 0);
2821 spin_unlock_irq(&gcwq->lock);
2822 del_timer_sync(&gcwq->idle_timer);
2823 spin_lock_irq(&gcwq->lock);
2826 * We're now in charge. Notify and proceed to drain. We need
2827 * to keep the gcwq running during the whole CPU down
2828 * procedure as other cpu hotunplug callbacks may need to
2829 * flush currently running tasks.
2831 gcwq->trustee_state = TRUSTEE_IN_CHARGE;
2832 wake_up_all(&gcwq->trustee_wait);
2835 * The original cpu is in the process of dying and may go away
2836 * anytime now. When that happens, we and all workers would
2837 * be migrated to other cpus. Try draining any left work. We
2838 * want to get it over with ASAP - spam rescuers, wake up as
2839 * many idlers as necessary and create new ones till the
2840 * worklist is empty. Note that if the gcwq is frozen, there
2841 * may be frozen works in freezeable cwqs. Don't declare
2842 * completion while frozen.
2844 while (gcwq->nr_workers != gcwq->nr_idle ||
2845 gcwq->flags & GCWQ_FREEZING ||
2846 gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
2849 list_for_each_entry(work, &gcwq->worklist, entry) {
2854 list_for_each_entry(worker, &gcwq->idle_list, entry) {
2857 wake_up_process(worker->task);
2860 if (need_to_create_worker(gcwq)) {
2861 spin_unlock_irq(&gcwq->lock);
2862 worker = create_worker(gcwq, false);
2863 spin_lock_irq(&gcwq->lock);
2865 worker_set_flags(worker, WORKER_ROGUE, false);
2866 start_worker(worker);
2870 /* give a breather */
2871 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
2876 * Either all works have been scheduled and cpu is down, or
2877 * cpu down has already been canceled. Wait for and butcher
2878 * all workers till we're canceled.
2881 rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
2882 while (!list_empty(&gcwq->idle_list))
2883 destroy_worker(list_first_entry(&gcwq->idle_list,
2884 struct worker, entry));
2885 } while (gcwq->nr_workers && rc >= 0);
2888 * At this point, either draining has completed and no worker
2889 * is left, or cpu down has been canceled or the cpu is being
2890 * brought back up. There shouldn't be any idle one left.
2891 * Tell the remaining busy ones to rebind once it finishes the
2892 * currently scheduled works by scheduling the rebind_work.
2894 WARN_ON(!list_empty(&gcwq->idle_list));
2896 for_each_busy_worker(worker, i, pos, gcwq) {
2897 struct work_struct *rebind_work = &worker->rebind_work;
2900 * Rebind_work may race with future cpu hotplug
2901 * operations. Use a separate flag to mark that
2902 * rebinding is scheduled.
2904 worker_set_flags(worker, WORKER_REBIND, false);
2905 worker_clr_flags(worker, WORKER_ROGUE);
2907 /* queue rebind_work, wq doesn't matter, use the default one */
2908 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
2909 work_data_bits(rebind_work)))
2912 debug_work_activate(rebind_work);
2913 insert_work(get_cwq(gcwq->cpu, keventd_wq), rebind_work,
2914 worker->scheduled.next,
2915 work_color_to_flags(WORK_NO_COLOR));
2918 /* relinquish manager role */
2919 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
2921 /* notify completion */
2922 gcwq->trustee = NULL;
2923 gcwq->trustee_state = TRUSTEE_DONE;
2924 wake_up_all(&gcwq->trustee_wait);
2925 spin_unlock_irq(&gcwq->lock);
2930 * wait_trustee_state - wait for trustee to enter the specified state
2931 * @gcwq: gcwq the trustee of interest belongs to
2932 * @state: target state to wait for
2934 * Wait for the trustee to reach @state. DONE is already matched.
2937 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2938 * multiple times. To be used by cpu_callback.
2940 static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
2942 if (!(gcwq->trustee_state == state ||
2943 gcwq->trustee_state == TRUSTEE_DONE)) {
2944 spin_unlock_irq(&gcwq->lock);
2945 __wait_event(gcwq->trustee_wait,
2946 gcwq->trustee_state == state ||
2947 gcwq->trustee_state == TRUSTEE_DONE);
2948 spin_lock_irq(&gcwq->lock);
2952 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
2953 unsigned long action,
2956 unsigned int cpu = (unsigned long)hcpu;
2957 struct global_cwq *gcwq = get_gcwq(cpu);
2958 struct task_struct *new_trustee = NULL;
2959 struct worker *uninitialized_var(new_worker);
2960 unsigned long flags;
2962 action &= ~CPU_TASKS_FROZEN;
2965 case CPU_DOWN_PREPARE:
2966 new_trustee = kthread_create(trustee_thread, gcwq,
2967 "workqueue_trustee/%d\n", cpu);
2968 if (IS_ERR(new_trustee))
2969 return notifier_from_errno(PTR_ERR(new_trustee));
2970 kthread_bind(new_trustee, cpu);
2972 case CPU_UP_PREPARE:
2973 BUG_ON(gcwq->first_idle);
2974 new_worker = create_worker(gcwq, false);
2977 kthread_stop(new_trustee);
2982 /* some are called w/ irq disabled, don't disturb irq status */
2983 spin_lock_irqsave(&gcwq->lock, flags);
2986 case CPU_DOWN_PREPARE:
2987 /* initialize trustee and tell it to acquire the gcwq */
2988 BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
2989 gcwq->trustee = new_trustee;
2990 gcwq->trustee_state = TRUSTEE_START;
2991 wake_up_process(gcwq->trustee);
2992 wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
2994 case CPU_UP_PREPARE:
2995 BUG_ON(gcwq->first_idle);
2996 gcwq->first_idle = new_worker;
3001 * Before this, the trustee and all workers except for
3002 * the ones which are still executing works from
3003 * before the last CPU down must be on the cpu. After
3004 * this, they'll all be diasporas.
3006 gcwq->flags |= GCWQ_DISASSOCIATED;
3010 gcwq->trustee_state = TRUSTEE_BUTCHER;
3012 case CPU_UP_CANCELED:
3013 destroy_worker(gcwq->first_idle);
3014 gcwq->first_idle = NULL;
3017 case CPU_DOWN_FAILED:
3019 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3020 if (gcwq->trustee_state != TRUSTEE_DONE) {
3021 gcwq->trustee_state = TRUSTEE_RELEASE;
3022 wake_up_process(gcwq->trustee);
3023 wait_trustee_state(gcwq, TRUSTEE_DONE);
3027 * Trustee is done and there might be no worker left.
3028 * Put the first_idle in and request a real manager to
3031 spin_unlock_irq(&gcwq->lock);
3032 kthread_bind(gcwq->first_idle->task, cpu);
3033 spin_lock_irq(&gcwq->lock);
3034 gcwq->flags |= GCWQ_MANAGE_WORKERS;
3035 start_worker(gcwq->first_idle);
3036 gcwq->first_idle = NULL;
3040 spin_unlock_irqrestore(&gcwq->lock, flags);
3042 return notifier_from_errno(0);
3047 struct work_for_cpu {
3048 struct completion completion;
3054 static int do_work_for_cpu(void *_wfc)
3056 struct work_for_cpu *wfc = _wfc;
3057 wfc->ret = wfc->fn(wfc->arg);
3058 complete(&wfc->completion);
3063 * work_on_cpu - run a function in user context on a particular cpu
3064 * @cpu: the cpu to run on
3065 * @fn: the function to run
3066 * @arg: the function arg
3068 * This will return the value @fn returns.
3069 * It is up to the caller to ensure that the cpu doesn't go offline.
3070 * The caller must not hold any locks which would prevent @fn from completing.
3072 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3074 struct task_struct *sub_thread;
3075 struct work_for_cpu wfc = {
3076 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3081 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3082 if (IS_ERR(sub_thread))
3083 return PTR_ERR(sub_thread);
3084 kthread_bind(sub_thread, cpu);
3085 wake_up_process(sub_thread);
3086 wait_for_completion(&wfc.completion);
3089 EXPORT_SYMBOL_GPL(work_on_cpu);
3090 #endif /* CONFIG_SMP */
3092 #ifdef CONFIG_FREEZER
3095 * freeze_workqueues_begin - begin freezing workqueues
3097 * Start freezing workqueues. After this function returns, all
3098 * freezeable workqueues will queue new works to their frozen_works
3099 * list instead of gcwq->worklist.
3102 * Grabs and releases workqueue_lock and gcwq->lock's.
3104 void freeze_workqueues_begin(void)
3106 struct workqueue_struct *wq;
3109 spin_lock(&workqueue_lock);
3111 BUG_ON(workqueue_freezing);
3112 workqueue_freezing = true;
3114 for_each_possible_cpu(cpu) {
3115 struct global_cwq *gcwq = get_gcwq(cpu);
3117 spin_lock_irq(&gcwq->lock);
3119 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3120 gcwq->flags |= GCWQ_FREEZING;
3122 list_for_each_entry(wq, &workqueues, list) {
3123 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3125 if (wq->flags & WQ_FREEZEABLE)
3126 cwq->max_active = 0;
3129 spin_unlock_irq(&gcwq->lock);
3132 spin_unlock(&workqueue_lock);
3136 * freeze_workqueues_busy - are freezeable workqueues still busy?
3138 * Check whether freezing is complete. This function must be called
3139 * between freeze_workqueues_begin() and thaw_workqueues().
3142 * Grabs and releases workqueue_lock.
3145 * %true if some freezeable workqueues are still busy. %false if
3146 * freezing is complete.
3148 bool freeze_workqueues_busy(void)
3150 struct workqueue_struct *wq;
3154 spin_lock(&workqueue_lock);
3156 BUG_ON(!workqueue_freezing);
3158 for_each_possible_cpu(cpu) {
3160 * nr_active is monotonically decreasing. It's safe
3161 * to peek without lock.
3163 list_for_each_entry(wq, &workqueues, list) {
3164 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3166 if (!(wq->flags & WQ_FREEZEABLE))
3169 BUG_ON(cwq->nr_active < 0);
3170 if (cwq->nr_active) {
3177 spin_unlock(&workqueue_lock);
3182 * thaw_workqueues - thaw workqueues
3184 * Thaw workqueues. Normal queueing is restored and all collected
3185 * frozen works are transferred to their respective gcwq worklists.
3188 * Grabs and releases workqueue_lock and gcwq->lock's.
3190 void thaw_workqueues(void)
3192 struct workqueue_struct *wq;
3195 spin_lock(&workqueue_lock);
3197 if (!workqueue_freezing)
3200 for_each_possible_cpu(cpu) {
3201 struct global_cwq *gcwq = get_gcwq(cpu);
3203 spin_lock_irq(&gcwq->lock);
3205 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3206 gcwq->flags &= ~GCWQ_FREEZING;
3208 list_for_each_entry(wq, &workqueues, list) {
3209 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3211 if (!(wq->flags & WQ_FREEZEABLE))
3214 /* restore max_active and repopulate worklist */
3215 cwq->max_active = wq->saved_max_active;
3217 while (!list_empty(&cwq->delayed_works) &&
3218 cwq->nr_active < cwq->max_active)
3219 cwq_activate_first_delayed(cwq);
3221 /* perform delayed unbind from single cpu if empty */
3222 if (wq->single_cpu == gcwq->cpu &&
3223 !cwq->nr_active && list_empty(&cwq->delayed_works))
3224 cwq_unbind_single_cpu(cwq);
3227 wake_up_worker(gcwq);
3229 spin_unlock_irq(&gcwq->lock);
3232 workqueue_freezing = false;
3234 spin_unlock(&workqueue_lock);
3236 #endif /* CONFIG_FREEZER */
3238 void __init init_workqueues(void)
3244 * The pointer part of work->data is either pointing to the
3245 * cwq or contains the cpu number the work ran last on. Make
3246 * sure cpu number won't overflow into kernel pointer area so
3247 * that they can be distinguished.
3249 BUILD_BUG_ON(NR_CPUS << WORK_STRUCT_FLAG_BITS >= PAGE_OFFSET);
3251 hotcpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
3253 /* initialize gcwqs */
3254 for_each_possible_cpu(cpu) {
3255 struct global_cwq *gcwq = get_gcwq(cpu);
3257 spin_lock_init(&gcwq->lock);
3258 INIT_LIST_HEAD(&gcwq->worklist);
3261 INIT_LIST_HEAD(&gcwq->idle_list);
3262 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3263 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3265 init_timer_deferrable(&gcwq->idle_timer);
3266 gcwq->idle_timer.function = idle_worker_timeout;
3267 gcwq->idle_timer.data = (unsigned long)gcwq;
3269 setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3270 (unsigned long)gcwq);
3272 ida_init(&gcwq->worker_ida);
3274 gcwq->trustee_state = TRUSTEE_DONE;
3275 init_waitqueue_head(&gcwq->trustee_wait);
3278 /* create the initial worker */
3279 for_each_online_cpu(cpu) {
3280 struct global_cwq *gcwq = get_gcwq(cpu);
3281 struct worker *worker;
3283 worker = create_worker(gcwq, true);
3285 spin_lock_irq(&gcwq->lock);
3286 start_worker(worker);
3287 spin_unlock_irq(&gcwq->lock);
3290 keventd_wq = __create_workqueue("events", 0, WQ_DFL_ACTIVE);
3291 BUG_ON(!keventd_wq);