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;
1072 local_irq_save(*flags);
1074 /* try to steal the timer if it exists */
1076 struct delayed_work *dwork = to_delayed_work(work);
1079 * dwork->timer is irqsafe. If del_timer() fails, it's
1080 * guaranteed that the timer is not queued anywhere and not
1081 * running on the local CPU.
1083 if (likely(del_timer(&dwork->timer)))
1087 /* try to claim PENDING the normal way */
1088 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1092 * The queueing is in progress, or it is already queued. Try to
1093 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1095 pool = get_work_pool(work);
1099 spin_lock(&pool->lock);
1100 if (!list_empty(&work->entry)) {
1102 * This work is queued, but perhaps we locked the wrong
1103 * pool. In that case we must see the new value after
1104 * rmb(), see insert_work()->wmb().
1107 if (pool == get_work_pool(work)) {
1108 debug_work_deactivate(work);
1111 * A delayed work item cannot be grabbed directly
1112 * because it might have linked NO_COLOR work items
1113 * which, if left on the delayed_list, will confuse
1114 * cwq->nr_active management later on and cause
1115 * stall. Make sure the work item is activated
1118 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1119 cwq_activate_delayed_work(work);
1121 list_del_init(&work->entry);
1122 cwq_dec_nr_in_flight(get_work_cwq(work),
1123 get_work_color(work));
1125 /* work->data points to cwq iff queued, point to pool */
1126 set_work_pool_and_keep_pending(work, pool->id);
1128 spin_unlock(&pool->lock);
1132 spin_unlock(&pool->lock);
1134 local_irq_restore(*flags);
1135 if (work_is_canceling(work))
1142 * insert_work - insert a work into a pool
1143 * @cwq: cwq @work belongs to
1144 * @work: work to insert
1145 * @head: insertion point
1146 * @extra_flags: extra WORK_STRUCT_* flags to set
1148 * Insert @work which belongs to @cwq after @head. @extra_flags is or'd to
1149 * work_struct flags.
1152 * spin_lock_irq(pool->lock).
1154 static void insert_work(struct cpu_workqueue_struct *cwq,
1155 struct work_struct *work, struct list_head *head,
1156 unsigned int extra_flags)
1158 struct worker_pool *pool = cwq->pool;
1160 /* we own @work, set data and link */
1161 set_work_cwq(work, cwq, extra_flags);
1164 * Ensure that we get the right work->data if we see the
1165 * result of list_add() below, see try_to_grab_pending().
1169 list_add_tail(&work->entry, head);
1172 * Ensure either worker_sched_deactivated() sees the above
1173 * list_add_tail() or we see zero nr_running to avoid workers
1174 * lying around lazily while there are works to be processed.
1178 if (__need_more_worker(pool))
1179 wake_up_worker(pool);
1183 * Test whether @work is being queued from another work executing on the
1184 * same workqueue. This is rather expensive and should only be used from
1187 static bool is_chained_work(struct workqueue_struct *wq)
1189 unsigned long flags;
1192 for_each_wq_cpu(cpu) {
1193 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
1194 struct worker_pool *pool = cwq->pool;
1195 struct worker *worker;
1196 struct hlist_node *pos;
1199 spin_lock_irqsave(&pool->lock, flags);
1200 for_each_busy_worker(worker, i, pos, pool) {
1201 if (worker->task != current)
1203 spin_unlock_irqrestore(&pool->lock, flags);
1205 * I'm @worker, no locking necessary. See if @work
1206 * is headed to the same workqueue.
1208 return worker->current_cwq->wq == wq;
1210 spin_unlock_irqrestore(&pool->lock, flags);
1215 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1216 struct work_struct *work)
1218 bool highpri = wq->flags & WQ_HIGHPRI;
1219 struct worker_pool *pool;
1220 struct cpu_workqueue_struct *cwq;
1221 struct list_head *worklist;
1222 unsigned int work_flags;
1223 unsigned int req_cpu = cpu;
1226 * While a work item is PENDING && off queue, a task trying to
1227 * steal the PENDING will busy-loop waiting for it to either get
1228 * queued or lose PENDING. Grabbing PENDING and queueing should
1229 * happen with IRQ disabled.
1231 WARN_ON_ONCE(!irqs_disabled());
1233 debug_work_activate(work);
1235 /* if dying, only works from the same workqueue are allowed */
1236 if (unlikely(wq->flags & WQ_DRAINING) &&
1237 WARN_ON_ONCE(!is_chained_work(wq)))
1240 /* determine pool to use */
1241 if (!(wq->flags & WQ_UNBOUND)) {
1242 struct worker_pool *last_pool;
1244 if (cpu == WORK_CPU_UNBOUND)
1245 cpu = raw_smp_processor_id();
1248 * It's multi cpu. If @work was previously on a different
1249 * cpu, it might still be running there, in which case the
1250 * work needs to be queued on that cpu to guarantee
1253 pool = get_std_worker_pool(cpu, highpri);
1254 last_pool = get_work_pool(work);
1256 if (last_pool && last_pool != pool) {
1257 struct worker *worker;
1259 spin_lock(&last_pool->lock);
1261 worker = find_worker_executing_work(last_pool, work);
1263 if (worker && worker->current_cwq->wq == wq)
1266 /* meh... not running there, queue here */
1267 spin_unlock(&last_pool->lock);
1268 spin_lock(&pool->lock);
1271 spin_lock(&pool->lock);
1274 pool = get_std_worker_pool(WORK_CPU_UNBOUND, highpri);
1275 spin_lock(&pool->lock);
1278 /* pool determined, get cwq and queue */
1279 cwq = get_cwq(pool->cpu, wq);
1280 trace_workqueue_queue_work(req_cpu, cwq, work);
1282 if (WARN_ON(!list_empty(&work->entry))) {
1283 spin_unlock(&pool->lock);
1287 cwq->nr_in_flight[cwq->work_color]++;
1288 work_flags = work_color_to_flags(cwq->work_color);
1290 if (likely(cwq->nr_active < cwq->max_active)) {
1291 trace_workqueue_activate_work(work);
1293 worklist = &cwq->pool->worklist;
1295 work_flags |= WORK_STRUCT_DELAYED;
1296 worklist = &cwq->delayed_works;
1299 insert_work(cwq, work, worklist, work_flags);
1301 spin_unlock(&pool->lock);
1305 * queue_work_on - queue work on specific cpu
1306 * @cpu: CPU number to execute work on
1307 * @wq: workqueue to use
1308 * @work: work to queue
1310 * Returns %false if @work was already on a queue, %true otherwise.
1312 * We queue the work to a specific CPU, the caller must ensure it
1315 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1316 struct work_struct *work)
1319 unsigned long flags;
1321 local_irq_save(flags);
1323 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1324 __queue_work(cpu, wq, work);
1328 local_irq_restore(flags);
1331 EXPORT_SYMBOL_GPL(queue_work_on);
1334 * queue_work - queue work on a workqueue
1335 * @wq: workqueue to use
1336 * @work: work to queue
1338 * Returns %false if @work was already on a queue, %true otherwise.
1340 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1341 * it can be processed by another CPU.
1343 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1345 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1347 EXPORT_SYMBOL_GPL(queue_work);
1349 void delayed_work_timer_fn(unsigned long __data)
1351 struct delayed_work *dwork = (struct delayed_work *)__data;
1353 /* should have been called from irqsafe timer with irq already off */
1354 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1356 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1358 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1359 struct delayed_work *dwork, unsigned long delay)
1361 struct timer_list *timer = &dwork->timer;
1362 struct work_struct *work = &dwork->work;
1364 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1365 timer->data != (unsigned long)dwork);
1366 WARN_ON_ONCE(timer_pending(timer));
1367 WARN_ON_ONCE(!list_empty(&work->entry));
1370 * If @delay is 0, queue @dwork->work immediately. This is for
1371 * both optimization and correctness. The earliest @timer can
1372 * expire is on the closest next tick and delayed_work users depend
1373 * on that there's no such delay when @delay is 0.
1376 __queue_work(cpu, wq, &dwork->work);
1380 timer_stats_timer_set_start_info(&dwork->timer);
1384 timer->expires = jiffies + delay;
1386 if (unlikely(cpu != WORK_CPU_UNBOUND))
1387 add_timer_on(timer, cpu);
1393 * queue_delayed_work_on - queue work on specific CPU after delay
1394 * @cpu: CPU number to execute work on
1395 * @wq: workqueue to use
1396 * @dwork: work to queue
1397 * @delay: number of jiffies to wait before queueing
1399 * Returns %false if @work was already on a queue, %true otherwise. If
1400 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1403 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1404 struct delayed_work *dwork, unsigned long delay)
1406 struct work_struct *work = &dwork->work;
1408 unsigned long flags;
1410 /* read the comment in __queue_work() */
1411 local_irq_save(flags);
1413 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1414 __queue_delayed_work(cpu, wq, dwork, delay);
1418 local_irq_restore(flags);
1421 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1424 * queue_delayed_work - queue work on a workqueue after delay
1425 * @wq: workqueue to use
1426 * @dwork: delayable work to queue
1427 * @delay: number of jiffies to wait before queueing
1429 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1431 bool queue_delayed_work(struct workqueue_struct *wq,
1432 struct delayed_work *dwork, unsigned long delay)
1434 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1436 EXPORT_SYMBOL_GPL(queue_delayed_work);
1439 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1440 * @cpu: CPU number to execute work on
1441 * @wq: workqueue to use
1442 * @dwork: work to queue
1443 * @delay: number of jiffies to wait before queueing
1445 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1446 * modify @dwork's timer so that it expires after @delay. If @delay is
1447 * zero, @work is guaranteed to be scheduled immediately regardless of its
1450 * Returns %false if @dwork was idle and queued, %true if @dwork was
1451 * pending and its timer was modified.
1453 * This function is safe to call from any context including IRQ handler.
1454 * See try_to_grab_pending() for details.
1456 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1457 struct delayed_work *dwork, unsigned long delay)
1459 unsigned long flags;
1463 ret = try_to_grab_pending(&dwork->work, true, &flags);
1464 } while (unlikely(ret == -EAGAIN));
1466 if (likely(ret >= 0)) {
1467 __queue_delayed_work(cpu, wq, dwork, delay);
1468 local_irq_restore(flags);
1471 /* -ENOENT from try_to_grab_pending() becomes %true */
1474 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1477 * mod_delayed_work - modify delay of or queue a delayed work
1478 * @wq: workqueue to use
1479 * @dwork: work to queue
1480 * @delay: number of jiffies to wait before queueing
1482 * mod_delayed_work_on() on local CPU.
1484 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1485 unsigned long delay)
1487 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1489 EXPORT_SYMBOL_GPL(mod_delayed_work);
1492 * worker_enter_idle - enter idle state
1493 * @worker: worker which is entering idle state
1495 * @worker is entering idle state. Update stats and idle timer if
1499 * spin_lock_irq(pool->lock).
1501 static void worker_enter_idle(struct worker *worker)
1503 struct worker_pool *pool = worker->pool;
1505 BUG_ON(worker->flags & WORKER_IDLE);
1506 BUG_ON(!list_empty(&worker->entry) &&
1507 (worker->hentry.next || worker->hentry.pprev));
1509 /* can't use worker_set_flags(), also called from start_worker() */
1510 worker->flags |= WORKER_IDLE;
1512 worker->last_active = jiffies;
1514 /* idle_list is LIFO */
1515 list_add(&worker->entry, &pool->idle_list);
1517 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1518 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1521 * Sanity check nr_running. Because wq_unbind_fn() releases
1522 * pool->lock between setting %WORKER_UNBOUND and zapping
1523 * nr_running, the warning may trigger spuriously. Check iff
1524 * unbind is not in progress.
1526 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1527 pool->nr_workers == pool->nr_idle &&
1528 atomic_read(get_pool_nr_running(pool)));
1532 * worker_leave_idle - leave idle state
1533 * @worker: worker which is leaving idle state
1535 * @worker is leaving idle state. Update stats.
1538 * spin_lock_irq(pool->lock).
1540 static void worker_leave_idle(struct worker *worker)
1542 struct worker_pool *pool = worker->pool;
1544 BUG_ON(!(worker->flags & WORKER_IDLE));
1545 worker_clr_flags(worker, WORKER_IDLE);
1547 list_del_init(&worker->entry);
1551 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock pool
1554 * Works which are scheduled while the cpu is online must at least be
1555 * scheduled to a worker which is bound to the cpu so that if they are
1556 * flushed from cpu callbacks while cpu is going down, they are
1557 * guaranteed to execute on the cpu.
1559 * This function is to be used by rogue workers and rescuers to bind
1560 * themselves to the target cpu and may race with cpu going down or
1561 * coming online. kthread_bind() can't be used because it may put the
1562 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1563 * verbatim as it's best effort and blocking and pool may be
1564 * [dis]associated in the meantime.
1566 * This function tries set_cpus_allowed() and locks pool and verifies the
1567 * binding against %POOL_DISASSOCIATED which is set during
1568 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1569 * enters idle state or fetches works without dropping lock, it can
1570 * guarantee the scheduling requirement described in the first paragraph.
1573 * Might sleep. Called without any lock but returns with pool->lock
1577 * %true if the associated pool is online (@worker is successfully
1578 * bound), %false if offline.
1580 static bool worker_maybe_bind_and_lock(struct worker *worker)
1581 __acquires(&pool->lock)
1583 struct worker_pool *pool = worker->pool;
1584 struct task_struct *task = worker->task;
1588 * The following call may fail, succeed or succeed
1589 * without actually migrating the task to the cpu if
1590 * it races with cpu hotunplug operation. Verify
1591 * against POOL_DISASSOCIATED.
1593 if (!(pool->flags & POOL_DISASSOCIATED))
1594 set_cpus_allowed_ptr(task, get_cpu_mask(pool->cpu));
1596 spin_lock_irq(&pool->lock);
1597 if (pool->flags & POOL_DISASSOCIATED)
1599 if (task_cpu(task) == pool->cpu &&
1600 cpumask_equal(¤t->cpus_allowed,
1601 get_cpu_mask(pool->cpu)))
1603 spin_unlock_irq(&pool->lock);
1606 * We've raced with CPU hot[un]plug. Give it a breather
1607 * and retry migration. cond_resched() is required here;
1608 * otherwise, we might deadlock against cpu_stop trying to
1609 * bring down the CPU on non-preemptive kernel.
1617 * Rebind an idle @worker to its CPU. worker_thread() will test
1618 * list_empty(@worker->entry) before leaving idle and call this function.
1620 static void idle_worker_rebind(struct worker *worker)
1622 /* CPU may go down again inbetween, clear UNBOUND only on success */
1623 if (worker_maybe_bind_and_lock(worker))
1624 worker_clr_flags(worker, WORKER_UNBOUND);
1626 /* rebind complete, become available again */
1627 list_add(&worker->entry, &worker->pool->idle_list);
1628 spin_unlock_irq(&worker->pool->lock);
1632 * Function for @worker->rebind.work used to rebind unbound busy workers to
1633 * the associated cpu which is coming back online. This is scheduled by
1634 * cpu up but can race with other cpu hotplug operations and may be
1635 * executed twice without intervening cpu down.
1637 static void busy_worker_rebind_fn(struct work_struct *work)
1639 struct worker *worker = container_of(work, struct worker, rebind_work);
1641 if (worker_maybe_bind_and_lock(worker))
1642 worker_clr_flags(worker, WORKER_UNBOUND);
1644 spin_unlock_irq(&worker->pool->lock);
1648 * rebind_workers - rebind all workers of a pool to the associated CPU
1649 * @pool: pool of interest
1651 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1652 * is different for idle and busy ones.
1654 * Idle ones will be removed from the idle_list and woken up. They will
1655 * add themselves back after completing rebind. This ensures that the
1656 * idle_list doesn't contain any unbound workers when re-bound busy workers
1657 * try to perform local wake-ups for concurrency management.
1659 * Busy workers can rebind after they finish their current work items.
1660 * Queueing the rebind work item at the head of the scheduled list is
1661 * enough. Note that nr_running will be properly bumped as busy workers
1664 * On return, all non-manager workers are scheduled for rebind - see
1665 * manage_workers() for the manager special case. Any idle worker
1666 * including the manager will not appear on @idle_list until rebind is
1667 * complete, making local wake-ups safe.
1669 static void rebind_workers(struct worker_pool *pool)
1671 struct worker *worker, *n;
1672 struct hlist_node *pos;
1675 lockdep_assert_held(&pool->assoc_mutex);
1676 lockdep_assert_held(&pool->lock);
1678 /* dequeue and kick idle ones */
1679 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1681 * idle workers should be off @pool->idle_list until rebind
1682 * is complete to avoid receiving premature local wake-ups.
1684 list_del_init(&worker->entry);
1687 * worker_thread() will see the above dequeuing and call
1688 * idle_worker_rebind().
1690 wake_up_process(worker->task);
1693 /* rebind busy workers */
1694 for_each_busy_worker(worker, i, pos, pool) {
1695 struct work_struct *rebind_work = &worker->rebind_work;
1696 struct workqueue_struct *wq;
1698 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1699 work_data_bits(rebind_work)))
1702 debug_work_activate(rebind_work);
1705 * wq doesn't really matter but let's keep @worker->pool
1706 * and @cwq->pool consistent for sanity.
1708 if (std_worker_pool_pri(worker->pool))
1709 wq = system_highpri_wq;
1713 insert_work(get_cwq(pool->cpu, wq), rebind_work,
1714 worker->scheduled.next,
1715 work_color_to_flags(WORK_NO_COLOR));
1719 static struct worker *alloc_worker(void)
1721 struct worker *worker;
1723 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1725 INIT_LIST_HEAD(&worker->entry);
1726 INIT_LIST_HEAD(&worker->scheduled);
1727 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1728 /* on creation a worker is in !idle && prep state */
1729 worker->flags = WORKER_PREP;
1735 * create_worker - create a new workqueue worker
1736 * @pool: pool the new worker will belong to
1738 * Create a new worker which is bound to @pool. The returned worker
1739 * can be started by calling start_worker() or destroyed using
1743 * Might sleep. Does GFP_KERNEL allocations.
1746 * Pointer to the newly created worker.
1748 static struct worker *create_worker(struct worker_pool *pool)
1750 const char *pri = std_worker_pool_pri(pool) ? "H" : "";
1751 struct worker *worker = NULL;
1754 spin_lock_irq(&pool->lock);
1755 while (ida_get_new(&pool->worker_ida, &id)) {
1756 spin_unlock_irq(&pool->lock);
1757 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1759 spin_lock_irq(&pool->lock);
1761 spin_unlock_irq(&pool->lock);
1763 worker = alloc_worker();
1767 worker->pool = pool;
1770 if (pool->cpu != WORK_CPU_UNBOUND)
1771 worker->task = kthread_create_on_node(worker_thread,
1772 worker, cpu_to_node(pool->cpu),
1773 "kworker/%u:%d%s", pool->cpu, id, pri);
1775 worker->task = kthread_create(worker_thread, worker,
1776 "kworker/u:%d%s", id, pri);
1777 if (IS_ERR(worker->task))
1780 if (std_worker_pool_pri(pool))
1781 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1784 * Determine CPU binding of the new worker depending on
1785 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1786 * flag remains stable across this function. See the comments
1787 * above the flag definition for details.
1789 * As an unbound worker may later become a regular one if CPU comes
1790 * online, make sure every worker has %PF_THREAD_BOUND set.
1792 if (!(pool->flags & POOL_DISASSOCIATED)) {
1793 kthread_bind(worker->task, pool->cpu);
1795 worker->task->flags |= PF_THREAD_BOUND;
1796 worker->flags |= WORKER_UNBOUND;
1802 spin_lock_irq(&pool->lock);
1803 ida_remove(&pool->worker_ida, id);
1804 spin_unlock_irq(&pool->lock);
1811 * start_worker - start a newly created worker
1812 * @worker: worker to start
1814 * Make the pool aware of @worker and start it.
1817 * spin_lock_irq(pool->lock).
1819 static void start_worker(struct worker *worker)
1821 worker->flags |= WORKER_STARTED;
1822 worker->pool->nr_workers++;
1823 worker_enter_idle(worker);
1824 wake_up_process(worker->task);
1828 * destroy_worker - destroy a workqueue worker
1829 * @worker: worker to be destroyed
1831 * Destroy @worker and adjust @pool stats accordingly.
1834 * spin_lock_irq(pool->lock) which is released and regrabbed.
1836 static void destroy_worker(struct worker *worker)
1838 struct worker_pool *pool = worker->pool;
1839 int id = worker->id;
1841 /* sanity check frenzy */
1842 BUG_ON(worker->current_work);
1843 BUG_ON(!list_empty(&worker->scheduled));
1845 if (worker->flags & WORKER_STARTED)
1847 if (worker->flags & WORKER_IDLE)
1850 list_del_init(&worker->entry);
1851 worker->flags |= WORKER_DIE;
1853 spin_unlock_irq(&pool->lock);
1855 kthread_stop(worker->task);
1858 spin_lock_irq(&pool->lock);
1859 ida_remove(&pool->worker_ida, id);
1862 static void idle_worker_timeout(unsigned long __pool)
1864 struct worker_pool *pool = (void *)__pool;
1866 spin_lock_irq(&pool->lock);
1868 if (too_many_workers(pool)) {
1869 struct worker *worker;
1870 unsigned long expires;
1872 /* idle_list is kept in LIFO order, check the last one */
1873 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1874 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1876 if (time_before(jiffies, expires))
1877 mod_timer(&pool->idle_timer, expires);
1879 /* it's been idle for too long, wake up manager */
1880 pool->flags |= POOL_MANAGE_WORKERS;
1881 wake_up_worker(pool);
1885 spin_unlock_irq(&pool->lock);
1888 static bool send_mayday(struct work_struct *work)
1890 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1891 struct workqueue_struct *wq = cwq->wq;
1894 if (!(wq->flags & WQ_RESCUER))
1897 /* mayday mayday mayday */
1898 cpu = cwq->pool->cpu;
1899 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1900 if (cpu == WORK_CPU_UNBOUND)
1902 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1903 wake_up_process(wq->rescuer->task);
1907 static void pool_mayday_timeout(unsigned long __pool)
1909 struct worker_pool *pool = (void *)__pool;
1910 struct work_struct *work;
1912 spin_lock_irq(&pool->lock);
1914 if (need_to_create_worker(pool)) {
1916 * We've been trying to create a new worker but
1917 * haven't been successful. We might be hitting an
1918 * allocation deadlock. Send distress signals to
1921 list_for_each_entry(work, &pool->worklist, entry)
1925 spin_unlock_irq(&pool->lock);
1927 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1931 * maybe_create_worker - create a new worker if necessary
1932 * @pool: pool to create a new worker for
1934 * Create a new worker for @pool if necessary. @pool is guaranteed to
1935 * have at least one idle worker on return from this function. If
1936 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1937 * sent to all rescuers with works scheduled on @pool to resolve
1938 * possible allocation deadlock.
1940 * On return, need_to_create_worker() is guaranteed to be false and
1941 * may_start_working() true.
1944 * spin_lock_irq(pool->lock) which may be released and regrabbed
1945 * multiple times. Does GFP_KERNEL allocations. Called only from
1949 * false if no action was taken and pool->lock stayed locked, true
1952 static bool maybe_create_worker(struct worker_pool *pool)
1953 __releases(&pool->lock)
1954 __acquires(&pool->lock)
1956 if (!need_to_create_worker(pool))
1959 spin_unlock_irq(&pool->lock);
1961 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1962 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1965 struct worker *worker;
1967 worker = create_worker(pool);
1969 del_timer_sync(&pool->mayday_timer);
1970 spin_lock_irq(&pool->lock);
1971 start_worker(worker);
1972 BUG_ON(need_to_create_worker(pool));
1976 if (!need_to_create_worker(pool))
1979 __set_current_state(TASK_INTERRUPTIBLE);
1980 schedule_timeout(CREATE_COOLDOWN);
1982 if (!need_to_create_worker(pool))
1986 del_timer_sync(&pool->mayday_timer);
1987 spin_lock_irq(&pool->lock);
1988 if (need_to_create_worker(pool))
1994 * maybe_destroy_worker - destroy workers which have been idle for a while
1995 * @pool: pool to destroy workers for
1997 * Destroy @pool workers which have been idle for longer than
1998 * IDLE_WORKER_TIMEOUT.
2001 * spin_lock_irq(pool->lock) which may be released and regrabbed
2002 * multiple times. Called only from manager.
2005 * false if no action was taken and pool->lock stayed locked, true
2008 static bool maybe_destroy_workers(struct worker_pool *pool)
2012 while (too_many_workers(pool)) {
2013 struct worker *worker;
2014 unsigned long expires;
2016 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2017 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2019 if (time_before(jiffies, expires)) {
2020 mod_timer(&pool->idle_timer, expires);
2024 destroy_worker(worker);
2032 * manage_workers - manage worker pool
2035 * Assume the manager role and manage the worker pool @worker belongs
2036 * to. At any given time, there can be only zero or one manager per
2037 * pool. The exclusion is handled automatically by this function.
2039 * The caller can safely start processing works on false return. On
2040 * true return, it's guaranteed that need_to_create_worker() is false
2041 * and may_start_working() is true.
2044 * spin_lock_irq(pool->lock) which may be released and regrabbed
2045 * multiple times. Does GFP_KERNEL allocations.
2048 * spin_lock_irq(pool->lock) which may be released and regrabbed
2049 * multiple times. Does GFP_KERNEL allocations.
2051 static bool manage_workers(struct worker *worker)
2053 struct worker_pool *pool = worker->pool;
2056 if (pool->flags & POOL_MANAGING_WORKERS)
2059 pool->flags |= POOL_MANAGING_WORKERS;
2062 * To simplify both worker management and CPU hotplug, hold off
2063 * management while hotplug is in progress. CPU hotplug path can't
2064 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2065 * lead to idle worker depletion (all become busy thinking someone
2066 * else is managing) which in turn can result in deadlock under
2067 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2068 * manager against CPU hotplug.
2070 * assoc_mutex would always be free unless CPU hotplug is in
2071 * progress. trylock first without dropping @pool->lock.
2073 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2074 spin_unlock_irq(&pool->lock);
2075 mutex_lock(&pool->assoc_mutex);
2077 * CPU hotplug could have happened while we were waiting
2078 * for assoc_mutex. Hotplug itself can't handle us
2079 * because manager isn't either on idle or busy list, and
2080 * @pool's state and ours could have deviated.
2082 * As hotplug is now excluded via assoc_mutex, we can
2083 * simply try to bind. It will succeed or fail depending
2084 * on @pool's current state. Try it and adjust
2085 * %WORKER_UNBOUND accordingly.
2087 if (worker_maybe_bind_and_lock(worker))
2088 worker->flags &= ~WORKER_UNBOUND;
2090 worker->flags |= WORKER_UNBOUND;
2095 pool->flags &= ~POOL_MANAGE_WORKERS;
2098 * Destroy and then create so that may_start_working() is true
2101 ret |= maybe_destroy_workers(pool);
2102 ret |= maybe_create_worker(pool);
2104 pool->flags &= ~POOL_MANAGING_WORKERS;
2105 mutex_unlock(&pool->assoc_mutex);
2110 * process_one_work - process single work
2112 * @work: work to process
2114 * Process @work. This function contains all the logics necessary to
2115 * process a single work including synchronization against and
2116 * interaction with other workers on the same cpu, queueing and
2117 * flushing. As long as context requirement is met, any worker can
2118 * call this function to process a work.
2121 * spin_lock_irq(pool->lock) which is released and regrabbed.
2123 static void process_one_work(struct worker *worker, struct work_struct *work)
2124 __releases(&pool->lock)
2125 __acquires(&pool->lock)
2127 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2128 struct worker_pool *pool = worker->pool;
2129 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2131 struct worker *collision;
2132 #ifdef CONFIG_LOCKDEP
2134 * It is permissible to free the struct work_struct from
2135 * inside the function that is called from it, this we need to
2136 * take into account for lockdep too. To avoid bogus "held
2137 * lock freed" warnings as well as problems when looking into
2138 * work->lockdep_map, make a copy and use that here.
2140 struct lockdep_map lockdep_map;
2142 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2145 * Ensure we're on the correct CPU. DISASSOCIATED test is
2146 * necessary to avoid spurious warnings from rescuers servicing the
2147 * unbound or a disassociated pool.
2149 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2150 !(pool->flags & POOL_DISASSOCIATED) &&
2151 raw_smp_processor_id() != pool->cpu);
2154 * A single work shouldn't be executed concurrently by
2155 * multiple workers on a single cpu. Check whether anyone is
2156 * already processing the work. If so, defer the work to the
2157 * currently executing one.
2159 collision = find_worker_executing_work(pool, work);
2160 if (unlikely(collision)) {
2161 move_linked_works(work, &collision->scheduled, NULL);
2165 /* claim and dequeue */
2166 debug_work_deactivate(work);
2167 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2168 worker->current_work = work;
2169 worker->current_func = work->func;
2170 worker->current_cwq = cwq;
2171 work_color = get_work_color(work);
2173 list_del_init(&work->entry);
2176 * CPU intensive works don't participate in concurrency
2177 * management. They're the scheduler's responsibility.
2179 if (unlikely(cpu_intensive))
2180 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2183 * Unbound pool isn't concurrency managed and work items should be
2184 * executed ASAP. Wake up another worker if necessary.
2186 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2187 wake_up_worker(pool);
2190 * Record the last pool and clear PENDING which should be the last
2191 * update to @work. Also, do this inside @pool->lock so that
2192 * PENDING and queued state changes happen together while IRQ is
2195 set_work_pool_and_clear_pending(work, pool->id);
2197 spin_unlock_irq(&pool->lock);
2199 lock_map_acquire_read(&cwq->wq->lockdep_map);
2200 lock_map_acquire(&lockdep_map);
2201 trace_workqueue_execute_start(work);
2202 worker->current_func(work);
2204 * While we must be careful to not use "work" after this, the trace
2205 * point will only record its address.
2207 trace_workqueue_execute_end(work);
2208 lock_map_release(&lockdep_map);
2209 lock_map_release(&cwq->wq->lockdep_map);
2211 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2212 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2213 " last function: %pf\n",
2214 current->comm, preempt_count(), task_pid_nr(current),
2215 worker->current_func);
2216 debug_show_held_locks(current);
2220 spin_lock_irq(&pool->lock);
2222 /* clear cpu intensive status */
2223 if (unlikely(cpu_intensive))
2224 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2226 /* we're done with it, release */
2227 hash_del(&worker->hentry);
2228 worker->current_work = NULL;
2229 worker->current_func = NULL;
2230 worker->current_cwq = NULL;
2231 cwq_dec_nr_in_flight(cwq, work_color);
2235 * process_scheduled_works - process scheduled works
2238 * Process all scheduled works. Please note that the scheduled list
2239 * may change while processing a work, so this function repeatedly
2240 * fetches a work from the top and executes it.
2243 * spin_lock_irq(pool->lock) which may be released and regrabbed
2246 static void process_scheduled_works(struct worker *worker)
2248 while (!list_empty(&worker->scheduled)) {
2249 struct work_struct *work = list_first_entry(&worker->scheduled,
2250 struct work_struct, entry);
2251 process_one_work(worker, work);
2256 * worker_thread - the worker thread function
2259 * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools
2260 * of these per each cpu. These workers process all works regardless of
2261 * their specific target workqueue. The only exception is works which
2262 * belong to workqueues with a rescuer which will be explained in
2265 static int worker_thread(void *__worker)
2267 struct worker *worker = __worker;
2268 struct worker_pool *pool = worker->pool;
2270 /* tell the scheduler that this is a workqueue worker */
2271 worker->task->flags |= PF_WQ_WORKER;
2273 spin_lock_irq(&pool->lock);
2275 /* we are off idle list if destruction or rebind is requested */
2276 if (unlikely(list_empty(&worker->entry))) {
2277 spin_unlock_irq(&pool->lock);
2279 /* if DIE is set, destruction is requested */
2280 if (worker->flags & WORKER_DIE) {
2281 worker->task->flags &= ~PF_WQ_WORKER;
2285 /* otherwise, rebind */
2286 idle_worker_rebind(worker);
2290 worker_leave_idle(worker);
2292 /* no more worker necessary? */
2293 if (!need_more_worker(pool))
2296 /* do we need to manage? */
2297 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2301 * ->scheduled list can only be filled while a worker is
2302 * preparing to process a work or actually processing it.
2303 * Make sure nobody diddled with it while I was sleeping.
2305 BUG_ON(!list_empty(&worker->scheduled));
2308 * When control reaches this point, we're guaranteed to have
2309 * at least one idle worker or that someone else has already
2310 * assumed the manager role.
2312 worker_clr_flags(worker, WORKER_PREP);
2315 struct work_struct *work =
2316 list_first_entry(&pool->worklist,
2317 struct work_struct, entry);
2319 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2320 /* optimization path, not strictly necessary */
2321 process_one_work(worker, work);
2322 if (unlikely(!list_empty(&worker->scheduled)))
2323 process_scheduled_works(worker);
2325 move_linked_works(work, &worker->scheduled, NULL);
2326 process_scheduled_works(worker);
2328 } while (keep_working(pool));
2330 worker_set_flags(worker, WORKER_PREP, false);
2332 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2336 * pool->lock is held and there's no work to process and no need to
2337 * manage, sleep. Workers are woken up only while holding
2338 * pool->lock or from local cpu, so setting the current state
2339 * before releasing pool->lock is enough to prevent losing any
2342 worker_enter_idle(worker);
2343 __set_current_state(TASK_INTERRUPTIBLE);
2344 spin_unlock_irq(&pool->lock);
2350 * rescuer_thread - the rescuer thread function
2353 * Workqueue rescuer thread function. There's one rescuer for each
2354 * workqueue which has WQ_RESCUER set.
2356 * Regular work processing on a pool may block trying to create a new
2357 * worker which uses GFP_KERNEL allocation which has slight chance of
2358 * developing into deadlock if some works currently on the same queue
2359 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2360 * the problem rescuer solves.
2362 * When such condition is possible, the pool summons rescuers of all
2363 * workqueues which have works queued on the pool and let them process
2364 * those works so that forward progress can be guaranteed.
2366 * This should happen rarely.
2368 static int rescuer_thread(void *__rescuer)
2370 struct worker *rescuer = __rescuer;
2371 struct workqueue_struct *wq = rescuer->rescue_wq;
2372 struct list_head *scheduled = &rescuer->scheduled;
2373 bool is_unbound = wq->flags & WQ_UNBOUND;
2376 set_user_nice(current, RESCUER_NICE_LEVEL);
2379 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2380 * doesn't participate in concurrency management.
2382 rescuer->task->flags |= PF_WQ_WORKER;
2384 set_current_state(TASK_INTERRUPTIBLE);
2386 if (kthread_should_stop()) {
2387 __set_current_state(TASK_RUNNING);
2388 rescuer->task->flags &= ~PF_WQ_WORKER;
2393 * See whether any cpu is asking for help. Unbounded
2394 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2396 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2397 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2398 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2399 struct worker_pool *pool = cwq->pool;
2400 struct work_struct *work, *n;
2402 __set_current_state(TASK_RUNNING);
2403 mayday_clear_cpu(cpu, wq->mayday_mask);
2405 /* migrate to the target cpu if possible */
2406 rescuer->pool = pool;
2407 worker_maybe_bind_and_lock(rescuer);
2410 * Slurp in all works issued via this workqueue and
2413 BUG_ON(!list_empty(&rescuer->scheduled));
2414 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2415 if (get_work_cwq(work) == cwq)
2416 move_linked_works(work, scheduled, &n);
2418 process_scheduled_works(rescuer);
2421 * Leave this pool. If keep_working() is %true, notify a
2422 * regular worker; otherwise, we end up with 0 concurrency
2423 * and stalling the execution.
2425 if (keep_working(pool))
2426 wake_up_worker(pool);
2428 spin_unlock_irq(&pool->lock);
2431 /* rescuers should never participate in concurrency management */
2432 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2438 struct work_struct work;
2439 struct completion done;
2442 static void wq_barrier_func(struct work_struct *work)
2444 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2445 complete(&barr->done);
2449 * insert_wq_barrier - insert a barrier work
2450 * @cwq: cwq to insert barrier into
2451 * @barr: wq_barrier to insert
2452 * @target: target work to attach @barr to
2453 * @worker: worker currently executing @target, NULL if @target is not executing
2455 * @barr is linked to @target such that @barr is completed only after
2456 * @target finishes execution. Please note that the ordering
2457 * guarantee is observed only with respect to @target and on the local
2460 * Currently, a queued barrier can't be canceled. This is because
2461 * try_to_grab_pending() can't determine whether the work to be
2462 * grabbed is at the head of the queue and thus can't clear LINKED
2463 * flag of the previous work while there must be a valid next work
2464 * after a work with LINKED flag set.
2466 * Note that when @worker is non-NULL, @target may be modified
2467 * underneath us, so we can't reliably determine cwq from @target.
2470 * spin_lock_irq(pool->lock).
2472 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2473 struct wq_barrier *barr,
2474 struct work_struct *target, struct worker *worker)
2476 struct list_head *head;
2477 unsigned int linked = 0;
2480 * debugobject calls are safe here even with pool->lock locked
2481 * as we know for sure that this will not trigger any of the
2482 * checks and call back into the fixup functions where we
2485 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2486 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2487 init_completion(&barr->done);
2490 * If @target is currently being executed, schedule the
2491 * barrier to the worker; otherwise, put it after @target.
2494 head = worker->scheduled.next;
2496 unsigned long *bits = work_data_bits(target);
2498 head = target->entry.next;
2499 /* there can already be other linked works, inherit and set */
2500 linked = *bits & WORK_STRUCT_LINKED;
2501 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2504 debug_work_activate(&barr->work);
2505 insert_work(cwq, &barr->work, head,
2506 work_color_to_flags(WORK_NO_COLOR) | linked);
2510 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2511 * @wq: workqueue being flushed
2512 * @flush_color: new flush color, < 0 for no-op
2513 * @work_color: new work color, < 0 for no-op
2515 * Prepare cwqs for workqueue flushing.
2517 * If @flush_color is non-negative, flush_color on all cwqs should be
2518 * -1. If no cwq has in-flight commands at the specified color, all
2519 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2520 * has in flight commands, its cwq->flush_color is set to
2521 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2522 * wakeup logic is armed and %true is returned.
2524 * The caller should have initialized @wq->first_flusher prior to
2525 * calling this function with non-negative @flush_color. If
2526 * @flush_color is negative, no flush color update is done and %false
2529 * If @work_color is non-negative, all cwqs should have the same
2530 * work_color which is previous to @work_color and all will be
2531 * advanced to @work_color.
2534 * mutex_lock(wq->flush_mutex).
2537 * %true if @flush_color >= 0 and there's something to flush. %false
2540 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2541 int flush_color, int work_color)
2546 if (flush_color >= 0) {
2547 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2548 atomic_set(&wq->nr_cwqs_to_flush, 1);
2551 for_each_cwq_cpu(cpu, wq) {
2552 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2553 struct worker_pool *pool = cwq->pool;
2555 spin_lock_irq(&pool->lock);
2557 if (flush_color >= 0) {
2558 BUG_ON(cwq->flush_color != -1);
2560 if (cwq->nr_in_flight[flush_color]) {
2561 cwq->flush_color = flush_color;
2562 atomic_inc(&wq->nr_cwqs_to_flush);
2567 if (work_color >= 0) {
2568 BUG_ON(work_color != work_next_color(cwq->work_color));
2569 cwq->work_color = work_color;
2572 spin_unlock_irq(&pool->lock);
2575 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2576 complete(&wq->first_flusher->done);
2582 * flush_workqueue - ensure that any scheduled work has run to completion.
2583 * @wq: workqueue to flush
2585 * Forces execution of the workqueue and blocks until its completion.
2586 * This is typically used in driver shutdown handlers.
2588 * We sleep until all works which were queued on entry have been handled,
2589 * but we are not livelocked by new incoming ones.
2591 void flush_workqueue(struct workqueue_struct *wq)
2593 struct wq_flusher this_flusher = {
2594 .list = LIST_HEAD_INIT(this_flusher.list),
2596 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2600 lock_map_acquire(&wq->lockdep_map);
2601 lock_map_release(&wq->lockdep_map);
2603 mutex_lock(&wq->flush_mutex);
2606 * Start-to-wait phase
2608 next_color = work_next_color(wq->work_color);
2610 if (next_color != wq->flush_color) {
2612 * Color space is not full. The current work_color
2613 * becomes our flush_color and work_color is advanced
2616 BUG_ON(!list_empty(&wq->flusher_overflow));
2617 this_flusher.flush_color = wq->work_color;
2618 wq->work_color = next_color;
2620 if (!wq->first_flusher) {
2621 /* no flush in progress, become the first flusher */
2622 BUG_ON(wq->flush_color != this_flusher.flush_color);
2624 wq->first_flusher = &this_flusher;
2626 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2628 /* nothing to flush, done */
2629 wq->flush_color = next_color;
2630 wq->first_flusher = NULL;
2635 BUG_ON(wq->flush_color == this_flusher.flush_color);
2636 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2637 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2641 * Oops, color space is full, wait on overflow queue.
2642 * The next flush completion will assign us
2643 * flush_color and transfer to flusher_queue.
2645 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2648 mutex_unlock(&wq->flush_mutex);
2650 wait_for_completion(&this_flusher.done);
2653 * Wake-up-and-cascade phase
2655 * First flushers are responsible for cascading flushes and
2656 * handling overflow. Non-first flushers can simply return.
2658 if (wq->first_flusher != &this_flusher)
2661 mutex_lock(&wq->flush_mutex);
2663 /* we might have raced, check again with mutex held */
2664 if (wq->first_flusher != &this_flusher)
2667 wq->first_flusher = NULL;
2669 BUG_ON(!list_empty(&this_flusher.list));
2670 BUG_ON(wq->flush_color != this_flusher.flush_color);
2673 struct wq_flusher *next, *tmp;
2675 /* complete all the flushers sharing the current flush color */
2676 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2677 if (next->flush_color != wq->flush_color)
2679 list_del_init(&next->list);
2680 complete(&next->done);
2683 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2684 wq->flush_color != work_next_color(wq->work_color));
2686 /* this flush_color is finished, advance by one */
2687 wq->flush_color = work_next_color(wq->flush_color);
2689 /* one color has been freed, handle overflow queue */
2690 if (!list_empty(&wq->flusher_overflow)) {
2692 * Assign the same color to all overflowed
2693 * flushers, advance work_color and append to
2694 * flusher_queue. This is the start-to-wait
2695 * phase for these overflowed flushers.
2697 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2698 tmp->flush_color = wq->work_color;
2700 wq->work_color = work_next_color(wq->work_color);
2702 list_splice_tail_init(&wq->flusher_overflow,
2703 &wq->flusher_queue);
2704 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2707 if (list_empty(&wq->flusher_queue)) {
2708 BUG_ON(wq->flush_color != wq->work_color);
2713 * Need to flush more colors. Make the next flusher
2714 * the new first flusher and arm cwqs.
2716 BUG_ON(wq->flush_color == wq->work_color);
2717 BUG_ON(wq->flush_color != next->flush_color);
2719 list_del_init(&next->list);
2720 wq->first_flusher = next;
2722 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2726 * Meh... this color is already done, clear first
2727 * flusher and repeat cascading.
2729 wq->first_flusher = NULL;
2733 mutex_unlock(&wq->flush_mutex);
2735 EXPORT_SYMBOL_GPL(flush_workqueue);
2738 * drain_workqueue - drain a workqueue
2739 * @wq: workqueue to drain
2741 * Wait until the workqueue becomes empty. While draining is in progress,
2742 * only chain queueing is allowed. IOW, only currently pending or running
2743 * work items on @wq can queue further work items on it. @wq is flushed
2744 * repeatedly until it becomes empty. The number of flushing is detemined
2745 * by the depth of chaining and should be relatively short. Whine if it
2748 void drain_workqueue(struct workqueue_struct *wq)
2750 unsigned int flush_cnt = 0;
2754 * __queue_work() needs to test whether there are drainers, is much
2755 * hotter than drain_workqueue() and already looks at @wq->flags.
2756 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2758 spin_lock(&workqueue_lock);
2759 if (!wq->nr_drainers++)
2760 wq->flags |= WQ_DRAINING;
2761 spin_unlock(&workqueue_lock);
2763 flush_workqueue(wq);
2765 for_each_cwq_cpu(cpu, wq) {
2766 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2769 spin_lock_irq(&cwq->pool->lock);
2770 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2771 spin_unlock_irq(&cwq->pool->lock);
2776 if (++flush_cnt == 10 ||
2777 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2778 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2779 wq->name, flush_cnt);
2783 spin_lock(&workqueue_lock);
2784 if (!--wq->nr_drainers)
2785 wq->flags &= ~WQ_DRAINING;
2786 spin_unlock(&workqueue_lock);
2788 EXPORT_SYMBOL_GPL(drain_workqueue);
2790 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2792 struct worker *worker = NULL;
2793 struct worker_pool *pool;
2794 struct cpu_workqueue_struct *cwq;
2797 pool = get_work_pool(work);
2801 spin_lock_irq(&pool->lock);
2802 if (!list_empty(&work->entry)) {
2804 * See the comment near try_to_grab_pending()->smp_rmb().
2805 * If it was re-queued to a different pool under us, we
2806 * are not going to wait.
2809 cwq = get_work_cwq(work);
2810 if (unlikely(!cwq || pool != cwq->pool))
2813 worker = find_worker_executing_work(pool, work);
2816 cwq = worker->current_cwq;
2819 insert_wq_barrier(cwq, barr, work, worker);
2820 spin_unlock_irq(&pool->lock);
2823 * If @max_active is 1 or rescuer is in use, flushing another work
2824 * item on the same workqueue may lead to deadlock. Make sure the
2825 * flusher is not running on the same workqueue by verifying write
2828 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2829 lock_map_acquire(&cwq->wq->lockdep_map);
2831 lock_map_acquire_read(&cwq->wq->lockdep_map);
2832 lock_map_release(&cwq->wq->lockdep_map);
2836 spin_unlock_irq(&pool->lock);
2841 * flush_work - wait for a work to finish executing the last queueing instance
2842 * @work: the work to flush
2844 * Wait until @work has finished execution. @work is guaranteed to be idle
2845 * on return if it hasn't been requeued since flush started.
2848 * %true if flush_work() waited for the work to finish execution,
2849 * %false if it was already idle.
2851 bool flush_work(struct work_struct *work)
2853 struct wq_barrier barr;
2855 lock_map_acquire(&work->lockdep_map);
2856 lock_map_release(&work->lockdep_map);
2858 if (start_flush_work(work, &barr)) {
2859 wait_for_completion(&barr.done);
2860 destroy_work_on_stack(&barr.work);
2866 EXPORT_SYMBOL_GPL(flush_work);
2868 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2870 unsigned long flags;
2874 ret = try_to_grab_pending(work, is_dwork, &flags);
2876 * If someone else is canceling, wait for the same event it
2877 * would be waiting for before retrying.
2879 if (unlikely(ret == -ENOENT))
2881 } while (unlikely(ret < 0));
2883 /* tell other tasks trying to grab @work to back off */
2884 mark_work_canceling(work);
2885 local_irq_restore(flags);
2888 clear_work_data(work);
2893 * cancel_work_sync - cancel a work and wait for it to finish
2894 * @work: the work to cancel
2896 * Cancel @work and wait for its execution to finish. This function
2897 * can be used even if the work re-queues itself or migrates to
2898 * another workqueue. On return from this function, @work is
2899 * guaranteed to be not pending or executing on any CPU.
2901 * cancel_work_sync(&delayed_work->work) must not be used for
2902 * delayed_work's. Use cancel_delayed_work_sync() instead.
2904 * The caller must ensure that the workqueue on which @work was last
2905 * queued can't be destroyed before this function returns.
2908 * %true if @work was pending, %false otherwise.
2910 bool cancel_work_sync(struct work_struct *work)
2912 return __cancel_work_timer(work, false);
2914 EXPORT_SYMBOL_GPL(cancel_work_sync);
2917 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2918 * @dwork: the delayed work to flush
2920 * Delayed timer is cancelled and the pending work is queued for
2921 * immediate execution. Like flush_work(), this function only
2922 * considers the last queueing instance of @dwork.
2925 * %true if flush_work() waited for the work to finish execution,
2926 * %false if it was already idle.
2928 bool flush_delayed_work(struct delayed_work *dwork)
2930 local_irq_disable();
2931 if (del_timer_sync(&dwork->timer))
2932 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2934 return flush_work(&dwork->work);
2936 EXPORT_SYMBOL(flush_delayed_work);
2939 * cancel_delayed_work - cancel a delayed work
2940 * @dwork: delayed_work to cancel
2942 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2943 * and canceled; %false if wasn't pending. Note that the work callback
2944 * function may still be running on return, unless it returns %true and the
2945 * work doesn't re-arm itself. Explicitly flush or use
2946 * cancel_delayed_work_sync() to wait on it.
2948 * This function is safe to call from any context including IRQ handler.
2950 bool cancel_delayed_work(struct delayed_work *dwork)
2952 unsigned long flags;
2956 ret = try_to_grab_pending(&dwork->work, true, &flags);
2957 } while (unlikely(ret == -EAGAIN));
2959 if (unlikely(ret < 0))
2962 set_work_pool_and_clear_pending(&dwork->work,
2963 get_work_pool_id(&dwork->work));
2964 local_irq_restore(flags);
2967 EXPORT_SYMBOL(cancel_delayed_work);
2970 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2971 * @dwork: the delayed work cancel
2973 * This is cancel_work_sync() for delayed works.
2976 * %true if @dwork was pending, %false otherwise.
2978 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2980 return __cancel_work_timer(&dwork->work, true);
2982 EXPORT_SYMBOL(cancel_delayed_work_sync);
2985 * schedule_work_on - put work task on a specific cpu
2986 * @cpu: cpu to put the work task on
2987 * @work: job to be done
2989 * This puts a job on a specific cpu
2991 bool schedule_work_on(int cpu, struct work_struct *work)
2993 return queue_work_on(cpu, system_wq, work);
2995 EXPORT_SYMBOL(schedule_work_on);
2998 * schedule_work - put work task in global workqueue
2999 * @work: job to be done
3001 * Returns %false if @work was already on the kernel-global workqueue and
3004 * This puts a job in the kernel-global workqueue if it was not already
3005 * queued and leaves it in the same position on the kernel-global
3006 * workqueue otherwise.
3008 bool schedule_work(struct work_struct *work)
3010 return queue_work(system_wq, work);
3012 EXPORT_SYMBOL(schedule_work);
3015 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3017 * @dwork: job to be done
3018 * @delay: number of jiffies to wait
3020 * After waiting for a given time this puts a job in the kernel-global
3021 * workqueue on the specified CPU.
3023 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3024 unsigned long delay)
3026 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3028 EXPORT_SYMBOL(schedule_delayed_work_on);
3031 * schedule_delayed_work - put work task in global workqueue after delay
3032 * @dwork: job to be done
3033 * @delay: number of jiffies to wait or 0 for immediate execution
3035 * After waiting for a given time this puts a job in the kernel-global
3038 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3040 return queue_delayed_work(system_wq, dwork, delay);
3042 EXPORT_SYMBOL(schedule_delayed_work);
3045 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3046 * @func: the function to call
3048 * schedule_on_each_cpu() executes @func on each online CPU using the
3049 * system workqueue and blocks until all CPUs have completed.
3050 * schedule_on_each_cpu() is very slow.
3053 * 0 on success, -errno on failure.
3055 int schedule_on_each_cpu(work_func_t func)
3058 struct work_struct __percpu *works;
3060 works = alloc_percpu(struct work_struct);
3066 for_each_online_cpu(cpu) {
3067 struct work_struct *work = per_cpu_ptr(works, cpu);
3069 INIT_WORK(work, func);
3070 schedule_work_on(cpu, work);
3073 for_each_online_cpu(cpu)
3074 flush_work(per_cpu_ptr(works, cpu));
3082 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3084 * Forces execution of the kernel-global workqueue and blocks until its
3087 * Think twice before calling this function! It's very easy to get into
3088 * trouble if you don't take great care. Either of the following situations
3089 * will lead to deadlock:
3091 * One of the work items currently on the workqueue needs to acquire
3092 * a lock held by your code or its caller.
3094 * Your code is running in the context of a work routine.
3096 * They will be detected by lockdep when they occur, but the first might not
3097 * occur very often. It depends on what work items are on the workqueue and
3098 * what locks they need, which you have no control over.
3100 * In most situations flushing the entire workqueue is overkill; you merely
3101 * need to know that a particular work item isn't queued and isn't running.
3102 * In such cases you should use cancel_delayed_work_sync() or
3103 * cancel_work_sync() instead.
3105 void flush_scheduled_work(void)
3107 flush_workqueue(system_wq);
3109 EXPORT_SYMBOL(flush_scheduled_work);
3112 * execute_in_process_context - reliably execute the routine with user context
3113 * @fn: the function to execute
3114 * @ew: guaranteed storage for the execute work structure (must
3115 * be available when the work executes)
3117 * Executes the function immediately if process context is available,
3118 * otherwise schedules the function for delayed execution.
3120 * Returns: 0 - function was executed
3121 * 1 - function was scheduled for execution
3123 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3125 if (!in_interrupt()) {
3130 INIT_WORK(&ew->work, fn);
3131 schedule_work(&ew->work);
3135 EXPORT_SYMBOL_GPL(execute_in_process_context);
3137 int keventd_up(void)
3139 return system_wq != NULL;
3142 static int alloc_cwqs(struct workqueue_struct *wq)
3145 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3146 * Make sure that the alignment isn't lower than that of
3147 * unsigned long long.
3149 const size_t size = sizeof(struct cpu_workqueue_struct);
3150 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3151 __alignof__(unsigned long long));
3153 if (!(wq->flags & WQ_UNBOUND))
3154 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3159 * Allocate enough room to align cwq and put an extra
3160 * pointer at the end pointing back to the originally
3161 * allocated pointer which will be used for free.
3163 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3165 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3166 *(void **)(wq->cpu_wq.single + 1) = ptr;
3170 /* just in case, make sure it's actually aligned */
3171 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3172 return wq->cpu_wq.v ? 0 : -ENOMEM;
3175 static void free_cwqs(struct workqueue_struct *wq)
3177 if (!(wq->flags & WQ_UNBOUND))
3178 free_percpu(wq->cpu_wq.pcpu);
3179 else if (wq->cpu_wq.single) {
3180 /* the pointer to free is stored right after the cwq */
3181 kfree(*(void **)(wq->cpu_wq.single + 1));
3185 static int wq_clamp_max_active(int max_active, unsigned int flags,
3188 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3190 if (max_active < 1 || max_active > lim)
3191 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3192 max_active, name, 1, lim);
3194 return clamp_val(max_active, 1, lim);
3197 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3200 struct lock_class_key *key,
3201 const char *lock_name, ...)
3203 va_list args, args1;
3204 struct workqueue_struct *wq;
3208 /* determine namelen, allocate wq and format name */
3209 va_start(args, lock_name);
3210 va_copy(args1, args);
3211 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3213 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3217 vsnprintf(wq->name, namelen, fmt, args1);
3222 * Workqueues which may be used during memory reclaim should
3223 * have a rescuer to guarantee forward progress.
3225 if (flags & WQ_MEM_RECLAIM)
3226 flags |= WQ_RESCUER;
3228 max_active = max_active ?: WQ_DFL_ACTIVE;
3229 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3233 wq->saved_max_active = max_active;
3234 mutex_init(&wq->flush_mutex);
3235 atomic_set(&wq->nr_cwqs_to_flush, 0);
3236 INIT_LIST_HEAD(&wq->flusher_queue);
3237 INIT_LIST_HEAD(&wq->flusher_overflow);
3239 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3240 INIT_LIST_HEAD(&wq->list);
3242 if (alloc_cwqs(wq) < 0)
3245 for_each_cwq_cpu(cpu, wq) {
3246 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3248 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3249 cwq->pool = get_std_worker_pool(cpu, flags & WQ_HIGHPRI);
3251 cwq->flush_color = -1;
3252 cwq->max_active = max_active;
3253 INIT_LIST_HEAD(&cwq->delayed_works);
3256 if (flags & WQ_RESCUER) {
3257 struct worker *rescuer;
3259 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3262 wq->rescuer = rescuer = alloc_worker();
3266 rescuer->rescue_wq = wq;
3267 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3269 if (IS_ERR(rescuer->task))
3272 rescuer->task->flags |= PF_THREAD_BOUND;
3273 wake_up_process(rescuer->task);
3277 * workqueue_lock protects global freeze state and workqueues
3278 * list. Grab it, set max_active accordingly and add the new
3279 * workqueue to workqueues list.
3281 spin_lock(&workqueue_lock);
3283 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3284 for_each_cwq_cpu(cpu, wq)
3285 get_cwq(cpu, wq)->max_active = 0;
3287 list_add(&wq->list, &workqueues);
3289 spin_unlock(&workqueue_lock);
3295 free_mayday_mask(wq->mayday_mask);
3301 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3304 * destroy_workqueue - safely terminate a workqueue
3305 * @wq: target workqueue
3307 * Safely destroy a workqueue. All work currently pending will be done first.
3309 void destroy_workqueue(struct workqueue_struct *wq)
3313 /* drain it before proceeding with destruction */
3314 drain_workqueue(wq);
3317 * wq list is used to freeze wq, remove from list after
3318 * flushing is complete in case freeze races us.
3320 spin_lock(&workqueue_lock);
3321 list_del(&wq->list);
3322 spin_unlock(&workqueue_lock);
3325 for_each_cwq_cpu(cpu, wq) {
3326 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3329 for (i = 0; i < WORK_NR_COLORS; i++)
3330 BUG_ON(cwq->nr_in_flight[i]);
3331 BUG_ON(cwq->nr_active);
3332 BUG_ON(!list_empty(&cwq->delayed_works));
3335 if (wq->flags & WQ_RESCUER) {
3336 kthread_stop(wq->rescuer->task);
3337 free_mayday_mask(wq->mayday_mask);
3344 EXPORT_SYMBOL_GPL(destroy_workqueue);
3347 * cwq_set_max_active - adjust max_active of a cwq
3348 * @cwq: target cpu_workqueue_struct
3349 * @max_active: new max_active value.
3351 * Set @cwq->max_active to @max_active and activate delayed works if
3355 * spin_lock_irq(pool->lock).
3357 static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active)
3359 cwq->max_active = max_active;
3361 while (!list_empty(&cwq->delayed_works) &&
3362 cwq->nr_active < cwq->max_active)
3363 cwq_activate_first_delayed(cwq);
3367 * workqueue_set_max_active - adjust max_active of a workqueue
3368 * @wq: target workqueue
3369 * @max_active: new max_active value.
3371 * Set max_active of @wq to @max_active.
3374 * Don't call from IRQ context.
3376 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3380 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3382 spin_lock(&workqueue_lock);
3384 wq->saved_max_active = max_active;
3386 for_each_cwq_cpu(cpu, wq) {
3387 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3388 struct worker_pool *pool = cwq->pool;
3390 spin_lock_irq(&pool->lock);
3392 if (!(wq->flags & WQ_FREEZABLE) ||
3393 !(pool->flags & POOL_FREEZING))
3394 cwq_set_max_active(cwq, max_active);
3396 spin_unlock_irq(&pool->lock);
3399 spin_unlock(&workqueue_lock);
3401 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3404 * workqueue_congested - test whether a workqueue is congested
3405 * @cpu: CPU in question
3406 * @wq: target workqueue
3408 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3409 * no synchronization around this function and the test result is
3410 * unreliable and only useful as advisory hints or for debugging.
3413 * %true if congested, %false otherwise.
3415 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3417 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3419 return !list_empty(&cwq->delayed_works);
3421 EXPORT_SYMBOL_GPL(workqueue_congested);
3424 * work_busy - test whether a work is currently pending or running
3425 * @work: the work to be tested
3427 * Test whether @work is currently pending or running. There is no
3428 * synchronization around this function and the test result is
3429 * unreliable and only useful as advisory hints or for debugging.
3432 * OR'd bitmask of WORK_BUSY_* bits.
3434 unsigned int work_busy(struct work_struct *work)
3436 struct worker_pool *pool = get_work_pool(work);
3437 unsigned long flags;
3438 unsigned int ret = 0;
3440 if (work_pending(work))
3441 ret |= WORK_BUSY_PENDING;
3444 spin_lock_irqsave(&pool->lock, flags);
3445 if (find_worker_executing_work(pool, work))
3446 ret |= WORK_BUSY_RUNNING;
3447 spin_unlock_irqrestore(&pool->lock, flags);
3452 EXPORT_SYMBOL_GPL(work_busy);
3457 * There are two challenges in supporting CPU hotplug. Firstly, there
3458 * are a lot of assumptions on strong associations among work, cwq and
3459 * pool which make migrating pending and scheduled works very
3460 * difficult to implement without impacting hot paths. Secondly,
3461 * worker pools serve mix of short, long and very long running works making
3462 * blocked draining impractical.
3464 * This is solved by allowing the pools to be disassociated from the CPU
3465 * running as an unbound one and allowing it to be reattached later if the
3466 * cpu comes back online.
3469 static void wq_unbind_fn(struct work_struct *work)
3471 int cpu = smp_processor_id();
3472 struct worker_pool *pool;
3473 struct worker *worker;
3474 struct hlist_node *pos;
3477 for_each_std_worker_pool(pool, cpu) {
3478 BUG_ON(cpu != smp_processor_id());
3480 mutex_lock(&pool->assoc_mutex);
3481 spin_lock_irq(&pool->lock);
3484 * We've claimed all manager positions. Make all workers
3485 * unbound and set DISASSOCIATED. Before this, all workers
3486 * except for the ones which are still executing works from
3487 * before the last CPU down must be on the cpu. After
3488 * this, they may become diasporas.
3490 list_for_each_entry(worker, &pool->idle_list, entry)
3491 worker->flags |= WORKER_UNBOUND;
3493 for_each_busy_worker(worker, i, pos, pool)
3494 worker->flags |= WORKER_UNBOUND;
3496 pool->flags |= POOL_DISASSOCIATED;
3498 spin_unlock_irq(&pool->lock);
3499 mutex_unlock(&pool->assoc_mutex);
3503 * Call schedule() so that we cross rq->lock and thus can guarantee
3504 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3505 * as scheduler callbacks may be invoked from other cpus.
3510 * Sched callbacks are disabled now. Zap nr_running. After this,
3511 * nr_running stays zero and need_more_worker() and keep_working()
3512 * are always true as long as the worklist is not empty. Pools on
3513 * @cpu now behave as unbound (in terms of concurrency management)
3514 * pools which are served by workers tied to the CPU.
3516 * On return from this function, the current worker would trigger
3517 * unbound chain execution of pending work items if other workers
3520 for_each_std_worker_pool(pool, cpu)
3521 atomic_set(get_pool_nr_running(pool), 0);
3525 * Workqueues should be brought up before normal priority CPU notifiers.
3526 * This will be registered high priority CPU notifier.
3528 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3529 unsigned long action,
3532 unsigned int cpu = (unsigned long)hcpu;
3533 struct worker_pool *pool;
3535 switch (action & ~CPU_TASKS_FROZEN) {
3536 case CPU_UP_PREPARE:
3537 for_each_std_worker_pool(pool, cpu) {
3538 struct worker *worker;
3540 if (pool->nr_workers)
3543 worker = create_worker(pool);
3547 spin_lock_irq(&pool->lock);
3548 start_worker(worker);
3549 spin_unlock_irq(&pool->lock);
3553 case CPU_DOWN_FAILED:
3555 for_each_std_worker_pool(pool, cpu) {
3556 mutex_lock(&pool->assoc_mutex);
3557 spin_lock_irq(&pool->lock);
3559 pool->flags &= ~POOL_DISASSOCIATED;
3560 rebind_workers(pool);
3562 spin_unlock_irq(&pool->lock);
3563 mutex_unlock(&pool->assoc_mutex);
3571 * Workqueues should be brought down after normal priority CPU notifiers.
3572 * This will be registered as low priority CPU notifier.
3574 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3575 unsigned long action,
3578 unsigned int cpu = (unsigned long)hcpu;
3579 struct work_struct unbind_work;
3581 switch (action & ~CPU_TASKS_FROZEN) {
3582 case CPU_DOWN_PREPARE:
3583 /* unbinding should happen on the local CPU */
3584 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
3585 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3586 flush_work(&unbind_work);
3594 struct work_for_cpu {
3595 struct work_struct work;
3601 static void work_for_cpu_fn(struct work_struct *work)
3603 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3605 wfc->ret = wfc->fn(wfc->arg);
3609 * work_on_cpu - run a function in user context on a particular cpu
3610 * @cpu: the cpu to run on
3611 * @fn: the function to run
3612 * @arg: the function arg
3614 * This will return the value @fn returns.
3615 * It is up to the caller to ensure that the cpu doesn't go offline.
3616 * The caller must not hold any locks which would prevent @fn from completing.
3618 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3620 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3622 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3623 schedule_work_on(cpu, &wfc.work);
3624 flush_work(&wfc.work);
3627 EXPORT_SYMBOL_GPL(work_on_cpu);
3628 #endif /* CONFIG_SMP */
3630 #ifdef CONFIG_FREEZER
3633 * freeze_workqueues_begin - begin freezing workqueues
3635 * Start freezing workqueues. After this function returns, all freezable
3636 * workqueues will queue new works to their frozen_works list instead of
3640 * Grabs and releases workqueue_lock and pool->lock's.
3642 void freeze_workqueues_begin(void)
3646 spin_lock(&workqueue_lock);
3648 BUG_ON(workqueue_freezing);
3649 workqueue_freezing = true;
3651 for_each_wq_cpu(cpu) {
3652 struct worker_pool *pool;
3653 struct workqueue_struct *wq;
3655 for_each_std_worker_pool(pool, cpu) {
3656 spin_lock_irq(&pool->lock);
3658 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
3659 pool->flags |= POOL_FREEZING;
3661 list_for_each_entry(wq, &workqueues, list) {
3662 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3664 if (cwq && cwq->pool == pool &&
3665 (wq->flags & WQ_FREEZABLE))
3666 cwq->max_active = 0;
3669 spin_unlock_irq(&pool->lock);
3673 spin_unlock(&workqueue_lock);
3677 * freeze_workqueues_busy - are freezable workqueues still busy?
3679 * Check whether freezing is complete. This function must be called
3680 * between freeze_workqueues_begin() and thaw_workqueues().
3683 * Grabs and releases workqueue_lock.
3686 * %true if some freezable workqueues are still busy. %false if freezing
3689 bool freeze_workqueues_busy(void)
3694 spin_lock(&workqueue_lock);
3696 BUG_ON(!workqueue_freezing);
3698 for_each_wq_cpu(cpu) {
3699 struct workqueue_struct *wq;
3701 * nr_active is monotonically decreasing. It's safe
3702 * to peek without lock.
3704 list_for_each_entry(wq, &workqueues, list) {
3705 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3707 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3710 BUG_ON(cwq->nr_active < 0);
3711 if (cwq->nr_active) {
3718 spin_unlock(&workqueue_lock);
3723 * thaw_workqueues - thaw workqueues
3725 * Thaw workqueues. Normal queueing is restored and all collected
3726 * frozen works are transferred to their respective pool worklists.
3729 * Grabs and releases workqueue_lock and pool->lock's.
3731 void thaw_workqueues(void)
3735 spin_lock(&workqueue_lock);
3737 if (!workqueue_freezing)
3740 for_each_wq_cpu(cpu) {
3741 struct worker_pool *pool;
3742 struct workqueue_struct *wq;
3744 for_each_std_worker_pool(pool, cpu) {
3745 spin_lock_irq(&pool->lock);
3747 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
3748 pool->flags &= ~POOL_FREEZING;
3750 list_for_each_entry(wq, &workqueues, list) {
3751 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3753 if (!cwq || cwq->pool != pool ||
3754 !(wq->flags & WQ_FREEZABLE))
3757 /* restore max_active and repopulate worklist */
3758 cwq_set_max_active(cwq, wq->saved_max_active);
3761 wake_up_worker(pool);
3763 spin_unlock_irq(&pool->lock);
3767 workqueue_freezing = false;
3769 spin_unlock(&workqueue_lock);
3771 #endif /* CONFIG_FREEZER */
3773 static int __init init_workqueues(void)
3777 /* make sure we have enough bits for OFFQ pool ID */
3778 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
3779 WORK_CPU_END * NR_STD_WORKER_POOLS);
3781 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3782 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3784 /* initialize CPU pools */
3785 for_each_wq_cpu(cpu) {
3786 struct worker_pool *pool;
3788 for_each_std_worker_pool(pool, cpu) {
3789 spin_lock_init(&pool->lock);
3791 pool->flags |= POOL_DISASSOCIATED;
3792 INIT_LIST_HEAD(&pool->worklist);
3793 INIT_LIST_HEAD(&pool->idle_list);
3794 hash_init(pool->busy_hash);
3796 init_timer_deferrable(&pool->idle_timer);
3797 pool->idle_timer.function = idle_worker_timeout;
3798 pool->idle_timer.data = (unsigned long)pool;
3800 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3801 (unsigned long)pool);
3803 mutex_init(&pool->assoc_mutex);
3804 ida_init(&pool->worker_ida);
3807 BUG_ON(worker_pool_assign_id(pool));
3811 /* create the initial worker */
3812 for_each_online_wq_cpu(cpu) {
3813 struct worker_pool *pool;
3815 for_each_std_worker_pool(pool, cpu) {
3816 struct worker *worker;
3818 if (cpu != WORK_CPU_UNBOUND)
3819 pool->flags &= ~POOL_DISASSOCIATED;
3821 worker = create_worker(pool);
3823 spin_lock_irq(&pool->lock);
3824 start_worker(worker);
3825 spin_unlock_irq(&pool->lock);
3829 system_wq = alloc_workqueue("events", 0, 0);
3830 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3831 system_long_wq = alloc_workqueue("events_long", 0, 0);
3832 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3833 WQ_UNBOUND_MAX_ACTIVE);
3834 system_freezable_wq = alloc_workqueue("events_freezable",
3836 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3837 !system_unbound_wq || !system_freezable_wq);
3840 early_initcall(init_workqueues);