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 */
147 } ____cacheline_aligned_in_smp;
150 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
151 * work_struct->data are used for flags and thus cwqs need to be
152 * aligned at two's power of the number of flag bits.
154 struct cpu_workqueue_struct {
155 struct worker_pool *pool; /* I: the associated pool */
156 struct workqueue_struct *wq; /* I: the owning workqueue */
157 int work_color; /* L: current color */
158 int flush_color; /* L: flushing color */
159 int nr_in_flight[WORK_NR_COLORS];
160 /* L: nr of in_flight works */
161 int nr_active; /* L: nr of active works */
162 int max_active; /* L: max active works */
163 struct list_head delayed_works; /* L: delayed works */
167 * Structure used to wait for workqueue flush.
170 struct list_head list; /* F: list of flushers */
171 int flush_color; /* F: flush color waiting for */
172 struct completion done; /* flush completion */
176 * All cpumasks are assumed to be always set on UP and thus can't be
177 * used to determine whether there's something to be done.
180 typedef cpumask_var_t mayday_mask_t;
181 #define mayday_test_and_set_cpu(cpu, mask) \
182 cpumask_test_and_set_cpu((cpu), (mask))
183 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
184 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
185 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
186 #define free_mayday_mask(mask) free_cpumask_var((mask))
188 typedef unsigned long mayday_mask_t;
189 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
190 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
191 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
192 #define alloc_mayday_mask(maskp, gfp) true
193 #define free_mayday_mask(mask) do { } while (0)
197 * The externally visible workqueue abstraction is an array of
198 * per-CPU workqueues:
200 struct workqueue_struct {
201 unsigned int flags; /* W: WQ_* flags */
203 struct cpu_workqueue_struct __percpu *pcpu;
204 struct cpu_workqueue_struct *single;
206 } cpu_wq; /* I: cwq's */
207 struct list_head list; /* W: list of all workqueues */
209 struct mutex flush_mutex; /* protects wq flushing */
210 int work_color; /* F: current work color */
211 int flush_color; /* F: current flush color */
212 atomic_t nr_cwqs_to_flush; /* flush in progress */
213 struct wq_flusher *first_flusher; /* F: first flusher */
214 struct list_head flusher_queue; /* F: flush waiters */
215 struct list_head flusher_overflow; /* F: flush overflow list */
217 mayday_mask_t mayday_mask; /* cpus requesting rescue */
218 struct worker *rescuer; /* I: rescue worker */
220 int nr_drainers; /* W: drain in progress */
221 int saved_max_active; /* W: saved cwq max_active */
222 #ifdef CONFIG_LOCKDEP
223 struct lockdep_map lockdep_map;
225 char name[]; /* I: workqueue name */
228 struct workqueue_struct *system_wq __read_mostly;
229 EXPORT_SYMBOL_GPL(system_wq);
230 struct workqueue_struct *system_highpri_wq __read_mostly;
231 EXPORT_SYMBOL_GPL(system_highpri_wq);
232 struct workqueue_struct *system_long_wq __read_mostly;
233 EXPORT_SYMBOL_GPL(system_long_wq);
234 struct workqueue_struct *system_unbound_wq __read_mostly;
235 EXPORT_SYMBOL_GPL(system_unbound_wq);
236 struct workqueue_struct *system_freezable_wq __read_mostly;
237 EXPORT_SYMBOL_GPL(system_freezable_wq);
239 #define CREATE_TRACE_POINTS
240 #include <trace/events/workqueue.h>
242 #define for_each_std_worker_pool(pool, cpu) \
243 for ((pool) = &std_worker_pools(cpu)[0]; \
244 (pool) < &std_worker_pools(cpu)[NR_STD_WORKER_POOLS]; (pool)++)
246 #define for_each_busy_worker(worker, i, pos, pool) \
247 hash_for_each(pool->busy_hash, i, pos, worker, hentry)
249 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
252 if (cpu < nr_cpu_ids) {
254 cpu = cpumask_next(cpu, mask);
255 if (cpu < nr_cpu_ids)
259 return WORK_CPU_UNBOUND;
264 static inline int __next_cwq_cpu(int cpu, const struct cpumask *mask,
265 struct workqueue_struct *wq)
267 return __next_wq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
273 * An extra cpu number is defined using an invalid cpu number
274 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
275 * specific CPU. The following iterators are similar to for_each_*_cpu()
276 * iterators but also considers the unbound CPU.
278 * for_each_wq_cpu() : possible CPUs + WORK_CPU_UNBOUND
279 * for_each_online_wq_cpu() : online CPUs + WORK_CPU_UNBOUND
280 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
281 * WORK_CPU_UNBOUND for unbound workqueues
283 #define for_each_wq_cpu(cpu) \
284 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, 3); \
285 (cpu) < WORK_CPU_END; \
286 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, 3))
288 #define for_each_online_wq_cpu(cpu) \
289 for ((cpu) = __next_wq_cpu(-1, cpu_online_mask, 3); \
290 (cpu) < WORK_CPU_END; \
291 (cpu) = __next_wq_cpu((cpu), cpu_online_mask, 3))
293 #define for_each_cwq_cpu(cpu, wq) \
294 for ((cpu) = __next_cwq_cpu(-1, cpu_possible_mask, (wq)); \
295 (cpu) < WORK_CPU_END; \
296 (cpu) = __next_cwq_cpu((cpu), cpu_possible_mask, (wq)))
298 #ifdef CONFIG_DEBUG_OBJECTS_WORK
300 static struct debug_obj_descr work_debug_descr;
302 static void *work_debug_hint(void *addr)
304 return ((struct work_struct *) addr)->func;
308 * fixup_init is called when:
309 * - an active object is initialized
311 static int work_fixup_init(void *addr, enum debug_obj_state state)
313 struct work_struct *work = addr;
316 case ODEBUG_STATE_ACTIVE:
317 cancel_work_sync(work);
318 debug_object_init(work, &work_debug_descr);
326 * fixup_activate is called when:
327 * - an active object is activated
328 * - an unknown object is activated (might be a statically initialized object)
330 static int work_fixup_activate(void *addr, enum debug_obj_state state)
332 struct work_struct *work = addr;
336 case ODEBUG_STATE_NOTAVAILABLE:
338 * This is not really a fixup. The work struct was
339 * statically initialized. We just make sure that it
340 * is tracked in the object tracker.
342 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
343 debug_object_init(work, &work_debug_descr);
344 debug_object_activate(work, &work_debug_descr);
350 case ODEBUG_STATE_ACTIVE:
359 * fixup_free is called when:
360 * - an active object is freed
362 static int work_fixup_free(void *addr, enum debug_obj_state state)
364 struct work_struct *work = addr;
367 case ODEBUG_STATE_ACTIVE:
368 cancel_work_sync(work);
369 debug_object_free(work, &work_debug_descr);
376 static struct debug_obj_descr work_debug_descr = {
377 .name = "work_struct",
378 .debug_hint = work_debug_hint,
379 .fixup_init = work_fixup_init,
380 .fixup_activate = work_fixup_activate,
381 .fixup_free = work_fixup_free,
384 static inline void debug_work_activate(struct work_struct *work)
386 debug_object_activate(work, &work_debug_descr);
389 static inline void debug_work_deactivate(struct work_struct *work)
391 debug_object_deactivate(work, &work_debug_descr);
394 void __init_work(struct work_struct *work, int onstack)
397 debug_object_init_on_stack(work, &work_debug_descr);
399 debug_object_init(work, &work_debug_descr);
401 EXPORT_SYMBOL_GPL(__init_work);
403 void destroy_work_on_stack(struct work_struct *work)
405 debug_object_free(work, &work_debug_descr);
407 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
410 static inline void debug_work_activate(struct work_struct *work) { }
411 static inline void debug_work_deactivate(struct work_struct *work) { }
414 /* Serializes the accesses to the list of workqueues. */
415 static DEFINE_SPINLOCK(workqueue_lock);
416 static LIST_HEAD(workqueues);
417 static bool workqueue_freezing; /* W: have wqs started freezing? */
420 * The CPU standard worker pools. nr_running is the only field which is
421 * expected to be used frequently by other cpus via try_to_wake_up(). Put
422 * it in a separate cacheline.
424 static DEFINE_PER_CPU(struct worker_pool [NR_STD_WORKER_POOLS],
425 cpu_std_worker_pools);
426 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t [NR_STD_WORKER_POOLS],
427 cpu_std_pool_nr_running);
430 * Standard worker pools and nr_running counter for unbound CPU. The pools
431 * have POOL_DISASSOCIATED set, and all workers have WORKER_UNBOUND set.
433 static struct worker_pool unbound_std_worker_pools[NR_STD_WORKER_POOLS];
434 static atomic_t unbound_std_pool_nr_running[NR_STD_WORKER_POOLS] = {
435 [0 ... NR_STD_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */
438 /* idr of all pools */
439 static DEFINE_MUTEX(worker_pool_idr_mutex);
440 static DEFINE_IDR(worker_pool_idr);
442 static int worker_thread(void *__worker);
444 static struct worker_pool *std_worker_pools(int cpu)
446 if (cpu != WORK_CPU_UNBOUND)
447 return per_cpu(cpu_std_worker_pools, cpu);
449 return unbound_std_worker_pools;
452 static int std_worker_pool_pri(struct worker_pool *pool)
454 return pool - std_worker_pools(pool->cpu);
457 /* allocate ID and assign it to @pool */
458 static int worker_pool_assign_id(struct worker_pool *pool)
462 mutex_lock(&worker_pool_idr_mutex);
463 idr_pre_get(&worker_pool_idr, GFP_KERNEL);
464 ret = idr_get_new(&worker_pool_idr, pool, &pool->id);
465 mutex_unlock(&worker_pool_idr_mutex);
471 * Lookup worker_pool by id. The idr currently is built during boot and
472 * never modified. Don't worry about locking for now.
474 static struct worker_pool *worker_pool_by_id(int pool_id)
476 return idr_find(&worker_pool_idr, pool_id);
479 static struct worker_pool *get_std_worker_pool(int cpu, bool highpri)
481 struct worker_pool *pools = std_worker_pools(cpu);
483 return &pools[highpri];
486 static atomic_t *get_pool_nr_running(struct worker_pool *pool)
489 int idx = std_worker_pool_pri(pool);
491 if (cpu != WORK_CPU_UNBOUND)
492 return &per_cpu(cpu_std_pool_nr_running, cpu)[idx];
494 return &unbound_std_pool_nr_running[idx];
497 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
498 struct workqueue_struct *wq)
500 if (!(wq->flags & WQ_UNBOUND)) {
501 if (likely(cpu < nr_cpu_ids))
502 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
503 } else if (likely(cpu == WORK_CPU_UNBOUND))
504 return wq->cpu_wq.single;
508 static unsigned int work_color_to_flags(int color)
510 return color << WORK_STRUCT_COLOR_SHIFT;
513 static int get_work_color(struct work_struct *work)
515 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
516 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
519 static int work_next_color(int color)
521 return (color + 1) % WORK_NR_COLORS;
525 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
526 * contain the pointer to the queued cwq. Once execution starts, the flag
527 * is cleared and the high bits contain OFFQ flags and pool ID.
529 * set_work_cwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
530 * and clear_work_data() can be used to set the cwq, pool or clear
531 * work->data. These functions should only be called while the work is
532 * owned - ie. while the PENDING bit is set.
534 * get_work_pool() and get_work_cwq() can be used to obtain the pool or cwq
535 * corresponding to a work. Pool is available once the work has been
536 * queued anywhere after initialization until it is sync canceled. cwq is
537 * available only while the work item is queued.
539 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
540 * canceled. While being canceled, a work item may have its PENDING set
541 * but stay off timer and worklist for arbitrarily long and nobody should
542 * try to steal the PENDING bit.
544 static inline void set_work_data(struct work_struct *work, unsigned long data,
547 BUG_ON(!work_pending(work));
548 atomic_long_set(&work->data, data | flags | work_static(work));
551 static void set_work_cwq(struct work_struct *work,
552 struct cpu_workqueue_struct *cwq,
553 unsigned long extra_flags)
555 set_work_data(work, (unsigned long)cwq,
556 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
559 static void set_work_pool_and_keep_pending(struct work_struct *work,
562 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
563 WORK_STRUCT_PENDING);
566 static void set_work_pool_and_clear_pending(struct work_struct *work,
570 * The following wmb is paired with the implied mb in
571 * test_and_set_bit(PENDING) and ensures all updates to @work made
572 * here are visible to and precede any updates by the next PENDING
576 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
579 static void clear_work_data(struct work_struct *work)
581 smp_wmb(); /* see set_work_pool_and_clear_pending() */
582 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
585 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
587 unsigned long data = atomic_long_read(&work->data);
589 if (data & WORK_STRUCT_CWQ)
590 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
596 * get_work_pool - return the worker_pool a given work was associated with
597 * @work: the work item of interest
599 * Return the worker_pool @work was last associated with. %NULL if none.
601 static struct worker_pool *get_work_pool(struct work_struct *work)
603 unsigned long data = atomic_long_read(&work->data);
604 struct worker_pool *pool;
607 if (data & WORK_STRUCT_CWQ)
608 return ((struct cpu_workqueue_struct *)
609 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
611 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
612 if (pool_id == WORK_OFFQ_POOL_NONE)
615 pool = worker_pool_by_id(pool_id);
621 * get_work_pool_id - return the worker pool ID a given work is associated with
622 * @work: the work item of interest
624 * Return the worker_pool ID @work was last associated with.
625 * %WORK_OFFQ_POOL_NONE if none.
627 static int get_work_pool_id(struct work_struct *work)
629 struct worker_pool *pool = get_work_pool(work);
631 return pool ? pool->id : WORK_OFFQ_POOL_NONE;
634 static void mark_work_canceling(struct work_struct *work)
636 unsigned long pool_id = get_work_pool_id(work);
638 pool_id <<= WORK_OFFQ_POOL_SHIFT;
639 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
642 static bool work_is_canceling(struct work_struct *work)
644 unsigned long data = atomic_long_read(&work->data);
646 return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING);
650 * Policy functions. These define the policies on how the global worker
651 * pools are managed. Unless noted otherwise, these functions assume that
652 * they're being called with pool->lock held.
655 static bool __need_more_worker(struct worker_pool *pool)
657 return !atomic_read(get_pool_nr_running(pool));
661 * Need to wake up a worker? Called from anything but currently
664 * Note that, because unbound workers never contribute to nr_running, this
665 * function will always return %true for unbound pools as long as the
666 * worklist isn't empty.
668 static bool need_more_worker(struct worker_pool *pool)
670 return !list_empty(&pool->worklist) && __need_more_worker(pool);
673 /* Can I start working? Called from busy but !running workers. */
674 static bool may_start_working(struct worker_pool *pool)
676 return pool->nr_idle;
679 /* Do I need to keep working? Called from currently running workers. */
680 static bool keep_working(struct worker_pool *pool)
682 atomic_t *nr_running = get_pool_nr_running(pool);
684 return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
687 /* Do we need a new worker? Called from manager. */
688 static bool need_to_create_worker(struct worker_pool *pool)
690 return need_more_worker(pool) && !may_start_working(pool);
693 /* Do I need to be the manager? */
694 static bool need_to_manage_workers(struct worker_pool *pool)
696 return need_to_create_worker(pool) ||
697 (pool->flags & POOL_MANAGE_WORKERS);
700 /* Do we have too many workers and should some go away? */
701 static bool too_many_workers(struct worker_pool *pool)
703 bool managing = pool->flags & POOL_MANAGING_WORKERS;
704 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
705 int nr_busy = pool->nr_workers - nr_idle;
708 * nr_idle and idle_list may disagree if idle rebinding is in
709 * progress. Never return %true if idle_list is empty.
711 if (list_empty(&pool->idle_list))
714 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
721 /* Return the first worker. Safe with preemption disabled */
722 static struct worker *first_worker(struct worker_pool *pool)
724 if (unlikely(list_empty(&pool->idle_list)))
727 return list_first_entry(&pool->idle_list, struct worker, entry);
731 * wake_up_worker - wake up an idle worker
732 * @pool: worker pool to wake worker from
734 * Wake up the first idle worker of @pool.
737 * spin_lock_irq(pool->lock).
739 static void wake_up_worker(struct worker_pool *pool)
741 struct worker *worker = first_worker(pool);
744 wake_up_process(worker->task);
748 * wq_worker_waking_up - a worker is waking up
749 * @task: task waking up
750 * @cpu: CPU @task is waking up to
752 * This function is called during try_to_wake_up() when a worker is
756 * spin_lock_irq(rq->lock)
758 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
760 struct worker *worker = kthread_data(task);
762 if (!(worker->flags & WORKER_NOT_RUNNING)) {
763 WARN_ON_ONCE(worker->pool->cpu != cpu);
764 atomic_inc(get_pool_nr_running(worker->pool));
769 * wq_worker_sleeping - a worker is going to sleep
770 * @task: task going to sleep
771 * @cpu: CPU in question, must be the current CPU number
773 * This function is called during schedule() when a busy worker is
774 * going to sleep. Worker on the same cpu can be woken up by
775 * returning pointer to its task.
778 * spin_lock_irq(rq->lock)
781 * Worker task on @cpu to wake up, %NULL if none.
783 struct task_struct *wq_worker_sleeping(struct task_struct *task,
786 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
787 struct worker_pool *pool;
788 atomic_t *nr_running;
791 * Rescuers, which may not have all the fields set up like normal
792 * workers, also reach here, let's not access anything before
793 * checking NOT_RUNNING.
795 if (worker->flags & WORKER_NOT_RUNNING)
799 nr_running = get_pool_nr_running(pool);
801 /* this can only happen on the local cpu */
802 BUG_ON(cpu != raw_smp_processor_id());
805 * The counterpart of the following dec_and_test, implied mb,
806 * worklist not empty test sequence is in insert_work().
807 * Please read comment there.
809 * NOT_RUNNING is clear. This means that we're bound to and
810 * running on the local cpu w/ rq lock held and preemption
811 * disabled, which in turn means that none else could be
812 * manipulating idle_list, so dereferencing idle_list without pool
815 if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
816 to_wakeup = first_worker(pool);
817 return to_wakeup ? to_wakeup->task : NULL;
821 * worker_set_flags - set worker flags and adjust nr_running accordingly
823 * @flags: flags to set
824 * @wakeup: wakeup an idle worker if necessary
826 * Set @flags in @worker->flags and adjust nr_running accordingly. If
827 * nr_running becomes zero and @wakeup is %true, an idle worker is
831 * spin_lock_irq(pool->lock)
833 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
836 struct worker_pool *pool = worker->pool;
838 WARN_ON_ONCE(worker->task != current);
841 * If transitioning into NOT_RUNNING, adjust nr_running and
842 * wake up an idle worker as necessary if requested by
845 if ((flags & WORKER_NOT_RUNNING) &&
846 !(worker->flags & WORKER_NOT_RUNNING)) {
847 atomic_t *nr_running = get_pool_nr_running(pool);
850 if (atomic_dec_and_test(nr_running) &&
851 !list_empty(&pool->worklist))
852 wake_up_worker(pool);
854 atomic_dec(nr_running);
857 worker->flags |= flags;
861 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
863 * @flags: flags to clear
865 * Clear @flags in @worker->flags and adjust nr_running accordingly.
868 * spin_lock_irq(pool->lock)
870 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
872 struct worker_pool *pool = worker->pool;
873 unsigned int oflags = worker->flags;
875 WARN_ON_ONCE(worker->task != current);
877 worker->flags &= ~flags;
880 * If transitioning out of NOT_RUNNING, increment nr_running. Note
881 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
882 * of multiple flags, not a single flag.
884 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
885 if (!(worker->flags & WORKER_NOT_RUNNING))
886 atomic_inc(get_pool_nr_running(pool));
890 * find_worker_executing_work - find worker which is executing a work
891 * @pool: pool of interest
892 * @work: work to find worker for
894 * Find a worker which is executing @work on @pool by searching
895 * @pool->busy_hash which is keyed by the address of @work. For a worker
896 * to match, its current execution should match the address of @work and
897 * its work function. This is to avoid unwanted dependency between
898 * unrelated work executions through a work item being recycled while still
901 * This is a bit tricky. A work item may be freed once its execution
902 * starts and nothing prevents the freed area from being recycled for
903 * another work item. If the same work item address ends up being reused
904 * before the original execution finishes, workqueue will identify the
905 * recycled work item as currently executing and make it wait until the
906 * current execution finishes, introducing an unwanted dependency.
908 * This function checks the work item address, work function and workqueue
909 * to avoid false positives. Note that this isn't complete as one may
910 * construct a work function which can introduce dependency onto itself
911 * through a recycled work item. Well, if somebody wants to shoot oneself
912 * in the foot that badly, there's only so much we can do, and if such
913 * deadlock actually occurs, it should be easy to locate the culprit work
917 * spin_lock_irq(pool->lock).
920 * Pointer to worker which is executing @work if found, NULL
923 static struct worker *find_worker_executing_work(struct worker_pool *pool,
924 struct work_struct *work)
926 struct worker *worker;
927 struct hlist_node *tmp;
929 hash_for_each_possible(pool->busy_hash, worker, tmp, hentry,
931 if (worker->current_work == work &&
932 worker->current_func == work->func)
939 * move_linked_works - move linked works to a list
940 * @work: start of series of works to be scheduled
941 * @head: target list to append @work to
942 * @nextp: out paramter for nested worklist walking
944 * Schedule linked works starting from @work to @head. Work series to
945 * be scheduled starts at @work and includes any consecutive work with
946 * WORK_STRUCT_LINKED set in its predecessor.
948 * If @nextp is not NULL, it's updated to point to the next work of
949 * the last scheduled work. This allows move_linked_works() to be
950 * nested inside outer list_for_each_entry_safe().
953 * spin_lock_irq(pool->lock).
955 static void move_linked_works(struct work_struct *work, struct list_head *head,
956 struct work_struct **nextp)
958 struct work_struct *n;
961 * Linked worklist will always end before the end of the list,
962 * use NULL for list head.
964 list_for_each_entry_safe_from(work, n, NULL, entry) {
965 list_move_tail(&work->entry, head);
966 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
971 * If we're already inside safe list traversal and have moved
972 * multiple works to the scheduled queue, the next position
973 * needs to be updated.
979 static void cwq_activate_delayed_work(struct work_struct *work)
981 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
983 trace_workqueue_activate_work(work);
984 move_linked_works(work, &cwq->pool->worklist, NULL);
985 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
989 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
991 struct work_struct *work = list_first_entry(&cwq->delayed_works,
992 struct work_struct, entry);
994 cwq_activate_delayed_work(work);
998 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
999 * @cwq: cwq of interest
1000 * @color: color of work which left the queue
1002 * A work either has completed or is removed from pending queue,
1003 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1006 * spin_lock_irq(pool->lock).
1008 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
1010 /* ignore uncolored works */
1011 if (color == WORK_NO_COLOR)
1014 cwq->nr_in_flight[color]--;
1017 if (!list_empty(&cwq->delayed_works)) {
1018 /* one down, submit a delayed one */
1019 if (cwq->nr_active < cwq->max_active)
1020 cwq_activate_first_delayed(cwq);
1023 /* is flush in progress and are we at the flushing tip? */
1024 if (likely(cwq->flush_color != color))
1027 /* are there still in-flight works? */
1028 if (cwq->nr_in_flight[color])
1031 /* this cwq is done, clear flush_color */
1032 cwq->flush_color = -1;
1035 * If this was the last cwq, wake up the first flusher. It
1036 * will handle the rest.
1038 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1039 complete(&cwq->wq->first_flusher->done);
1043 * try_to_grab_pending - steal work item from worklist and disable irq
1044 * @work: work item to steal
1045 * @is_dwork: @work is a delayed_work
1046 * @flags: place to store irq state
1048 * Try to grab PENDING bit of @work. This function can handle @work in any
1049 * stable state - idle, on timer or on worklist. Return values are
1051 * 1 if @work was pending and we successfully stole PENDING
1052 * 0 if @work was idle and we claimed PENDING
1053 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1054 * -ENOENT if someone else is canceling @work, this state may persist
1055 * for arbitrarily long
1057 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1058 * interrupted while holding PENDING and @work off queue, irq must be
1059 * disabled on entry. This, combined with delayed_work->timer being
1060 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1062 * On successful return, >= 0, irq is disabled and the caller is
1063 * responsible for releasing it using local_irq_restore(*@flags).
1065 * This function is safe to call from any context including IRQ handler.
1067 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1068 unsigned long *flags)
1070 struct worker_pool *pool;
1071 struct cpu_workqueue_struct *cwq;
1073 local_irq_save(*flags);
1075 /* try to steal the timer if it exists */
1077 struct delayed_work *dwork = to_delayed_work(work);
1080 * dwork->timer is irqsafe. If del_timer() fails, it's
1081 * guaranteed that the timer is not queued anywhere and not
1082 * running on the local CPU.
1084 if (likely(del_timer(&dwork->timer)))
1088 /* try to claim PENDING the normal way */
1089 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1093 * The queueing is in progress, or it is already queued. Try to
1094 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1096 pool = get_work_pool(work);
1100 spin_lock(&pool->lock);
1102 * work->data is guaranteed to point to cwq only while the work
1103 * item is queued on cwq->wq, and both updating work->data to point
1104 * to cwq on queueing and to pool on dequeueing are done under
1105 * cwq->pool->lock. This in turn guarantees that, if work->data
1106 * points to cwq which is associated with a locked pool, the work
1107 * item is currently queued on that pool.
1109 cwq = get_work_cwq(work);
1111 if (cwq->pool == pool) {
1112 debug_work_deactivate(work);
1115 * A delayed work item cannot be grabbed directly
1116 * because it might have linked NO_COLOR work items
1117 * which, if left on the delayed_list, will confuse
1118 * cwq->nr_active management later on and cause
1119 * stall. Make sure the work item is activated
1122 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1123 cwq_activate_delayed_work(work);
1125 list_del_init(&work->entry);
1126 cwq_dec_nr_in_flight(get_work_cwq(work),
1127 get_work_color(work));
1129 /* work->data points to cwq iff queued, point to pool */
1130 set_work_pool_and_keep_pending(work, pool->id);
1132 spin_unlock(&pool->lock);
1136 spin_unlock(&pool->lock);
1138 local_irq_restore(*flags);
1139 if (work_is_canceling(work))
1146 * insert_work - insert a work into a pool
1147 * @cwq: cwq @work belongs to
1148 * @work: work to insert
1149 * @head: insertion point
1150 * @extra_flags: extra WORK_STRUCT_* flags to set
1152 * Insert @work which belongs to @cwq after @head. @extra_flags is or'd to
1153 * work_struct flags.
1156 * spin_lock_irq(pool->lock).
1158 static void insert_work(struct cpu_workqueue_struct *cwq,
1159 struct work_struct *work, struct list_head *head,
1160 unsigned int extra_flags)
1162 struct worker_pool *pool = cwq->pool;
1164 /* we own @work, set data and link */
1165 set_work_cwq(work, cwq, extra_flags);
1166 list_add_tail(&work->entry, head);
1169 * Ensure either worker_sched_deactivated() sees the above
1170 * list_add_tail() or we see zero nr_running to avoid workers
1171 * lying around lazily while there are works to be processed.
1175 if (__need_more_worker(pool))
1176 wake_up_worker(pool);
1180 * Test whether @work is being queued from another work executing on the
1181 * same workqueue. This is rather expensive and should only be used from
1184 static bool is_chained_work(struct workqueue_struct *wq)
1186 unsigned long flags;
1189 for_each_wq_cpu(cpu) {
1190 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
1191 struct worker_pool *pool = cwq->pool;
1192 struct worker *worker;
1193 struct hlist_node *pos;
1196 spin_lock_irqsave(&pool->lock, flags);
1197 for_each_busy_worker(worker, i, pos, pool) {
1198 if (worker->task != current)
1200 spin_unlock_irqrestore(&pool->lock, flags);
1202 * I'm @worker, no locking necessary. See if @work
1203 * is headed to the same workqueue.
1205 return worker->current_cwq->wq == wq;
1207 spin_unlock_irqrestore(&pool->lock, flags);
1212 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1213 struct work_struct *work)
1215 bool highpri = wq->flags & WQ_HIGHPRI;
1216 struct worker_pool *pool;
1217 struct cpu_workqueue_struct *cwq;
1218 struct list_head *worklist;
1219 unsigned int work_flags;
1220 unsigned int req_cpu = cpu;
1223 * While a work item is PENDING && off queue, a task trying to
1224 * steal the PENDING will busy-loop waiting for it to either get
1225 * queued or lose PENDING. Grabbing PENDING and queueing should
1226 * happen with IRQ disabled.
1228 WARN_ON_ONCE(!irqs_disabled());
1230 debug_work_activate(work);
1232 /* if dying, only works from the same workqueue are allowed */
1233 if (unlikely(wq->flags & WQ_DRAINING) &&
1234 WARN_ON_ONCE(!is_chained_work(wq)))
1237 /* determine pool to use */
1238 if (!(wq->flags & WQ_UNBOUND)) {
1239 struct worker_pool *last_pool;
1241 if (cpu == WORK_CPU_UNBOUND)
1242 cpu = raw_smp_processor_id();
1245 * It's multi cpu. If @work was previously on a different
1246 * cpu, it might still be running there, in which case the
1247 * work needs to be queued on that cpu to guarantee
1250 pool = get_std_worker_pool(cpu, highpri);
1251 last_pool = get_work_pool(work);
1253 if (last_pool && last_pool != pool) {
1254 struct worker *worker;
1256 spin_lock(&last_pool->lock);
1258 worker = find_worker_executing_work(last_pool, work);
1260 if (worker && worker->current_cwq->wq == wq)
1263 /* meh... not running there, queue here */
1264 spin_unlock(&last_pool->lock);
1265 spin_lock(&pool->lock);
1268 spin_lock(&pool->lock);
1271 pool = get_std_worker_pool(WORK_CPU_UNBOUND, highpri);
1272 spin_lock(&pool->lock);
1275 /* pool determined, get cwq and queue */
1276 cwq = get_cwq(pool->cpu, wq);
1277 trace_workqueue_queue_work(req_cpu, cwq, work);
1279 if (WARN_ON(!list_empty(&work->entry))) {
1280 spin_unlock(&pool->lock);
1284 cwq->nr_in_flight[cwq->work_color]++;
1285 work_flags = work_color_to_flags(cwq->work_color);
1287 if (likely(cwq->nr_active < cwq->max_active)) {
1288 trace_workqueue_activate_work(work);
1290 worklist = &cwq->pool->worklist;
1292 work_flags |= WORK_STRUCT_DELAYED;
1293 worklist = &cwq->delayed_works;
1296 insert_work(cwq, work, worklist, work_flags);
1298 spin_unlock(&pool->lock);
1302 * queue_work_on - queue work on specific cpu
1303 * @cpu: CPU number to execute work on
1304 * @wq: workqueue to use
1305 * @work: work to queue
1307 * Returns %false if @work was already on a queue, %true otherwise.
1309 * We queue the work to a specific CPU, the caller must ensure it
1312 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1313 struct work_struct *work)
1316 unsigned long flags;
1318 local_irq_save(flags);
1320 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1321 __queue_work(cpu, wq, work);
1325 local_irq_restore(flags);
1328 EXPORT_SYMBOL_GPL(queue_work_on);
1331 * queue_work - queue work on a workqueue
1332 * @wq: workqueue to use
1333 * @work: work to queue
1335 * Returns %false if @work was already on a queue, %true otherwise.
1337 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1338 * it can be processed by another CPU.
1340 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1342 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1344 EXPORT_SYMBOL_GPL(queue_work);
1346 void delayed_work_timer_fn(unsigned long __data)
1348 struct delayed_work *dwork = (struct delayed_work *)__data;
1350 /* should have been called from irqsafe timer with irq already off */
1351 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1353 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1355 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1356 struct delayed_work *dwork, unsigned long delay)
1358 struct timer_list *timer = &dwork->timer;
1359 struct work_struct *work = &dwork->work;
1361 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1362 timer->data != (unsigned long)dwork);
1363 WARN_ON_ONCE(timer_pending(timer));
1364 WARN_ON_ONCE(!list_empty(&work->entry));
1367 * If @delay is 0, queue @dwork->work immediately. This is for
1368 * both optimization and correctness. The earliest @timer can
1369 * expire is on the closest next tick and delayed_work users depend
1370 * on that there's no such delay when @delay is 0.
1373 __queue_work(cpu, wq, &dwork->work);
1377 timer_stats_timer_set_start_info(&dwork->timer);
1381 timer->expires = jiffies + delay;
1383 if (unlikely(cpu != WORK_CPU_UNBOUND))
1384 add_timer_on(timer, cpu);
1390 * queue_delayed_work_on - queue work on specific CPU after delay
1391 * @cpu: CPU number to execute work on
1392 * @wq: workqueue to use
1393 * @dwork: work to queue
1394 * @delay: number of jiffies to wait before queueing
1396 * Returns %false if @work was already on a queue, %true otherwise. If
1397 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1400 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1401 struct delayed_work *dwork, unsigned long delay)
1403 struct work_struct *work = &dwork->work;
1405 unsigned long flags;
1407 /* read the comment in __queue_work() */
1408 local_irq_save(flags);
1410 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1411 __queue_delayed_work(cpu, wq, dwork, delay);
1415 local_irq_restore(flags);
1418 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1421 * queue_delayed_work - queue work on a workqueue after delay
1422 * @wq: workqueue to use
1423 * @dwork: delayable work to queue
1424 * @delay: number of jiffies to wait before queueing
1426 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1428 bool queue_delayed_work(struct workqueue_struct *wq,
1429 struct delayed_work *dwork, unsigned long delay)
1431 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1433 EXPORT_SYMBOL_GPL(queue_delayed_work);
1436 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1437 * @cpu: CPU number to execute work on
1438 * @wq: workqueue to use
1439 * @dwork: work to queue
1440 * @delay: number of jiffies to wait before queueing
1442 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1443 * modify @dwork's timer so that it expires after @delay. If @delay is
1444 * zero, @work is guaranteed to be scheduled immediately regardless of its
1447 * Returns %false if @dwork was idle and queued, %true if @dwork was
1448 * pending and its timer was modified.
1450 * This function is safe to call from any context including IRQ handler.
1451 * See try_to_grab_pending() for details.
1453 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1454 struct delayed_work *dwork, unsigned long delay)
1456 unsigned long flags;
1460 ret = try_to_grab_pending(&dwork->work, true, &flags);
1461 } while (unlikely(ret == -EAGAIN));
1463 if (likely(ret >= 0)) {
1464 __queue_delayed_work(cpu, wq, dwork, delay);
1465 local_irq_restore(flags);
1468 /* -ENOENT from try_to_grab_pending() becomes %true */
1471 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1474 * mod_delayed_work - modify delay of or queue a delayed work
1475 * @wq: workqueue to use
1476 * @dwork: work to queue
1477 * @delay: number of jiffies to wait before queueing
1479 * mod_delayed_work_on() on local CPU.
1481 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1482 unsigned long delay)
1484 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1486 EXPORT_SYMBOL_GPL(mod_delayed_work);
1489 * worker_enter_idle - enter idle state
1490 * @worker: worker which is entering idle state
1492 * @worker is entering idle state. Update stats and idle timer if
1496 * spin_lock_irq(pool->lock).
1498 static void worker_enter_idle(struct worker *worker)
1500 struct worker_pool *pool = worker->pool;
1502 BUG_ON(worker->flags & WORKER_IDLE);
1503 BUG_ON(!list_empty(&worker->entry) &&
1504 (worker->hentry.next || worker->hentry.pprev));
1506 /* can't use worker_set_flags(), also called from start_worker() */
1507 worker->flags |= WORKER_IDLE;
1509 worker->last_active = jiffies;
1511 /* idle_list is LIFO */
1512 list_add(&worker->entry, &pool->idle_list);
1514 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1515 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1518 * Sanity check nr_running. Because wq_unbind_fn() releases
1519 * pool->lock between setting %WORKER_UNBOUND and zapping
1520 * nr_running, the warning may trigger spuriously. Check iff
1521 * unbind is not in progress.
1523 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1524 pool->nr_workers == pool->nr_idle &&
1525 atomic_read(get_pool_nr_running(pool)));
1529 * worker_leave_idle - leave idle state
1530 * @worker: worker which is leaving idle state
1532 * @worker is leaving idle state. Update stats.
1535 * spin_lock_irq(pool->lock).
1537 static void worker_leave_idle(struct worker *worker)
1539 struct worker_pool *pool = worker->pool;
1541 BUG_ON(!(worker->flags & WORKER_IDLE));
1542 worker_clr_flags(worker, WORKER_IDLE);
1544 list_del_init(&worker->entry);
1548 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock pool
1551 * Works which are scheduled while the cpu is online must at least be
1552 * scheduled to a worker which is bound to the cpu so that if they are
1553 * flushed from cpu callbacks while cpu is going down, they are
1554 * guaranteed to execute on the cpu.
1556 * This function is to be used by rogue workers and rescuers to bind
1557 * themselves to the target cpu and may race with cpu going down or
1558 * coming online. kthread_bind() can't be used because it may put the
1559 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1560 * verbatim as it's best effort and blocking and pool may be
1561 * [dis]associated in the meantime.
1563 * This function tries set_cpus_allowed() and locks pool and verifies the
1564 * binding against %POOL_DISASSOCIATED which is set during
1565 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1566 * enters idle state or fetches works without dropping lock, it can
1567 * guarantee the scheduling requirement described in the first paragraph.
1570 * Might sleep. Called without any lock but returns with pool->lock
1574 * %true if the associated pool is online (@worker is successfully
1575 * bound), %false if offline.
1577 static bool worker_maybe_bind_and_lock(struct worker *worker)
1578 __acquires(&pool->lock)
1580 struct worker_pool *pool = worker->pool;
1581 struct task_struct *task = worker->task;
1585 * The following call may fail, succeed or succeed
1586 * without actually migrating the task to the cpu if
1587 * it races with cpu hotunplug operation. Verify
1588 * against POOL_DISASSOCIATED.
1590 if (!(pool->flags & POOL_DISASSOCIATED))
1591 set_cpus_allowed_ptr(task, get_cpu_mask(pool->cpu));
1593 spin_lock_irq(&pool->lock);
1594 if (pool->flags & POOL_DISASSOCIATED)
1596 if (task_cpu(task) == pool->cpu &&
1597 cpumask_equal(¤t->cpus_allowed,
1598 get_cpu_mask(pool->cpu)))
1600 spin_unlock_irq(&pool->lock);
1603 * We've raced with CPU hot[un]plug. Give it a breather
1604 * and retry migration. cond_resched() is required here;
1605 * otherwise, we might deadlock against cpu_stop trying to
1606 * bring down the CPU on non-preemptive kernel.
1614 * Rebind an idle @worker to its CPU. worker_thread() will test
1615 * list_empty(@worker->entry) before leaving idle and call this function.
1617 static void idle_worker_rebind(struct worker *worker)
1619 /* CPU may go down again inbetween, clear UNBOUND only on success */
1620 if (worker_maybe_bind_and_lock(worker))
1621 worker_clr_flags(worker, WORKER_UNBOUND);
1623 /* rebind complete, become available again */
1624 list_add(&worker->entry, &worker->pool->idle_list);
1625 spin_unlock_irq(&worker->pool->lock);
1629 * Function for @worker->rebind.work used to rebind unbound busy workers to
1630 * the associated cpu which is coming back online. This is scheduled by
1631 * cpu up but can race with other cpu hotplug operations and may be
1632 * executed twice without intervening cpu down.
1634 static void busy_worker_rebind_fn(struct work_struct *work)
1636 struct worker *worker = container_of(work, struct worker, rebind_work);
1638 if (worker_maybe_bind_and_lock(worker))
1639 worker_clr_flags(worker, WORKER_UNBOUND);
1641 spin_unlock_irq(&worker->pool->lock);
1645 * rebind_workers - rebind all workers of a pool to the associated CPU
1646 * @pool: pool of interest
1648 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1649 * is different for idle and busy ones.
1651 * Idle ones will be removed from the idle_list and woken up. They will
1652 * add themselves back after completing rebind. This ensures that the
1653 * idle_list doesn't contain any unbound workers when re-bound busy workers
1654 * try to perform local wake-ups for concurrency management.
1656 * Busy workers can rebind after they finish their current work items.
1657 * Queueing the rebind work item at the head of the scheduled list is
1658 * enough. Note that nr_running will be properly bumped as busy workers
1661 * On return, all non-manager workers are scheduled for rebind - see
1662 * manage_workers() for the manager special case. Any idle worker
1663 * including the manager will not appear on @idle_list until rebind is
1664 * complete, making local wake-ups safe.
1666 static void rebind_workers(struct worker_pool *pool)
1668 struct worker *worker, *n;
1669 struct hlist_node *pos;
1672 lockdep_assert_held(&pool->assoc_mutex);
1673 lockdep_assert_held(&pool->lock);
1675 /* dequeue and kick idle ones */
1676 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1678 * idle workers should be off @pool->idle_list until rebind
1679 * is complete to avoid receiving premature local wake-ups.
1681 list_del_init(&worker->entry);
1684 * worker_thread() will see the above dequeuing and call
1685 * idle_worker_rebind().
1687 wake_up_process(worker->task);
1690 /* rebind busy workers */
1691 for_each_busy_worker(worker, i, pos, pool) {
1692 struct work_struct *rebind_work = &worker->rebind_work;
1693 struct workqueue_struct *wq;
1695 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1696 work_data_bits(rebind_work)))
1699 debug_work_activate(rebind_work);
1702 * wq doesn't really matter but let's keep @worker->pool
1703 * and @cwq->pool consistent for sanity.
1705 if (std_worker_pool_pri(worker->pool))
1706 wq = system_highpri_wq;
1710 insert_work(get_cwq(pool->cpu, wq), rebind_work,
1711 worker->scheduled.next,
1712 work_color_to_flags(WORK_NO_COLOR));
1716 static struct worker *alloc_worker(void)
1718 struct worker *worker;
1720 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1722 INIT_LIST_HEAD(&worker->entry);
1723 INIT_LIST_HEAD(&worker->scheduled);
1724 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1725 /* on creation a worker is in !idle && prep state */
1726 worker->flags = WORKER_PREP;
1732 * create_worker - create a new workqueue worker
1733 * @pool: pool the new worker will belong to
1735 * Create a new worker which is bound to @pool. The returned worker
1736 * can be started by calling start_worker() or destroyed using
1740 * Might sleep. Does GFP_KERNEL allocations.
1743 * Pointer to the newly created worker.
1745 static struct worker *create_worker(struct worker_pool *pool)
1747 const char *pri = std_worker_pool_pri(pool) ? "H" : "";
1748 struct worker *worker = NULL;
1751 spin_lock_irq(&pool->lock);
1752 while (ida_get_new(&pool->worker_ida, &id)) {
1753 spin_unlock_irq(&pool->lock);
1754 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1756 spin_lock_irq(&pool->lock);
1758 spin_unlock_irq(&pool->lock);
1760 worker = alloc_worker();
1764 worker->pool = pool;
1767 if (pool->cpu != WORK_CPU_UNBOUND)
1768 worker->task = kthread_create_on_node(worker_thread,
1769 worker, cpu_to_node(pool->cpu),
1770 "kworker/%u:%d%s", pool->cpu, id, pri);
1772 worker->task = kthread_create(worker_thread, worker,
1773 "kworker/u:%d%s", id, pri);
1774 if (IS_ERR(worker->task))
1777 if (std_worker_pool_pri(pool))
1778 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1781 * Determine CPU binding of the new worker depending on
1782 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1783 * flag remains stable across this function. See the comments
1784 * above the flag definition for details.
1786 * As an unbound worker may later become a regular one if CPU comes
1787 * online, make sure every worker has %PF_THREAD_BOUND set.
1789 if (!(pool->flags & POOL_DISASSOCIATED)) {
1790 kthread_bind(worker->task, pool->cpu);
1792 worker->task->flags |= PF_THREAD_BOUND;
1793 worker->flags |= WORKER_UNBOUND;
1799 spin_lock_irq(&pool->lock);
1800 ida_remove(&pool->worker_ida, id);
1801 spin_unlock_irq(&pool->lock);
1808 * start_worker - start a newly created worker
1809 * @worker: worker to start
1811 * Make the pool aware of @worker and start it.
1814 * spin_lock_irq(pool->lock).
1816 static void start_worker(struct worker *worker)
1818 worker->flags |= WORKER_STARTED;
1819 worker->pool->nr_workers++;
1820 worker_enter_idle(worker);
1821 wake_up_process(worker->task);
1825 * destroy_worker - destroy a workqueue worker
1826 * @worker: worker to be destroyed
1828 * Destroy @worker and adjust @pool stats accordingly.
1831 * spin_lock_irq(pool->lock) which is released and regrabbed.
1833 static void destroy_worker(struct worker *worker)
1835 struct worker_pool *pool = worker->pool;
1836 int id = worker->id;
1838 /* sanity check frenzy */
1839 BUG_ON(worker->current_work);
1840 BUG_ON(!list_empty(&worker->scheduled));
1842 if (worker->flags & WORKER_STARTED)
1844 if (worker->flags & WORKER_IDLE)
1847 list_del_init(&worker->entry);
1848 worker->flags |= WORKER_DIE;
1850 spin_unlock_irq(&pool->lock);
1852 kthread_stop(worker->task);
1855 spin_lock_irq(&pool->lock);
1856 ida_remove(&pool->worker_ida, id);
1859 static void idle_worker_timeout(unsigned long __pool)
1861 struct worker_pool *pool = (void *)__pool;
1863 spin_lock_irq(&pool->lock);
1865 if (too_many_workers(pool)) {
1866 struct worker *worker;
1867 unsigned long expires;
1869 /* idle_list is kept in LIFO order, check the last one */
1870 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1871 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1873 if (time_before(jiffies, expires))
1874 mod_timer(&pool->idle_timer, expires);
1876 /* it's been idle for too long, wake up manager */
1877 pool->flags |= POOL_MANAGE_WORKERS;
1878 wake_up_worker(pool);
1882 spin_unlock_irq(&pool->lock);
1885 static bool send_mayday(struct work_struct *work)
1887 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1888 struct workqueue_struct *wq = cwq->wq;
1891 if (!(wq->flags & WQ_RESCUER))
1894 /* mayday mayday mayday */
1895 cpu = cwq->pool->cpu;
1896 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1897 if (cpu == WORK_CPU_UNBOUND)
1899 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1900 wake_up_process(wq->rescuer->task);
1904 static void pool_mayday_timeout(unsigned long __pool)
1906 struct worker_pool *pool = (void *)__pool;
1907 struct work_struct *work;
1909 spin_lock_irq(&pool->lock);
1911 if (need_to_create_worker(pool)) {
1913 * We've been trying to create a new worker but
1914 * haven't been successful. We might be hitting an
1915 * allocation deadlock. Send distress signals to
1918 list_for_each_entry(work, &pool->worklist, entry)
1922 spin_unlock_irq(&pool->lock);
1924 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1928 * maybe_create_worker - create a new worker if necessary
1929 * @pool: pool to create a new worker for
1931 * Create a new worker for @pool if necessary. @pool is guaranteed to
1932 * have at least one idle worker on return from this function. If
1933 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1934 * sent to all rescuers with works scheduled on @pool to resolve
1935 * possible allocation deadlock.
1937 * On return, need_to_create_worker() is guaranteed to be false and
1938 * may_start_working() true.
1941 * spin_lock_irq(pool->lock) which may be released and regrabbed
1942 * multiple times. Does GFP_KERNEL allocations. Called only from
1946 * false if no action was taken and pool->lock stayed locked, true
1949 static bool maybe_create_worker(struct worker_pool *pool)
1950 __releases(&pool->lock)
1951 __acquires(&pool->lock)
1953 if (!need_to_create_worker(pool))
1956 spin_unlock_irq(&pool->lock);
1958 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1959 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1962 struct worker *worker;
1964 worker = create_worker(pool);
1966 del_timer_sync(&pool->mayday_timer);
1967 spin_lock_irq(&pool->lock);
1968 start_worker(worker);
1969 BUG_ON(need_to_create_worker(pool));
1973 if (!need_to_create_worker(pool))
1976 __set_current_state(TASK_INTERRUPTIBLE);
1977 schedule_timeout(CREATE_COOLDOWN);
1979 if (!need_to_create_worker(pool))
1983 del_timer_sync(&pool->mayday_timer);
1984 spin_lock_irq(&pool->lock);
1985 if (need_to_create_worker(pool))
1991 * maybe_destroy_worker - destroy workers which have been idle for a while
1992 * @pool: pool to destroy workers for
1994 * Destroy @pool workers which have been idle for longer than
1995 * IDLE_WORKER_TIMEOUT.
1998 * spin_lock_irq(pool->lock) which may be released and regrabbed
1999 * multiple times. Called only from manager.
2002 * false if no action was taken and pool->lock stayed locked, true
2005 static bool maybe_destroy_workers(struct worker_pool *pool)
2009 while (too_many_workers(pool)) {
2010 struct worker *worker;
2011 unsigned long expires;
2013 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2014 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2016 if (time_before(jiffies, expires)) {
2017 mod_timer(&pool->idle_timer, expires);
2021 destroy_worker(worker);
2029 * manage_workers - manage worker pool
2032 * Assume the manager role and manage the worker pool @worker belongs
2033 * to. At any given time, there can be only zero or one manager per
2034 * pool. The exclusion is handled automatically by this function.
2036 * The caller can safely start processing works on false return. On
2037 * true return, it's guaranteed that need_to_create_worker() is false
2038 * and may_start_working() is true.
2041 * spin_lock_irq(pool->lock) which may be released and regrabbed
2042 * multiple times. Does GFP_KERNEL allocations.
2045 * spin_lock_irq(pool->lock) which may be released and regrabbed
2046 * multiple times. Does GFP_KERNEL allocations.
2048 static bool manage_workers(struct worker *worker)
2050 struct worker_pool *pool = worker->pool;
2053 if (pool->flags & POOL_MANAGING_WORKERS)
2056 pool->flags |= POOL_MANAGING_WORKERS;
2059 * To simplify both worker management and CPU hotplug, hold off
2060 * management while hotplug is in progress. CPU hotplug path can't
2061 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2062 * lead to idle worker depletion (all become busy thinking someone
2063 * else is managing) which in turn can result in deadlock under
2064 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2065 * manager against CPU hotplug.
2067 * assoc_mutex would always be free unless CPU hotplug is in
2068 * progress. trylock first without dropping @pool->lock.
2070 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2071 spin_unlock_irq(&pool->lock);
2072 mutex_lock(&pool->assoc_mutex);
2074 * CPU hotplug could have happened while we were waiting
2075 * for assoc_mutex. Hotplug itself can't handle us
2076 * because manager isn't either on idle or busy list, and
2077 * @pool's state and ours could have deviated.
2079 * As hotplug is now excluded via assoc_mutex, we can
2080 * simply try to bind. It will succeed or fail depending
2081 * on @pool's current state. Try it and adjust
2082 * %WORKER_UNBOUND accordingly.
2084 if (worker_maybe_bind_and_lock(worker))
2085 worker->flags &= ~WORKER_UNBOUND;
2087 worker->flags |= WORKER_UNBOUND;
2092 pool->flags &= ~POOL_MANAGE_WORKERS;
2095 * Destroy and then create so that may_start_working() is true
2098 ret |= maybe_destroy_workers(pool);
2099 ret |= maybe_create_worker(pool);
2101 pool->flags &= ~POOL_MANAGING_WORKERS;
2102 mutex_unlock(&pool->assoc_mutex);
2107 * process_one_work - process single work
2109 * @work: work to process
2111 * Process @work. This function contains all the logics necessary to
2112 * process a single work including synchronization against and
2113 * interaction with other workers on the same cpu, queueing and
2114 * flushing. As long as context requirement is met, any worker can
2115 * call this function to process a work.
2118 * spin_lock_irq(pool->lock) which is released and regrabbed.
2120 static void process_one_work(struct worker *worker, struct work_struct *work)
2121 __releases(&pool->lock)
2122 __acquires(&pool->lock)
2124 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2125 struct worker_pool *pool = worker->pool;
2126 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2128 struct worker *collision;
2129 #ifdef CONFIG_LOCKDEP
2131 * It is permissible to free the struct work_struct from
2132 * inside the function that is called from it, this we need to
2133 * take into account for lockdep too. To avoid bogus "held
2134 * lock freed" warnings as well as problems when looking into
2135 * work->lockdep_map, make a copy and use that here.
2137 struct lockdep_map lockdep_map;
2139 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2142 * Ensure we're on the correct CPU. DISASSOCIATED test is
2143 * necessary to avoid spurious warnings from rescuers servicing the
2144 * unbound or a disassociated pool.
2146 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2147 !(pool->flags & POOL_DISASSOCIATED) &&
2148 raw_smp_processor_id() != pool->cpu);
2151 * A single work shouldn't be executed concurrently by
2152 * multiple workers on a single cpu. Check whether anyone is
2153 * already processing the work. If so, defer the work to the
2154 * currently executing one.
2156 collision = find_worker_executing_work(pool, work);
2157 if (unlikely(collision)) {
2158 move_linked_works(work, &collision->scheduled, NULL);
2162 /* claim and dequeue */
2163 debug_work_deactivate(work);
2164 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2165 worker->current_work = work;
2166 worker->current_func = work->func;
2167 worker->current_cwq = cwq;
2168 work_color = get_work_color(work);
2170 list_del_init(&work->entry);
2173 * CPU intensive works don't participate in concurrency
2174 * management. They're the scheduler's responsibility.
2176 if (unlikely(cpu_intensive))
2177 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2180 * Unbound pool isn't concurrency managed and work items should be
2181 * executed ASAP. Wake up another worker if necessary.
2183 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2184 wake_up_worker(pool);
2187 * Record the last pool and clear PENDING which should be the last
2188 * update to @work. Also, do this inside @pool->lock so that
2189 * PENDING and queued state changes happen together while IRQ is
2192 set_work_pool_and_clear_pending(work, pool->id);
2194 spin_unlock_irq(&pool->lock);
2196 lock_map_acquire_read(&cwq->wq->lockdep_map);
2197 lock_map_acquire(&lockdep_map);
2198 trace_workqueue_execute_start(work);
2199 worker->current_func(work);
2201 * While we must be careful to not use "work" after this, the trace
2202 * point will only record its address.
2204 trace_workqueue_execute_end(work);
2205 lock_map_release(&lockdep_map);
2206 lock_map_release(&cwq->wq->lockdep_map);
2208 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2209 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2210 " last function: %pf\n",
2211 current->comm, preempt_count(), task_pid_nr(current),
2212 worker->current_func);
2213 debug_show_held_locks(current);
2217 spin_lock_irq(&pool->lock);
2219 /* clear cpu intensive status */
2220 if (unlikely(cpu_intensive))
2221 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2223 /* we're done with it, release */
2224 hash_del(&worker->hentry);
2225 worker->current_work = NULL;
2226 worker->current_func = NULL;
2227 worker->current_cwq = NULL;
2228 cwq_dec_nr_in_flight(cwq, work_color);
2232 * process_scheduled_works - process scheduled works
2235 * Process all scheduled works. Please note that the scheduled list
2236 * may change while processing a work, so this function repeatedly
2237 * fetches a work from the top and executes it.
2240 * spin_lock_irq(pool->lock) which may be released and regrabbed
2243 static void process_scheduled_works(struct worker *worker)
2245 while (!list_empty(&worker->scheduled)) {
2246 struct work_struct *work = list_first_entry(&worker->scheduled,
2247 struct work_struct, entry);
2248 process_one_work(worker, work);
2253 * worker_thread - the worker thread function
2256 * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools
2257 * of these per each cpu. These workers process all works regardless of
2258 * their specific target workqueue. The only exception is works which
2259 * belong to workqueues with a rescuer which will be explained in
2262 static int worker_thread(void *__worker)
2264 struct worker *worker = __worker;
2265 struct worker_pool *pool = worker->pool;
2267 /* tell the scheduler that this is a workqueue worker */
2268 worker->task->flags |= PF_WQ_WORKER;
2270 spin_lock_irq(&pool->lock);
2272 /* we are off idle list if destruction or rebind is requested */
2273 if (unlikely(list_empty(&worker->entry))) {
2274 spin_unlock_irq(&pool->lock);
2276 /* if DIE is set, destruction is requested */
2277 if (worker->flags & WORKER_DIE) {
2278 worker->task->flags &= ~PF_WQ_WORKER;
2282 /* otherwise, rebind */
2283 idle_worker_rebind(worker);
2287 worker_leave_idle(worker);
2289 /* no more worker necessary? */
2290 if (!need_more_worker(pool))
2293 /* do we need to manage? */
2294 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2298 * ->scheduled list can only be filled while a worker is
2299 * preparing to process a work or actually processing it.
2300 * Make sure nobody diddled with it while I was sleeping.
2302 BUG_ON(!list_empty(&worker->scheduled));
2305 * When control reaches this point, we're guaranteed to have
2306 * at least one idle worker or that someone else has already
2307 * assumed the manager role.
2309 worker_clr_flags(worker, WORKER_PREP);
2312 struct work_struct *work =
2313 list_first_entry(&pool->worklist,
2314 struct work_struct, entry);
2316 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2317 /* optimization path, not strictly necessary */
2318 process_one_work(worker, work);
2319 if (unlikely(!list_empty(&worker->scheduled)))
2320 process_scheduled_works(worker);
2322 move_linked_works(work, &worker->scheduled, NULL);
2323 process_scheduled_works(worker);
2325 } while (keep_working(pool));
2327 worker_set_flags(worker, WORKER_PREP, false);
2329 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2333 * pool->lock is held and there's no work to process and no need to
2334 * manage, sleep. Workers are woken up only while holding
2335 * pool->lock or from local cpu, so setting the current state
2336 * before releasing pool->lock is enough to prevent losing any
2339 worker_enter_idle(worker);
2340 __set_current_state(TASK_INTERRUPTIBLE);
2341 spin_unlock_irq(&pool->lock);
2347 * rescuer_thread - the rescuer thread function
2350 * Workqueue rescuer thread function. There's one rescuer for each
2351 * workqueue which has WQ_RESCUER set.
2353 * Regular work processing on a pool may block trying to create a new
2354 * worker which uses GFP_KERNEL allocation which has slight chance of
2355 * developing into deadlock if some works currently on the same queue
2356 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2357 * the problem rescuer solves.
2359 * When such condition is possible, the pool summons rescuers of all
2360 * workqueues which have works queued on the pool and let them process
2361 * those works so that forward progress can be guaranteed.
2363 * This should happen rarely.
2365 static int rescuer_thread(void *__rescuer)
2367 struct worker *rescuer = __rescuer;
2368 struct workqueue_struct *wq = rescuer->rescue_wq;
2369 struct list_head *scheduled = &rescuer->scheduled;
2370 bool is_unbound = wq->flags & WQ_UNBOUND;
2373 set_user_nice(current, RESCUER_NICE_LEVEL);
2376 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2377 * doesn't participate in concurrency management.
2379 rescuer->task->flags |= PF_WQ_WORKER;
2381 set_current_state(TASK_INTERRUPTIBLE);
2383 if (kthread_should_stop()) {
2384 __set_current_state(TASK_RUNNING);
2385 rescuer->task->flags &= ~PF_WQ_WORKER;
2390 * See whether any cpu is asking for help. Unbounded
2391 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2393 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2394 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2395 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2396 struct worker_pool *pool = cwq->pool;
2397 struct work_struct *work, *n;
2399 __set_current_state(TASK_RUNNING);
2400 mayday_clear_cpu(cpu, wq->mayday_mask);
2402 /* migrate to the target cpu if possible */
2403 rescuer->pool = pool;
2404 worker_maybe_bind_and_lock(rescuer);
2407 * Slurp in all works issued via this workqueue and
2410 BUG_ON(!list_empty(&rescuer->scheduled));
2411 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2412 if (get_work_cwq(work) == cwq)
2413 move_linked_works(work, scheduled, &n);
2415 process_scheduled_works(rescuer);
2418 * Leave this pool. If keep_working() is %true, notify a
2419 * regular worker; otherwise, we end up with 0 concurrency
2420 * and stalling the execution.
2422 if (keep_working(pool))
2423 wake_up_worker(pool);
2425 spin_unlock_irq(&pool->lock);
2428 /* rescuers should never participate in concurrency management */
2429 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2435 struct work_struct work;
2436 struct completion done;
2439 static void wq_barrier_func(struct work_struct *work)
2441 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2442 complete(&barr->done);
2446 * insert_wq_barrier - insert a barrier work
2447 * @cwq: cwq to insert barrier into
2448 * @barr: wq_barrier to insert
2449 * @target: target work to attach @barr to
2450 * @worker: worker currently executing @target, NULL if @target is not executing
2452 * @barr is linked to @target such that @barr is completed only after
2453 * @target finishes execution. Please note that the ordering
2454 * guarantee is observed only with respect to @target and on the local
2457 * Currently, a queued barrier can't be canceled. This is because
2458 * try_to_grab_pending() can't determine whether the work to be
2459 * grabbed is at the head of the queue and thus can't clear LINKED
2460 * flag of the previous work while there must be a valid next work
2461 * after a work with LINKED flag set.
2463 * Note that when @worker is non-NULL, @target may be modified
2464 * underneath us, so we can't reliably determine cwq from @target.
2467 * spin_lock_irq(pool->lock).
2469 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2470 struct wq_barrier *barr,
2471 struct work_struct *target, struct worker *worker)
2473 struct list_head *head;
2474 unsigned int linked = 0;
2477 * debugobject calls are safe here even with pool->lock locked
2478 * as we know for sure that this will not trigger any of the
2479 * checks and call back into the fixup functions where we
2482 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2483 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2484 init_completion(&barr->done);
2487 * If @target is currently being executed, schedule the
2488 * barrier to the worker; otherwise, put it after @target.
2491 head = worker->scheduled.next;
2493 unsigned long *bits = work_data_bits(target);
2495 head = target->entry.next;
2496 /* there can already be other linked works, inherit and set */
2497 linked = *bits & WORK_STRUCT_LINKED;
2498 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2501 debug_work_activate(&barr->work);
2502 insert_work(cwq, &barr->work, head,
2503 work_color_to_flags(WORK_NO_COLOR) | linked);
2507 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2508 * @wq: workqueue being flushed
2509 * @flush_color: new flush color, < 0 for no-op
2510 * @work_color: new work color, < 0 for no-op
2512 * Prepare cwqs for workqueue flushing.
2514 * If @flush_color is non-negative, flush_color on all cwqs should be
2515 * -1. If no cwq has in-flight commands at the specified color, all
2516 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2517 * has in flight commands, its cwq->flush_color is set to
2518 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2519 * wakeup logic is armed and %true is returned.
2521 * The caller should have initialized @wq->first_flusher prior to
2522 * calling this function with non-negative @flush_color. If
2523 * @flush_color is negative, no flush color update is done and %false
2526 * If @work_color is non-negative, all cwqs should have the same
2527 * work_color which is previous to @work_color and all will be
2528 * advanced to @work_color.
2531 * mutex_lock(wq->flush_mutex).
2534 * %true if @flush_color >= 0 and there's something to flush. %false
2537 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2538 int flush_color, int work_color)
2543 if (flush_color >= 0) {
2544 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2545 atomic_set(&wq->nr_cwqs_to_flush, 1);
2548 for_each_cwq_cpu(cpu, wq) {
2549 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2550 struct worker_pool *pool = cwq->pool;
2552 spin_lock_irq(&pool->lock);
2554 if (flush_color >= 0) {
2555 BUG_ON(cwq->flush_color != -1);
2557 if (cwq->nr_in_flight[flush_color]) {
2558 cwq->flush_color = flush_color;
2559 atomic_inc(&wq->nr_cwqs_to_flush);
2564 if (work_color >= 0) {
2565 BUG_ON(work_color != work_next_color(cwq->work_color));
2566 cwq->work_color = work_color;
2569 spin_unlock_irq(&pool->lock);
2572 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2573 complete(&wq->first_flusher->done);
2579 * flush_workqueue - ensure that any scheduled work has run to completion.
2580 * @wq: workqueue to flush
2582 * Forces execution of the workqueue and blocks until its completion.
2583 * This is typically used in driver shutdown handlers.
2585 * We sleep until all works which were queued on entry have been handled,
2586 * but we are not livelocked by new incoming ones.
2588 void flush_workqueue(struct workqueue_struct *wq)
2590 struct wq_flusher this_flusher = {
2591 .list = LIST_HEAD_INIT(this_flusher.list),
2593 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2597 lock_map_acquire(&wq->lockdep_map);
2598 lock_map_release(&wq->lockdep_map);
2600 mutex_lock(&wq->flush_mutex);
2603 * Start-to-wait phase
2605 next_color = work_next_color(wq->work_color);
2607 if (next_color != wq->flush_color) {
2609 * Color space is not full. The current work_color
2610 * becomes our flush_color and work_color is advanced
2613 BUG_ON(!list_empty(&wq->flusher_overflow));
2614 this_flusher.flush_color = wq->work_color;
2615 wq->work_color = next_color;
2617 if (!wq->first_flusher) {
2618 /* no flush in progress, become the first flusher */
2619 BUG_ON(wq->flush_color != this_flusher.flush_color);
2621 wq->first_flusher = &this_flusher;
2623 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2625 /* nothing to flush, done */
2626 wq->flush_color = next_color;
2627 wq->first_flusher = NULL;
2632 BUG_ON(wq->flush_color == this_flusher.flush_color);
2633 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2634 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2638 * Oops, color space is full, wait on overflow queue.
2639 * The next flush completion will assign us
2640 * flush_color and transfer to flusher_queue.
2642 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2645 mutex_unlock(&wq->flush_mutex);
2647 wait_for_completion(&this_flusher.done);
2650 * Wake-up-and-cascade phase
2652 * First flushers are responsible for cascading flushes and
2653 * handling overflow. Non-first flushers can simply return.
2655 if (wq->first_flusher != &this_flusher)
2658 mutex_lock(&wq->flush_mutex);
2660 /* we might have raced, check again with mutex held */
2661 if (wq->first_flusher != &this_flusher)
2664 wq->first_flusher = NULL;
2666 BUG_ON(!list_empty(&this_flusher.list));
2667 BUG_ON(wq->flush_color != this_flusher.flush_color);
2670 struct wq_flusher *next, *tmp;
2672 /* complete all the flushers sharing the current flush color */
2673 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2674 if (next->flush_color != wq->flush_color)
2676 list_del_init(&next->list);
2677 complete(&next->done);
2680 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2681 wq->flush_color != work_next_color(wq->work_color));
2683 /* this flush_color is finished, advance by one */
2684 wq->flush_color = work_next_color(wq->flush_color);
2686 /* one color has been freed, handle overflow queue */
2687 if (!list_empty(&wq->flusher_overflow)) {
2689 * Assign the same color to all overflowed
2690 * flushers, advance work_color and append to
2691 * flusher_queue. This is the start-to-wait
2692 * phase for these overflowed flushers.
2694 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2695 tmp->flush_color = wq->work_color;
2697 wq->work_color = work_next_color(wq->work_color);
2699 list_splice_tail_init(&wq->flusher_overflow,
2700 &wq->flusher_queue);
2701 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2704 if (list_empty(&wq->flusher_queue)) {
2705 BUG_ON(wq->flush_color != wq->work_color);
2710 * Need to flush more colors. Make the next flusher
2711 * the new first flusher and arm cwqs.
2713 BUG_ON(wq->flush_color == wq->work_color);
2714 BUG_ON(wq->flush_color != next->flush_color);
2716 list_del_init(&next->list);
2717 wq->first_flusher = next;
2719 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2723 * Meh... this color is already done, clear first
2724 * flusher and repeat cascading.
2726 wq->first_flusher = NULL;
2730 mutex_unlock(&wq->flush_mutex);
2732 EXPORT_SYMBOL_GPL(flush_workqueue);
2735 * drain_workqueue - drain a workqueue
2736 * @wq: workqueue to drain
2738 * Wait until the workqueue becomes empty. While draining is in progress,
2739 * only chain queueing is allowed. IOW, only currently pending or running
2740 * work items on @wq can queue further work items on it. @wq is flushed
2741 * repeatedly until it becomes empty. The number of flushing is detemined
2742 * by the depth of chaining and should be relatively short. Whine if it
2745 void drain_workqueue(struct workqueue_struct *wq)
2747 unsigned int flush_cnt = 0;
2751 * __queue_work() needs to test whether there are drainers, is much
2752 * hotter than drain_workqueue() and already looks at @wq->flags.
2753 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2755 spin_lock(&workqueue_lock);
2756 if (!wq->nr_drainers++)
2757 wq->flags |= WQ_DRAINING;
2758 spin_unlock(&workqueue_lock);
2760 flush_workqueue(wq);
2762 for_each_cwq_cpu(cpu, wq) {
2763 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2766 spin_lock_irq(&cwq->pool->lock);
2767 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2768 spin_unlock_irq(&cwq->pool->lock);
2773 if (++flush_cnt == 10 ||
2774 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2775 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2776 wq->name, flush_cnt);
2780 spin_lock(&workqueue_lock);
2781 if (!--wq->nr_drainers)
2782 wq->flags &= ~WQ_DRAINING;
2783 spin_unlock(&workqueue_lock);
2785 EXPORT_SYMBOL_GPL(drain_workqueue);
2787 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2789 struct worker *worker = NULL;
2790 struct worker_pool *pool;
2791 struct cpu_workqueue_struct *cwq;
2794 pool = get_work_pool(work);
2798 spin_lock_irq(&pool->lock);
2799 /* see the comment in try_to_grab_pending() with the same code */
2800 cwq = get_work_cwq(work);
2802 if (unlikely(cwq->pool != pool))
2805 worker = find_worker_executing_work(pool, work);
2808 cwq = worker->current_cwq;
2811 insert_wq_barrier(cwq, barr, work, worker);
2812 spin_unlock_irq(&pool->lock);
2815 * If @max_active is 1 or rescuer is in use, flushing another work
2816 * item on the same workqueue may lead to deadlock. Make sure the
2817 * flusher is not running on the same workqueue by verifying write
2820 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2821 lock_map_acquire(&cwq->wq->lockdep_map);
2823 lock_map_acquire_read(&cwq->wq->lockdep_map);
2824 lock_map_release(&cwq->wq->lockdep_map);
2828 spin_unlock_irq(&pool->lock);
2833 * flush_work - wait for a work to finish executing the last queueing instance
2834 * @work: the work to flush
2836 * Wait until @work has finished execution. @work is guaranteed to be idle
2837 * on return if it hasn't been requeued since flush started.
2840 * %true if flush_work() waited for the work to finish execution,
2841 * %false if it was already idle.
2843 bool flush_work(struct work_struct *work)
2845 struct wq_barrier barr;
2847 lock_map_acquire(&work->lockdep_map);
2848 lock_map_release(&work->lockdep_map);
2850 if (start_flush_work(work, &barr)) {
2851 wait_for_completion(&barr.done);
2852 destroy_work_on_stack(&barr.work);
2858 EXPORT_SYMBOL_GPL(flush_work);
2860 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2862 unsigned long flags;
2866 ret = try_to_grab_pending(work, is_dwork, &flags);
2868 * If someone else is canceling, wait for the same event it
2869 * would be waiting for before retrying.
2871 if (unlikely(ret == -ENOENT))
2873 } while (unlikely(ret < 0));
2875 /* tell other tasks trying to grab @work to back off */
2876 mark_work_canceling(work);
2877 local_irq_restore(flags);
2880 clear_work_data(work);
2885 * cancel_work_sync - cancel a work and wait for it to finish
2886 * @work: the work to cancel
2888 * Cancel @work and wait for its execution to finish. This function
2889 * can be used even if the work re-queues itself or migrates to
2890 * another workqueue. On return from this function, @work is
2891 * guaranteed to be not pending or executing on any CPU.
2893 * cancel_work_sync(&delayed_work->work) must not be used for
2894 * delayed_work's. Use cancel_delayed_work_sync() instead.
2896 * The caller must ensure that the workqueue on which @work was last
2897 * queued can't be destroyed before this function returns.
2900 * %true if @work was pending, %false otherwise.
2902 bool cancel_work_sync(struct work_struct *work)
2904 return __cancel_work_timer(work, false);
2906 EXPORT_SYMBOL_GPL(cancel_work_sync);
2909 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2910 * @dwork: the delayed work to flush
2912 * Delayed timer is cancelled and the pending work is queued for
2913 * immediate execution. Like flush_work(), this function only
2914 * considers the last queueing instance of @dwork.
2917 * %true if flush_work() waited for the work to finish execution,
2918 * %false if it was already idle.
2920 bool flush_delayed_work(struct delayed_work *dwork)
2922 local_irq_disable();
2923 if (del_timer_sync(&dwork->timer))
2924 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2926 return flush_work(&dwork->work);
2928 EXPORT_SYMBOL(flush_delayed_work);
2931 * cancel_delayed_work - cancel a delayed work
2932 * @dwork: delayed_work to cancel
2934 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2935 * and canceled; %false if wasn't pending. Note that the work callback
2936 * function may still be running on return, unless it returns %true and the
2937 * work doesn't re-arm itself. Explicitly flush or use
2938 * cancel_delayed_work_sync() to wait on it.
2940 * This function is safe to call from any context including IRQ handler.
2942 bool cancel_delayed_work(struct delayed_work *dwork)
2944 unsigned long flags;
2948 ret = try_to_grab_pending(&dwork->work, true, &flags);
2949 } while (unlikely(ret == -EAGAIN));
2951 if (unlikely(ret < 0))
2954 set_work_pool_and_clear_pending(&dwork->work,
2955 get_work_pool_id(&dwork->work));
2956 local_irq_restore(flags);
2959 EXPORT_SYMBOL(cancel_delayed_work);
2962 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2963 * @dwork: the delayed work cancel
2965 * This is cancel_work_sync() for delayed works.
2968 * %true if @dwork was pending, %false otherwise.
2970 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2972 return __cancel_work_timer(&dwork->work, true);
2974 EXPORT_SYMBOL(cancel_delayed_work_sync);
2977 * schedule_work_on - put work task on a specific cpu
2978 * @cpu: cpu to put the work task on
2979 * @work: job to be done
2981 * This puts a job on a specific cpu
2983 bool schedule_work_on(int cpu, struct work_struct *work)
2985 return queue_work_on(cpu, system_wq, work);
2987 EXPORT_SYMBOL(schedule_work_on);
2990 * schedule_work - put work task in global workqueue
2991 * @work: job to be done
2993 * Returns %false if @work was already on the kernel-global workqueue and
2996 * This puts a job in the kernel-global workqueue if it was not already
2997 * queued and leaves it in the same position on the kernel-global
2998 * workqueue otherwise.
3000 bool schedule_work(struct work_struct *work)
3002 return queue_work(system_wq, work);
3004 EXPORT_SYMBOL(schedule_work);
3007 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3009 * @dwork: job to be done
3010 * @delay: number of jiffies to wait
3012 * After waiting for a given time this puts a job in the kernel-global
3013 * workqueue on the specified CPU.
3015 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3016 unsigned long delay)
3018 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3020 EXPORT_SYMBOL(schedule_delayed_work_on);
3023 * schedule_delayed_work - put work task in global workqueue after delay
3024 * @dwork: job to be done
3025 * @delay: number of jiffies to wait or 0 for immediate execution
3027 * After waiting for a given time this puts a job in the kernel-global
3030 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3032 return queue_delayed_work(system_wq, dwork, delay);
3034 EXPORT_SYMBOL(schedule_delayed_work);
3037 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3038 * @func: the function to call
3040 * schedule_on_each_cpu() executes @func on each online CPU using the
3041 * system workqueue and blocks until all CPUs have completed.
3042 * schedule_on_each_cpu() is very slow.
3045 * 0 on success, -errno on failure.
3047 int schedule_on_each_cpu(work_func_t func)
3050 struct work_struct __percpu *works;
3052 works = alloc_percpu(struct work_struct);
3058 for_each_online_cpu(cpu) {
3059 struct work_struct *work = per_cpu_ptr(works, cpu);
3061 INIT_WORK(work, func);
3062 schedule_work_on(cpu, work);
3065 for_each_online_cpu(cpu)
3066 flush_work(per_cpu_ptr(works, cpu));
3074 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3076 * Forces execution of the kernel-global workqueue and blocks until its
3079 * Think twice before calling this function! It's very easy to get into
3080 * trouble if you don't take great care. Either of the following situations
3081 * will lead to deadlock:
3083 * One of the work items currently on the workqueue needs to acquire
3084 * a lock held by your code or its caller.
3086 * Your code is running in the context of a work routine.
3088 * They will be detected by lockdep when they occur, but the first might not
3089 * occur very often. It depends on what work items are on the workqueue and
3090 * what locks they need, which you have no control over.
3092 * In most situations flushing the entire workqueue is overkill; you merely
3093 * need to know that a particular work item isn't queued and isn't running.
3094 * In such cases you should use cancel_delayed_work_sync() or
3095 * cancel_work_sync() instead.
3097 void flush_scheduled_work(void)
3099 flush_workqueue(system_wq);
3101 EXPORT_SYMBOL(flush_scheduled_work);
3104 * execute_in_process_context - reliably execute the routine with user context
3105 * @fn: the function to execute
3106 * @ew: guaranteed storage for the execute work structure (must
3107 * be available when the work executes)
3109 * Executes the function immediately if process context is available,
3110 * otherwise schedules the function for delayed execution.
3112 * Returns: 0 - function was executed
3113 * 1 - function was scheduled for execution
3115 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3117 if (!in_interrupt()) {
3122 INIT_WORK(&ew->work, fn);
3123 schedule_work(&ew->work);
3127 EXPORT_SYMBOL_GPL(execute_in_process_context);
3129 int keventd_up(void)
3131 return system_wq != NULL;
3134 static int alloc_cwqs(struct workqueue_struct *wq)
3137 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3138 * Make sure that the alignment isn't lower than that of
3139 * unsigned long long.
3141 const size_t size = sizeof(struct cpu_workqueue_struct);
3142 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3143 __alignof__(unsigned long long));
3145 if (!(wq->flags & WQ_UNBOUND))
3146 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3151 * Allocate enough room to align cwq and put an extra
3152 * pointer at the end pointing back to the originally
3153 * allocated pointer which will be used for free.
3155 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3157 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3158 *(void **)(wq->cpu_wq.single + 1) = ptr;
3162 /* just in case, make sure it's actually aligned */
3163 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3164 return wq->cpu_wq.v ? 0 : -ENOMEM;
3167 static void free_cwqs(struct workqueue_struct *wq)
3169 if (!(wq->flags & WQ_UNBOUND))
3170 free_percpu(wq->cpu_wq.pcpu);
3171 else if (wq->cpu_wq.single) {
3172 /* the pointer to free is stored right after the cwq */
3173 kfree(*(void **)(wq->cpu_wq.single + 1));
3177 static int wq_clamp_max_active(int max_active, unsigned int flags,
3180 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3182 if (max_active < 1 || max_active > lim)
3183 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3184 max_active, name, 1, lim);
3186 return clamp_val(max_active, 1, lim);
3189 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3192 struct lock_class_key *key,
3193 const char *lock_name, ...)
3195 va_list args, args1;
3196 struct workqueue_struct *wq;
3200 /* determine namelen, allocate wq and format name */
3201 va_start(args, lock_name);
3202 va_copy(args1, args);
3203 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3205 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3209 vsnprintf(wq->name, namelen, fmt, args1);
3214 * Workqueues which may be used during memory reclaim should
3215 * have a rescuer to guarantee forward progress.
3217 if (flags & WQ_MEM_RECLAIM)
3218 flags |= WQ_RESCUER;
3220 max_active = max_active ?: WQ_DFL_ACTIVE;
3221 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3225 wq->saved_max_active = max_active;
3226 mutex_init(&wq->flush_mutex);
3227 atomic_set(&wq->nr_cwqs_to_flush, 0);
3228 INIT_LIST_HEAD(&wq->flusher_queue);
3229 INIT_LIST_HEAD(&wq->flusher_overflow);
3231 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3232 INIT_LIST_HEAD(&wq->list);
3234 if (alloc_cwqs(wq) < 0)
3237 for_each_cwq_cpu(cpu, wq) {
3238 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3240 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3241 cwq->pool = get_std_worker_pool(cpu, flags & WQ_HIGHPRI);
3243 cwq->flush_color = -1;
3244 cwq->max_active = max_active;
3245 INIT_LIST_HEAD(&cwq->delayed_works);
3248 if (flags & WQ_RESCUER) {
3249 struct worker *rescuer;
3251 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3254 wq->rescuer = rescuer = alloc_worker();
3258 rescuer->rescue_wq = wq;
3259 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3261 if (IS_ERR(rescuer->task))
3264 rescuer->task->flags |= PF_THREAD_BOUND;
3265 wake_up_process(rescuer->task);
3269 * workqueue_lock protects global freeze state and workqueues
3270 * list. Grab it, set max_active accordingly and add the new
3271 * workqueue to workqueues list.
3273 spin_lock(&workqueue_lock);
3275 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3276 for_each_cwq_cpu(cpu, wq)
3277 get_cwq(cpu, wq)->max_active = 0;
3279 list_add(&wq->list, &workqueues);
3281 spin_unlock(&workqueue_lock);
3287 free_mayday_mask(wq->mayday_mask);
3293 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3296 * destroy_workqueue - safely terminate a workqueue
3297 * @wq: target workqueue
3299 * Safely destroy a workqueue. All work currently pending will be done first.
3301 void destroy_workqueue(struct workqueue_struct *wq)
3305 /* drain it before proceeding with destruction */
3306 drain_workqueue(wq);
3309 * wq list is used to freeze wq, remove from list after
3310 * flushing is complete in case freeze races us.
3312 spin_lock(&workqueue_lock);
3313 list_del(&wq->list);
3314 spin_unlock(&workqueue_lock);
3317 for_each_cwq_cpu(cpu, wq) {
3318 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3321 for (i = 0; i < WORK_NR_COLORS; i++)
3322 BUG_ON(cwq->nr_in_flight[i]);
3323 BUG_ON(cwq->nr_active);
3324 BUG_ON(!list_empty(&cwq->delayed_works));
3327 if (wq->flags & WQ_RESCUER) {
3328 kthread_stop(wq->rescuer->task);
3329 free_mayday_mask(wq->mayday_mask);
3336 EXPORT_SYMBOL_GPL(destroy_workqueue);
3339 * cwq_set_max_active - adjust max_active of a cwq
3340 * @cwq: target cpu_workqueue_struct
3341 * @max_active: new max_active value.
3343 * Set @cwq->max_active to @max_active and activate delayed works if
3347 * spin_lock_irq(pool->lock).
3349 static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active)
3351 cwq->max_active = max_active;
3353 while (!list_empty(&cwq->delayed_works) &&
3354 cwq->nr_active < cwq->max_active)
3355 cwq_activate_first_delayed(cwq);
3359 * workqueue_set_max_active - adjust max_active of a workqueue
3360 * @wq: target workqueue
3361 * @max_active: new max_active value.
3363 * Set max_active of @wq to @max_active.
3366 * Don't call from IRQ context.
3368 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3372 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3374 spin_lock(&workqueue_lock);
3376 wq->saved_max_active = max_active;
3378 for_each_cwq_cpu(cpu, wq) {
3379 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3380 struct worker_pool *pool = cwq->pool;
3382 spin_lock_irq(&pool->lock);
3384 if (!(wq->flags & WQ_FREEZABLE) ||
3385 !(pool->flags & POOL_FREEZING))
3386 cwq_set_max_active(cwq, max_active);
3388 spin_unlock_irq(&pool->lock);
3391 spin_unlock(&workqueue_lock);
3393 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3396 * workqueue_congested - test whether a workqueue is congested
3397 * @cpu: CPU in question
3398 * @wq: target workqueue
3400 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3401 * no synchronization around this function and the test result is
3402 * unreliable and only useful as advisory hints or for debugging.
3405 * %true if congested, %false otherwise.
3407 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3409 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3411 return !list_empty(&cwq->delayed_works);
3413 EXPORT_SYMBOL_GPL(workqueue_congested);
3416 * work_busy - test whether a work is currently pending or running
3417 * @work: the work to be tested
3419 * Test whether @work is currently pending or running. There is no
3420 * synchronization around this function and the test result is
3421 * unreliable and only useful as advisory hints or for debugging.
3424 * OR'd bitmask of WORK_BUSY_* bits.
3426 unsigned int work_busy(struct work_struct *work)
3428 struct worker_pool *pool = get_work_pool(work);
3429 unsigned long flags;
3430 unsigned int ret = 0;
3432 if (work_pending(work))
3433 ret |= WORK_BUSY_PENDING;
3436 spin_lock_irqsave(&pool->lock, flags);
3437 if (find_worker_executing_work(pool, work))
3438 ret |= WORK_BUSY_RUNNING;
3439 spin_unlock_irqrestore(&pool->lock, flags);
3444 EXPORT_SYMBOL_GPL(work_busy);
3449 * There are two challenges in supporting CPU hotplug. Firstly, there
3450 * are a lot of assumptions on strong associations among work, cwq and
3451 * pool which make migrating pending and scheduled works very
3452 * difficult to implement without impacting hot paths. Secondly,
3453 * worker pools serve mix of short, long and very long running works making
3454 * blocked draining impractical.
3456 * This is solved by allowing the pools to be disassociated from the CPU
3457 * running as an unbound one and allowing it to be reattached later if the
3458 * cpu comes back online.
3461 static void wq_unbind_fn(struct work_struct *work)
3463 int cpu = smp_processor_id();
3464 struct worker_pool *pool;
3465 struct worker *worker;
3466 struct hlist_node *pos;
3469 for_each_std_worker_pool(pool, cpu) {
3470 BUG_ON(cpu != smp_processor_id());
3472 mutex_lock(&pool->assoc_mutex);
3473 spin_lock_irq(&pool->lock);
3476 * We've claimed all manager positions. Make all workers
3477 * unbound and set DISASSOCIATED. Before this, all workers
3478 * except for the ones which are still executing works from
3479 * before the last CPU down must be on the cpu. After
3480 * this, they may become diasporas.
3482 list_for_each_entry(worker, &pool->idle_list, entry)
3483 worker->flags |= WORKER_UNBOUND;
3485 for_each_busy_worker(worker, i, pos, pool)
3486 worker->flags |= WORKER_UNBOUND;
3488 pool->flags |= POOL_DISASSOCIATED;
3490 spin_unlock_irq(&pool->lock);
3491 mutex_unlock(&pool->assoc_mutex);
3495 * Call schedule() so that we cross rq->lock and thus can guarantee
3496 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3497 * as scheduler callbacks may be invoked from other cpus.
3502 * Sched callbacks are disabled now. Zap nr_running. After this,
3503 * nr_running stays zero and need_more_worker() and keep_working()
3504 * are always true as long as the worklist is not empty. Pools on
3505 * @cpu now behave as unbound (in terms of concurrency management)
3506 * pools which are served by workers tied to the CPU.
3508 * On return from this function, the current worker would trigger
3509 * unbound chain execution of pending work items if other workers
3512 for_each_std_worker_pool(pool, cpu)
3513 atomic_set(get_pool_nr_running(pool), 0);
3517 * Workqueues should be brought up before normal priority CPU notifiers.
3518 * This will be registered high priority CPU notifier.
3520 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3521 unsigned long action,
3524 unsigned int cpu = (unsigned long)hcpu;
3525 struct worker_pool *pool;
3527 switch (action & ~CPU_TASKS_FROZEN) {
3528 case CPU_UP_PREPARE:
3529 for_each_std_worker_pool(pool, cpu) {
3530 struct worker *worker;
3532 if (pool->nr_workers)
3535 worker = create_worker(pool);
3539 spin_lock_irq(&pool->lock);
3540 start_worker(worker);
3541 spin_unlock_irq(&pool->lock);
3545 case CPU_DOWN_FAILED:
3547 for_each_std_worker_pool(pool, cpu) {
3548 mutex_lock(&pool->assoc_mutex);
3549 spin_lock_irq(&pool->lock);
3551 pool->flags &= ~POOL_DISASSOCIATED;
3552 rebind_workers(pool);
3554 spin_unlock_irq(&pool->lock);
3555 mutex_unlock(&pool->assoc_mutex);
3563 * Workqueues should be brought down after normal priority CPU notifiers.
3564 * This will be registered as low priority CPU notifier.
3566 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3567 unsigned long action,
3570 unsigned int cpu = (unsigned long)hcpu;
3571 struct work_struct unbind_work;
3573 switch (action & ~CPU_TASKS_FROZEN) {
3574 case CPU_DOWN_PREPARE:
3575 /* unbinding should happen on the local CPU */
3576 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
3577 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3578 flush_work(&unbind_work);
3586 struct work_for_cpu {
3587 struct work_struct work;
3593 static void work_for_cpu_fn(struct work_struct *work)
3595 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3597 wfc->ret = wfc->fn(wfc->arg);
3601 * work_on_cpu - run a function in user context on a particular cpu
3602 * @cpu: the cpu to run on
3603 * @fn: the function to run
3604 * @arg: the function arg
3606 * This will return the value @fn returns.
3607 * It is up to the caller to ensure that the cpu doesn't go offline.
3608 * The caller must not hold any locks which would prevent @fn from completing.
3610 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3612 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3614 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3615 schedule_work_on(cpu, &wfc.work);
3616 flush_work(&wfc.work);
3619 EXPORT_SYMBOL_GPL(work_on_cpu);
3620 #endif /* CONFIG_SMP */
3622 #ifdef CONFIG_FREEZER
3625 * freeze_workqueues_begin - begin freezing workqueues
3627 * Start freezing workqueues. After this function returns, all freezable
3628 * workqueues will queue new works to their frozen_works list instead of
3632 * Grabs and releases workqueue_lock and pool->lock's.
3634 void freeze_workqueues_begin(void)
3638 spin_lock(&workqueue_lock);
3640 BUG_ON(workqueue_freezing);
3641 workqueue_freezing = true;
3643 for_each_wq_cpu(cpu) {
3644 struct worker_pool *pool;
3645 struct workqueue_struct *wq;
3647 for_each_std_worker_pool(pool, cpu) {
3648 spin_lock_irq(&pool->lock);
3650 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
3651 pool->flags |= POOL_FREEZING;
3653 list_for_each_entry(wq, &workqueues, list) {
3654 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3656 if (cwq && cwq->pool == pool &&
3657 (wq->flags & WQ_FREEZABLE))
3658 cwq->max_active = 0;
3661 spin_unlock_irq(&pool->lock);
3665 spin_unlock(&workqueue_lock);
3669 * freeze_workqueues_busy - are freezable workqueues still busy?
3671 * Check whether freezing is complete. This function must be called
3672 * between freeze_workqueues_begin() and thaw_workqueues().
3675 * Grabs and releases workqueue_lock.
3678 * %true if some freezable workqueues are still busy. %false if freezing
3681 bool freeze_workqueues_busy(void)
3686 spin_lock(&workqueue_lock);
3688 BUG_ON(!workqueue_freezing);
3690 for_each_wq_cpu(cpu) {
3691 struct workqueue_struct *wq;
3693 * nr_active is monotonically decreasing. It's safe
3694 * to peek without lock.
3696 list_for_each_entry(wq, &workqueues, list) {
3697 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3699 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3702 BUG_ON(cwq->nr_active < 0);
3703 if (cwq->nr_active) {
3710 spin_unlock(&workqueue_lock);
3715 * thaw_workqueues - thaw workqueues
3717 * Thaw workqueues. Normal queueing is restored and all collected
3718 * frozen works are transferred to their respective pool worklists.
3721 * Grabs and releases workqueue_lock and pool->lock's.
3723 void thaw_workqueues(void)
3727 spin_lock(&workqueue_lock);
3729 if (!workqueue_freezing)
3732 for_each_wq_cpu(cpu) {
3733 struct worker_pool *pool;
3734 struct workqueue_struct *wq;
3736 for_each_std_worker_pool(pool, cpu) {
3737 spin_lock_irq(&pool->lock);
3739 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
3740 pool->flags &= ~POOL_FREEZING;
3742 list_for_each_entry(wq, &workqueues, list) {
3743 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3745 if (!cwq || cwq->pool != pool ||
3746 !(wq->flags & WQ_FREEZABLE))
3749 /* restore max_active and repopulate worklist */
3750 cwq_set_max_active(cwq, wq->saved_max_active);
3753 wake_up_worker(pool);
3755 spin_unlock_irq(&pool->lock);
3759 workqueue_freezing = false;
3761 spin_unlock(&workqueue_lock);
3763 #endif /* CONFIG_FREEZER */
3765 static int __init init_workqueues(void)
3769 /* make sure we have enough bits for OFFQ pool ID */
3770 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
3771 WORK_CPU_END * NR_STD_WORKER_POOLS);
3773 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3774 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3776 /* initialize CPU pools */
3777 for_each_wq_cpu(cpu) {
3778 struct worker_pool *pool;
3780 for_each_std_worker_pool(pool, cpu) {
3781 spin_lock_init(&pool->lock);
3783 pool->flags |= POOL_DISASSOCIATED;
3784 INIT_LIST_HEAD(&pool->worklist);
3785 INIT_LIST_HEAD(&pool->idle_list);
3786 hash_init(pool->busy_hash);
3788 init_timer_deferrable(&pool->idle_timer);
3789 pool->idle_timer.function = idle_worker_timeout;
3790 pool->idle_timer.data = (unsigned long)pool;
3792 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3793 (unsigned long)pool);
3795 mutex_init(&pool->assoc_mutex);
3796 ida_init(&pool->worker_ida);
3799 BUG_ON(worker_pool_assign_id(pool));
3803 /* create the initial worker */
3804 for_each_online_wq_cpu(cpu) {
3805 struct worker_pool *pool;
3807 for_each_std_worker_pool(pool, cpu) {
3808 struct worker *worker;
3810 if (cpu != WORK_CPU_UNBOUND)
3811 pool->flags &= ~POOL_DISASSOCIATED;
3813 worker = create_worker(pool);
3815 spin_lock_irq(&pool->lock);
3816 start_worker(worker);
3817 spin_unlock_irq(&pool->lock);
3821 system_wq = alloc_workqueue("events", 0, 0);
3822 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3823 system_long_wq = alloc_workqueue("events_long", 0, 0);
3824 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3825 WQ_UNBOUND_MAX_ACTIVE);
3826 system_freezable_wq = alloc_workqueue("events_freezable",
3828 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3829 !system_unbound_wq || !system_freezable_wq);
3832 early_initcall(init_workqueues);