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 unsigned 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 } __aligned(1 << WORK_STRUCT_FLAG_BITS);
176 * Structure used to wait for workqueue flush.
179 struct list_head list; /* F: list of flushers */
180 int flush_color; /* F: flush color waiting for */
181 struct completion done; /* flush completion */
185 * All cpumasks are assumed to be always set on UP and thus can't be
186 * used to determine whether there's something to be done.
189 typedef cpumask_var_t mayday_mask_t;
190 #define mayday_test_and_set_cpu(cpu, mask) \
191 cpumask_test_and_set_cpu((cpu), (mask))
192 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
193 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
194 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
195 #define free_mayday_mask(mask) free_cpumask_var((mask))
197 typedef unsigned long mayday_mask_t;
198 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
199 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
200 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
201 #define alloc_mayday_mask(maskp, gfp) true
202 #define free_mayday_mask(mask) do { } while (0)
206 * The externally visible workqueue abstraction is an array of
207 * per-CPU workqueues:
209 struct workqueue_struct {
210 unsigned int flags; /* W: WQ_* flags */
212 struct pool_workqueue __percpu *pcpu;
213 struct pool_workqueue *single;
215 } pool_wq; /* I: pwq's */
216 struct list_head pwqs; /* I: all pwqs of this wq */
217 struct list_head list; /* W: list of all workqueues */
219 struct mutex flush_mutex; /* protects wq flushing */
220 int work_color; /* F: current work color */
221 int flush_color; /* F: current flush color */
222 atomic_t nr_pwqs_to_flush; /* flush in progress */
223 struct wq_flusher *first_flusher; /* F: first flusher */
224 struct list_head flusher_queue; /* F: flush waiters */
225 struct list_head flusher_overflow; /* F: flush overflow list */
227 mayday_mask_t mayday_mask; /* cpus requesting rescue */
228 struct worker *rescuer; /* I: rescue worker */
230 int nr_drainers; /* W: drain in progress */
231 int saved_max_active; /* W: saved pwq max_active */
232 #ifdef CONFIG_LOCKDEP
233 struct lockdep_map lockdep_map;
235 char name[]; /* I: workqueue name */
238 static struct kmem_cache *pwq_cache;
240 struct workqueue_struct *system_wq __read_mostly;
241 EXPORT_SYMBOL_GPL(system_wq);
242 struct workqueue_struct *system_highpri_wq __read_mostly;
243 EXPORT_SYMBOL_GPL(system_highpri_wq);
244 struct workqueue_struct *system_long_wq __read_mostly;
245 EXPORT_SYMBOL_GPL(system_long_wq);
246 struct workqueue_struct *system_unbound_wq __read_mostly;
247 EXPORT_SYMBOL_GPL(system_unbound_wq);
248 struct workqueue_struct *system_freezable_wq __read_mostly;
249 EXPORT_SYMBOL_GPL(system_freezable_wq);
251 #define CREATE_TRACE_POINTS
252 #include <trace/events/workqueue.h>
254 #define for_each_std_worker_pool(pool, cpu) \
255 for ((pool) = &std_worker_pools(cpu)[0]; \
256 (pool) < &std_worker_pools(cpu)[NR_STD_WORKER_POOLS]; (pool)++)
258 #define for_each_busy_worker(worker, i, pool) \
259 hash_for_each(pool->busy_hash, i, worker, hentry)
261 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
264 if (cpu < nr_cpu_ids) {
266 cpu = cpumask_next(cpu, mask);
267 if (cpu < nr_cpu_ids)
271 return WORK_CPU_UNBOUND;
276 static inline int __next_pwq_cpu(int cpu, const struct cpumask *mask,
277 struct workqueue_struct *wq)
279 return __next_wq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
285 * An extra cpu number is defined using an invalid cpu number
286 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
287 * specific CPU. The following iterators are similar to for_each_*_cpu()
288 * iterators but also considers the unbound CPU.
290 * for_each_wq_cpu() : possible CPUs + WORK_CPU_UNBOUND
291 * for_each_online_wq_cpu() : online CPUs + WORK_CPU_UNBOUND
292 * for_each_pwq_cpu() : possible CPUs for bound workqueues,
293 * WORK_CPU_UNBOUND for unbound workqueues
295 #define for_each_wq_cpu(cpu) \
296 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, 3); \
297 (cpu) < WORK_CPU_END; \
298 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, 3))
300 #define for_each_online_wq_cpu(cpu) \
301 for ((cpu) = __next_wq_cpu(-1, cpu_online_mask, 3); \
302 (cpu) < WORK_CPU_END; \
303 (cpu) = __next_wq_cpu((cpu), cpu_online_mask, 3))
305 #define for_each_pwq_cpu(cpu, wq) \
306 for ((cpu) = __next_pwq_cpu(-1, cpu_possible_mask, (wq)); \
307 (cpu) < WORK_CPU_END; \
308 (cpu) = __next_pwq_cpu((cpu), cpu_possible_mask, (wq)))
310 #ifdef CONFIG_DEBUG_OBJECTS_WORK
312 static struct debug_obj_descr work_debug_descr;
314 static void *work_debug_hint(void *addr)
316 return ((struct work_struct *) addr)->func;
320 * fixup_init is called when:
321 * - an active object is initialized
323 static int work_fixup_init(void *addr, enum debug_obj_state state)
325 struct work_struct *work = addr;
328 case ODEBUG_STATE_ACTIVE:
329 cancel_work_sync(work);
330 debug_object_init(work, &work_debug_descr);
338 * fixup_activate is called when:
339 * - an active object is activated
340 * - an unknown object is activated (might be a statically initialized object)
342 static int work_fixup_activate(void *addr, enum debug_obj_state state)
344 struct work_struct *work = addr;
348 case ODEBUG_STATE_NOTAVAILABLE:
350 * This is not really a fixup. The work struct was
351 * statically initialized. We just make sure that it
352 * is tracked in the object tracker.
354 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
355 debug_object_init(work, &work_debug_descr);
356 debug_object_activate(work, &work_debug_descr);
362 case ODEBUG_STATE_ACTIVE:
371 * fixup_free is called when:
372 * - an active object is freed
374 static int work_fixup_free(void *addr, enum debug_obj_state state)
376 struct work_struct *work = addr;
379 case ODEBUG_STATE_ACTIVE:
380 cancel_work_sync(work);
381 debug_object_free(work, &work_debug_descr);
388 static struct debug_obj_descr work_debug_descr = {
389 .name = "work_struct",
390 .debug_hint = work_debug_hint,
391 .fixup_init = work_fixup_init,
392 .fixup_activate = work_fixup_activate,
393 .fixup_free = work_fixup_free,
396 static inline void debug_work_activate(struct work_struct *work)
398 debug_object_activate(work, &work_debug_descr);
401 static inline void debug_work_deactivate(struct work_struct *work)
403 debug_object_deactivate(work, &work_debug_descr);
406 void __init_work(struct work_struct *work, int onstack)
409 debug_object_init_on_stack(work, &work_debug_descr);
411 debug_object_init(work, &work_debug_descr);
413 EXPORT_SYMBOL_GPL(__init_work);
415 void destroy_work_on_stack(struct work_struct *work)
417 debug_object_free(work, &work_debug_descr);
419 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
422 static inline void debug_work_activate(struct work_struct *work) { }
423 static inline void debug_work_deactivate(struct work_struct *work) { }
426 /* Serializes the accesses to the list of workqueues. */
427 static DEFINE_SPINLOCK(workqueue_lock);
428 static LIST_HEAD(workqueues);
429 static bool workqueue_freezing; /* W: have wqs started freezing? */
432 * The CPU and unbound standard worker pools. The unbound ones have
433 * POOL_DISASSOCIATED set, and their workers have WORKER_UNBOUND set.
435 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
436 cpu_std_worker_pools);
437 static struct worker_pool unbound_std_worker_pools[NR_STD_WORKER_POOLS];
439 /* idr of all pools */
440 static DEFINE_MUTEX(worker_pool_idr_mutex);
441 static DEFINE_IDR(worker_pool_idr);
443 static int worker_thread(void *__worker);
445 static struct worker_pool *std_worker_pools(int cpu)
447 if (cpu != WORK_CPU_UNBOUND)
448 return per_cpu(cpu_std_worker_pools, cpu);
450 return unbound_std_worker_pools;
453 static int std_worker_pool_pri(struct worker_pool *pool)
455 return pool - std_worker_pools(pool->cpu);
458 /* allocate ID and assign it to @pool */
459 static int worker_pool_assign_id(struct worker_pool *pool)
463 mutex_lock(&worker_pool_idr_mutex);
464 idr_pre_get(&worker_pool_idr, GFP_KERNEL);
465 ret = idr_get_new(&worker_pool_idr, pool, &pool->id);
466 mutex_unlock(&worker_pool_idr_mutex);
472 * Lookup worker_pool by id. The idr currently is built during boot and
473 * never modified. Don't worry about locking for now.
475 static struct worker_pool *worker_pool_by_id(int pool_id)
477 return idr_find(&worker_pool_idr, pool_id);
480 static struct worker_pool *get_std_worker_pool(int cpu, bool highpri)
482 struct worker_pool *pools = std_worker_pools(cpu);
484 return &pools[highpri];
487 static struct pool_workqueue *get_pwq(unsigned int cpu,
488 struct workqueue_struct *wq)
490 if (!(wq->flags & WQ_UNBOUND)) {
491 if (likely(cpu < nr_cpu_ids))
492 return per_cpu_ptr(wq->pool_wq.pcpu, cpu);
493 } else if (likely(cpu == WORK_CPU_UNBOUND))
494 return wq->pool_wq.single;
498 static unsigned int work_color_to_flags(int color)
500 return color << WORK_STRUCT_COLOR_SHIFT;
503 static int get_work_color(struct work_struct *work)
505 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
506 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
509 static int work_next_color(int color)
511 return (color + 1) % WORK_NR_COLORS;
515 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
516 * contain the pointer to the queued pwq. Once execution starts, the flag
517 * is cleared and the high bits contain OFFQ flags and pool ID.
519 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
520 * and clear_work_data() can be used to set the pwq, pool or clear
521 * work->data. These functions should only be called while the work is
522 * owned - ie. while the PENDING bit is set.
524 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
525 * corresponding to a work. Pool is available once the work has been
526 * queued anywhere after initialization until it is sync canceled. pwq is
527 * available only while the work item is queued.
529 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
530 * canceled. While being canceled, a work item may have its PENDING set
531 * but stay off timer and worklist for arbitrarily long and nobody should
532 * try to steal the PENDING bit.
534 static inline void set_work_data(struct work_struct *work, unsigned long data,
537 WARN_ON_ONCE(!work_pending(work));
538 atomic_long_set(&work->data, data | flags | work_static(work));
541 static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
542 unsigned long extra_flags)
544 set_work_data(work, (unsigned long)pwq,
545 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
548 static void set_work_pool_and_keep_pending(struct work_struct *work,
551 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
552 WORK_STRUCT_PENDING);
555 static void set_work_pool_and_clear_pending(struct work_struct *work,
559 * The following wmb is paired with the implied mb in
560 * test_and_set_bit(PENDING) and ensures all updates to @work made
561 * here are visible to and precede any updates by the next PENDING
565 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
568 static void clear_work_data(struct work_struct *work)
570 smp_wmb(); /* see set_work_pool_and_clear_pending() */
571 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
574 static struct pool_workqueue *get_work_pwq(struct work_struct *work)
576 unsigned long data = atomic_long_read(&work->data);
578 if (data & WORK_STRUCT_PWQ)
579 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
585 * get_work_pool - return the worker_pool a given work was associated with
586 * @work: the work item of interest
588 * Return the worker_pool @work was last associated with. %NULL if none.
590 static struct worker_pool *get_work_pool(struct work_struct *work)
592 unsigned long data = atomic_long_read(&work->data);
593 struct worker_pool *pool;
596 if (data & WORK_STRUCT_PWQ)
597 return ((struct pool_workqueue *)
598 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
600 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
601 if (pool_id == WORK_OFFQ_POOL_NONE)
604 pool = worker_pool_by_id(pool_id);
610 * get_work_pool_id - return the worker pool ID a given work is associated with
611 * @work: the work item of interest
613 * Return the worker_pool ID @work was last associated with.
614 * %WORK_OFFQ_POOL_NONE if none.
616 static int get_work_pool_id(struct work_struct *work)
618 unsigned long data = atomic_long_read(&work->data);
620 if (data & WORK_STRUCT_PWQ)
621 return ((struct pool_workqueue *)
622 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
624 return data >> WORK_OFFQ_POOL_SHIFT;
627 static void mark_work_canceling(struct work_struct *work)
629 unsigned long pool_id = get_work_pool_id(work);
631 pool_id <<= WORK_OFFQ_POOL_SHIFT;
632 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
635 static bool work_is_canceling(struct work_struct *work)
637 unsigned long data = atomic_long_read(&work->data);
639 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
643 * Policy functions. These define the policies on how the global worker
644 * pools are managed. Unless noted otherwise, these functions assume that
645 * they're being called with pool->lock held.
648 static bool __need_more_worker(struct worker_pool *pool)
650 return !atomic_read(&pool->nr_running);
654 * Need to wake up a worker? Called from anything but currently
657 * Note that, because unbound workers never contribute to nr_running, this
658 * function will always return %true for unbound pools as long as the
659 * worklist isn't empty.
661 static bool need_more_worker(struct worker_pool *pool)
663 return !list_empty(&pool->worklist) && __need_more_worker(pool);
666 /* Can I start working? Called from busy but !running workers. */
667 static bool may_start_working(struct worker_pool *pool)
669 return pool->nr_idle;
672 /* Do I need to keep working? Called from currently running workers. */
673 static bool keep_working(struct worker_pool *pool)
675 return !list_empty(&pool->worklist) &&
676 atomic_read(&pool->nr_running) <= 1;
679 /* Do we need a new worker? Called from manager. */
680 static bool need_to_create_worker(struct worker_pool *pool)
682 return need_more_worker(pool) && !may_start_working(pool);
685 /* Do I need to be the manager? */
686 static bool need_to_manage_workers(struct worker_pool *pool)
688 return need_to_create_worker(pool) ||
689 (pool->flags & POOL_MANAGE_WORKERS);
692 /* Do we have too many workers and should some go away? */
693 static bool too_many_workers(struct worker_pool *pool)
695 bool managing = pool->flags & POOL_MANAGING_WORKERS;
696 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
697 int nr_busy = pool->nr_workers - nr_idle;
700 * nr_idle and idle_list may disagree if idle rebinding is in
701 * progress. Never return %true if idle_list is empty.
703 if (list_empty(&pool->idle_list))
706 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
713 /* Return the first worker. Safe with preemption disabled */
714 static struct worker *first_worker(struct worker_pool *pool)
716 if (unlikely(list_empty(&pool->idle_list)))
719 return list_first_entry(&pool->idle_list, struct worker, entry);
723 * wake_up_worker - wake up an idle worker
724 * @pool: worker pool to wake worker from
726 * Wake up the first idle worker of @pool.
729 * spin_lock_irq(pool->lock).
731 static void wake_up_worker(struct worker_pool *pool)
733 struct worker *worker = first_worker(pool);
736 wake_up_process(worker->task);
740 * wq_worker_waking_up - a worker is waking up
741 * @task: task waking up
742 * @cpu: CPU @task is waking up to
744 * This function is called during try_to_wake_up() when a worker is
748 * spin_lock_irq(rq->lock)
750 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
752 struct worker *worker = kthread_data(task);
754 if (!(worker->flags & WORKER_NOT_RUNNING)) {
755 WARN_ON_ONCE(worker->pool->cpu != cpu);
756 atomic_inc(&worker->pool->nr_running);
761 * wq_worker_sleeping - a worker is going to sleep
762 * @task: task going to sleep
763 * @cpu: CPU in question, must be the current CPU number
765 * This function is called during schedule() when a busy worker is
766 * going to sleep. Worker on the same cpu can be woken up by
767 * returning pointer to its task.
770 * spin_lock_irq(rq->lock)
773 * Worker task on @cpu to wake up, %NULL if none.
775 struct task_struct *wq_worker_sleeping(struct task_struct *task,
778 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
779 struct worker_pool *pool;
782 * Rescuers, which may not have all the fields set up like normal
783 * workers, also reach here, let's not access anything before
784 * checking NOT_RUNNING.
786 if (worker->flags & WORKER_NOT_RUNNING)
791 /* this can only happen on the local cpu */
792 if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
796 * The counterpart of the following dec_and_test, implied mb,
797 * worklist not empty test sequence is in insert_work().
798 * Please read comment there.
800 * NOT_RUNNING is clear. This means that we're bound to and
801 * running on the local cpu w/ rq lock held and preemption
802 * disabled, which in turn means that none else could be
803 * manipulating idle_list, so dereferencing idle_list without pool
806 if (atomic_dec_and_test(&pool->nr_running) &&
807 !list_empty(&pool->worklist))
808 to_wakeup = first_worker(pool);
809 return to_wakeup ? to_wakeup->task : NULL;
813 * worker_set_flags - set worker flags and adjust nr_running accordingly
815 * @flags: flags to set
816 * @wakeup: wakeup an idle worker if necessary
818 * Set @flags in @worker->flags and adjust nr_running accordingly. If
819 * nr_running becomes zero and @wakeup is %true, an idle worker is
823 * spin_lock_irq(pool->lock)
825 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
828 struct worker_pool *pool = worker->pool;
830 WARN_ON_ONCE(worker->task != current);
833 * If transitioning into NOT_RUNNING, adjust nr_running and
834 * wake up an idle worker as necessary if requested by
837 if ((flags & WORKER_NOT_RUNNING) &&
838 !(worker->flags & WORKER_NOT_RUNNING)) {
840 if (atomic_dec_and_test(&pool->nr_running) &&
841 !list_empty(&pool->worklist))
842 wake_up_worker(pool);
844 atomic_dec(&pool->nr_running);
847 worker->flags |= flags;
851 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
853 * @flags: flags to clear
855 * Clear @flags in @worker->flags and adjust nr_running accordingly.
858 * spin_lock_irq(pool->lock)
860 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
862 struct worker_pool *pool = worker->pool;
863 unsigned int oflags = worker->flags;
865 WARN_ON_ONCE(worker->task != current);
867 worker->flags &= ~flags;
870 * If transitioning out of NOT_RUNNING, increment nr_running. Note
871 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
872 * of multiple flags, not a single flag.
874 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
875 if (!(worker->flags & WORKER_NOT_RUNNING))
876 atomic_inc(&pool->nr_running);
880 * find_worker_executing_work - find worker which is executing a work
881 * @pool: pool of interest
882 * @work: work to find worker for
884 * Find a worker which is executing @work on @pool by searching
885 * @pool->busy_hash which is keyed by the address of @work. For a worker
886 * to match, its current execution should match the address of @work and
887 * its work function. This is to avoid unwanted dependency between
888 * unrelated work executions through a work item being recycled while still
891 * This is a bit tricky. A work item may be freed once its execution
892 * starts and nothing prevents the freed area from being recycled for
893 * another work item. If the same work item address ends up being reused
894 * before the original execution finishes, workqueue will identify the
895 * recycled work item as currently executing and make it wait until the
896 * current execution finishes, introducing an unwanted dependency.
898 * This function checks the work item address, work function and workqueue
899 * to avoid false positives. Note that this isn't complete as one may
900 * construct a work function which can introduce dependency onto itself
901 * through a recycled work item. Well, if somebody wants to shoot oneself
902 * in the foot that badly, there's only so much we can do, and if such
903 * deadlock actually occurs, it should be easy to locate the culprit work
907 * spin_lock_irq(pool->lock).
910 * Pointer to worker which is executing @work if found, NULL
913 static struct worker *find_worker_executing_work(struct worker_pool *pool,
914 struct work_struct *work)
916 struct worker *worker;
918 hash_for_each_possible(pool->busy_hash, worker, hentry,
920 if (worker->current_work == work &&
921 worker->current_func == work->func)
928 * move_linked_works - move linked works to a list
929 * @work: start of series of works to be scheduled
930 * @head: target list to append @work to
931 * @nextp: out paramter for nested worklist walking
933 * Schedule linked works starting from @work to @head. Work series to
934 * be scheduled starts at @work and includes any consecutive work with
935 * WORK_STRUCT_LINKED set in its predecessor.
937 * If @nextp is not NULL, it's updated to point to the next work of
938 * the last scheduled work. This allows move_linked_works() to be
939 * nested inside outer list_for_each_entry_safe().
942 * spin_lock_irq(pool->lock).
944 static void move_linked_works(struct work_struct *work, struct list_head *head,
945 struct work_struct **nextp)
947 struct work_struct *n;
950 * Linked worklist will always end before the end of the list,
951 * use NULL for list head.
953 list_for_each_entry_safe_from(work, n, NULL, entry) {
954 list_move_tail(&work->entry, head);
955 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
960 * If we're already inside safe list traversal and have moved
961 * multiple works to the scheduled queue, the next position
962 * needs to be updated.
968 static void pwq_activate_delayed_work(struct work_struct *work)
970 struct pool_workqueue *pwq = get_work_pwq(work);
972 trace_workqueue_activate_work(work);
973 move_linked_works(work, &pwq->pool->worklist, NULL);
974 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
978 static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
980 struct work_struct *work = list_first_entry(&pwq->delayed_works,
981 struct work_struct, entry);
983 pwq_activate_delayed_work(work);
987 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
988 * @pwq: pwq of interest
989 * @color: color of work which left the queue
991 * A work either has completed or is removed from pending queue,
992 * decrement nr_in_flight of its pwq and handle workqueue flushing.
995 * spin_lock_irq(pool->lock).
997 static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
999 /* ignore uncolored works */
1000 if (color == WORK_NO_COLOR)
1003 pwq->nr_in_flight[color]--;
1006 if (!list_empty(&pwq->delayed_works)) {
1007 /* one down, submit a delayed one */
1008 if (pwq->nr_active < pwq->max_active)
1009 pwq_activate_first_delayed(pwq);
1012 /* is flush in progress and are we at the flushing tip? */
1013 if (likely(pwq->flush_color != color))
1016 /* are there still in-flight works? */
1017 if (pwq->nr_in_flight[color])
1020 /* this pwq is done, clear flush_color */
1021 pwq->flush_color = -1;
1024 * If this was the last pwq, wake up the first flusher. It
1025 * will handle the rest.
1027 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1028 complete(&pwq->wq->first_flusher->done);
1032 * try_to_grab_pending - steal work item from worklist and disable irq
1033 * @work: work item to steal
1034 * @is_dwork: @work is a delayed_work
1035 * @flags: place to store irq state
1037 * Try to grab PENDING bit of @work. This function can handle @work in any
1038 * stable state - idle, on timer or on worklist. Return values are
1040 * 1 if @work was pending and we successfully stole PENDING
1041 * 0 if @work was idle and we claimed PENDING
1042 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1043 * -ENOENT if someone else is canceling @work, this state may persist
1044 * for arbitrarily long
1046 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1047 * interrupted while holding PENDING and @work off queue, irq must be
1048 * disabled on entry. This, combined with delayed_work->timer being
1049 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1051 * On successful return, >= 0, irq is disabled and the caller is
1052 * responsible for releasing it using local_irq_restore(*@flags).
1054 * This function is safe to call from any context including IRQ handler.
1056 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1057 unsigned long *flags)
1059 struct worker_pool *pool;
1060 struct pool_workqueue *pwq;
1062 local_irq_save(*flags);
1064 /* try to steal the timer if it exists */
1066 struct delayed_work *dwork = to_delayed_work(work);
1069 * dwork->timer is irqsafe. If del_timer() fails, it's
1070 * guaranteed that the timer is not queued anywhere and not
1071 * running on the local CPU.
1073 if (likely(del_timer(&dwork->timer)))
1077 /* try to claim PENDING the normal way */
1078 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1082 * The queueing is in progress, or it is already queued. Try to
1083 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1085 pool = get_work_pool(work);
1089 spin_lock(&pool->lock);
1091 * work->data is guaranteed to point to pwq only while the work
1092 * item is queued on pwq->wq, and both updating work->data to point
1093 * to pwq on queueing and to pool on dequeueing are done under
1094 * pwq->pool->lock. This in turn guarantees that, if work->data
1095 * points to pwq which is associated with a locked pool, the work
1096 * item is currently queued on that pool.
1098 pwq = get_work_pwq(work);
1099 if (pwq && pwq->pool == pool) {
1100 debug_work_deactivate(work);
1103 * A delayed work item cannot be grabbed directly because
1104 * it might have linked NO_COLOR work items which, if left
1105 * on the delayed_list, will confuse pwq->nr_active
1106 * management later on and cause stall. Make sure the work
1107 * item is activated before grabbing.
1109 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1110 pwq_activate_delayed_work(work);
1112 list_del_init(&work->entry);
1113 pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1115 /* work->data points to pwq iff queued, point to pool */
1116 set_work_pool_and_keep_pending(work, pool->id);
1118 spin_unlock(&pool->lock);
1121 spin_unlock(&pool->lock);
1123 local_irq_restore(*flags);
1124 if (work_is_canceling(work))
1131 * insert_work - insert a work into a pool
1132 * @pwq: pwq @work belongs to
1133 * @work: work to insert
1134 * @head: insertion point
1135 * @extra_flags: extra WORK_STRUCT_* flags to set
1137 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1138 * work_struct flags.
1141 * spin_lock_irq(pool->lock).
1143 static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1144 struct list_head *head, unsigned int extra_flags)
1146 struct worker_pool *pool = pwq->pool;
1148 /* we own @work, set data and link */
1149 set_work_pwq(work, pwq, extra_flags);
1150 list_add_tail(&work->entry, head);
1153 * Ensure either worker_sched_deactivated() sees the above
1154 * list_add_tail() or we see zero nr_running to avoid workers
1155 * lying around lazily while there are works to be processed.
1159 if (__need_more_worker(pool))
1160 wake_up_worker(pool);
1164 * Test whether @work is being queued from another work executing on the
1167 static bool is_chained_work(struct workqueue_struct *wq)
1169 struct worker *worker;
1171 worker = current_wq_worker();
1173 * Return %true iff I'm a worker execuing a work item on @wq. If
1174 * I'm @worker, it's safe to dereference it without locking.
1176 return worker && worker->current_pwq->wq == wq;
1179 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1180 struct work_struct *work)
1182 struct pool_workqueue *pwq;
1183 struct list_head *worklist;
1184 unsigned int work_flags;
1185 unsigned int req_cpu = cpu;
1188 * While a work item is PENDING && off queue, a task trying to
1189 * steal the PENDING will busy-loop waiting for it to either get
1190 * queued or lose PENDING. Grabbing PENDING and queueing should
1191 * happen with IRQ disabled.
1193 WARN_ON_ONCE(!irqs_disabled());
1195 debug_work_activate(work);
1197 /* if dying, only works from the same workqueue are allowed */
1198 if (unlikely(wq->flags & WQ_DRAINING) &&
1199 WARN_ON_ONCE(!is_chained_work(wq)))
1202 /* determine the pwq to use */
1203 if (!(wq->flags & WQ_UNBOUND)) {
1204 struct worker_pool *last_pool;
1206 if (cpu == WORK_CPU_UNBOUND)
1207 cpu = raw_smp_processor_id();
1210 * It's multi cpu. If @work was previously on a different
1211 * cpu, it might still be running there, in which case the
1212 * work needs to be queued on that cpu to guarantee
1215 pwq = get_pwq(cpu, wq);
1216 last_pool = get_work_pool(work);
1218 if (last_pool && last_pool != pwq->pool) {
1219 struct worker *worker;
1221 spin_lock(&last_pool->lock);
1223 worker = find_worker_executing_work(last_pool, work);
1225 if (worker && worker->current_pwq->wq == wq) {
1226 pwq = get_pwq(last_pool->cpu, wq);
1228 /* meh... not running there, queue here */
1229 spin_unlock(&last_pool->lock);
1230 spin_lock(&pwq->pool->lock);
1233 spin_lock(&pwq->pool->lock);
1236 pwq = get_pwq(WORK_CPU_UNBOUND, wq);
1237 spin_lock(&pwq->pool->lock);
1240 /* pwq determined, queue */
1241 trace_workqueue_queue_work(req_cpu, pwq, work);
1243 if (WARN_ON(!list_empty(&work->entry))) {
1244 spin_unlock(&pwq->pool->lock);
1248 pwq->nr_in_flight[pwq->work_color]++;
1249 work_flags = work_color_to_flags(pwq->work_color);
1251 if (likely(pwq->nr_active < pwq->max_active)) {
1252 trace_workqueue_activate_work(work);
1254 worklist = &pwq->pool->worklist;
1256 work_flags |= WORK_STRUCT_DELAYED;
1257 worklist = &pwq->delayed_works;
1260 insert_work(pwq, work, worklist, work_flags);
1262 spin_unlock(&pwq->pool->lock);
1266 * queue_work_on - queue work on specific cpu
1267 * @cpu: CPU number to execute work on
1268 * @wq: workqueue to use
1269 * @work: work to queue
1271 * Returns %false if @work was already on a queue, %true otherwise.
1273 * We queue the work to a specific CPU, the caller must ensure it
1276 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1277 struct work_struct *work)
1280 unsigned long flags;
1282 local_irq_save(flags);
1284 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1285 __queue_work(cpu, wq, work);
1289 local_irq_restore(flags);
1292 EXPORT_SYMBOL_GPL(queue_work_on);
1295 * queue_work - queue work on a workqueue
1296 * @wq: workqueue to use
1297 * @work: work to queue
1299 * Returns %false if @work was already on a queue, %true otherwise.
1301 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1302 * it can be processed by another CPU.
1304 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1306 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1308 EXPORT_SYMBOL_GPL(queue_work);
1310 void delayed_work_timer_fn(unsigned long __data)
1312 struct delayed_work *dwork = (struct delayed_work *)__data;
1314 /* should have been called from irqsafe timer with irq already off */
1315 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1317 EXPORT_SYMBOL(delayed_work_timer_fn);
1319 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1320 struct delayed_work *dwork, unsigned long delay)
1322 struct timer_list *timer = &dwork->timer;
1323 struct work_struct *work = &dwork->work;
1325 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1326 timer->data != (unsigned long)dwork);
1327 WARN_ON_ONCE(timer_pending(timer));
1328 WARN_ON_ONCE(!list_empty(&work->entry));
1331 * If @delay is 0, queue @dwork->work immediately. This is for
1332 * both optimization and correctness. The earliest @timer can
1333 * expire is on the closest next tick and delayed_work users depend
1334 * on that there's no such delay when @delay is 0.
1337 __queue_work(cpu, wq, &dwork->work);
1341 timer_stats_timer_set_start_info(&dwork->timer);
1345 timer->expires = jiffies + delay;
1347 if (unlikely(cpu != WORK_CPU_UNBOUND))
1348 add_timer_on(timer, cpu);
1354 * queue_delayed_work_on - queue work on specific CPU after delay
1355 * @cpu: CPU number to execute work on
1356 * @wq: workqueue to use
1357 * @dwork: work to queue
1358 * @delay: number of jiffies to wait before queueing
1360 * Returns %false if @work was already on a queue, %true otherwise. If
1361 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1364 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1365 struct delayed_work *dwork, unsigned long delay)
1367 struct work_struct *work = &dwork->work;
1369 unsigned long flags;
1371 /* read the comment in __queue_work() */
1372 local_irq_save(flags);
1374 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1375 __queue_delayed_work(cpu, wq, dwork, delay);
1379 local_irq_restore(flags);
1382 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1385 * queue_delayed_work - queue work on a workqueue after delay
1386 * @wq: workqueue to use
1387 * @dwork: delayable work to queue
1388 * @delay: number of jiffies to wait before queueing
1390 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1392 bool queue_delayed_work(struct workqueue_struct *wq,
1393 struct delayed_work *dwork, unsigned long delay)
1395 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1397 EXPORT_SYMBOL_GPL(queue_delayed_work);
1400 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1401 * @cpu: CPU number to execute work on
1402 * @wq: workqueue to use
1403 * @dwork: work to queue
1404 * @delay: number of jiffies to wait before queueing
1406 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1407 * modify @dwork's timer so that it expires after @delay. If @delay is
1408 * zero, @work is guaranteed to be scheduled immediately regardless of its
1411 * Returns %false if @dwork was idle and queued, %true if @dwork was
1412 * pending and its timer was modified.
1414 * This function is safe to call from any context including IRQ handler.
1415 * See try_to_grab_pending() for details.
1417 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1418 struct delayed_work *dwork, unsigned long delay)
1420 unsigned long flags;
1424 ret = try_to_grab_pending(&dwork->work, true, &flags);
1425 } while (unlikely(ret == -EAGAIN));
1427 if (likely(ret >= 0)) {
1428 __queue_delayed_work(cpu, wq, dwork, delay);
1429 local_irq_restore(flags);
1432 /* -ENOENT from try_to_grab_pending() becomes %true */
1435 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1438 * mod_delayed_work - modify delay of or queue a delayed work
1439 * @wq: workqueue to use
1440 * @dwork: work to queue
1441 * @delay: number of jiffies to wait before queueing
1443 * mod_delayed_work_on() on local CPU.
1445 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1446 unsigned long delay)
1448 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1450 EXPORT_SYMBOL_GPL(mod_delayed_work);
1453 * worker_enter_idle - enter idle state
1454 * @worker: worker which is entering idle state
1456 * @worker is entering idle state. Update stats and idle timer if
1460 * spin_lock_irq(pool->lock).
1462 static void worker_enter_idle(struct worker *worker)
1464 struct worker_pool *pool = worker->pool;
1466 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1467 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1468 (worker->hentry.next || worker->hentry.pprev)))
1471 /* can't use worker_set_flags(), also called from start_worker() */
1472 worker->flags |= WORKER_IDLE;
1474 worker->last_active = jiffies;
1476 /* idle_list is LIFO */
1477 list_add(&worker->entry, &pool->idle_list);
1479 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1480 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1483 * Sanity check nr_running. Because wq_unbind_fn() releases
1484 * pool->lock between setting %WORKER_UNBOUND and zapping
1485 * nr_running, the warning may trigger spuriously. Check iff
1486 * unbind is not in progress.
1488 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1489 pool->nr_workers == pool->nr_idle &&
1490 atomic_read(&pool->nr_running));
1494 * worker_leave_idle - leave idle state
1495 * @worker: worker which is leaving idle state
1497 * @worker is leaving idle state. Update stats.
1500 * spin_lock_irq(pool->lock).
1502 static void worker_leave_idle(struct worker *worker)
1504 struct worker_pool *pool = worker->pool;
1506 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1508 worker_clr_flags(worker, WORKER_IDLE);
1510 list_del_init(&worker->entry);
1514 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1515 * @pool: target worker_pool
1517 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1519 * Works which are scheduled while the cpu is online must at least be
1520 * scheduled to a worker which is bound to the cpu so that if they are
1521 * flushed from cpu callbacks while cpu is going down, they are
1522 * guaranteed to execute on the cpu.
1524 * This function is to be used by unbound workers and rescuers to bind
1525 * themselves to the target cpu and may race with cpu going down or
1526 * coming online. kthread_bind() can't be used because it may put the
1527 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1528 * verbatim as it's best effort and blocking and pool may be
1529 * [dis]associated in the meantime.
1531 * This function tries set_cpus_allowed() and locks pool and verifies the
1532 * binding against %POOL_DISASSOCIATED which is set during
1533 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1534 * enters idle state or fetches works without dropping lock, it can
1535 * guarantee the scheduling requirement described in the first paragraph.
1538 * Might sleep. Called without any lock but returns with pool->lock
1542 * %true if the associated pool is online (@worker is successfully
1543 * bound), %false if offline.
1545 static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1546 __acquires(&pool->lock)
1550 * The following call may fail, succeed or succeed
1551 * without actually migrating the task to the cpu if
1552 * it races with cpu hotunplug operation. Verify
1553 * against POOL_DISASSOCIATED.
1555 if (!(pool->flags & POOL_DISASSOCIATED))
1556 set_cpus_allowed_ptr(current, get_cpu_mask(pool->cpu));
1558 spin_lock_irq(&pool->lock);
1559 if (pool->flags & POOL_DISASSOCIATED)
1561 if (task_cpu(current) == pool->cpu &&
1562 cpumask_equal(¤t->cpus_allowed,
1563 get_cpu_mask(pool->cpu)))
1565 spin_unlock_irq(&pool->lock);
1568 * We've raced with CPU hot[un]plug. Give it a breather
1569 * and retry migration. cond_resched() is required here;
1570 * otherwise, we might deadlock against cpu_stop trying to
1571 * bring down the CPU on non-preemptive kernel.
1579 * Rebind an idle @worker to its CPU. worker_thread() will test
1580 * list_empty(@worker->entry) before leaving idle and call this function.
1582 static void idle_worker_rebind(struct worker *worker)
1584 /* CPU may go down again inbetween, clear UNBOUND only on success */
1585 if (worker_maybe_bind_and_lock(worker->pool))
1586 worker_clr_flags(worker, WORKER_UNBOUND);
1588 /* rebind complete, become available again */
1589 list_add(&worker->entry, &worker->pool->idle_list);
1590 spin_unlock_irq(&worker->pool->lock);
1594 * Function for @worker->rebind.work used to rebind unbound busy workers to
1595 * the associated cpu which is coming back online. This is scheduled by
1596 * cpu up but can race with other cpu hotplug operations and may be
1597 * executed twice without intervening cpu down.
1599 static void busy_worker_rebind_fn(struct work_struct *work)
1601 struct worker *worker = container_of(work, struct worker, rebind_work);
1603 if (worker_maybe_bind_and_lock(worker->pool))
1604 worker_clr_flags(worker, WORKER_UNBOUND);
1606 spin_unlock_irq(&worker->pool->lock);
1610 * rebind_workers - rebind all workers of a pool to the associated CPU
1611 * @pool: pool of interest
1613 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1614 * is different for idle and busy ones.
1616 * Idle ones will be removed from the idle_list and woken up. They will
1617 * add themselves back after completing rebind. This ensures that the
1618 * idle_list doesn't contain any unbound workers when re-bound busy workers
1619 * try to perform local wake-ups for concurrency management.
1621 * Busy workers can rebind after they finish their current work items.
1622 * Queueing the rebind work item at the head of the scheduled list is
1623 * enough. Note that nr_running will be properly bumped as busy workers
1626 * On return, all non-manager workers are scheduled for rebind - see
1627 * manage_workers() for the manager special case. Any idle worker
1628 * including the manager will not appear on @idle_list until rebind is
1629 * complete, making local wake-ups safe.
1631 static void rebind_workers(struct worker_pool *pool)
1633 struct worker *worker, *n;
1636 lockdep_assert_held(&pool->assoc_mutex);
1637 lockdep_assert_held(&pool->lock);
1639 /* dequeue and kick idle ones */
1640 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1642 * idle workers should be off @pool->idle_list until rebind
1643 * is complete to avoid receiving premature local wake-ups.
1645 list_del_init(&worker->entry);
1648 * worker_thread() will see the above dequeuing and call
1649 * idle_worker_rebind().
1651 wake_up_process(worker->task);
1654 /* rebind busy workers */
1655 for_each_busy_worker(worker, i, pool) {
1656 struct work_struct *rebind_work = &worker->rebind_work;
1657 struct workqueue_struct *wq;
1659 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1660 work_data_bits(rebind_work)))
1663 debug_work_activate(rebind_work);
1666 * wq doesn't really matter but let's keep @worker->pool
1667 * and @pwq->pool consistent for sanity.
1669 if (std_worker_pool_pri(worker->pool))
1670 wq = system_highpri_wq;
1674 insert_work(get_pwq(pool->cpu, wq), rebind_work,
1675 worker->scheduled.next,
1676 work_color_to_flags(WORK_NO_COLOR));
1680 static struct worker *alloc_worker(void)
1682 struct worker *worker;
1684 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1686 INIT_LIST_HEAD(&worker->entry);
1687 INIT_LIST_HEAD(&worker->scheduled);
1688 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1689 /* on creation a worker is in !idle && prep state */
1690 worker->flags = WORKER_PREP;
1696 * create_worker - create a new workqueue worker
1697 * @pool: pool the new worker will belong to
1699 * Create a new worker which is bound to @pool. The returned worker
1700 * can be started by calling start_worker() or destroyed using
1704 * Might sleep. Does GFP_KERNEL allocations.
1707 * Pointer to the newly created worker.
1709 static struct worker *create_worker(struct worker_pool *pool)
1711 const char *pri = std_worker_pool_pri(pool) ? "H" : "";
1712 struct worker *worker = NULL;
1715 spin_lock_irq(&pool->lock);
1716 while (ida_get_new(&pool->worker_ida, &id)) {
1717 spin_unlock_irq(&pool->lock);
1718 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1720 spin_lock_irq(&pool->lock);
1722 spin_unlock_irq(&pool->lock);
1724 worker = alloc_worker();
1728 worker->pool = pool;
1731 if (pool->cpu != WORK_CPU_UNBOUND)
1732 worker->task = kthread_create_on_node(worker_thread,
1733 worker, cpu_to_node(pool->cpu),
1734 "kworker/%u:%d%s", pool->cpu, id, pri);
1736 worker->task = kthread_create(worker_thread, worker,
1737 "kworker/u:%d%s", id, pri);
1738 if (IS_ERR(worker->task))
1741 if (std_worker_pool_pri(pool))
1742 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1745 * Determine CPU binding of the new worker depending on
1746 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1747 * flag remains stable across this function. See the comments
1748 * above the flag definition for details.
1750 * As an unbound worker may later become a regular one if CPU comes
1751 * online, make sure every worker has %PF_THREAD_BOUND set.
1753 if (!(pool->flags & POOL_DISASSOCIATED)) {
1754 kthread_bind(worker->task, pool->cpu);
1756 worker->task->flags |= PF_THREAD_BOUND;
1757 worker->flags |= WORKER_UNBOUND;
1763 spin_lock_irq(&pool->lock);
1764 ida_remove(&pool->worker_ida, id);
1765 spin_unlock_irq(&pool->lock);
1772 * start_worker - start a newly created worker
1773 * @worker: worker to start
1775 * Make the pool aware of @worker and start it.
1778 * spin_lock_irq(pool->lock).
1780 static void start_worker(struct worker *worker)
1782 worker->flags |= WORKER_STARTED;
1783 worker->pool->nr_workers++;
1784 worker_enter_idle(worker);
1785 wake_up_process(worker->task);
1789 * destroy_worker - destroy a workqueue worker
1790 * @worker: worker to be destroyed
1792 * Destroy @worker and adjust @pool stats accordingly.
1795 * spin_lock_irq(pool->lock) which is released and regrabbed.
1797 static void destroy_worker(struct worker *worker)
1799 struct worker_pool *pool = worker->pool;
1800 int id = worker->id;
1802 /* sanity check frenzy */
1803 if (WARN_ON(worker->current_work) ||
1804 WARN_ON(!list_empty(&worker->scheduled)))
1807 if (worker->flags & WORKER_STARTED)
1809 if (worker->flags & WORKER_IDLE)
1812 list_del_init(&worker->entry);
1813 worker->flags |= WORKER_DIE;
1815 spin_unlock_irq(&pool->lock);
1817 kthread_stop(worker->task);
1820 spin_lock_irq(&pool->lock);
1821 ida_remove(&pool->worker_ida, id);
1824 static void idle_worker_timeout(unsigned long __pool)
1826 struct worker_pool *pool = (void *)__pool;
1828 spin_lock_irq(&pool->lock);
1830 if (too_many_workers(pool)) {
1831 struct worker *worker;
1832 unsigned long expires;
1834 /* idle_list is kept in LIFO order, check the last one */
1835 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1836 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1838 if (time_before(jiffies, expires))
1839 mod_timer(&pool->idle_timer, expires);
1841 /* it's been idle for too long, wake up manager */
1842 pool->flags |= POOL_MANAGE_WORKERS;
1843 wake_up_worker(pool);
1847 spin_unlock_irq(&pool->lock);
1850 static bool send_mayday(struct work_struct *work)
1852 struct pool_workqueue *pwq = get_work_pwq(work);
1853 struct workqueue_struct *wq = pwq->wq;
1856 if (!(wq->flags & WQ_RESCUER))
1859 /* mayday mayday mayday */
1860 cpu = pwq->pool->cpu;
1861 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1862 if (cpu == WORK_CPU_UNBOUND)
1864 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1865 wake_up_process(wq->rescuer->task);
1869 static void pool_mayday_timeout(unsigned long __pool)
1871 struct worker_pool *pool = (void *)__pool;
1872 struct work_struct *work;
1874 spin_lock_irq(&pool->lock);
1876 if (need_to_create_worker(pool)) {
1878 * We've been trying to create a new worker but
1879 * haven't been successful. We might be hitting an
1880 * allocation deadlock. Send distress signals to
1883 list_for_each_entry(work, &pool->worklist, entry)
1887 spin_unlock_irq(&pool->lock);
1889 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1893 * maybe_create_worker - create a new worker if necessary
1894 * @pool: pool to create a new worker for
1896 * Create a new worker for @pool if necessary. @pool is guaranteed to
1897 * have at least one idle worker on return from this function. If
1898 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1899 * sent to all rescuers with works scheduled on @pool to resolve
1900 * possible allocation deadlock.
1902 * On return, need_to_create_worker() is guaranteed to be false and
1903 * may_start_working() true.
1906 * spin_lock_irq(pool->lock) which may be released and regrabbed
1907 * multiple times. Does GFP_KERNEL allocations. Called only from
1911 * false if no action was taken and pool->lock stayed locked, true
1914 static bool maybe_create_worker(struct worker_pool *pool)
1915 __releases(&pool->lock)
1916 __acquires(&pool->lock)
1918 if (!need_to_create_worker(pool))
1921 spin_unlock_irq(&pool->lock);
1923 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1924 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1927 struct worker *worker;
1929 worker = create_worker(pool);
1931 del_timer_sync(&pool->mayday_timer);
1932 spin_lock_irq(&pool->lock);
1933 start_worker(worker);
1934 if (WARN_ON_ONCE(need_to_create_worker(pool)))
1939 if (!need_to_create_worker(pool))
1942 __set_current_state(TASK_INTERRUPTIBLE);
1943 schedule_timeout(CREATE_COOLDOWN);
1945 if (!need_to_create_worker(pool))
1949 del_timer_sync(&pool->mayday_timer);
1950 spin_lock_irq(&pool->lock);
1951 if (need_to_create_worker(pool))
1957 * maybe_destroy_worker - destroy workers which have been idle for a while
1958 * @pool: pool to destroy workers for
1960 * Destroy @pool workers which have been idle for longer than
1961 * IDLE_WORKER_TIMEOUT.
1964 * spin_lock_irq(pool->lock) which may be released and regrabbed
1965 * multiple times. Called only from manager.
1968 * false if no action was taken and pool->lock stayed locked, true
1971 static bool maybe_destroy_workers(struct worker_pool *pool)
1975 while (too_many_workers(pool)) {
1976 struct worker *worker;
1977 unsigned long expires;
1979 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1980 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1982 if (time_before(jiffies, expires)) {
1983 mod_timer(&pool->idle_timer, expires);
1987 destroy_worker(worker);
1995 * manage_workers - manage worker pool
1998 * Assume the manager role and manage the worker pool @worker belongs
1999 * to. At any given time, there can be only zero or one manager per
2000 * pool. The exclusion is handled automatically by this function.
2002 * The caller can safely start processing works on false return. On
2003 * true return, it's guaranteed that need_to_create_worker() is false
2004 * and may_start_working() is true.
2007 * spin_lock_irq(pool->lock) which may be released and regrabbed
2008 * multiple times. Does GFP_KERNEL allocations.
2011 * spin_lock_irq(pool->lock) which may be released and regrabbed
2012 * multiple times. Does GFP_KERNEL allocations.
2014 static bool manage_workers(struct worker *worker)
2016 struct worker_pool *pool = worker->pool;
2019 if (pool->flags & POOL_MANAGING_WORKERS)
2022 pool->flags |= POOL_MANAGING_WORKERS;
2025 * To simplify both worker management and CPU hotplug, hold off
2026 * management while hotplug is in progress. CPU hotplug path can't
2027 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2028 * lead to idle worker depletion (all become busy thinking someone
2029 * else is managing) which in turn can result in deadlock under
2030 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2031 * manager against CPU hotplug.
2033 * assoc_mutex would always be free unless CPU hotplug is in
2034 * progress. trylock first without dropping @pool->lock.
2036 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2037 spin_unlock_irq(&pool->lock);
2038 mutex_lock(&pool->assoc_mutex);
2040 * CPU hotplug could have happened while we were waiting
2041 * for assoc_mutex. Hotplug itself can't handle us
2042 * because manager isn't either on idle or busy list, and
2043 * @pool's state and ours could have deviated.
2045 * As hotplug is now excluded via assoc_mutex, we can
2046 * simply try to bind. It will succeed or fail depending
2047 * on @pool's current state. Try it and adjust
2048 * %WORKER_UNBOUND accordingly.
2050 if (worker_maybe_bind_and_lock(pool))
2051 worker->flags &= ~WORKER_UNBOUND;
2053 worker->flags |= WORKER_UNBOUND;
2058 pool->flags &= ~POOL_MANAGE_WORKERS;
2061 * Destroy and then create so that may_start_working() is true
2064 ret |= maybe_destroy_workers(pool);
2065 ret |= maybe_create_worker(pool);
2067 pool->flags &= ~POOL_MANAGING_WORKERS;
2068 mutex_unlock(&pool->assoc_mutex);
2073 * process_one_work - process single work
2075 * @work: work to process
2077 * Process @work. This function contains all the logics necessary to
2078 * process a single work including synchronization against and
2079 * interaction with other workers on the same cpu, queueing and
2080 * flushing. As long as context requirement is met, any worker can
2081 * call this function to process a work.
2084 * spin_lock_irq(pool->lock) which is released and regrabbed.
2086 static void process_one_work(struct worker *worker, struct work_struct *work)
2087 __releases(&pool->lock)
2088 __acquires(&pool->lock)
2090 struct pool_workqueue *pwq = get_work_pwq(work);
2091 struct worker_pool *pool = worker->pool;
2092 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2094 struct worker *collision;
2095 #ifdef CONFIG_LOCKDEP
2097 * It is permissible to free the struct work_struct from
2098 * inside the function that is called from it, this we need to
2099 * take into account for lockdep too. To avoid bogus "held
2100 * lock freed" warnings as well as problems when looking into
2101 * work->lockdep_map, make a copy and use that here.
2103 struct lockdep_map lockdep_map;
2105 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2108 * Ensure we're on the correct CPU. DISASSOCIATED test is
2109 * necessary to avoid spurious warnings from rescuers servicing the
2110 * unbound or a disassociated pool.
2112 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2113 !(pool->flags & POOL_DISASSOCIATED) &&
2114 raw_smp_processor_id() != pool->cpu);
2117 * A single work shouldn't be executed concurrently by
2118 * multiple workers on a single cpu. Check whether anyone is
2119 * already processing the work. If so, defer the work to the
2120 * currently executing one.
2122 collision = find_worker_executing_work(pool, work);
2123 if (unlikely(collision)) {
2124 move_linked_works(work, &collision->scheduled, NULL);
2128 /* claim and dequeue */
2129 debug_work_deactivate(work);
2130 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2131 worker->current_work = work;
2132 worker->current_func = work->func;
2133 worker->current_pwq = pwq;
2134 work_color = get_work_color(work);
2136 list_del_init(&work->entry);
2139 * CPU intensive works don't participate in concurrency
2140 * management. They're the scheduler's responsibility.
2142 if (unlikely(cpu_intensive))
2143 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2146 * Unbound pool isn't concurrency managed and work items should be
2147 * executed ASAP. Wake up another worker if necessary.
2149 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2150 wake_up_worker(pool);
2153 * Record the last pool and clear PENDING which should be the last
2154 * update to @work. Also, do this inside @pool->lock so that
2155 * PENDING and queued state changes happen together while IRQ is
2158 set_work_pool_and_clear_pending(work, pool->id);
2160 spin_unlock_irq(&pool->lock);
2162 lock_map_acquire_read(&pwq->wq->lockdep_map);
2163 lock_map_acquire(&lockdep_map);
2164 trace_workqueue_execute_start(work);
2165 worker->current_func(work);
2167 * While we must be careful to not use "work" after this, the trace
2168 * point will only record its address.
2170 trace_workqueue_execute_end(work);
2171 lock_map_release(&lockdep_map);
2172 lock_map_release(&pwq->wq->lockdep_map);
2174 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2175 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2176 " last function: %pf\n",
2177 current->comm, preempt_count(), task_pid_nr(current),
2178 worker->current_func);
2179 debug_show_held_locks(current);
2183 spin_lock_irq(&pool->lock);
2185 /* clear cpu intensive status */
2186 if (unlikely(cpu_intensive))
2187 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2189 /* we're done with it, release */
2190 hash_del(&worker->hentry);
2191 worker->current_work = NULL;
2192 worker->current_func = NULL;
2193 worker->current_pwq = NULL;
2194 pwq_dec_nr_in_flight(pwq, work_color);
2198 * process_scheduled_works - process scheduled works
2201 * Process all scheduled works. Please note that the scheduled list
2202 * may change while processing a work, so this function repeatedly
2203 * fetches a work from the top and executes it.
2206 * spin_lock_irq(pool->lock) which may be released and regrabbed
2209 static void process_scheduled_works(struct worker *worker)
2211 while (!list_empty(&worker->scheduled)) {
2212 struct work_struct *work = list_first_entry(&worker->scheduled,
2213 struct work_struct, entry);
2214 process_one_work(worker, work);
2219 * worker_thread - the worker thread function
2222 * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools
2223 * of these per each cpu. These workers process all works regardless of
2224 * their specific target workqueue. The only exception is works which
2225 * belong to workqueues with a rescuer which will be explained in
2228 static int worker_thread(void *__worker)
2230 struct worker *worker = __worker;
2231 struct worker_pool *pool = worker->pool;
2233 /* tell the scheduler that this is a workqueue worker */
2234 worker->task->flags |= PF_WQ_WORKER;
2236 spin_lock_irq(&pool->lock);
2238 /* we are off idle list if destruction or rebind is requested */
2239 if (unlikely(list_empty(&worker->entry))) {
2240 spin_unlock_irq(&pool->lock);
2242 /* if DIE is set, destruction is requested */
2243 if (worker->flags & WORKER_DIE) {
2244 worker->task->flags &= ~PF_WQ_WORKER;
2248 /* otherwise, rebind */
2249 idle_worker_rebind(worker);
2253 worker_leave_idle(worker);
2255 /* no more worker necessary? */
2256 if (!need_more_worker(pool))
2259 /* do we need to manage? */
2260 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2264 * ->scheduled list can only be filled while a worker is
2265 * preparing to process a work or actually processing it.
2266 * Make sure nobody diddled with it while I was sleeping.
2268 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2271 * When control reaches this point, we're guaranteed to have
2272 * at least one idle worker or that someone else has already
2273 * assumed the manager role.
2275 worker_clr_flags(worker, WORKER_PREP);
2278 struct work_struct *work =
2279 list_first_entry(&pool->worklist,
2280 struct work_struct, entry);
2282 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2283 /* optimization path, not strictly necessary */
2284 process_one_work(worker, work);
2285 if (unlikely(!list_empty(&worker->scheduled)))
2286 process_scheduled_works(worker);
2288 move_linked_works(work, &worker->scheduled, NULL);
2289 process_scheduled_works(worker);
2291 } while (keep_working(pool));
2293 worker_set_flags(worker, WORKER_PREP, false);
2295 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2299 * pool->lock is held and there's no work to process and no need to
2300 * manage, sleep. Workers are woken up only while holding
2301 * pool->lock or from local cpu, so setting the current state
2302 * before releasing pool->lock is enough to prevent losing any
2305 worker_enter_idle(worker);
2306 __set_current_state(TASK_INTERRUPTIBLE);
2307 spin_unlock_irq(&pool->lock);
2313 * rescuer_thread - the rescuer thread function
2316 * Workqueue rescuer thread function. There's one rescuer for each
2317 * workqueue which has WQ_RESCUER set.
2319 * Regular work processing on a pool may block trying to create a new
2320 * worker which uses GFP_KERNEL allocation which has slight chance of
2321 * developing into deadlock if some works currently on the same queue
2322 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2323 * the problem rescuer solves.
2325 * When such condition is possible, the pool summons rescuers of all
2326 * workqueues which have works queued on the pool and let them process
2327 * those works so that forward progress can be guaranteed.
2329 * This should happen rarely.
2331 static int rescuer_thread(void *__rescuer)
2333 struct worker *rescuer = __rescuer;
2334 struct workqueue_struct *wq = rescuer->rescue_wq;
2335 struct list_head *scheduled = &rescuer->scheduled;
2336 bool is_unbound = wq->flags & WQ_UNBOUND;
2339 set_user_nice(current, RESCUER_NICE_LEVEL);
2342 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2343 * doesn't participate in concurrency management.
2345 rescuer->task->flags |= PF_WQ_WORKER;
2347 set_current_state(TASK_INTERRUPTIBLE);
2349 if (kthread_should_stop()) {
2350 __set_current_state(TASK_RUNNING);
2351 rescuer->task->flags &= ~PF_WQ_WORKER;
2356 * See whether any cpu is asking for help. Unbounded
2357 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2359 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2360 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2361 struct pool_workqueue *pwq = get_pwq(tcpu, wq);
2362 struct worker_pool *pool = pwq->pool;
2363 struct work_struct *work, *n;
2365 __set_current_state(TASK_RUNNING);
2366 mayday_clear_cpu(cpu, wq->mayday_mask);
2368 /* migrate to the target cpu if possible */
2369 worker_maybe_bind_and_lock(pool);
2370 rescuer->pool = pool;
2373 * Slurp in all works issued via this workqueue and
2376 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2377 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2378 if (get_work_pwq(work) == pwq)
2379 move_linked_works(work, scheduled, &n);
2381 process_scheduled_works(rescuer);
2384 * Leave this pool. If keep_working() is %true, notify a
2385 * regular worker; otherwise, we end up with 0 concurrency
2386 * and stalling the execution.
2388 if (keep_working(pool))
2389 wake_up_worker(pool);
2391 rescuer->pool = NULL;
2392 spin_unlock_irq(&pool->lock);
2395 /* rescuers should never participate in concurrency management */
2396 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2402 struct work_struct work;
2403 struct completion done;
2406 static void wq_barrier_func(struct work_struct *work)
2408 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2409 complete(&barr->done);
2413 * insert_wq_barrier - insert a barrier work
2414 * @pwq: pwq to insert barrier into
2415 * @barr: wq_barrier to insert
2416 * @target: target work to attach @barr to
2417 * @worker: worker currently executing @target, NULL if @target is not executing
2419 * @barr is linked to @target such that @barr is completed only after
2420 * @target finishes execution. Please note that the ordering
2421 * guarantee is observed only with respect to @target and on the local
2424 * Currently, a queued barrier can't be canceled. This is because
2425 * try_to_grab_pending() can't determine whether the work to be
2426 * grabbed is at the head of the queue and thus can't clear LINKED
2427 * flag of the previous work while there must be a valid next work
2428 * after a work with LINKED flag set.
2430 * Note that when @worker is non-NULL, @target may be modified
2431 * underneath us, so we can't reliably determine pwq from @target.
2434 * spin_lock_irq(pool->lock).
2436 static void insert_wq_barrier(struct pool_workqueue *pwq,
2437 struct wq_barrier *barr,
2438 struct work_struct *target, struct worker *worker)
2440 struct list_head *head;
2441 unsigned int linked = 0;
2444 * debugobject calls are safe here even with pool->lock locked
2445 * as we know for sure that this will not trigger any of the
2446 * checks and call back into the fixup functions where we
2449 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2450 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2451 init_completion(&barr->done);
2454 * If @target is currently being executed, schedule the
2455 * barrier to the worker; otherwise, put it after @target.
2458 head = worker->scheduled.next;
2460 unsigned long *bits = work_data_bits(target);
2462 head = target->entry.next;
2463 /* there can already be other linked works, inherit and set */
2464 linked = *bits & WORK_STRUCT_LINKED;
2465 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2468 debug_work_activate(&barr->work);
2469 insert_work(pwq, &barr->work, head,
2470 work_color_to_flags(WORK_NO_COLOR) | linked);
2474 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2475 * @wq: workqueue being flushed
2476 * @flush_color: new flush color, < 0 for no-op
2477 * @work_color: new work color, < 0 for no-op
2479 * Prepare pwqs for workqueue flushing.
2481 * If @flush_color is non-negative, flush_color on all pwqs should be
2482 * -1. If no pwq has in-flight commands at the specified color, all
2483 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2484 * has in flight commands, its pwq->flush_color is set to
2485 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2486 * wakeup logic is armed and %true is returned.
2488 * The caller should have initialized @wq->first_flusher prior to
2489 * calling this function with non-negative @flush_color. If
2490 * @flush_color is negative, no flush color update is done and %false
2493 * If @work_color is non-negative, all pwqs should have the same
2494 * work_color which is previous to @work_color and all will be
2495 * advanced to @work_color.
2498 * mutex_lock(wq->flush_mutex).
2501 * %true if @flush_color >= 0 and there's something to flush. %false
2504 static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2505 int flush_color, int work_color)
2510 if (flush_color >= 0) {
2511 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2512 atomic_set(&wq->nr_pwqs_to_flush, 1);
2515 for_each_pwq_cpu(cpu, wq) {
2516 struct pool_workqueue *pwq = get_pwq(cpu, wq);
2517 struct worker_pool *pool = pwq->pool;
2519 spin_lock_irq(&pool->lock);
2521 if (flush_color >= 0) {
2522 WARN_ON_ONCE(pwq->flush_color != -1);
2524 if (pwq->nr_in_flight[flush_color]) {
2525 pwq->flush_color = flush_color;
2526 atomic_inc(&wq->nr_pwqs_to_flush);
2531 if (work_color >= 0) {
2532 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2533 pwq->work_color = work_color;
2536 spin_unlock_irq(&pool->lock);
2539 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2540 complete(&wq->first_flusher->done);
2546 * flush_workqueue - ensure that any scheduled work has run to completion.
2547 * @wq: workqueue to flush
2549 * Forces execution of the workqueue and blocks until its completion.
2550 * This is typically used in driver shutdown handlers.
2552 * We sleep until all works which were queued on entry have been handled,
2553 * but we are not livelocked by new incoming ones.
2555 void flush_workqueue(struct workqueue_struct *wq)
2557 struct wq_flusher this_flusher = {
2558 .list = LIST_HEAD_INIT(this_flusher.list),
2560 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2564 lock_map_acquire(&wq->lockdep_map);
2565 lock_map_release(&wq->lockdep_map);
2567 mutex_lock(&wq->flush_mutex);
2570 * Start-to-wait phase
2572 next_color = work_next_color(wq->work_color);
2574 if (next_color != wq->flush_color) {
2576 * Color space is not full. The current work_color
2577 * becomes our flush_color and work_color is advanced
2580 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2581 this_flusher.flush_color = wq->work_color;
2582 wq->work_color = next_color;
2584 if (!wq->first_flusher) {
2585 /* no flush in progress, become the first flusher */
2586 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2588 wq->first_flusher = &this_flusher;
2590 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2592 /* nothing to flush, done */
2593 wq->flush_color = next_color;
2594 wq->first_flusher = NULL;
2599 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2600 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2601 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2605 * Oops, color space is full, wait on overflow queue.
2606 * The next flush completion will assign us
2607 * flush_color and transfer to flusher_queue.
2609 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2612 mutex_unlock(&wq->flush_mutex);
2614 wait_for_completion(&this_flusher.done);
2617 * Wake-up-and-cascade phase
2619 * First flushers are responsible for cascading flushes and
2620 * handling overflow. Non-first flushers can simply return.
2622 if (wq->first_flusher != &this_flusher)
2625 mutex_lock(&wq->flush_mutex);
2627 /* we might have raced, check again with mutex held */
2628 if (wq->first_flusher != &this_flusher)
2631 wq->first_flusher = NULL;
2633 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2634 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2637 struct wq_flusher *next, *tmp;
2639 /* complete all the flushers sharing the current flush color */
2640 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2641 if (next->flush_color != wq->flush_color)
2643 list_del_init(&next->list);
2644 complete(&next->done);
2647 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2648 wq->flush_color != work_next_color(wq->work_color));
2650 /* this flush_color is finished, advance by one */
2651 wq->flush_color = work_next_color(wq->flush_color);
2653 /* one color has been freed, handle overflow queue */
2654 if (!list_empty(&wq->flusher_overflow)) {
2656 * Assign the same color to all overflowed
2657 * flushers, advance work_color and append to
2658 * flusher_queue. This is the start-to-wait
2659 * phase for these overflowed flushers.
2661 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2662 tmp->flush_color = wq->work_color;
2664 wq->work_color = work_next_color(wq->work_color);
2666 list_splice_tail_init(&wq->flusher_overflow,
2667 &wq->flusher_queue);
2668 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2671 if (list_empty(&wq->flusher_queue)) {
2672 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2677 * Need to flush more colors. Make the next flusher
2678 * the new first flusher and arm pwqs.
2680 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2681 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2683 list_del_init(&next->list);
2684 wq->first_flusher = next;
2686 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2690 * Meh... this color is already done, clear first
2691 * flusher and repeat cascading.
2693 wq->first_flusher = NULL;
2697 mutex_unlock(&wq->flush_mutex);
2699 EXPORT_SYMBOL_GPL(flush_workqueue);
2702 * drain_workqueue - drain a workqueue
2703 * @wq: workqueue to drain
2705 * Wait until the workqueue becomes empty. While draining is in progress,
2706 * only chain queueing is allowed. IOW, only currently pending or running
2707 * work items on @wq can queue further work items on it. @wq is flushed
2708 * repeatedly until it becomes empty. The number of flushing is detemined
2709 * by the depth of chaining and should be relatively short. Whine if it
2712 void drain_workqueue(struct workqueue_struct *wq)
2714 unsigned int flush_cnt = 0;
2718 * __queue_work() needs to test whether there are drainers, is much
2719 * hotter than drain_workqueue() and already looks at @wq->flags.
2720 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2722 spin_lock_irq(&workqueue_lock);
2723 if (!wq->nr_drainers++)
2724 wq->flags |= WQ_DRAINING;
2725 spin_unlock_irq(&workqueue_lock);
2727 flush_workqueue(wq);
2729 for_each_pwq_cpu(cpu, wq) {
2730 struct pool_workqueue *pwq = get_pwq(cpu, wq);
2733 spin_lock_irq(&pwq->pool->lock);
2734 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2735 spin_unlock_irq(&pwq->pool->lock);
2740 if (++flush_cnt == 10 ||
2741 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2742 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2743 wq->name, flush_cnt);
2747 spin_lock_irq(&workqueue_lock);
2748 if (!--wq->nr_drainers)
2749 wq->flags &= ~WQ_DRAINING;
2750 spin_unlock_irq(&workqueue_lock);
2752 EXPORT_SYMBOL_GPL(drain_workqueue);
2754 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2756 struct worker *worker = NULL;
2757 struct worker_pool *pool;
2758 struct pool_workqueue *pwq;
2761 pool = get_work_pool(work);
2765 spin_lock_irq(&pool->lock);
2766 /* see the comment in try_to_grab_pending() with the same code */
2767 pwq = get_work_pwq(work);
2769 if (unlikely(pwq->pool != pool))
2772 worker = find_worker_executing_work(pool, work);
2775 pwq = worker->current_pwq;
2778 insert_wq_barrier(pwq, barr, work, worker);
2779 spin_unlock_irq(&pool->lock);
2782 * If @max_active is 1 or rescuer is in use, flushing another work
2783 * item on the same workqueue may lead to deadlock. Make sure the
2784 * flusher is not running on the same workqueue by verifying write
2787 if (pwq->wq->saved_max_active == 1 || pwq->wq->flags & WQ_RESCUER)
2788 lock_map_acquire(&pwq->wq->lockdep_map);
2790 lock_map_acquire_read(&pwq->wq->lockdep_map);
2791 lock_map_release(&pwq->wq->lockdep_map);
2795 spin_unlock_irq(&pool->lock);
2800 * flush_work - wait for a work to finish executing the last queueing instance
2801 * @work: the work to flush
2803 * Wait until @work has finished execution. @work is guaranteed to be idle
2804 * on return if it hasn't been requeued since flush started.
2807 * %true if flush_work() waited for the work to finish execution,
2808 * %false if it was already idle.
2810 bool flush_work(struct work_struct *work)
2812 struct wq_barrier barr;
2814 lock_map_acquire(&work->lockdep_map);
2815 lock_map_release(&work->lockdep_map);
2817 if (start_flush_work(work, &barr)) {
2818 wait_for_completion(&barr.done);
2819 destroy_work_on_stack(&barr.work);
2825 EXPORT_SYMBOL_GPL(flush_work);
2827 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2829 unsigned long flags;
2833 ret = try_to_grab_pending(work, is_dwork, &flags);
2835 * If someone else is canceling, wait for the same event it
2836 * would be waiting for before retrying.
2838 if (unlikely(ret == -ENOENT))
2840 } while (unlikely(ret < 0));
2842 /* tell other tasks trying to grab @work to back off */
2843 mark_work_canceling(work);
2844 local_irq_restore(flags);
2847 clear_work_data(work);
2852 * cancel_work_sync - cancel a work and wait for it to finish
2853 * @work: the work to cancel
2855 * Cancel @work and wait for its execution to finish. This function
2856 * can be used even if the work re-queues itself or migrates to
2857 * another workqueue. On return from this function, @work is
2858 * guaranteed to be not pending or executing on any CPU.
2860 * cancel_work_sync(&delayed_work->work) must not be used for
2861 * delayed_work's. Use cancel_delayed_work_sync() instead.
2863 * The caller must ensure that the workqueue on which @work was last
2864 * queued can't be destroyed before this function returns.
2867 * %true if @work was pending, %false otherwise.
2869 bool cancel_work_sync(struct work_struct *work)
2871 return __cancel_work_timer(work, false);
2873 EXPORT_SYMBOL_GPL(cancel_work_sync);
2876 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2877 * @dwork: the delayed work to flush
2879 * Delayed timer is cancelled and the pending work is queued for
2880 * immediate execution. Like flush_work(), this function only
2881 * considers the last queueing instance of @dwork.
2884 * %true if flush_work() waited for the work to finish execution,
2885 * %false if it was already idle.
2887 bool flush_delayed_work(struct delayed_work *dwork)
2889 local_irq_disable();
2890 if (del_timer_sync(&dwork->timer))
2891 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2893 return flush_work(&dwork->work);
2895 EXPORT_SYMBOL(flush_delayed_work);
2898 * cancel_delayed_work - cancel a delayed work
2899 * @dwork: delayed_work to cancel
2901 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2902 * and canceled; %false if wasn't pending. Note that the work callback
2903 * function may still be running on return, unless it returns %true and the
2904 * work doesn't re-arm itself. Explicitly flush or use
2905 * cancel_delayed_work_sync() to wait on it.
2907 * This function is safe to call from any context including IRQ handler.
2909 bool cancel_delayed_work(struct delayed_work *dwork)
2911 unsigned long flags;
2915 ret = try_to_grab_pending(&dwork->work, true, &flags);
2916 } while (unlikely(ret == -EAGAIN));
2918 if (unlikely(ret < 0))
2921 set_work_pool_and_clear_pending(&dwork->work,
2922 get_work_pool_id(&dwork->work));
2923 local_irq_restore(flags);
2926 EXPORT_SYMBOL(cancel_delayed_work);
2929 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2930 * @dwork: the delayed work cancel
2932 * This is cancel_work_sync() for delayed works.
2935 * %true if @dwork was pending, %false otherwise.
2937 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2939 return __cancel_work_timer(&dwork->work, true);
2941 EXPORT_SYMBOL(cancel_delayed_work_sync);
2944 * schedule_work_on - put work task on a specific cpu
2945 * @cpu: cpu to put the work task on
2946 * @work: job to be done
2948 * This puts a job on a specific cpu
2950 bool schedule_work_on(int cpu, struct work_struct *work)
2952 return queue_work_on(cpu, system_wq, work);
2954 EXPORT_SYMBOL(schedule_work_on);
2957 * schedule_work - put work task in global workqueue
2958 * @work: job to be done
2960 * Returns %false if @work was already on the kernel-global workqueue and
2963 * This puts a job in the kernel-global workqueue if it was not already
2964 * queued and leaves it in the same position on the kernel-global
2965 * workqueue otherwise.
2967 bool schedule_work(struct work_struct *work)
2969 return queue_work(system_wq, work);
2971 EXPORT_SYMBOL(schedule_work);
2974 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2976 * @dwork: job to be done
2977 * @delay: number of jiffies to wait
2979 * After waiting for a given time this puts a job in the kernel-global
2980 * workqueue on the specified CPU.
2982 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
2983 unsigned long delay)
2985 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2987 EXPORT_SYMBOL(schedule_delayed_work_on);
2990 * schedule_delayed_work - put work task in global workqueue after delay
2991 * @dwork: job to be done
2992 * @delay: number of jiffies to wait or 0 for immediate execution
2994 * After waiting for a given time this puts a job in the kernel-global
2997 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
2999 return queue_delayed_work(system_wq, dwork, delay);
3001 EXPORT_SYMBOL(schedule_delayed_work);
3004 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3005 * @func: the function to call
3007 * schedule_on_each_cpu() executes @func on each online CPU using the
3008 * system workqueue and blocks until all CPUs have completed.
3009 * schedule_on_each_cpu() is very slow.
3012 * 0 on success, -errno on failure.
3014 int schedule_on_each_cpu(work_func_t func)
3017 struct work_struct __percpu *works;
3019 works = alloc_percpu(struct work_struct);
3025 for_each_online_cpu(cpu) {
3026 struct work_struct *work = per_cpu_ptr(works, cpu);
3028 INIT_WORK(work, func);
3029 schedule_work_on(cpu, work);
3032 for_each_online_cpu(cpu)
3033 flush_work(per_cpu_ptr(works, cpu));
3041 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3043 * Forces execution of the kernel-global workqueue and blocks until its
3046 * Think twice before calling this function! It's very easy to get into
3047 * trouble if you don't take great care. Either of the following situations
3048 * will lead to deadlock:
3050 * One of the work items currently on the workqueue needs to acquire
3051 * a lock held by your code or its caller.
3053 * Your code is running in the context of a work routine.
3055 * They will be detected by lockdep when they occur, but the first might not
3056 * occur very often. It depends on what work items are on the workqueue and
3057 * what locks they need, which you have no control over.
3059 * In most situations flushing the entire workqueue is overkill; you merely
3060 * need to know that a particular work item isn't queued and isn't running.
3061 * In such cases you should use cancel_delayed_work_sync() or
3062 * cancel_work_sync() instead.
3064 void flush_scheduled_work(void)
3066 flush_workqueue(system_wq);
3068 EXPORT_SYMBOL(flush_scheduled_work);
3071 * execute_in_process_context - reliably execute the routine with user context
3072 * @fn: the function to execute
3073 * @ew: guaranteed storage for the execute work structure (must
3074 * be available when the work executes)
3076 * Executes the function immediately if process context is available,
3077 * otherwise schedules the function for delayed execution.
3079 * Returns: 0 - function was executed
3080 * 1 - function was scheduled for execution
3082 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3084 if (!in_interrupt()) {
3089 INIT_WORK(&ew->work, fn);
3090 schedule_work(&ew->work);
3094 EXPORT_SYMBOL_GPL(execute_in_process_context);
3096 int keventd_up(void)
3098 return system_wq != NULL;
3101 static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3105 if (!(wq->flags & WQ_UNBOUND)) {
3106 wq->pool_wq.pcpu = alloc_percpu(struct pool_workqueue);
3107 if (!wq->pool_wq.pcpu)
3110 for_each_possible_cpu(cpu) {
3111 struct pool_workqueue *pwq = get_pwq(cpu, wq);
3113 list_add_tail(&pwq->pwqs_node, &wq->pwqs);
3116 struct pool_workqueue *pwq;
3118 pwq = kmem_cache_zalloc(pwq_cache, GFP_KERNEL);
3122 wq->pool_wq.single = pwq;
3123 list_add_tail(&pwq->pwqs_node, &wq->pwqs);
3129 static void free_pwqs(struct workqueue_struct *wq)
3131 if (!(wq->flags & WQ_UNBOUND))
3132 free_percpu(wq->pool_wq.pcpu);
3134 kmem_cache_free(pwq_cache, wq->pool_wq.single);
3137 static int wq_clamp_max_active(int max_active, unsigned int flags,
3140 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3142 if (max_active < 1 || max_active > lim)
3143 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3144 max_active, name, 1, lim);
3146 return clamp_val(max_active, 1, lim);
3149 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3152 struct lock_class_key *key,
3153 const char *lock_name, ...)
3155 va_list args, args1;
3156 struct workqueue_struct *wq;
3160 /* determine namelen, allocate wq and format name */
3161 va_start(args, lock_name);
3162 va_copy(args1, args);
3163 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3165 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3169 vsnprintf(wq->name, namelen, fmt, args1);
3174 * Workqueues which may be used during memory reclaim should
3175 * have a rescuer to guarantee forward progress.
3177 if (flags & WQ_MEM_RECLAIM)
3178 flags |= WQ_RESCUER;
3180 max_active = max_active ?: WQ_DFL_ACTIVE;
3181 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3185 wq->saved_max_active = max_active;
3186 mutex_init(&wq->flush_mutex);
3187 atomic_set(&wq->nr_pwqs_to_flush, 0);
3188 INIT_LIST_HEAD(&wq->pwqs);
3189 INIT_LIST_HEAD(&wq->flusher_queue);
3190 INIT_LIST_HEAD(&wq->flusher_overflow);
3192 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3193 INIT_LIST_HEAD(&wq->list);
3195 if (alloc_and_link_pwqs(wq) < 0)
3198 for_each_pwq_cpu(cpu, wq) {
3199 struct pool_workqueue *pwq = get_pwq(cpu, wq);
3201 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3202 pwq->pool = get_std_worker_pool(cpu, flags & WQ_HIGHPRI);
3204 pwq->flush_color = -1;
3205 pwq->max_active = max_active;
3206 INIT_LIST_HEAD(&pwq->delayed_works);
3209 if (flags & WQ_RESCUER) {
3210 struct worker *rescuer;
3212 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3215 wq->rescuer = rescuer = alloc_worker();
3219 rescuer->rescue_wq = wq;
3220 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3222 if (IS_ERR(rescuer->task))
3225 rescuer->task->flags |= PF_THREAD_BOUND;
3226 wake_up_process(rescuer->task);
3230 * workqueue_lock protects global freeze state and workqueues
3231 * list. Grab it, set max_active accordingly and add the new
3232 * workqueue to workqueues list.
3234 spin_lock_irq(&workqueue_lock);
3236 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3237 for_each_pwq_cpu(cpu, wq)
3238 get_pwq(cpu, wq)->max_active = 0;
3240 list_add(&wq->list, &workqueues);
3242 spin_unlock_irq(&workqueue_lock);
3248 free_mayday_mask(wq->mayday_mask);
3254 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3257 * destroy_workqueue - safely terminate a workqueue
3258 * @wq: target workqueue
3260 * Safely destroy a workqueue. All work currently pending will be done first.
3262 void destroy_workqueue(struct workqueue_struct *wq)
3266 /* drain it before proceeding with destruction */
3267 drain_workqueue(wq);
3270 for_each_pwq_cpu(cpu, wq) {
3271 struct pool_workqueue *pwq = get_pwq(cpu, wq);
3274 for (i = 0; i < WORK_NR_COLORS; i++)
3275 if (WARN_ON(pwq->nr_in_flight[i]))
3277 if (WARN_ON(pwq->nr_active) ||
3278 WARN_ON(!list_empty(&pwq->delayed_works)))
3283 * wq list is used to freeze wq, remove from list after
3284 * flushing is complete in case freeze races us.
3286 spin_lock_irq(&workqueue_lock);
3287 list_del(&wq->list);
3288 spin_unlock_irq(&workqueue_lock);
3290 if (wq->flags & WQ_RESCUER) {
3291 kthread_stop(wq->rescuer->task);
3292 free_mayday_mask(wq->mayday_mask);
3299 EXPORT_SYMBOL_GPL(destroy_workqueue);
3302 * pwq_set_max_active - adjust max_active of a pwq
3303 * @pwq: target pool_workqueue
3304 * @max_active: new max_active value.
3306 * Set @pwq->max_active to @max_active and activate delayed works if
3310 * spin_lock_irq(pool->lock).
3312 static void pwq_set_max_active(struct pool_workqueue *pwq, int max_active)
3314 pwq->max_active = max_active;
3316 while (!list_empty(&pwq->delayed_works) &&
3317 pwq->nr_active < pwq->max_active)
3318 pwq_activate_first_delayed(pwq);
3322 * workqueue_set_max_active - adjust max_active of a workqueue
3323 * @wq: target workqueue
3324 * @max_active: new max_active value.
3326 * Set max_active of @wq to @max_active.
3329 * Don't call from IRQ context.
3331 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3335 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3337 spin_lock_irq(&workqueue_lock);
3339 wq->saved_max_active = max_active;
3341 for_each_pwq_cpu(cpu, wq) {
3342 struct pool_workqueue *pwq = get_pwq(cpu, wq);
3343 struct worker_pool *pool = pwq->pool;
3345 spin_lock(&pool->lock);
3347 if (!(wq->flags & WQ_FREEZABLE) ||
3348 !(pool->flags & POOL_FREEZING))
3349 pwq_set_max_active(pwq, max_active);
3351 spin_unlock(&pool->lock);
3354 spin_unlock_irq(&workqueue_lock);
3356 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3359 * workqueue_congested - test whether a workqueue is congested
3360 * @cpu: CPU in question
3361 * @wq: target workqueue
3363 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3364 * no synchronization around this function and the test result is
3365 * unreliable and only useful as advisory hints or for debugging.
3368 * %true if congested, %false otherwise.
3370 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3372 struct pool_workqueue *pwq = get_pwq(cpu, wq);
3374 return !list_empty(&pwq->delayed_works);
3376 EXPORT_SYMBOL_GPL(workqueue_congested);
3379 * work_busy - test whether a work is currently pending or running
3380 * @work: the work to be tested
3382 * Test whether @work is currently pending or running. There is no
3383 * synchronization around this function and the test result is
3384 * unreliable and only useful as advisory hints or for debugging.
3387 * OR'd bitmask of WORK_BUSY_* bits.
3389 unsigned int work_busy(struct work_struct *work)
3391 struct worker_pool *pool = get_work_pool(work);
3392 unsigned long flags;
3393 unsigned int ret = 0;
3395 if (work_pending(work))
3396 ret |= WORK_BUSY_PENDING;
3399 spin_lock_irqsave(&pool->lock, flags);
3400 if (find_worker_executing_work(pool, work))
3401 ret |= WORK_BUSY_RUNNING;
3402 spin_unlock_irqrestore(&pool->lock, flags);
3407 EXPORT_SYMBOL_GPL(work_busy);
3412 * There are two challenges in supporting CPU hotplug. Firstly, there
3413 * are a lot of assumptions on strong associations among work, pwq and
3414 * pool which make migrating pending and scheduled works very
3415 * difficult to implement without impacting hot paths. Secondly,
3416 * worker pools serve mix of short, long and very long running works making
3417 * blocked draining impractical.
3419 * This is solved by allowing the pools to be disassociated from the CPU
3420 * running as an unbound one and allowing it to be reattached later if the
3421 * cpu comes back online.
3424 static void wq_unbind_fn(struct work_struct *work)
3426 int cpu = smp_processor_id();
3427 struct worker_pool *pool;
3428 struct worker *worker;
3431 for_each_std_worker_pool(pool, cpu) {
3432 WARN_ON_ONCE(cpu != smp_processor_id());
3434 mutex_lock(&pool->assoc_mutex);
3435 spin_lock_irq(&pool->lock);
3438 * We've claimed all manager positions. Make all workers
3439 * unbound and set DISASSOCIATED. Before this, all workers
3440 * except for the ones which are still executing works from
3441 * before the last CPU down must be on the cpu. After
3442 * this, they may become diasporas.
3444 list_for_each_entry(worker, &pool->idle_list, entry)
3445 worker->flags |= WORKER_UNBOUND;
3447 for_each_busy_worker(worker, i, pool)
3448 worker->flags |= WORKER_UNBOUND;
3450 pool->flags |= POOL_DISASSOCIATED;
3452 spin_unlock_irq(&pool->lock);
3453 mutex_unlock(&pool->assoc_mutex);
3457 * Call schedule() so that we cross rq->lock and thus can guarantee
3458 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3459 * as scheduler callbacks may be invoked from other cpus.
3464 * Sched callbacks are disabled now. Zap nr_running. After this,
3465 * nr_running stays zero and need_more_worker() and keep_working()
3466 * are always true as long as the worklist is not empty. Pools on
3467 * @cpu now behave as unbound (in terms of concurrency management)
3468 * pools which are served by workers tied to the CPU.
3470 * On return from this function, the current worker would trigger
3471 * unbound chain execution of pending work items if other workers
3474 for_each_std_worker_pool(pool, cpu)
3475 atomic_set(&pool->nr_running, 0);
3479 * Workqueues should be brought up before normal priority CPU notifiers.
3480 * This will be registered high priority CPU notifier.
3482 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3483 unsigned long action,
3486 unsigned int cpu = (unsigned long)hcpu;
3487 struct worker_pool *pool;
3489 switch (action & ~CPU_TASKS_FROZEN) {
3490 case CPU_UP_PREPARE:
3491 for_each_std_worker_pool(pool, cpu) {
3492 struct worker *worker;
3494 if (pool->nr_workers)
3497 worker = create_worker(pool);
3501 spin_lock_irq(&pool->lock);
3502 start_worker(worker);
3503 spin_unlock_irq(&pool->lock);
3507 case CPU_DOWN_FAILED:
3509 for_each_std_worker_pool(pool, cpu) {
3510 mutex_lock(&pool->assoc_mutex);
3511 spin_lock_irq(&pool->lock);
3513 pool->flags &= ~POOL_DISASSOCIATED;
3514 rebind_workers(pool);
3516 spin_unlock_irq(&pool->lock);
3517 mutex_unlock(&pool->assoc_mutex);
3525 * Workqueues should be brought down after normal priority CPU notifiers.
3526 * This will be registered as low priority CPU notifier.
3528 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3529 unsigned long action,
3532 unsigned int cpu = (unsigned long)hcpu;
3533 struct work_struct unbind_work;
3535 switch (action & ~CPU_TASKS_FROZEN) {
3536 case CPU_DOWN_PREPARE:
3537 /* unbinding should happen on the local CPU */
3538 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
3539 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3540 flush_work(&unbind_work);
3548 struct work_for_cpu {
3549 struct work_struct work;
3555 static void work_for_cpu_fn(struct work_struct *work)
3557 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3559 wfc->ret = wfc->fn(wfc->arg);
3563 * work_on_cpu - run a function in user context on a particular cpu
3564 * @cpu: the cpu to run on
3565 * @fn: the function to run
3566 * @arg: the function arg
3568 * This will return the value @fn returns.
3569 * It is up to the caller to ensure that the cpu doesn't go offline.
3570 * The caller must not hold any locks which would prevent @fn from completing.
3572 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3574 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3576 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3577 schedule_work_on(cpu, &wfc.work);
3578 flush_work(&wfc.work);
3581 EXPORT_SYMBOL_GPL(work_on_cpu);
3582 #endif /* CONFIG_SMP */
3584 #ifdef CONFIG_FREEZER
3587 * freeze_workqueues_begin - begin freezing workqueues
3589 * Start freezing workqueues. After this function returns, all freezable
3590 * workqueues will queue new works to their frozen_works list instead of
3594 * Grabs and releases workqueue_lock and pool->lock's.
3596 void freeze_workqueues_begin(void)
3600 spin_lock_irq(&workqueue_lock);
3602 WARN_ON_ONCE(workqueue_freezing);
3603 workqueue_freezing = true;
3605 for_each_wq_cpu(cpu) {
3606 struct worker_pool *pool;
3607 struct workqueue_struct *wq;
3609 for_each_std_worker_pool(pool, cpu) {
3610 spin_lock(&pool->lock);
3612 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
3613 pool->flags |= POOL_FREEZING;
3615 list_for_each_entry(wq, &workqueues, list) {
3616 struct pool_workqueue *pwq = get_pwq(cpu, wq);
3618 if (pwq && pwq->pool == pool &&
3619 (wq->flags & WQ_FREEZABLE))
3620 pwq->max_active = 0;
3623 spin_unlock(&pool->lock);
3627 spin_unlock_irq(&workqueue_lock);
3631 * freeze_workqueues_busy - are freezable workqueues still busy?
3633 * Check whether freezing is complete. This function must be called
3634 * between freeze_workqueues_begin() and thaw_workqueues().
3637 * Grabs and releases workqueue_lock.
3640 * %true if some freezable workqueues are still busy. %false if freezing
3643 bool freeze_workqueues_busy(void)
3648 spin_lock_irq(&workqueue_lock);
3650 WARN_ON_ONCE(!workqueue_freezing);
3652 for_each_wq_cpu(cpu) {
3653 struct workqueue_struct *wq;
3655 * nr_active is monotonically decreasing. It's safe
3656 * to peek without lock.
3658 list_for_each_entry(wq, &workqueues, list) {
3659 struct pool_workqueue *pwq = get_pwq(cpu, wq);
3661 if (!pwq || !(wq->flags & WQ_FREEZABLE))
3664 WARN_ON_ONCE(pwq->nr_active < 0);
3665 if (pwq->nr_active) {
3672 spin_unlock_irq(&workqueue_lock);
3677 * thaw_workqueues - thaw workqueues
3679 * Thaw workqueues. Normal queueing is restored and all collected
3680 * frozen works are transferred to their respective pool worklists.
3683 * Grabs and releases workqueue_lock and pool->lock's.
3685 void thaw_workqueues(void)
3689 spin_lock_irq(&workqueue_lock);
3691 if (!workqueue_freezing)
3694 for_each_wq_cpu(cpu) {
3695 struct worker_pool *pool;
3696 struct workqueue_struct *wq;
3698 for_each_std_worker_pool(pool, cpu) {
3699 spin_lock(&pool->lock);
3701 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
3702 pool->flags &= ~POOL_FREEZING;
3704 list_for_each_entry(wq, &workqueues, list) {
3705 struct pool_workqueue *pwq = get_pwq(cpu, wq);
3707 if (!pwq || pwq->pool != pool ||
3708 !(wq->flags & WQ_FREEZABLE))
3711 /* restore max_active and repopulate worklist */
3712 pwq_set_max_active(pwq, wq->saved_max_active);
3715 wake_up_worker(pool);
3717 spin_unlock(&pool->lock);
3721 workqueue_freezing = false;
3723 spin_unlock_irq(&workqueue_lock);
3725 #endif /* CONFIG_FREEZER */
3727 static int __init init_workqueues(void)
3731 /* make sure we have enough bits for OFFQ pool ID */
3732 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
3733 WORK_CPU_END * NR_STD_WORKER_POOLS);
3735 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
3737 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
3739 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3740 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3742 /* initialize CPU pools */
3743 for_each_wq_cpu(cpu) {
3744 struct worker_pool *pool;
3746 for_each_std_worker_pool(pool, cpu) {
3747 spin_lock_init(&pool->lock);
3749 pool->flags |= POOL_DISASSOCIATED;
3750 INIT_LIST_HEAD(&pool->worklist);
3751 INIT_LIST_HEAD(&pool->idle_list);
3752 hash_init(pool->busy_hash);
3754 init_timer_deferrable(&pool->idle_timer);
3755 pool->idle_timer.function = idle_worker_timeout;
3756 pool->idle_timer.data = (unsigned long)pool;
3758 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3759 (unsigned long)pool);
3761 mutex_init(&pool->assoc_mutex);
3762 ida_init(&pool->worker_ida);
3765 BUG_ON(worker_pool_assign_id(pool));
3769 /* create the initial worker */
3770 for_each_online_wq_cpu(cpu) {
3771 struct worker_pool *pool;
3773 for_each_std_worker_pool(pool, cpu) {
3774 struct worker *worker;
3776 if (cpu != WORK_CPU_UNBOUND)
3777 pool->flags &= ~POOL_DISASSOCIATED;
3779 worker = create_worker(pool);
3781 spin_lock_irq(&pool->lock);
3782 start_worker(worker);
3783 spin_unlock_irq(&pool->lock);
3787 system_wq = alloc_workqueue("events", 0, 0);
3788 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3789 system_long_wq = alloc_workqueue("events_long", 0, 0);
3790 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3791 WQ_UNBOUND_MAX_ACTIVE);
3792 system_freezable_wq = alloc_workqueue("events_freezable",
3794 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3795 !system_unbound_wq || !system_freezable_wq);
3798 early_initcall(init_workqueues);