2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/hashtable.h>
46 #include "workqueue_internal.h"
52 * A bound pool is either associated or disassociated with its CPU.
53 * While associated (!DISASSOCIATED), all workers are bound to the
54 * CPU and none has %WORKER_UNBOUND set and concurrency management
57 * While DISASSOCIATED, the cpu may be offline and all workers have
58 * %WORKER_UNBOUND set and concurrency management disabled, and may
59 * be executing on any CPU. The pool behaves as an unbound one.
61 * Note that DISASSOCIATED can be flipped only while holding
62 * assoc_mutex to avoid changing binding state while
63 * create_worker() is in progress.
65 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
66 POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */
67 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
68 POOL_FREEZING = 1 << 3, /* freeze in progress */
71 WORKER_STARTED = 1 << 0, /* started */
72 WORKER_DIE = 1 << 1, /* die die die */
73 WORKER_IDLE = 1 << 2, /* is idle */
74 WORKER_PREP = 1 << 3, /* preparing to run works */
75 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
76 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
78 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND |
81 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
83 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
85 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
86 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
88 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
89 /* call for help after 10ms
91 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
92 CREATE_COOLDOWN = HZ, /* time to breath after fail */
95 * Rescue workers are used only on emergencies and shared by
98 RESCUER_NICE_LEVEL = -20,
99 HIGHPRI_NICE_LEVEL = -20,
103 * Structure fields follow one of the following exclusion rules.
105 * I: Modifiable by initialization/destruction paths and read-only for
108 * P: Preemption protected. Disabling preemption is enough and should
109 * only be modified and accessed from the local cpu.
111 * L: pool->lock protected. Access with pool->lock held.
113 * X: During normal operation, modification requires pool->lock and should
114 * be done only from local cpu. Either disabling preemption on local
115 * cpu or grabbing pool->lock is enough for read access. If
116 * POOL_DISASSOCIATED is set, it's identical to L.
118 * F: wq->flush_mutex protected.
120 * W: workqueue_lock protected.
123 /* struct worker is defined in workqueue_internal.h */
126 spinlock_t lock; /* the pool lock */
127 int cpu; /* I: the associated cpu */
128 int id; /* I: pool ID */
129 unsigned int flags; /* X: flags */
131 struct list_head worklist; /* L: list of pending works */
132 int nr_workers; /* L: total number of workers */
134 /* nr_idle includes the ones off idle_list for rebinding */
135 int nr_idle; /* L: currently idle ones */
137 struct list_head idle_list; /* X: list of idle workers */
138 struct timer_list idle_timer; /* L: worker idle timeout */
139 struct timer_list mayday_timer; /* L: SOS timer for workers */
141 /* workers are chained either in busy_hash or idle_list */
142 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
143 /* L: hash of busy workers */
145 struct mutex assoc_mutex; /* protect POOL_DISASSOCIATED */
146 struct ida worker_ida; /* L: for worker IDs */
149 * The current concurrency level. As it's likely to be accessed
150 * from other CPUs during try_to_wake_up(), put it in a separate
153 atomic_t nr_running ____cacheline_aligned_in_smp;
154 } ____cacheline_aligned_in_smp;
157 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
158 * of work_struct->data are used for flags and the remaining high bits
159 * point to the pwq; thus, pwqs need to be aligned at two's power of the
160 * number of flag bits.
162 struct pool_workqueue {
163 struct worker_pool *pool; /* I: the associated pool */
164 struct workqueue_struct *wq; /* I: the owning workqueue */
165 int work_color; /* L: current color */
166 int flush_color; /* L: flushing color */
167 int nr_in_flight[WORK_NR_COLORS];
168 /* L: nr of in_flight works */
169 int nr_active; /* L: nr of active works */
170 int max_active; /* L: max active works */
171 struct list_head delayed_works; /* L: delayed works */
172 struct list_head pwqs_node; /* I: node on wq->pwqs */
173 struct list_head mayday_node; /* W: node on wq->maydays */
174 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
177 * Structure used to wait for workqueue flush.
180 struct list_head list; /* F: list of flushers */
181 int flush_color; /* F: flush color waiting for */
182 struct completion done; /* flush completion */
186 * The externally visible workqueue abstraction is an array of
187 * per-CPU workqueues:
189 struct workqueue_struct {
190 unsigned int flags; /* W: WQ_* flags */
191 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwq's */
192 struct list_head pwqs; /* I: all pwqs of this wq */
193 struct list_head list; /* W: list of all workqueues */
195 struct mutex flush_mutex; /* protects wq flushing */
196 int work_color; /* F: current work color */
197 int flush_color; /* F: current flush color */
198 atomic_t nr_pwqs_to_flush; /* flush in progress */
199 struct wq_flusher *first_flusher; /* F: first flusher */
200 struct list_head flusher_queue; /* F: flush waiters */
201 struct list_head flusher_overflow; /* F: flush overflow list */
203 struct list_head maydays; /* W: pwqs requesting rescue */
204 struct worker *rescuer; /* I: rescue worker */
206 int nr_drainers; /* W: drain in progress */
207 int saved_max_active; /* W: saved pwq max_active */
208 #ifdef CONFIG_LOCKDEP
209 struct lockdep_map lockdep_map;
211 char name[]; /* I: workqueue name */
214 static struct kmem_cache *pwq_cache;
216 struct workqueue_struct *system_wq __read_mostly;
217 EXPORT_SYMBOL_GPL(system_wq);
218 struct workqueue_struct *system_highpri_wq __read_mostly;
219 EXPORT_SYMBOL_GPL(system_highpri_wq);
220 struct workqueue_struct *system_long_wq __read_mostly;
221 EXPORT_SYMBOL_GPL(system_long_wq);
222 struct workqueue_struct *system_unbound_wq __read_mostly;
223 EXPORT_SYMBOL_GPL(system_unbound_wq);
224 struct workqueue_struct *system_freezable_wq __read_mostly;
225 EXPORT_SYMBOL_GPL(system_freezable_wq);
227 #define CREATE_TRACE_POINTS
228 #include <trace/events/workqueue.h>
230 #define for_each_std_worker_pool(pool, cpu) \
231 for ((pool) = &std_worker_pools(cpu)[0]; \
232 (pool) < &std_worker_pools(cpu)[NR_STD_WORKER_POOLS]; (pool)++)
234 #define for_each_busy_worker(worker, i, pool) \
235 hash_for_each(pool->busy_hash, i, worker, hentry)
237 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
240 if (cpu < nr_cpu_ids) {
242 cpu = cpumask_next(cpu, mask);
243 if (cpu < nr_cpu_ids)
247 return WORK_CPU_UNBOUND;
255 * An extra cpu number is defined using an invalid cpu number
256 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
257 * specific CPU. The following iterators are similar to for_each_*_cpu()
258 * iterators but also considers the unbound CPU.
260 * for_each_wq_cpu() : possible CPUs + WORK_CPU_UNBOUND
261 * for_each_online_wq_cpu() : online CPUs + WORK_CPU_UNBOUND
263 #define for_each_wq_cpu(cpu) \
264 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, 3); \
265 (cpu) < WORK_CPU_END; \
266 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, 3))
268 #define for_each_online_wq_cpu(cpu) \
269 for ((cpu) = __next_wq_cpu(-1, cpu_online_mask, 3); \
270 (cpu) < WORK_CPU_END; \
271 (cpu) = __next_wq_cpu((cpu), cpu_online_mask, 3))
274 * for_each_pool - iterate through all worker_pools in the system
275 * @pool: iteration cursor
276 * @id: integer used for iteration
278 #define for_each_pool(pool, id) \
279 idr_for_each_entry(&worker_pool_idr, pool, id)
282 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
283 * @pwq: iteration cursor
284 * @wq: the target workqueue
286 #define for_each_pwq(pwq, wq) \
287 list_for_each_entry((pwq), &(wq)->pwqs, pwqs_node)
289 #ifdef CONFIG_DEBUG_OBJECTS_WORK
291 static struct debug_obj_descr work_debug_descr;
293 static void *work_debug_hint(void *addr)
295 return ((struct work_struct *) addr)->func;
299 * fixup_init is called when:
300 * - an active object is initialized
302 static int work_fixup_init(void *addr, enum debug_obj_state state)
304 struct work_struct *work = addr;
307 case ODEBUG_STATE_ACTIVE:
308 cancel_work_sync(work);
309 debug_object_init(work, &work_debug_descr);
317 * fixup_activate is called when:
318 * - an active object is activated
319 * - an unknown object is activated (might be a statically initialized object)
321 static int work_fixup_activate(void *addr, enum debug_obj_state state)
323 struct work_struct *work = addr;
327 case ODEBUG_STATE_NOTAVAILABLE:
329 * This is not really a fixup. The work struct was
330 * statically initialized. We just make sure that it
331 * is tracked in the object tracker.
333 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
334 debug_object_init(work, &work_debug_descr);
335 debug_object_activate(work, &work_debug_descr);
341 case ODEBUG_STATE_ACTIVE:
350 * fixup_free is called when:
351 * - an active object is freed
353 static int work_fixup_free(void *addr, enum debug_obj_state state)
355 struct work_struct *work = addr;
358 case ODEBUG_STATE_ACTIVE:
359 cancel_work_sync(work);
360 debug_object_free(work, &work_debug_descr);
367 static struct debug_obj_descr work_debug_descr = {
368 .name = "work_struct",
369 .debug_hint = work_debug_hint,
370 .fixup_init = work_fixup_init,
371 .fixup_activate = work_fixup_activate,
372 .fixup_free = work_fixup_free,
375 static inline void debug_work_activate(struct work_struct *work)
377 debug_object_activate(work, &work_debug_descr);
380 static inline void debug_work_deactivate(struct work_struct *work)
382 debug_object_deactivate(work, &work_debug_descr);
385 void __init_work(struct work_struct *work, int onstack)
388 debug_object_init_on_stack(work, &work_debug_descr);
390 debug_object_init(work, &work_debug_descr);
392 EXPORT_SYMBOL_GPL(__init_work);
394 void destroy_work_on_stack(struct work_struct *work)
396 debug_object_free(work, &work_debug_descr);
398 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
401 static inline void debug_work_activate(struct work_struct *work) { }
402 static inline void debug_work_deactivate(struct work_struct *work) { }
405 /* Serializes the accesses to the list of workqueues. */
406 static DEFINE_SPINLOCK(workqueue_lock);
407 static LIST_HEAD(workqueues);
408 static bool workqueue_freezing; /* W: have wqs started freezing? */
411 * The CPU and unbound standard worker pools. The unbound ones have
412 * POOL_DISASSOCIATED set, and their workers have WORKER_UNBOUND set.
414 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
415 cpu_std_worker_pools);
416 static struct worker_pool unbound_std_worker_pools[NR_STD_WORKER_POOLS];
418 /* idr of all pools */
419 static DEFINE_MUTEX(worker_pool_idr_mutex);
420 static DEFINE_IDR(worker_pool_idr);
422 static int worker_thread(void *__worker);
424 static struct worker_pool *std_worker_pools(int cpu)
426 if (cpu != WORK_CPU_UNBOUND)
427 return per_cpu(cpu_std_worker_pools, cpu);
429 return unbound_std_worker_pools;
432 static int std_worker_pool_pri(struct worker_pool *pool)
434 return pool - std_worker_pools(pool->cpu);
437 /* allocate ID and assign it to @pool */
438 static int worker_pool_assign_id(struct worker_pool *pool)
442 mutex_lock(&worker_pool_idr_mutex);
443 idr_pre_get(&worker_pool_idr, GFP_KERNEL);
444 ret = idr_get_new(&worker_pool_idr, pool, &pool->id);
445 mutex_unlock(&worker_pool_idr_mutex);
451 * Lookup worker_pool by id. The idr currently is built during boot and
452 * never modified. Don't worry about locking for now.
454 static struct worker_pool *worker_pool_by_id(int pool_id)
456 return idr_find(&worker_pool_idr, pool_id);
459 static struct worker_pool *get_std_worker_pool(int cpu, bool highpri)
461 struct worker_pool *pools = std_worker_pools(cpu);
463 return &pools[highpri];
466 static struct pool_workqueue *get_pwq(int cpu, struct workqueue_struct *wq)
468 if (!(wq->flags & WQ_UNBOUND)) {
469 if (likely(cpu < nr_cpu_ids))
470 return per_cpu_ptr(wq->cpu_pwqs, cpu);
471 } else if (likely(cpu == WORK_CPU_UNBOUND)) {
472 return list_first_entry(&wq->pwqs, struct pool_workqueue,
478 static unsigned int work_color_to_flags(int color)
480 return color << WORK_STRUCT_COLOR_SHIFT;
483 static int get_work_color(struct work_struct *work)
485 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
486 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
489 static int work_next_color(int color)
491 return (color + 1) % WORK_NR_COLORS;
495 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
496 * contain the pointer to the queued pwq. Once execution starts, the flag
497 * is cleared and the high bits contain OFFQ flags and pool ID.
499 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
500 * and clear_work_data() can be used to set the pwq, pool or clear
501 * work->data. These functions should only be called while the work is
502 * owned - ie. while the PENDING bit is set.
504 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
505 * corresponding to a work. Pool is available once the work has been
506 * queued anywhere after initialization until it is sync canceled. pwq is
507 * available only while the work item is queued.
509 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
510 * canceled. While being canceled, a work item may have its PENDING set
511 * but stay off timer and worklist for arbitrarily long and nobody should
512 * try to steal the PENDING bit.
514 static inline void set_work_data(struct work_struct *work, unsigned long data,
517 WARN_ON_ONCE(!work_pending(work));
518 atomic_long_set(&work->data, data | flags | work_static(work));
521 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
522 unsigned long extra_flags)
524 set_work_data(work, (unsigned long)pwq,
525 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
528 static void set_work_pool_and_keep_pending(struct work_struct *work,
531 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
532 WORK_STRUCT_PENDING);
535 static void set_work_pool_and_clear_pending(struct work_struct *work,
539 * The following wmb is paired with the implied mb in
540 * test_and_set_bit(PENDING) and ensures all updates to @work made
541 * here are visible to and precede any updates by the next PENDING
545 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
548 static void clear_work_data(struct work_struct *work)
550 smp_wmb(); /* see set_work_pool_and_clear_pending() */
551 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
554 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
556 unsigned long data = atomic_long_read(&work->data);
558 if (data & WORK_STRUCT_PWQ)
559 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
565 * get_work_pool - return the worker_pool a given work was associated with
566 * @work: the work item of interest
568 * Return the worker_pool @work was last associated with. %NULL if none.
570 static struct worker_pool *get_work_pool(struct work_struct *work)
572 unsigned long data = atomic_long_read(&work->data);
573 struct worker_pool *pool;
576 if (data & WORK_STRUCT_PWQ)
577 return ((struct pool_workqueue *)
578 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
580 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
581 if (pool_id == WORK_OFFQ_POOL_NONE)
584 pool = worker_pool_by_id(pool_id);
590 * get_work_pool_id - return the worker pool ID a given work is associated with
591 * @work: the work item of interest
593 * Return the worker_pool ID @work was last associated with.
594 * %WORK_OFFQ_POOL_NONE if none.
596 static int get_work_pool_id(struct work_struct *work)
598 unsigned long data = atomic_long_read(&work->data);
600 if (data & WORK_STRUCT_PWQ)
601 return ((struct pool_workqueue *)
602 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
604 return data >> WORK_OFFQ_POOL_SHIFT;
607 static void mark_work_canceling(struct work_struct *work)
609 unsigned long pool_id = get_work_pool_id(work);
611 pool_id <<= WORK_OFFQ_POOL_SHIFT;
612 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
615 static bool work_is_canceling(struct work_struct *work)
617 unsigned long data = atomic_long_read(&work->data);
619 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
623 * Policy functions. These define the policies on how the global worker
624 * pools are managed. Unless noted otherwise, these functions assume that
625 * they're being called with pool->lock held.
628 static bool __need_more_worker(struct worker_pool *pool)
630 return !atomic_read(&pool->nr_running);
634 * Need to wake up a worker? Called from anything but currently
637 * Note that, because unbound workers never contribute to nr_running, this
638 * function will always return %true for unbound pools as long as the
639 * worklist isn't empty.
641 static bool need_more_worker(struct worker_pool *pool)
643 return !list_empty(&pool->worklist) && __need_more_worker(pool);
646 /* Can I start working? Called from busy but !running workers. */
647 static bool may_start_working(struct worker_pool *pool)
649 return pool->nr_idle;
652 /* Do I need to keep working? Called from currently running workers. */
653 static bool keep_working(struct worker_pool *pool)
655 return !list_empty(&pool->worklist) &&
656 atomic_read(&pool->nr_running) <= 1;
659 /* Do we need a new worker? Called from manager. */
660 static bool need_to_create_worker(struct worker_pool *pool)
662 return need_more_worker(pool) && !may_start_working(pool);
665 /* Do I need to be the manager? */
666 static bool need_to_manage_workers(struct worker_pool *pool)
668 return need_to_create_worker(pool) ||
669 (pool->flags & POOL_MANAGE_WORKERS);
672 /* Do we have too many workers and should some go away? */
673 static bool too_many_workers(struct worker_pool *pool)
675 bool managing = pool->flags & POOL_MANAGING_WORKERS;
676 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
677 int nr_busy = pool->nr_workers - nr_idle;
680 * nr_idle and idle_list may disagree if idle rebinding is in
681 * progress. Never return %true if idle_list is empty.
683 if (list_empty(&pool->idle_list))
686 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
693 /* Return the first worker. Safe with preemption disabled */
694 static struct worker *first_worker(struct worker_pool *pool)
696 if (unlikely(list_empty(&pool->idle_list)))
699 return list_first_entry(&pool->idle_list, struct worker, entry);
703 * wake_up_worker - wake up an idle worker
704 * @pool: worker pool to wake worker from
706 * Wake up the first idle worker of @pool.
709 * spin_lock_irq(pool->lock).
711 static void wake_up_worker(struct worker_pool *pool)
713 struct worker *worker = first_worker(pool);
716 wake_up_process(worker->task);
720 * wq_worker_waking_up - a worker is waking up
721 * @task: task waking up
722 * @cpu: CPU @task is waking up to
724 * This function is called during try_to_wake_up() when a worker is
728 * spin_lock_irq(rq->lock)
730 void wq_worker_waking_up(struct task_struct *task, int cpu)
732 struct worker *worker = kthread_data(task);
734 if (!(worker->flags & WORKER_NOT_RUNNING)) {
735 WARN_ON_ONCE(worker->pool->cpu != cpu);
736 atomic_inc(&worker->pool->nr_running);
741 * wq_worker_sleeping - a worker is going to sleep
742 * @task: task going to sleep
743 * @cpu: CPU in question, must be the current CPU number
745 * This function is called during schedule() when a busy worker is
746 * going to sleep. Worker on the same cpu can be woken up by
747 * returning pointer to its task.
750 * spin_lock_irq(rq->lock)
753 * Worker task on @cpu to wake up, %NULL if none.
755 struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
757 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
758 struct worker_pool *pool;
761 * Rescuers, which may not have all the fields set up like normal
762 * workers, also reach here, let's not access anything before
763 * checking NOT_RUNNING.
765 if (worker->flags & WORKER_NOT_RUNNING)
770 /* this can only happen on the local cpu */
771 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
775 * The counterpart of the following dec_and_test, implied mb,
776 * worklist not empty test sequence is in insert_work().
777 * Please read comment there.
779 * NOT_RUNNING is clear. This means that we're bound to and
780 * running on the local cpu w/ rq lock held and preemption
781 * disabled, which in turn means that none else could be
782 * manipulating idle_list, so dereferencing idle_list without pool
785 if (atomic_dec_and_test(&pool->nr_running) &&
786 !list_empty(&pool->worklist))
787 to_wakeup = first_worker(pool);
788 return to_wakeup ? to_wakeup->task : NULL;
792 * worker_set_flags - set worker flags and adjust nr_running accordingly
794 * @flags: flags to set
795 * @wakeup: wakeup an idle worker if necessary
797 * Set @flags in @worker->flags and adjust nr_running accordingly. If
798 * nr_running becomes zero and @wakeup is %true, an idle worker is
802 * spin_lock_irq(pool->lock)
804 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
807 struct worker_pool *pool = worker->pool;
809 WARN_ON_ONCE(worker->task != current);
812 * If transitioning into NOT_RUNNING, adjust nr_running and
813 * wake up an idle worker as necessary if requested by
816 if ((flags & WORKER_NOT_RUNNING) &&
817 !(worker->flags & WORKER_NOT_RUNNING)) {
819 if (atomic_dec_and_test(&pool->nr_running) &&
820 !list_empty(&pool->worklist))
821 wake_up_worker(pool);
823 atomic_dec(&pool->nr_running);
826 worker->flags |= flags;
830 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
832 * @flags: flags to clear
834 * Clear @flags in @worker->flags and adjust nr_running accordingly.
837 * spin_lock_irq(pool->lock)
839 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
841 struct worker_pool *pool = worker->pool;
842 unsigned int oflags = worker->flags;
844 WARN_ON_ONCE(worker->task != current);
846 worker->flags &= ~flags;
849 * If transitioning out of NOT_RUNNING, increment nr_running. Note
850 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
851 * of multiple flags, not a single flag.
853 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
854 if (!(worker->flags & WORKER_NOT_RUNNING))
855 atomic_inc(&pool->nr_running);
859 * find_worker_executing_work - find worker which is executing a work
860 * @pool: pool of interest
861 * @work: work to find worker for
863 * Find a worker which is executing @work on @pool by searching
864 * @pool->busy_hash which is keyed by the address of @work. For a worker
865 * to match, its current execution should match the address of @work and
866 * its work function. This is to avoid unwanted dependency between
867 * unrelated work executions through a work item being recycled while still
870 * This is a bit tricky. A work item may be freed once its execution
871 * starts and nothing prevents the freed area from being recycled for
872 * another work item. If the same work item address ends up being reused
873 * before the original execution finishes, workqueue will identify the
874 * recycled work item as currently executing and make it wait until the
875 * current execution finishes, introducing an unwanted dependency.
877 * This function checks the work item address, work function and workqueue
878 * to avoid false positives. Note that this isn't complete as one may
879 * construct a work function which can introduce dependency onto itself
880 * through a recycled work item. Well, if somebody wants to shoot oneself
881 * in the foot that badly, there's only so much we can do, and if such
882 * deadlock actually occurs, it should be easy to locate the culprit work
886 * spin_lock_irq(pool->lock).
889 * Pointer to worker which is executing @work if found, NULL
892 static struct worker *find_worker_executing_work(struct worker_pool *pool,
893 struct work_struct *work)
895 struct worker *worker;
897 hash_for_each_possible(pool->busy_hash, worker, hentry,
899 if (worker->current_work == work &&
900 worker->current_func == work->func)
907 * move_linked_works - move linked works to a list
908 * @work: start of series of works to be scheduled
909 * @head: target list to append @work to
910 * @nextp: out paramter for nested worklist walking
912 * Schedule linked works starting from @work to @head. Work series to
913 * be scheduled starts at @work and includes any consecutive work with
914 * WORK_STRUCT_LINKED set in its predecessor.
916 * If @nextp is not NULL, it's updated to point to the next work of
917 * the last scheduled work. This allows move_linked_works() to be
918 * nested inside outer list_for_each_entry_safe().
921 * spin_lock_irq(pool->lock).
923 static void move_linked_works(struct work_struct *work, struct list_head *head,
924 struct work_struct **nextp)
926 struct work_struct *n;
929 * Linked worklist will always end before the end of the list,
930 * use NULL for list head.
932 list_for_each_entry_safe_from(work, n, NULL, entry) {
933 list_move_tail(&work->entry, head);
934 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
939 * If we're already inside safe list traversal and have moved
940 * multiple works to the scheduled queue, the next position
941 * needs to be updated.
947 static void pwq_activate_delayed_work(struct work_struct *work)
949 struct pool_workqueue *pwq = get_work_pwq(work);
951 trace_workqueue_activate_work(work);
952 move_linked_works(work, &pwq->pool->worklist, NULL);
953 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
957 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
959 struct work_struct *work = list_first_entry(&pwq->delayed_works,
960 struct work_struct, entry);
962 pwq_activate_delayed_work(work);
966 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
967 * @pwq: pwq of interest
968 * @color: color of work which left the queue
970 * A work either has completed or is removed from pending queue,
971 * decrement nr_in_flight of its pwq and handle workqueue flushing.
974 * spin_lock_irq(pool->lock).
976 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
978 /* ignore uncolored works */
979 if (color == WORK_NO_COLOR)
982 pwq->nr_in_flight[color]--;
985 if (!list_empty(&pwq->delayed_works)) {
986 /* one down, submit a delayed one */
987 if (pwq->nr_active < pwq->max_active)
988 pwq_activate_first_delayed(pwq);
991 /* is flush in progress and are we at the flushing tip? */
992 if (likely(pwq->flush_color != color))
995 /* are there still in-flight works? */
996 if (pwq->nr_in_flight[color])
999 /* this pwq is done, clear flush_color */
1000 pwq->flush_color = -1;
1003 * If this was the last pwq, wake up the first flusher. It
1004 * will handle the rest.
1006 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1007 complete(&pwq->wq->first_flusher->done);
1011 * try_to_grab_pending - steal work item from worklist and disable irq
1012 * @work: work item to steal
1013 * @is_dwork: @work is a delayed_work
1014 * @flags: place to store irq state
1016 * Try to grab PENDING bit of @work. This function can handle @work in any
1017 * stable state - idle, on timer or on worklist. Return values are
1019 * 1 if @work was pending and we successfully stole PENDING
1020 * 0 if @work was idle and we claimed PENDING
1021 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1022 * -ENOENT if someone else is canceling @work, this state may persist
1023 * for arbitrarily long
1025 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1026 * interrupted while holding PENDING and @work off queue, irq must be
1027 * disabled on entry. This, combined with delayed_work->timer being
1028 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1030 * On successful return, >= 0, irq is disabled and the caller is
1031 * responsible for releasing it using local_irq_restore(*@flags).
1033 * This function is safe to call from any context including IRQ handler.
1035 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1036 unsigned long *flags)
1038 struct worker_pool *pool;
1039 struct pool_workqueue *pwq;
1041 local_irq_save(*flags);
1043 /* try to steal the timer if it exists */
1045 struct delayed_work *dwork = to_delayed_work(work);
1048 * dwork->timer is irqsafe. If del_timer() fails, it's
1049 * guaranteed that the timer is not queued anywhere and not
1050 * running on the local CPU.
1052 if (likely(del_timer(&dwork->timer)))
1056 /* try to claim PENDING the normal way */
1057 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1061 * The queueing is in progress, or it is already queued. Try to
1062 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1064 pool = get_work_pool(work);
1068 spin_lock(&pool->lock);
1070 * work->data is guaranteed to point to pwq only while the work
1071 * item is queued on pwq->wq, and both updating work->data to point
1072 * to pwq on queueing and to pool on dequeueing are done under
1073 * pwq->pool->lock. This in turn guarantees that, if work->data
1074 * points to pwq which is associated with a locked pool, the work
1075 * item is currently queued on that pool.
1077 pwq = get_work_pwq(work);
1078 if (pwq && pwq->pool == pool) {
1079 debug_work_deactivate(work);
1082 * A delayed work item cannot be grabbed directly because
1083 * it might have linked NO_COLOR work items which, if left
1084 * on the delayed_list, will confuse pwq->nr_active
1085 * management later on and cause stall. Make sure the work
1086 * item is activated before grabbing.
1088 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1089 pwq_activate_delayed_work(work);
1091 list_del_init(&work->entry);
1092 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1094 /* work->data points to pwq iff queued, point to pool */
1095 set_work_pool_and_keep_pending(work, pool->id);
1097 spin_unlock(&pool->lock);
1100 spin_unlock(&pool->lock);
1102 local_irq_restore(*flags);
1103 if (work_is_canceling(work))
1110 * insert_work - insert a work into a pool
1111 * @pwq: pwq @work belongs to
1112 * @work: work to insert
1113 * @head: insertion point
1114 * @extra_flags: extra WORK_STRUCT_* flags to set
1116 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1117 * work_struct flags.
1120 * spin_lock_irq(pool->lock).
1122 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1123 struct list_head *head, unsigned int extra_flags)
1125 struct worker_pool *pool = pwq->pool;
1127 /* we own @work, set data and link */
1128 set_work_pwq(work, pwq, extra_flags);
1129 list_add_tail(&work->entry, head);
1132 * Ensure either worker_sched_deactivated() sees the above
1133 * list_add_tail() or we see zero nr_running to avoid workers
1134 * lying around lazily while there are works to be processed.
1138 if (__need_more_worker(pool))
1139 wake_up_worker(pool);
1143 * Test whether @work is being queued from another work executing on the
1146 static bool is_chained_work(struct workqueue_struct *wq)
1148 struct worker *worker;
1150 worker = current_wq_worker();
1152 * Return %true iff I'm a worker execuing a work item on @wq. If
1153 * I'm @worker, it's safe to dereference it without locking.
1155 return worker && worker->current_pwq->wq == wq;
1158 static void __queue_work(int cpu, struct workqueue_struct *wq,
1159 struct work_struct *work)
1161 struct pool_workqueue *pwq;
1162 struct list_head *worklist;
1163 unsigned int work_flags;
1164 unsigned int req_cpu = cpu;
1167 * While a work item is PENDING && off queue, a task trying to
1168 * steal the PENDING will busy-loop waiting for it to either get
1169 * queued or lose PENDING. Grabbing PENDING and queueing should
1170 * happen with IRQ disabled.
1172 WARN_ON_ONCE(!irqs_disabled());
1174 debug_work_activate(work);
1176 /* if dying, only works from the same workqueue are allowed */
1177 if (unlikely(wq->flags & WQ_DRAINING) &&
1178 WARN_ON_ONCE(!is_chained_work(wq)))
1181 /* determine the pwq to use */
1182 if (!(wq->flags & WQ_UNBOUND)) {
1183 struct worker_pool *last_pool;
1185 if (cpu == WORK_CPU_UNBOUND)
1186 cpu = raw_smp_processor_id();
1189 * It's multi cpu. If @work was previously on a different
1190 * cpu, it might still be running there, in which case the
1191 * work needs to be queued on that cpu to guarantee
1194 pwq = get_pwq(cpu, wq);
1195 last_pool = get_work_pool(work);
1197 if (last_pool && last_pool != pwq->pool) {
1198 struct worker *worker;
1200 spin_lock(&last_pool->lock);
1202 worker = find_worker_executing_work(last_pool, work);
1204 if (worker && worker->current_pwq->wq == wq) {
1205 pwq = get_pwq(last_pool->cpu, wq);
1207 /* meh... not running there, queue here */
1208 spin_unlock(&last_pool->lock);
1209 spin_lock(&pwq->pool->lock);
1212 spin_lock(&pwq->pool->lock);
1215 pwq = get_pwq(WORK_CPU_UNBOUND, wq);
1216 spin_lock(&pwq->pool->lock);
1219 /* pwq determined, queue */
1220 trace_workqueue_queue_work(req_cpu, pwq, work);
1222 if (WARN_ON(!list_empty(&work->entry))) {
1223 spin_unlock(&pwq->pool->lock);
1227 pwq->nr_in_flight[pwq->work_color]++;
1228 work_flags = work_color_to_flags(pwq->work_color);
1230 if (likely(pwq->nr_active < pwq->max_active)) {
1231 trace_workqueue_activate_work(work);
1233 worklist = &pwq->pool->worklist;
1235 work_flags |= WORK_STRUCT_DELAYED;
1236 worklist = &pwq->delayed_works;
1239 insert_work(pwq, work, worklist, work_flags);
1241 spin_unlock(&pwq->pool->lock);
1245 * queue_work_on - queue work on specific cpu
1246 * @cpu: CPU number to execute work on
1247 * @wq: workqueue to use
1248 * @work: work to queue
1250 * Returns %false if @work was already on a queue, %true otherwise.
1252 * We queue the work to a specific CPU, the caller must ensure it
1255 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1256 struct work_struct *work)
1259 unsigned long flags;
1261 local_irq_save(flags);
1263 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1264 __queue_work(cpu, wq, work);
1268 local_irq_restore(flags);
1271 EXPORT_SYMBOL_GPL(queue_work_on);
1274 * queue_work - queue work on a workqueue
1275 * @wq: workqueue to use
1276 * @work: work to queue
1278 * Returns %false if @work was already on a queue, %true otherwise.
1280 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1281 * it can be processed by another CPU.
1283 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1285 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1287 EXPORT_SYMBOL_GPL(queue_work);
1289 void delayed_work_timer_fn(unsigned long __data)
1291 struct delayed_work *dwork = (struct delayed_work *)__data;
1293 /* should have been called from irqsafe timer with irq already off */
1294 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1296 EXPORT_SYMBOL(delayed_work_timer_fn);
1298 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1299 struct delayed_work *dwork, unsigned long delay)
1301 struct timer_list *timer = &dwork->timer;
1302 struct work_struct *work = &dwork->work;
1304 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1305 timer->data != (unsigned long)dwork);
1306 WARN_ON_ONCE(timer_pending(timer));
1307 WARN_ON_ONCE(!list_empty(&work->entry));
1310 * If @delay is 0, queue @dwork->work immediately. This is for
1311 * both optimization and correctness. The earliest @timer can
1312 * expire is on the closest next tick and delayed_work users depend
1313 * on that there's no such delay when @delay is 0.
1316 __queue_work(cpu, wq, &dwork->work);
1320 timer_stats_timer_set_start_info(&dwork->timer);
1324 timer->expires = jiffies + delay;
1326 if (unlikely(cpu != WORK_CPU_UNBOUND))
1327 add_timer_on(timer, cpu);
1333 * queue_delayed_work_on - queue work on specific CPU after delay
1334 * @cpu: CPU number to execute work on
1335 * @wq: workqueue to use
1336 * @dwork: work to queue
1337 * @delay: number of jiffies to wait before queueing
1339 * Returns %false if @work was already on a queue, %true otherwise. If
1340 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1343 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1344 struct delayed_work *dwork, unsigned long delay)
1346 struct work_struct *work = &dwork->work;
1348 unsigned long flags;
1350 /* read the comment in __queue_work() */
1351 local_irq_save(flags);
1353 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1354 __queue_delayed_work(cpu, wq, dwork, delay);
1358 local_irq_restore(flags);
1361 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1364 * queue_delayed_work - queue work on a workqueue after delay
1365 * @wq: workqueue to use
1366 * @dwork: delayable work to queue
1367 * @delay: number of jiffies to wait before queueing
1369 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1371 bool queue_delayed_work(struct workqueue_struct *wq,
1372 struct delayed_work *dwork, unsigned long delay)
1374 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1376 EXPORT_SYMBOL_GPL(queue_delayed_work);
1379 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1380 * @cpu: CPU number to execute work on
1381 * @wq: workqueue to use
1382 * @dwork: work to queue
1383 * @delay: number of jiffies to wait before queueing
1385 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1386 * modify @dwork's timer so that it expires after @delay. If @delay is
1387 * zero, @work is guaranteed to be scheduled immediately regardless of its
1390 * Returns %false if @dwork was idle and queued, %true if @dwork was
1391 * pending and its timer was modified.
1393 * This function is safe to call from any context including IRQ handler.
1394 * See try_to_grab_pending() for details.
1396 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1397 struct delayed_work *dwork, unsigned long delay)
1399 unsigned long flags;
1403 ret = try_to_grab_pending(&dwork->work, true, &flags);
1404 } while (unlikely(ret == -EAGAIN));
1406 if (likely(ret >= 0)) {
1407 __queue_delayed_work(cpu, wq, dwork, delay);
1408 local_irq_restore(flags);
1411 /* -ENOENT from try_to_grab_pending() becomes %true */
1414 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1417 * mod_delayed_work - modify delay of or queue a delayed work
1418 * @wq: workqueue to use
1419 * @dwork: work to queue
1420 * @delay: number of jiffies to wait before queueing
1422 * mod_delayed_work_on() on local CPU.
1424 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1425 unsigned long delay)
1427 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1429 EXPORT_SYMBOL_GPL(mod_delayed_work);
1432 * worker_enter_idle - enter idle state
1433 * @worker: worker which is entering idle state
1435 * @worker is entering idle state. Update stats and idle timer if
1439 * spin_lock_irq(pool->lock).
1441 static void worker_enter_idle(struct worker *worker)
1443 struct worker_pool *pool = worker->pool;
1445 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1446 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1447 (worker->hentry.next || worker->hentry.pprev)))
1450 /* can't use worker_set_flags(), also called from start_worker() */
1451 worker->flags |= WORKER_IDLE;
1453 worker->last_active = jiffies;
1455 /* idle_list is LIFO */
1456 list_add(&worker->entry, &pool->idle_list);
1458 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1459 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1462 * Sanity check nr_running. Because wq_unbind_fn() releases
1463 * pool->lock between setting %WORKER_UNBOUND and zapping
1464 * nr_running, the warning may trigger spuriously. Check iff
1465 * unbind is not in progress.
1467 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1468 pool->nr_workers == pool->nr_idle &&
1469 atomic_read(&pool->nr_running));
1473 * worker_leave_idle - leave idle state
1474 * @worker: worker which is leaving idle state
1476 * @worker is leaving idle state. Update stats.
1479 * spin_lock_irq(pool->lock).
1481 static void worker_leave_idle(struct worker *worker)
1483 struct worker_pool *pool = worker->pool;
1485 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1487 worker_clr_flags(worker, WORKER_IDLE);
1489 list_del_init(&worker->entry);
1493 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1494 * @pool: target worker_pool
1496 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1498 * Works which are scheduled while the cpu is online must at least be
1499 * scheduled to a worker which is bound to the cpu so that if they are
1500 * flushed from cpu callbacks while cpu is going down, they are
1501 * guaranteed to execute on the cpu.
1503 * This function is to be used by unbound workers and rescuers to bind
1504 * themselves to the target cpu and may race with cpu going down or
1505 * coming online. kthread_bind() can't be used because it may put the
1506 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1507 * verbatim as it's best effort and blocking and pool may be
1508 * [dis]associated in the meantime.
1510 * This function tries set_cpus_allowed() and locks pool and verifies the
1511 * binding against %POOL_DISASSOCIATED which is set during
1512 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1513 * enters idle state or fetches works without dropping lock, it can
1514 * guarantee the scheduling requirement described in the first paragraph.
1517 * Might sleep. Called without any lock but returns with pool->lock
1521 * %true if the associated pool is online (@worker is successfully
1522 * bound), %false if offline.
1524 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1525 __acquires(&pool->lock)
1529 * The following call may fail, succeed or succeed
1530 * without actually migrating the task to the cpu if
1531 * it races with cpu hotunplug operation. Verify
1532 * against POOL_DISASSOCIATED.
1534 if (!(pool->flags & POOL_DISASSOCIATED))
1535 set_cpus_allowed_ptr(current, get_cpu_mask(pool->cpu));
1537 spin_lock_irq(&pool->lock);
1538 if (pool->flags & POOL_DISASSOCIATED)
1540 if (task_cpu(current) == pool->cpu &&
1541 cpumask_equal(¤t->cpus_allowed,
1542 get_cpu_mask(pool->cpu)))
1544 spin_unlock_irq(&pool->lock);
1547 * We've raced with CPU hot[un]plug. Give it a breather
1548 * and retry migration. cond_resched() is required here;
1549 * otherwise, we might deadlock against cpu_stop trying to
1550 * bring down the CPU on non-preemptive kernel.
1558 * Rebind an idle @worker to its CPU. worker_thread() will test
1559 * list_empty(@worker->entry) before leaving idle and call this function.
1561 static void idle_worker_rebind(struct worker *worker)
1563 /* CPU may go down again inbetween, clear UNBOUND only on success */
1564 if (worker_maybe_bind_and_lock(worker->pool))
1565 worker_clr_flags(worker, WORKER_UNBOUND);
1567 /* rebind complete, become available again */
1568 list_add(&worker->entry, &worker->pool->idle_list);
1569 spin_unlock_irq(&worker->pool->lock);
1573 * Function for @worker->rebind.work used to rebind unbound busy workers to
1574 * the associated cpu which is coming back online. This is scheduled by
1575 * cpu up but can race with other cpu hotplug operations and may be
1576 * executed twice without intervening cpu down.
1578 static void busy_worker_rebind_fn(struct work_struct *work)
1580 struct worker *worker = container_of(work, struct worker, rebind_work);
1582 if (worker_maybe_bind_and_lock(worker->pool))
1583 worker_clr_flags(worker, WORKER_UNBOUND);
1585 spin_unlock_irq(&worker->pool->lock);
1589 * rebind_workers - rebind all workers of a pool to the associated CPU
1590 * @pool: pool of interest
1592 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1593 * is different for idle and busy ones.
1595 * Idle ones will be removed from the idle_list and woken up. They will
1596 * add themselves back after completing rebind. This ensures that the
1597 * idle_list doesn't contain any unbound workers when re-bound busy workers
1598 * try to perform local wake-ups for concurrency management.
1600 * Busy workers can rebind after they finish their current work items.
1601 * Queueing the rebind work item at the head of the scheduled list is
1602 * enough. Note that nr_running will be properly bumped as busy workers
1605 * On return, all non-manager workers are scheduled for rebind - see
1606 * manage_workers() for the manager special case. Any idle worker
1607 * including the manager will not appear on @idle_list until rebind is
1608 * complete, making local wake-ups safe.
1610 static void rebind_workers(struct worker_pool *pool)
1612 struct worker *worker, *n;
1615 lockdep_assert_held(&pool->assoc_mutex);
1616 lockdep_assert_held(&pool->lock);
1618 /* dequeue and kick idle ones */
1619 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1621 * idle workers should be off @pool->idle_list until rebind
1622 * is complete to avoid receiving premature local wake-ups.
1624 list_del_init(&worker->entry);
1627 * worker_thread() will see the above dequeuing and call
1628 * idle_worker_rebind().
1630 wake_up_process(worker->task);
1633 /* rebind busy workers */
1634 for_each_busy_worker(worker, i, pool) {
1635 struct work_struct *rebind_work = &worker->rebind_work;
1636 struct workqueue_struct *wq;
1638 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1639 work_data_bits(rebind_work)))
1642 debug_work_activate(rebind_work);
1645 * wq doesn't really matter but let's keep @worker->pool
1646 * and @pwq->pool consistent for sanity.
1648 if (std_worker_pool_pri(worker->pool))
1649 wq = system_highpri_wq;
1653 insert_work(get_pwq(pool->cpu, wq), rebind_work,
1654 worker->scheduled.next,
1655 work_color_to_flags(WORK_NO_COLOR));
1659 static struct worker *alloc_worker(void)
1661 struct worker *worker;
1663 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1665 INIT_LIST_HEAD(&worker->entry);
1666 INIT_LIST_HEAD(&worker->scheduled);
1667 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1668 /* on creation a worker is in !idle && prep state */
1669 worker->flags = WORKER_PREP;
1675 * create_worker - create a new workqueue worker
1676 * @pool: pool the new worker will belong to
1678 * Create a new worker which is bound to @pool. The returned worker
1679 * can be started by calling start_worker() or destroyed using
1683 * Might sleep. Does GFP_KERNEL allocations.
1686 * Pointer to the newly created worker.
1688 static struct worker *create_worker(struct worker_pool *pool)
1690 const char *pri = std_worker_pool_pri(pool) ? "H" : "";
1691 struct worker *worker = NULL;
1694 spin_lock_irq(&pool->lock);
1695 while (ida_get_new(&pool->worker_ida, &id)) {
1696 spin_unlock_irq(&pool->lock);
1697 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1699 spin_lock_irq(&pool->lock);
1701 spin_unlock_irq(&pool->lock);
1703 worker = alloc_worker();
1707 worker->pool = pool;
1710 if (pool->cpu != WORK_CPU_UNBOUND)
1711 worker->task = kthread_create_on_node(worker_thread,
1712 worker, cpu_to_node(pool->cpu),
1713 "kworker/%d:%d%s", pool->cpu, id, pri);
1715 worker->task = kthread_create(worker_thread, worker,
1716 "kworker/u:%d%s", id, pri);
1717 if (IS_ERR(worker->task))
1720 if (std_worker_pool_pri(pool))
1721 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1724 * Determine CPU binding of the new worker depending on
1725 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1726 * flag remains stable across this function. See the comments
1727 * above the flag definition for details.
1729 * As an unbound worker may later become a regular one if CPU comes
1730 * online, make sure every worker has %PF_THREAD_BOUND set.
1732 if (!(pool->flags & POOL_DISASSOCIATED)) {
1733 kthread_bind(worker->task, pool->cpu);
1735 worker->task->flags |= PF_THREAD_BOUND;
1736 worker->flags |= WORKER_UNBOUND;
1742 spin_lock_irq(&pool->lock);
1743 ida_remove(&pool->worker_ida, id);
1744 spin_unlock_irq(&pool->lock);
1751 * start_worker - start a newly created worker
1752 * @worker: worker to start
1754 * Make the pool aware of @worker and start it.
1757 * spin_lock_irq(pool->lock).
1759 static void start_worker(struct worker *worker)
1761 worker->flags |= WORKER_STARTED;
1762 worker->pool->nr_workers++;
1763 worker_enter_idle(worker);
1764 wake_up_process(worker->task);
1768 * destroy_worker - destroy a workqueue worker
1769 * @worker: worker to be destroyed
1771 * Destroy @worker and adjust @pool stats accordingly.
1774 * spin_lock_irq(pool->lock) which is released and regrabbed.
1776 static void destroy_worker(struct worker *worker)
1778 struct worker_pool *pool = worker->pool;
1779 int id = worker->id;
1781 /* sanity check frenzy */
1782 if (WARN_ON(worker->current_work) ||
1783 WARN_ON(!list_empty(&worker->scheduled)))
1786 if (worker->flags & WORKER_STARTED)
1788 if (worker->flags & WORKER_IDLE)
1791 list_del_init(&worker->entry);
1792 worker->flags |= WORKER_DIE;
1794 spin_unlock_irq(&pool->lock);
1796 kthread_stop(worker->task);
1799 spin_lock_irq(&pool->lock);
1800 ida_remove(&pool->worker_ida, id);
1803 static void idle_worker_timeout(unsigned long __pool)
1805 struct worker_pool *pool = (void *)__pool;
1807 spin_lock_irq(&pool->lock);
1809 if (too_many_workers(pool)) {
1810 struct worker *worker;
1811 unsigned long expires;
1813 /* idle_list is kept in LIFO order, check the last one */
1814 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1815 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1817 if (time_before(jiffies, expires))
1818 mod_timer(&pool->idle_timer, expires);
1820 /* it's been idle for too long, wake up manager */
1821 pool->flags |= POOL_MANAGE_WORKERS;
1822 wake_up_worker(pool);
1826 spin_unlock_irq(&pool->lock);
1829 static void send_mayday(struct work_struct *work)
1831 struct pool_workqueue *pwq = get_work_pwq(work);
1832 struct workqueue_struct *wq = pwq->wq;
1834 lockdep_assert_held(&workqueue_lock);
1836 if (!(wq->flags & WQ_RESCUER))
1839 /* mayday mayday mayday */
1840 if (list_empty(&pwq->mayday_node)) {
1841 list_add_tail(&pwq->mayday_node, &wq->maydays);
1842 wake_up_process(wq->rescuer->task);
1846 static void pool_mayday_timeout(unsigned long __pool)
1848 struct worker_pool *pool = (void *)__pool;
1849 struct work_struct *work;
1851 spin_lock_irq(&workqueue_lock); /* for wq->maydays */
1852 spin_lock(&pool->lock);
1854 if (need_to_create_worker(pool)) {
1856 * We've been trying to create a new worker but
1857 * haven't been successful. We might be hitting an
1858 * allocation deadlock. Send distress signals to
1861 list_for_each_entry(work, &pool->worklist, entry)
1865 spin_unlock(&pool->lock);
1866 spin_unlock_irq(&workqueue_lock);
1868 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1872 * maybe_create_worker - create a new worker if necessary
1873 * @pool: pool to create a new worker for
1875 * Create a new worker for @pool if necessary. @pool is guaranteed to
1876 * have at least one idle worker on return from this function. If
1877 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1878 * sent to all rescuers with works scheduled on @pool to resolve
1879 * possible allocation deadlock.
1881 * On return, need_to_create_worker() is guaranteed to be false and
1882 * may_start_working() true.
1885 * spin_lock_irq(pool->lock) which may be released and regrabbed
1886 * multiple times. Does GFP_KERNEL allocations. Called only from
1890 * false if no action was taken and pool->lock stayed locked, true
1893 static bool maybe_create_worker(struct worker_pool *pool)
1894 __releases(&pool->lock)
1895 __acquires(&pool->lock)
1897 if (!need_to_create_worker(pool))
1900 spin_unlock_irq(&pool->lock);
1902 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1903 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1906 struct worker *worker;
1908 worker = create_worker(pool);
1910 del_timer_sync(&pool->mayday_timer);
1911 spin_lock_irq(&pool->lock);
1912 start_worker(worker);
1913 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1918 if (!need_to_create_worker(pool))
1921 __set_current_state(TASK_INTERRUPTIBLE);
1922 schedule_timeout(CREATE_COOLDOWN);
1924 if (!need_to_create_worker(pool))
1928 del_timer_sync(&pool->mayday_timer);
1929 spin_lock_irq(&pool->lock);
1930 if (need_to_create_worker(pool))
1936 * maybe_destroy_worker - destroy workers which have been idle for a while
1937 * @pool: pool to destroy workers for
1939 * Destroy @pool workers which have been idle for longer than
1940 * IDLE_WORKER_TIMEOUT.
1943 * spin_lock_irq(pool->lock) which may be released and regrabbed
1944 * multiple times. Called only from manager.
1947 * false if no action was taken and pool->lock stayed locked, true
1950 static bool maybe_destroy_workers(struct worker_pool *pool)
1954 while (too_many_workers(pool)) {
1955 struct worker *worker;
1956 unsigned long expires;
1958 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1959 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1961 if (time_before(jiffies, expires)) {
1962 mod_timer(&pool->idle_timer, expires);
1966 destroy_worker(worker);
1974 * manage_workers - manage worker pool
1977 * Assume the manager role and manage the worker pool @worker belongs
1978 * to. At any given time, there can be only zero or one manager per
1979 * pool. The exclusion is handled automatically by this function.
1981 * The caller can safely start processing works on false return. On
1982 * true return, it's guaranteed that need_to_create_worker() is false
1983 * and may_start_working() is true.
1986 * spin_lock_irq(pool->lock) which may be released and regrabbed
1987 * multiple times. Does GFP_KERNEL allocations.
1990 * spin_lock_irq(pool->lock) which may be released and regrabbed
1991 * multiple times. Does GFP_KERNEL allocations.
1993 static bool manage_workers(struct worker *worker)
1995 struct worker_pool *pool = worker->pool;
1998 if (pool->flags & POOL_MANAGING_WORKERS)
2001 pool->flags |= POOL_MANAGING_WORKERS;
2004 * To simplify both worker management and CPU hotplug, hold off
2005 * management while hotplug is in progress. CPU hotplug path can't
2006 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2007 * lead to idle worker depletion (all become busy thinking someone
2008 * else is managing) which in turn can result in deadlock under
2009 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2010 * manager against CPU hotplug.
2012 * assoc_mutex would always be free unless CPU hotplug is in
2013 * progress. trylock first without dropping @pool->lock.
2015 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2016 spin_unlock_irq(&pool->lock);
2017 mutex_lock(&pool->assoc_mutex);
2019 * CPU hotplug could have happened while we were waiting
2020 * for assoc_mutex. Hotplug itself can't handle us
2021 * because manager isn't either on idle or busy list, and
2022 * @pool's state and ours could have deviated.
2024 * As hotplug is now excluded via assoc_mutex, we can
2025 * simply try to bind. It will succeed or fail depending
2026 * on @pool's current state. Try it and adjust
2027 * %WORKER_UNBOUND accordingly.
2029 if (worker_maybe_bind_and_lock(pool))
2030 worker->flags &= ~WORKER_UNBOUND;
2032 worker->flags |= WORKER_UNBOUND;
2037 pool->flags &= ~POOL_MANAGE_WORKERS;
2040 * Destroy and then create so that may_start_working() is true
2043 ret |= maybe_destroy_workers(pool);
2044 ret |= maybe_create_worker(pool);
2046 pool->flags &= ~POOL_MANAGING_WORKERS;
2047 mutex_unlock(&pool->assoc_mutex);
2052 * process_one_work - process single work
2054 * @work: work to process
2056 * Process @work. This function contains all the logics necessary to
2057 * process a single work including synchronization against and
2058 * interaction with other workers on the same cpu, queueing and
2059 * flushing. As long as context requirement is met, any worker can
2060 * call this function to process a work.
2063 * spin_lock_irq(pool->lock) which is released and regrabbed.
2065 static void process_one_work(struct worker *worker, struct work_struct *work)
2066 __releases(&pool->lock)
2067 __acquires(&pool->lock)
2069 struct pool_workqueue *pwq = get_work_pwq(work);
2070 struct worker_pool *pool = worker->pool;
2071 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2073 struct worker *collision;
2074 #ifdef CONFIG_LOCKDEP
2076 * It is permissible to free the struct work_struct from
2077 * inside the function that is called from it, this we need to
2078 * take into account for lockdep too. To avoid bogus "held
2079 * lock freed" warnings as well as problems when looking into
2080 * work->lockdep_map, make a copy and use that here.
2082 struct lockdep_map lockdep_map;
2084 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2087 * Ensure we're on the correct CPU. DISASSOCIATED test is
2088 * necessary to avoid spurious warnings from rescuers servicing the
2089 * unbound or a disassociated pool.
2091 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2092 !(pool->flags & POOL_DISASSOCIATED) &&
2093 raw_smp_processor_id() != pool->cpu);
2096 * A single work shouldn't be executed concurrently by
2097 * multiple workers on a single cpu. Check whether anyone is
2098 * already processing the work. If so, defer the work to the
2099 * currently executing one.
2101 collision = find_worker_executing_work(pool, work);
2102 if (unlikely(collision)) {
2103 move_linked_works(work, &collision->scheduled, NULL);
2107 /* claim and dequeue */
2108 debug_work_deactivate(work);
2109 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2110 worker->current_work = work;
2111 worker->current_func = work->func;
2112 worker->current_pwq = pwq;
2113 work_color = get_work_color(work);
2115 list_del_init(&work->entry);
2118 * CPU intensive works don't participate in concurrency
2119 * management. They're the scheduler's responsibility.
2121 if (unlikely(cpu_intensive))
2122 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2125 * Unbound pool isn't concurrency managed and work items should be
2126 * executed ASAP. Wake up another worker if necessary.
2128 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2129 wake_up_worker(pool);
2132 * Record the last pool and clear PENDING which should be the last
2133 * update to @work. Also, do this inside @pool->lock so that
2134 * PENDING and queued state changes happen together while IRQ is
2137 set_work_pool_and_clear_pending(work, pool->id);
2139 spin_unlock_irq(&pool->lock);
2141 lock_map_acquire_read(&pwq->wq->lockdep_map);
2142 lock_map_acquire(&lockdep_map);
2143 trace_workqueue_execute_start(work);
2144 worker->current_func(work);
2146 * While we must be careful to not use "work" after this, the trace
2147 * point will only record its address.
2149 trace_workqueue_execute_end(work);
2150 lock_map_release(&lockdep_map);
2151 lock_map_release(&pwq->wq->lockdep_map);
2153 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2154 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2155 " last function: %pf\n",
2156 current->comm, preempt_count(), task_pid_nr(current),
2157 worker->current_func);
2158 debug_show_held_locks(current);
2162 spin_lock_irq(&pool->lock);
2164 /* clear cpu intensive status */
2165 if (unlikely(cpu_intensive))
2166 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2168 /* we're done with it, release */
2169 hash_del(&worker->hentry);
2170 worker->current_work = NULL;
2171 worker->current_func = NULL;
2172 worker->current_pwq = NULL;
2173 pwq_dec_nr_in_flight(pwq, work_color);
2177 * process_scheduled_works - process scheduled works
2180 * Process all scheduled works. Please note that the scheduled list
2181 * may change while processing a work, so this function repeatedly
2182 * fetches a work from the top and executes it.
2185 * spin_lock_irq(pool->lock) which may be released and regrabbed
2188 static void process_scheduled_works(struct worker *worker)
2190 while (!list_empty(&worker->scheduled)) {
2191 struct work_struct *work = list_first_entry(&worker->scheduled,
2192 struct work_struct, entry);
2193 process_one_work(worker, work);
2198 * worker_thread - the worker thread function
2201 * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools
2202 * of these per each cpu. These workers process all works regardless of
2203 * their specific target workqueue. The only exception is works which
2204 * belong to workqueues with a rescuer which will be explained in
2207 static int worker_thread(void *__worker)
2209 struct worker *worker = __worker;
2210 struct worker_pool *pool = worker->pool;
2212 /* tell the scheduler that this is a workqueue worker */
2213 worker->task->flags |= PF_WQ_WORKER;
2215 spin_lock_irq(&pool->lock);
2217 /* we are off idle list if destruction or rebind is requested */
2218 if (unlikely(list_empty(&worker->entry))) {
2219 spin_unlock_irq(&pool->lock);
2221 /* if DIE is set, destruction is requested */
2222 if (worker->flags & WORKER_DIE) {
2223 worker->task->flags &= ~PF_WQ_WORKER;
2227 /* otherwise, rebind */
2228 idle_worker_rebind(worker);
2232 worker_leave_idle(worker);
2234 /* no more worker necessary? */
2235 if (!need_more_worker(pool))
2238 /* do we need to manage? */
2239 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2243 * ->scheduled list can only be filled while a worker is
2244 * preparing to process a work or actually processing it.
2245 * Make sure nobody diddled with it while I was sleeping.
2247 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2250 * When control reaches this point, we're guaranteed to have
2251 * at least one idle worker or that someone else has already
2252 * assumed the manager role.
2254 worker_clr_flags(worker, WORKER_PREP);
2257 struct work_struct *work =
2258 list_first_entry(&pool->worklist,
2259 struct work_struct, entry);
2261 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2262 /* optimization path, not strictly necessary */
2263 process_one_work(worker, work);
2264 if (unlikely(!list_empty(&worker->scheduled)))
2265 process_scheduled_works(worker);
2267 move_linked_works(work, &worker->scheduled, NULL);
2268 process_scheduled_works(worker);
2270 } while (keep_working(pool));
2272 worker_set_flags(worker, WORKER_PREP, false);
2274 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2278 * pool->lock is held and there's no work to process and no need to
2279 * manage, sleep. Workers are woken up only while holding
2280 * pool->lock or from local cpu, so setting the current state
2281 * before releasing pool->lock is enough to prevent losing any
2284 worker_enter_idle(worker);
2285 __set_current_state(TASK_INTERRUPTIBLE);
2286 spin_unlock_irq(&pool->lock);
2292 * rescuer_thread - the rescuer thread function
2295 * Workqueue rescuer thread function. There's one rescuer for each
2296 * workqueue which has WQ_RESCUER set.
2298 * Regular work processing on a pool may block trying to create a new
2299 * worker which uses GFP_KERNEL allocation which has slight chance of
2300 * developing into deadlock if some works currently on the same queue
2301 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2302 * the problem rescuer solves.
2304 * When such condition is possible, the pool summons rescuers of all
2305 * workqueues which have works queued on the pool and let them process
2306 * those works so that forward progress can be guaranteed.
2308 * This should happen rarely.
2310 static int rescuer_thread(void *__rescuer)
2312 struct worker *rescuer = __rescuer;
2313 struct workqueue_struct *wq = rescuer->rescue_wq;
2314 struct list_head *scheduled = &rescuer->scheduled;
2316 set_user_nice(current, RESCUER_NICE_LEVEL);
2319 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2320 * doesn't participate in concurrency management.
2322 rescuer->task->flags |= PF_WQ_WORKER;
2324 set_current_state(TASK_INTERRUPTIBLE);
2326 if (kthread_should_stop()) {
2327 __set_current_state(TASK_RUNNING);
2328 rescuer->task->flags &= ~PF_WQ_WORKER;
2332 /* see whether any pwq is asking for help */
2333 spin_lock_irq(&workqueue_lock);
2335 while (!list_empty(&wq->maydays)) {
2336 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2337 struct pool_workqueue, mayday_node);
2338 struct worker_pool *pool = pwq->pool;
2339 struct work_struct *work, *n;
2341 __set_current_state(TASK_RUNNING);
2342 list_del_init(&pwq->mayday_node);
2344 spin_unlock_irq(&workqueue_lock);
2346 /* migrate to the target cpu if possible */
2347 worker_maybe_bind_and_lock(pool);
2348 rescuer->pool = pool;
2351 * Slurp in all works issued via this workqueue and
2354 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2355 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2356 if (get_work_pwq(work) == pwq)
2357 move_linked_works(work, scheduled, &n);
2359 process_scheduled_works(rescuer);
2362 * Leave this pool. If keep_working() is %true, notify a
2363 * regular worker; otherwise, we end up with 0 concurrency
2364 * and stalling the execution.
2366 if (keep_working(pool))
2367 wake_up_worker(pool);
2369 rescuer->pool = NULL;
2370 spin_unlock(&pool->lock);
2371 spin_lock(&workqueue_lock);
2374 spin_unlock_irq(&workqueue_lock);
2376 /* rescuers should never participate in concurrency management */
2377 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2383 struct work_struct work;
2384 struct completion done;
2387 static void wq_barrier_func(struct work_struct *work)
2389 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2390 complete(&barr->done);
2394 * insert_wq_barrier - insert a barrier work
2395 * @pwq: pwq to insert barrier into
2396 * @barr: wq_barrier to insert
2397 * @target: target work to attach @barr to
2398 * @worker: worker currently executing @target, NULL if @target is not executing
2400 * @barr is linked to @target such that @barr is completed only after
2401 * @target finishes execution. Please note that the ordering
2402 * guarantee is observed only with respect to @target and on the local
2405 * Currently, a queued barrier can't be canceled. This is because
2406 * try_to_grab_pending() can't determine whether the work to be
2407 * grabbed is at the head of the queue and thus can't clear LINKED
2408 * flag of the previous work while there must be a valid next work
2409 * after a work with LINKED flag set.
2411 * Note that when @worker is non-NULL, @target may be modified
2412 * underneath us, so we can't reliably determine pwq from @target.
2415 * spin_lock_irq(pool->lock).
2417 static void insert_wq_barrier(struct pool_workqueue *pwq,
2418 struct wq_barrier *barr,
2419 struct work_struct *target, struct worker *worker)
2421 struct list_head *head;
2422 unsigned int linked = 0;
2425 * debugobject calls are safe here even with pool->lock locked
2426 * as we know for sure that this will not trigger any of the
2427 * checks and call back into the fixup functions where we
2430 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2431 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2432 init_completion(&barr->done);
2435 * If @target is currently being executed, schedule the
2436 * barrier to the worker; otherwise, put it after @target.
2439 head = worker->scheduled.next;
2441 unsigned long *bits = work_data_bits(target);
2443 head = target->entry.next;
2444 /* there can already be other linked works, inherit and set */
2445 linked = *bits & WORK_STRUCT_LINKED;
2446 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2449 debug_work_activate(&barr->work);
2450 insert_work(pwq, &barr->work, head,
2451 work_color_to_flags(WORK_NO_COLOR) | linked);
2455 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2456 * @wq: workqueue being flushed
2457 * @flush_color: new flush color, < 0 for no-op
2458 * @work_color: new work color, < 0 for no-op
2460 * Prepare pwqs for workqueue flushing.
2462 * If @flush_color is non-negative, flush_color on all pwqs should be
2463 * -1. If no pwq has in-flight commands at the specified color, all
2464 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2465 * has in flight commands, its pwq->flush_color is set to
2466 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2467 * wakeup logic is armed and %true is returned.
2469 * The caller should have initialized @wq->first_flusher prior to
2470 * calling this function with non-negative @flush_color. If
2471 * @flush_color is negative, no flush color update is done and %false
2474 * If @work_color is non-negative, all pwqs should have the same
2475 * work_color which is previous to @work_color and all will be
2476 * advanced to @work_color.
2479 * mutex_lock(wq->flush_mutex).
2482 * %true if @flush_color >= 0 and there's something to flush. %false
2485 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2486 int flush_color, int work_color)
2489 struct pool_workqueue *pwq;
2491 if (flush_color >= 0) {
2492 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2493 atomic_set(&wq->nr_pwqs_to_flush, 1);
2496 for_each_pwq(pwq, wq) {
2497 struct worker_pool *pool = pwq->pool;
2499 spin_lock_irq(&pool->lock);
2501 if (flush_color >= 0) {
2502 WARN_ON_ONCE(pwq->flush_color != -1);
2504 if (pwq->nr_in_flight[flush_color]) {
2505 pwq->flush_color = flush_color;
2506 atomic_inc(&wq->nr_pwqs_to_flush);
2511 if (work_color >= 0) {
2512 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2513 pwq->work_color = work_color;
2516 spin_unlock_irq(&pool->lock);
2519 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2520 complete(&wq->first_flusher->done);
2526 * flush_workqueue - ensure that any scheduled work has run to completion.
2527 * @wq: workqueue to flush
2529 * Forces execution of the workqueue and blocks until its completion.
2530 * This is typically used in driver shutdown handlers.
2532 * We sleep until all works which were queued on entry have been handled,
2533 * but we are not livelocked by new incoming ones.
2535 void flush_workqueue(struct workqueue_struct *wq)
2537 struct wq_flusher this_flusher = {
2538 .list = LIST_HEAD_INIT(this_flusher.list),
2540 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2544 lock_map_acquire(&wq->lockdep_map);
2545 lock_map_release(&wq->lockdep_map);
2547 mutex_lock(&wq->flush_mutex);
2550 * Start-to-wait phase
2552 next_color = work_next_color(wq->work_color);
2554 if (next_color != wq->flush_color) {
2556 * Color space is not full. The current work_color
2557 * becomes our flush_color and work_color is advanced
2560 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2561 this_flusher.flush_color = wq->work_color;
2562 wq->work_color = next_color;
2564 if (!wq->first_flusher) {
2565 /* no flush in progress, become the first flusher */
2566 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2568 wq->first_flusher = &this_flusher;
2570 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2572 /* nothing to flush, done */
2573 wq->flush_color = next_color;
2574 wq->first_flusher = NULL;
2579 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2580 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2581 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2585 * Oops, color space is full, wait on overflow queue.
2586 * The next flush completion will assign us
2587 * flush_color and transfer to flusher_queue.
2589 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2592 mutex_unlock(&wq->flush_mutex);
2594 wait_for_completion(&this_flusher.done);
2597 * Wake-up-and-cascade phase
2599 * First flushers are responsible for cascading flushes and
2600 * handling overflow. Non-first flushers can simply return.
2602 if (wq->first_flusher != &this_flusher)
2605 mutex_lock(&wq->flush_mutex);
2607 /* we might have raced, check again with mutex held */
2608 if (wq->first_flusher != &this_flusher)
2611 wq->first_flusher = NULL;
2613 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2614 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2617 struct wq_flusher *next, *tmp;
2619 /* complete all the flushers sharing the current flush color */
2620 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2621 if (next->flush_color != wq->flush_color)
2623 list_del_init(&next->list);
2624 complete(&next->done);
2627 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2628 wq->flush_color != work_next_color(wq->work_color));
2630 /* this flush_color is finished, advance by one */
2631 wq->flush_color = work_next_color(wq->flush_color);
2633 /* one color has been freed, handle overflow queue */
2634 if (!list_empty(&wq->flusher_overflow)) {
2636 * Assign the same color to all overflowed
2637 * flushers, advance work_color and append to
2638 * flusher_queue. This is the start-to-wait
2639 * phase for these overflowed flushers.
2641 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2642 tmp->flush_color = wq->work_color;
2644 wq->work_color = work_next_color(wq->work_color);
2646 list_splice_tail_init(&wq->flusher_overflow,
2647 &wq->flusher_queue);
2648 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2651 if (list_empty(&wq->flusher_queue)) {
2652 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2657 * Need to flush more colors. Make the next flusher
2658 * the new first flusher and arm pwqs.
2660 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2661 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2663 list_del_init(&next->list);
2664 wq->first_flusher = next;
2666 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2670 * Meh... this color is already done, clear first
2671 * flusher and repeat cascading.
2673 wq->first_flusher = NULL;
2677 mutex_unlock(&wq->flush_mutex);
2679 EXPORT_SYMBOL_GPL(flush_workqueue);
2682 * drain_workqueue - drain a workqueue
2683 * @wq: workqueue to drain
2685 * Wait until the workqueue becomes empty. While draining is in progress,
2686 * only chain queueing is allowed. IOW, only currently pending or running
2687 * work items on @wq can queue further work items on it. @wq is flushed
2688 * repeatedly until it becomes empty. The number of flushing is detemined
2689 * by the depth of chaining and should be relatively short. Whine if it
2692 void drain_workqueue(struct workqueue_struct *wq)
2694 unsigned int flush_cnt = 0;
2695 struct pool_workqueue *pwq;
2698 * __queue_work() needs to test whether there are drainers, is much
2699 * hotter than drain_workqueue() and already looks at @wq->flags.
2700 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2702 spin_lock_irq(&workqueue_lock);
2703 if (!wq->nr_drainers++)
2704 wq->flags |= WQ_DRAINING;
2705 spin_unlock_irq(&workqueue_lock);
2707 flush_workqueue(wq);
2709 for_each_pwq(pwq, wq) {
2712 spin_lock_irq(&pwq->pool->lock);
2713 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2714 spin_unlock_irq(&pwq->pool->lock);
2719 if (++flush_cnt == 10 ||
2720 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2721 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2722 wq->name, flush_cnt);
2726 spin_lock_irq(&workqueue_lock);
2727 if (!--wq->nr_drainers)
2728 wq->flags &= ~WQ_DRAINING;
2729 spin_unlock_irq(&workqueue_lock);
2731 EXPORT_SYMBOL_GPL(drain_workqueue);
2733 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2735 struct worker *worker = NULL;
2736 struct worker_pool *pool;
2737 struct pool_workqueue *pwq;
2740 pool = get_work_pool(work);
2744 spin_lock_irq(&pool->lock);
2745 /* see the comment in try_to_grab_pending() with the same code */
2746 pwq = get_work_pwq(work);
2748 if (unlikely(pwq->pool != pool))
2751 worker = find_worker_executing_work(pool, work);
2754 pwq = worker->current_pwq;
2757 insert_wq_barrier(pwq, barr, work, worker);
2758 spin_unlock_irq(&pool->lock);
2761 * If @max_active is 1 or rescuer is in use, flushing another work
2762 * item on the same workqueue may lead to deadlock. Make sure the
2763 * flusher is not running on the same workqueue by verifying write
2766 if (pwq->wq->saved_max_active == 1 || pwq->wq->flags & WQ_RESCUER)
2767 lock_map_acquire(&pwq->wq->lockdep_map);
2769 lock_map_acquire_read(&pwq->wq->lockdep_map);
2770 lock_map_release(&pwq->wq->lockdep_map);
2774 spin_unlock_irq(&pool->lock);
2779 * flush_work - wait for a work to finish executing the last queueing instance
2780 * @work: the work to flush
2782 * Wait until @work has finished execution. @work is guaranteed to be idle
2783 * on return if it hasn't been requeued since flush started.
2786 * %true if flush_work() waited for the work to finish execution,
2787 * %false if it was already idle.
2789 bool flush_work(struct work_struct *work)
2791 struct wq_barrier barr;
2793 lock_map_acquire(&work->lockdep_map);
2794 lock_map_release(&work->lockdep_map);
2796 if (start_flush_work(work, &barr)) {
2797 wait_for_completion(&barr.done);
2798 destroy_work_on_stack(&barr.work);
2804 EXPORT_SYMBOL_GPL(flush_work);
2806 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2808 unsigned long flags;
2812 ret = try_to_grab_pending(work, is_dwork, &flags);
2814 * If someone else is canceling, wait for the same event it
2815 * would be waiting for before retrying.
2817 if (unlikely(ret == -ENOENT))
2819 } while (unlikely(ret < 0));
2821 /* tell other tasks trying to grab @work to back off */
2822 mark_work_canceling(work);
2823 local_irq_restore(flags);
2826 clear_work_data(work);
2831 * cancel_work_sync - cancel a work and wait for it to finish
2832 * @work: the work to cancel
2834 * Cancel @work and wait for its execution to finish. This function
2835 * can be used even if the work re-queues itself or migrates to
2836 * another workqueue. On return from this function, @work is
2837 * guaranteed to be not pending or executing on any CPU.
2839 * cancel_work_sync(&delayed_work->work) must not be used for
2840 * delayed_work's. Use cancel_delayed_work_sync() instead.
2842 * The caller must ensure that the workqueue on which @work was last
2843 * queued can't be destroyed before this function returns.
2846 * %true if @work was pending, %false otherwise.
2848 bool cancel_work_sync(struct work_struct *work)
2850 return __cancel_work_timer(work, false);
2852 EXPORT_SYMBOL_GPL(cancel_work_sync);
2855 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2856 * @dwork: the delayed work to flush
2858 * Delayed timer is cancelled and the pending work is queued for
2859 * immediate execution. Like flush_work(), this function only
2860 * considers the last queueing instance of @dwork.
2863 * %true if flush_work() waited for the work to finish execution,
2864 * %false if it was already idle.
2866 bool flush_delayed_work(struct delayed_work *dwork)
2868 local_irq_disable();
2869 if (del_timer_sync(&dwork->timer))
2870 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2872 return flush_work(&dwork->work);
2874 EXPORT_SYMBOL(flush_delayed_work);
2877 * cancel_delayed_work - cancel a delayed work
2878 * @dwork: delayed_work to cancel
2880 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2881 * and canceled; %false if wasn't pending. Note that the work callback
2882 * function may still be running on return, unless it returns %true and the
2883 * work doesn't re-arm itself. Explicitly flush or use
2884 * cancel_delayed_work_sync() to wait on it.
2886 * This function is safe to call from any context including IRQ handler.
2888 bool cancel_delayed_work(struct delayed_work *dwork)
2890 unsigned long flags;
2894 ret = try_to_grab_pending(&dwork->work, true, &flags);
2895 } while (unlikely(ret == -EAGAIN));
2897 if (unlikely(ret < 0))
2900 set_work_pool_and_clear_pending(&dwork->work,
2901 get_work_pool_id(&dwork->work));
2902 local_irq_restore(flags);
2905 EXPORT_SYMBOL(cancel_delayed_work);
2908 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2909 * @dwork: the delayed work cancel
2911 * This is cancel_work_sync() for delayed works.
2914 * %true if @dwork was pending, %false otherwise.
2916 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2918 return __cancel_work_timer(&dwork->work, true);
2920 EXPORT_SYMBOL(cancel_delayed_work_sync);
2923 * schedule_work_on - put work task on a specific cpu
2924 * @cpu: cpu to put the work task on
2925 * @work: job to be done
2927 * This puts a job on a specific cpu
2929 bool schedule_work_on(int cpu, struct work_struct *work)
2931 return queue_work_on(cpu, system_wq, work);
2933 EXPORT_SYMBOL(schedule_work_on);
2936 * schedule_work - put work task in global workqueue
2937 * @work: job to be done
2939 * Returns %false if @work was already on the kernel-global workqueue and
2942 * This puts a job in the kernel-global workqueue if it was not already
2943 * queued and leaves it in the same position on the kernel-global
2944 * workqueue otherwise.
2946 bool schedule_work(struct work_struct *work)
2948 return queue_work(system_wq, work);
2950 EXPORT_SYMBOL(schedule_work);
2953 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2955 * @dwork: job to be done
2956 * @delay: number of jiffies to wait
2958 * After waiting for a given time this puts a job in the kernel-global
2959 * workqueue on the specified CPU.
2961 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
2962 unsigned long delay)
2964 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2966 EXPORT_SYMBOL(schedule_delayed_work_on);
2969 * schedule_delayed_work - put work task in global workqueue after delay
2970 * @dwork: job to be done
2971 * @delay: number of jiffies to wait or 0 for immediate execution
2973 * After waiting for a given time this puts a job in the kernel-global
2976 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
2978 return queue_delayed_work(system_wq, dwork, delay);
2980 EXPORT_SYMBOL(schedule_delayed_work);
2983 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2984 * @func: the function to call
2986 * schedule_on_each_cpu() executes @func on each online CPU using the
2987 * system workqueue and blocks until all CPUs have completed.
2988 * schedule_on_each_cpu() is very slow.
2991 * 0 on success, -errno on failure.
2993 int schedule_on_each_cpu(work_func_t func)
2996 struct work_struct __percpu *works;
2998 works = alloc_percpu(struct work_struct);
3004 for_each_online_cpu(cpu) {
3005 struct work_struct *work = per_cpu_ptr(works, cpu);
3007 INIT_WORK(work, func);
3008 schedule_work_on(cpu, work);
3011 for_each_online_cpu(cpu)
3012 flush_work(per_cpu_ptr(works, cpu));
3020 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3022 * Forces execution of the kernel-global workqueue and blocks until its
3025 * Think twice before calling this function! It's very easy to get into
3026 * trouble if you don't take great care. Either of the following situations
3027 * will lead to deadlock:
3029 * One of the work items currently on the workqueue needs to acquire
3030 * a lock held by your code or its caller.
3032 * Your code is running in the context of a work routine.
3034 * They will be detected by lockdep when they occur, but the first might not
3035 * occur very often. It depends on what work items are on the workqueue and
3036 * what locks they need, which you have no control over.
3038 * In most situations flushing the entire workqueue is overkill; you merely
3039 * need to know that a particular work item isn't queued and isn't running.
3040 * In such cases you should use cancel_delayed_work_sync() or
3041 * cancel_work_sync() instead.
3043 void flush_scheduled_work(void)
3045 flush_workqueue(system_wq);
3047 EXPORT_SYMBOL(flush_scheduled_work);
3050 * execute_in_process_context - reliably execute the routine with user context
3051 * @fn: the function to execute
3052 * @ew: guaranteed storage for the execute work structure (must
3053 * be available when the work executes)
3055 * Executes the function immediately if process context is available,
3056 * otherwise schedules the function for delayed execution.
3058 * Returns: 0 - function was executed
3059 * 1 - function was scheduled for execution
3061 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3063 if (!in_interrupt()) {
3068 INIT_WORK(&ew->work, fn);
3069 schedule_work(&ew->work);
3073 EXPORT_SYMBOL_GPL(execute_in_process_context);
3075 int keventd_up(void)
3077 return system_wq != NULL;
3080 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3082 bool highpri = wq->flags & WQ_HIGHPRI;
3085 if (!(wq->flags & WQ_UNBOUND)) {
3086 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3090 for_each_possible_cpu(cpu) {
3091 struct pool_workqueue *pwq = get_pwq(cpu, wq);
3093 pwq->pool = get_std_worker_pool(cpu, highpri);
3094 list_add_tail(&pwq->pwqs_node, &wq->pwqs);
3097 struct pool_workqueue *pwq;
3099 pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL);
3103 pwq->pool = get_std_worker_pool(WORK_CPU_UNBOUND, highpri);
3104 list_add_tail(&pwq->pwqs_node, &wq->pwqs);
3110 static void free_pwqs(struct workqueue_struct *wq)
3112 if (!(wq->flags & WQ_UNBOUND))
3113 free_percpu(wq->cpu_pwqs);
3114 else if (!list_empty(&wq->pwqs))
3115 kmem_cache_free(pwq_cache, list_first_entry(&wq->pwqs,
3116 struct pool_workqueue, pwqs_node));
3119 static int wq_clamp_max_active(int max_active, unsigned int flags,
3122 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3124 if (max_active < 1 || max_active > lim)
3125 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3126 max_active, name, 1, lim);
3128 return clamp_val(max_active, 1, lim);
3131 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3134 struct lock_class_key *key,
3135 const char *lock_name, ...)
3137 va_list args, args1;
3138 struct workqueue_struct *wq;
3139 struct pool_workqueue *pwq;
3142 /* determine namelen, allocate wq and format name */
3143 va_start(args, lock_name);
3144 va_copy(args1, args);
3145 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3147 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3151 vsnprintf(wq->name, namelen, fmt, args1);
3156 * Workqueues which may be used during memory reclaim should
3157 * have a rescuer to guarantee forward progress.
3159 if (flags & WQ_MEM_RECLAIM)
3160 flags |= WQ_RESCUER;
3162 max_active = max_active ?: WQ_DFL_ACTIVE;
3163 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3167 wq->saved_max_active = max_active;
3168 mutex_init(&wq->flush_mutex);
3169 atomic_set(&wq->nr_pwqs_to_flush, 0);
3170 INIT_LIST_HEAD(&wq->pwqs);
3171 INIT_LIST_HEAD(&wq->flusher_queue);
3172 INIT_LIST_HEAD(&wq->flusher_overflow);
3173 INIT_LIST_HEAD(&wq->maydays);
3175 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3176 INIT_LIST_HEAD(&wq->list);
3178 if (alloc_and_link_pwqs(wq) < 0)
3181 for_each_pwq(pwq, wq) {
3182 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3184 pwq->flush_color = -1;
3185 pwq->max_active = max_active;
3186 INIT_LIST_HEAD(&pwq->delayed_works);
3187 INIT_LIST_HEAD(&pwq->mayday_node);
3190 if (flags & WQ_RESCUER) {
3191 struct worker *rescuer;
3193 wq->rescuer = rescuer = alloc_worker();
3197 rescuer->rescue_wq = wq;
3198 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3200 if (IS_ERR(rescuer->task))
3203 rescuer->task->flags |= PF_THREAD_BOUND;
3204 wake_up_process(rescuer->task);
3208 * workqueue_lock protects global freeze state and workqueues
3209 * list. Grab it, set max_active accordingly and add the new
3210 * workqueue to workqueues list.
3212 spin_lock_irq(&workqueue_lock);
3214 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3215 for_each_pwq(pwq, wq)
3216 pwq->max_active = 0;
3218 list_add(&wq->list, &workqueues);
3220 spin_unlock_irq(&workqueue_lock);
3231 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3234 * destroy_workqueue - safely terminate a workqueue
3235 * @wq: target workqueue
3237 * Safely destroy a workqueue. All work currently pending will be done first.
3239 void destroy_workqueue(struct workqueue_struct *wq)
3241 struct pool_workqueue *pwq;
3243 /* drain it before proceeding with destruction */
3244 drain_workqueue(wq);
3247 for_each_pwq(pwq, wq) {
3250 for (i = 0; i < WORK_NR_COLORS; i++)
3251 if (WARN_ON(pwq->nr_in_flight[i]))
3253 if (WARN_ON(pwq->nr_active) ||
3254 WARN_ON(!list_empty(&pwq->delayed_works)))
3259 * wq list is used to freeze wq, remove from list after
3260 * flushing is complete in case freeze races us.
3262 spin_lock_irq(&workqueue_lock);
3263 list_del(&wq->list);
3264 spin_unlock_irq(&workqueue_lock);
3266 if (wq->flags & WQ_RESCUER) {
3267 kthread_stop(wq->rescuer->task);
3274 EXPORT_SYMBOL_GPL(destroy_workqueue);
3277 * pwq_set_max_active - adjust max_active of a pwq
3278 * @pwq: target pool_workqueue
3279 * @max_active: new max_active value.
3281 * Set @pwq->max_active to @max_active and activate delayed works if
3285 * spin_lock_irq(pool->lock).
3287 static void pwq_set_max_active(struct pool_workqueue *pwq, int max_active)
3289 pwq->max_active = max_active;
3291 while (!list_empty(&pwq->delayed_works) &&
3292 pwq->nr_active < pwq->max_active)
3293 pwq_activate_first_delayed(pwq);
3297 * workqueue_set_max_active - adjust max_active of a workqueue
3298 * @wq: target workqueue
3299 * @max_active: new max_active value.
3301 * Set max_active of @wq to @max_active.
3304 * Don't call from IRQ context.
3306 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3308 struct pool_workqueue *pwq;
3310 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3312 spin_lock_irq(&workqueue_lock);
3314 wq->saved_max_active = max_active;
3316 for_each_pwq(pwq, wq) {
3317 struct worker_pool *pool = pwq->pool;
3319 spin_lock(&pool->lock);
3321 if (!(wq->flags & WQ_FREEZABLE) ||
3322 !(pool->flags & POOL_FREEZING))
3323 pwq_set_max_active(pwq, max_active);
3325 spin_unlock(&pool->lock);
3328 spin_unlock_irq(&workqueue_lock);
3330 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3333 * workqueue_congested - test whether a workqueue is congested
3334 * @cpu: CPU in question
3335 * @wq: target workqueue
3337 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3338 * no synchronization around this function and the test result is
3339 * unreliable and only useful as advisory hints or for debugging.
3342 * %true if congested, %false otherwise.
3344 bool workqueue_congested(int cpu, struct workqueue_struct *wq)
3346 struct pool_workqueue *pwq = get_pwq(cpu, wq);
3348 return !list_empty(&pwq->delayed_works);
3350 EXPORT_SYMBOL_GPL(workqueue_congested);
3353 * work_busy - test whether a work is currently pending or running
3354 * @work: the work to be tested
3356 * Test whether @work is currently pending or running. There is no
3357 * synchronization around this function and the test result is
3358 * unreliable and only useful as advisory hints or for debugging.
3361 * OR'd bitmask of WORK_BUSY_* bits.
3363 unsigned int work_busy(struct work_struct *work)
3365 struct worker_pool *pool = get_work_pool(work);
3366 unsigned long flags;
3367 unsigned int ret = 0;
3369 if (work_pending(work))
3370 ret |= WORK_BUSY_PENDING;
3373 spin_lock_irqsave(&pool->lock, flags);
3374 if (find_worker_executing_work(pool, work))
3375 ret |= WORK_BUSY_RUNNING;
3376 spin_unlock_irqrestore(&pool->lock, flags);
3381 EXPORT_SYMBOL_GPL(work_busy);
3386 * There are two challenges in supporting CPU hotplug. Firstly, there
3387 * are a lot of assumptions on strong associations among work, pwq and
3388 * pool which make migrating pending and scheduled works very
3389 * difficult to implement without impacting hot paths. Secondly,
3390 * worker pools serve mix of short, long and very long running works making
3391 * blocked draining impractical.
3393 * This is solved by allowing the pools to be disassociated from the CPU
3394 * running as an unbound one and allowing it to be reattached later if the
3395 * cpu comes back online.
3398 static void wq_unbind_fn(struct work_struct *work)
3400 int cpu = smp_processor_id();
3401 struct worker_pool *pool;
3402 struct worker *worker;
3405 for_each_std_worker_pool(pool, cpu) {
3406 WARN_ON_ONCE(cpu != smp_processor_id());
3408 mutex_lock(&pool->assoc_mutex);
3409 spin_lock_irq(&pool->lock);
3412 * We've claimed all manager positions. Make all workers
3413 * unbound and set DISASSOCIATED. Before this, all workers
3414 * except for the ones which are still executing works from
3415 * before the last CPU down must be on the cpu. After
3416 * this, they may become diasporas.
3418 list_for_each_entry(worker, &pool->idle_list, entry)
3419 worker->flags |= WORKER_UNBOUND;
3421 for_each_busy_worker(worker, i, pool)
3422 worker->flags |= WORKER_UNBOUND;
3424 pool->flags |= POOL_DISASSOCIATED;
3426 spin_unlock_irq(&pool->lock);
3427 mutex_unlock(&pool->assoc_mutex);
3431 * Call schedule() so that we cross rq->lock and thus can guarantee
3432 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3433 * as scheduler callbacks may be invoked from other cpus.
3438 * Sched callbacks are disabled now. Zap nr_running. After this,
3439 * nr_running stays zero and need_more_worker() and keep_working()
3440 * are always true as long as the worklist is not empty. Pools on
3441 * @cpu now behave as unbound (in terms of concurrency management)
3442 * pools which are served by workers tied to the CPU.
3444 * On return from this function, the current worker would trigger
3445 * unbound chain execution of pending work items if other workers
3448 for_each_std_worker_pool(pool, cpu)
3449 atomic_set(&pool->nr_running, 0);
3453 * Workqueues should be brought up before normal priority CPU notifiers.
3454 * This will be registered high priority CPU notifier.
3456 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3457 unsigned long action,
3460 int cpu = (unsigned long)hcpu;
3461 struct worker_pool *pool;
3463 switch (action & ~CPU_TASKS_FROZEN) {
3464 case CPU_UP_PREPARE:
3465 for_each_std_worker_pool(pool, cpu) {
3466 struct worker *worker;
3468 if (pool->nr_workers)
3471 worker = create_worker(pool);
3475 spin_lock_irq(&pool->lock);
3476 start_worker(worker);
3477 spin_unlock_irq(&pool->lock);
3481 case CPU_DOWN_FAILED:
3483 for_each_std_worker_pool(pool, cpu) {
3484 mutex_lock(&pool->assoc_mutex);
3485 spin_lock_irq(&pool->lock);
3487 pool->flags &= ~POOL_DISASSOCIATED;
3488 rebind_workers(pool);
3490 spin_unlock_irq(&pool->lock);
3491 mutex_unlock(&pool->assoc_mutex);
3499 * Workqueues should be brought down after normal priority CPU notifiers.
3500 * This will be registered as low priority CPU notifier.
3502 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3503 unsigned long action,
3506 int cpu = (unsigned long)hcpu;
3507 struct work_struct unbind_work;
3509 switch (action & ~CPU_TASKS_FROZEN) {
3510 case CPU_DOWN_PREPARE:
3511 /* unbinding should happen on the local CPU */
3512 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
3513 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3514 flush_work(&unbind_work);
3522 struct work_for_cpu {
3523 struct work_struct work;
3529 static void work_for_cpu_fn(struct work_struct *work)
3531 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3533 wfc->ret = wfc->fn(wfc->arg);
3537 * work_on_cpu - run a function in user context on a particular cpu
3538 * @cpu: the cpu to run on
3539 * @fn: the function to run
3540 * @arg: the function arg
3542 * This will return the value @fn returns.
3543 * It is up to the caller to ensure that the cpu doesn't go offline.
3544 * The caller must not hold any locks which would prevent @fn from completing.
3546 long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
3548 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3550 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3551 schedule_work_on(cpu, &wfc.work);
3552 flush_work(&wfc.work);
3555 EXPORT_SYMBOL_GPL(work_on_cpu);
3556 #endif /* CONFIG_SMP */
3558 #ifdef CONFIG_FREEZER
3561 * freeze_workqueues_begin - begin freezing workqueues
3563 * Start freezing workqueues. After this function returns, all freezable
3564 * workqueues will queue new works to their frozen_works list instead of
3568 * Grabs and releases workqueue_lock and pool->lock's.
3570 void freeze_workqueues_begin(void)
3572 struct worker_pool *pool;
3573 struct workqueue_struct *wq;
3574 struct pool_workqueue *pwq;
3577 spin_lock_irq(&workqueue_lock);
3579 WARN_ON_ONCE(workqueue_freezing);
3580 workqueue_freezing = true;
3583 for_each_pool(pool, id) {
3584 spin_lock(&pool->lock);
3585 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
3586 pool->flags |= POOL_FREEZING;
3587 spin_unlock(&pool->lock);
3590 /* suppress further executions by setting max_active to zero */
3591 list_for_each_entry(wq, &workqueues, list) {
3592 if (!(wq->flags & WQ_FREEZABLE))
3595 for_each_pwq(pwq, wq) {
3596 spin_lock(&pwq->pool->lock);
3597 pwq->max_active = 0;
3598 spin_unlock(&pwq->pool->lock);
3602 spin_unlock_irq(&workqueue_lock);
3606 * freeze_workqueues_busy - are freezable workqueues still busy?
3608 * Check whether freezing is complete. This function must be called
3609 * between freeze_workqueues_begin() and thaw_workqueues().
3612 * Grabs and releases workqueue_lock.
3615 * %true if some freezable workqueues are still busy. %false if freezing
3618 bool freeze_workqueues_busy(void)
3621 struct workqueue_struct *wq;
3622 struct pool_workqueue *pwq;
3624 spin_lock_irq(&workqueue_lock);
3626 WARN_ON_ONCE(!workqueue_freezing);
3628 list_for_each_entry(wq, &workqueues, list) {
3629 if (!(wq->flags & WQ_FREEZABLE))
3632 * nr_active is monotonically decreasing. It's safe
3633 * to peek without lock.
3635 for_each_pwq(pwq, wq) {
3636 WARN_ON_ONCE(pwq->nr_active < 0);
3637 if (pwq->nr_active) {
3644 spin_unlock_irq(&workqueue_lock);
3649 * thaw_workqueues - thaw workqueues
3651 * Thaw workqueues. Normal queueing is restored and all collected
3652 * frozen works are transferred to their respective pool worklists.
3655 * Grabs and releases workqueue_lock and pool->lock's.
3657 void thaw_workqueues(void)
3659 struct workqueue_struct *wq;
3660 struct pool_workqueue *pwq;
3661 struct worker_pool *pool;
3664 spin_lock_irq(&workqueue_lock);
3666 if (!workqueue_freezing)
3669 /* clear FREEZING */
3670 for_each_pool(pool, id) {
3671 spin_lock(&pool->lock);
3672 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
3673 pool->flags &= ~POOL_FREEZING;
3674 spin_unlock(&pool->lock);
3677 /* restore max_active and repopulate worklist */
3678 list_for_each_entry(wq, &workqueues, list) {
3679 if (!(wq->flags & WQ_FREEZABLE))
3682 for_each_pwq(pwq, wq) {
3683 spin_lock(&pwq->pool->lock);
3684 pwq_set_max_active(pwq, wq->saved_max_active);
3685 spin_unlock(&pwq->pool->lock);
3690 for_each_pool(pool, id) {
3691 spin_lock(&pool->lock);
3692 wake_up_worker(pool);
3693 spin_unlock(&pool->lock);
3696 workqueue_freezing = false;
3698 spin_unlock_irq(&workqueue_lock);
3700 #endif /* CONFIG_FREEZER */
3702 static int __init init_workqueues(void)
3706 /* make sure we have enough bits for OFFQ pool ID */
3707 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
3708 WORK_CPU_END * NR_STD_WORKER_POOLS);
3710 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
3712 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
3714 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3715 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3717 /* initialize CPU pools */
3718 for_each_wq_cpu(cpu) {
3719 struct worker_pool *pool;
3721 for_each_std_worker_pool(pool, cpu) {
3722 spin_lock_init(&pool->lock);
3724 pool->flags |= POOL_DISASSOCIATED;
3725 INIT_LIST_HEAD(&pool->worklist);
3726 INIT_LIST_HEAD(&pool->idle_list);
3727 hash_init(pool->busy_hash);
3729 init_timer_deferrable(&pool->idle_timer);
3730 pool->idle_timer.function = idle_worker_timeout;
3731 pool->idle_timer.data = (unsigned long)pool;
3733 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3734 (unsigned long)pool);
3736 mutex_init(&pool->assoc_mutex);
3737 ida_init(&pool->worker_ida);
3740 BUG_ON(worker_pool_assign_id(pool));
3744 /* create the initial worker */
3745 for_each_online_wq_cpu(cpu) {
3746 struct worker_pool *pool;
3748 for_each_std_worker_pool(pool, cpu) {
3749 struct worker *worker;
3751 if (cpu != WORK_CPU_UNBOUND)
3752 pool->flags &= ~POOL_DISASSOCIATED;
3754 worker = create_worker(pool);
3756 spin_lock_irq(&pool->lock);
3757 start_worker(worker);
3758 spin_unlock_irq(&pool->lock);
3762 system_wq = alloc_workqueue("events", 0, 0);
3763 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3764 system_long_wq = alloc_workqueue("events_long", 0, 0);
3765 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3766 WQ_UNBOUND_MAX_ACTIVE);
3767 system_freezable_wq = alloc_workqueue("events_freezable",
3769 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3770 !system_unbound_wq || !system_freezable_wq);
3773 early_initcall(init_workqueues);