2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/hashtable.h>
46 #include "workqueue_internal.h"
52 * A bound pool is either associated or disassociated with its CPU.
53 * While associated (!DISASSOCIATED), all workers are bound to the
54 * CPU and none has %WORKER_UNBOUND set and concurrency management
57 * While DISASSOCIATED, the cpu may be offline and all workers have
58 * %WORKER_UNBOUND set and concurrency management disabled, and may
59 * be executing on any CPU. The pool behaves as an unbound one.
61 * Note that DISASSOCIATED can be flipped only while holding
62 * assoc_mutex to avoid changing binding state while
63 * create_worker() is in progress.
65 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
66 POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */
67 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
68 POOL_FREEZING = 1 << 3, /* freeze in progress */
71 WORKER_STARTED = 1 << 0, /* started */
72 WORKER_DIE = 1 << 1, /* die die die */
73 WORKER_IDLE = 1 << 2, /* is idle */
74 WORKER_PREP = 1 << 3, /* preparing to run works */
75 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
76 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
78 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND |
81 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
83 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
85 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
86 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
88 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
89 /* call for help after 10ms
91 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
92 CREATE_COOLDOWN = HZ, /* time to breath after fail */
95 * Rescue workers are used only on emergencies and shared by
98 RESCUER_NICE_LEVEL = -20,
99 HIGHPRI_NICE_LEVEL = -20,
103 * Structure fields follow one of the following exclusion rules.
105 * I: Modifiable by initialization/destruction paths and read-only for
108 * P: Preemption protected. Disabling preemption is enough and should
109 * only be modified and accessed from the local cpu.
111 * L: pool->lock protected. Access with pool->lock held.
113 * X: During normal operation, modification requires pool->lock and should
114 * be done only from local cpu. Either disabling preemption on local
115 * cpu or grabbing pool->lock is enough for read access. If
116 * POOL_DISASSOCIATED is set, it's identical to L.
118 * F: wq->flush_mutex protected.
120 * W: workqueue_lock protected.
123 /* struct worker is defined in workqueue_internal.h */
126 spinlock_t lock; /* the pool lock */
127 unsigned int cpu; /* I: the associated cpu */
128 int id; /* I: pool ID */
129 unsigned int flags; /* X: flags */
131 struct list_head worklist; /* L: list of pending works */
132 int nr_workers; /* L: total number of workers */
134 /* nr_idle includes the ones off idle_list for rebinding */
135 int nr_idle; /* L: currently idle ones */
137 struct list_head idle_list; /* X: list of idle workers */
138 struct timer_list idle_timer; /* L: worker idle timeout */
139 struct timer_list mayday_timer; /* L: SOS timer for workers */
141 /* workers are chained either in busy_hash or idle_list */
142 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
143 /* L: hash of busy workers */
145 struct mutex assoc_mutex; /* protect POOL_DISASSOCIATED */
146 struct ida worker_ida; /* L: for worker IDs */
147 } ____cacheline_aligned_in_smp;
150 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
151 * work_struct->data are used for flags and thus cwqs need to be
152 * aligned at two's power of the number of flag bits.
154 struct cpu_workqueue_struct {
155 struct worker_pool *pool; /* I: the associated pool */
156 struct workqueue_struct *wq; /* I: the owning workqueue */
157 int work_color; /* L: current color */
158 int flush_color; /* L: flushing color */
159 int nr_in_flight[WORK_NR_COLORS];
160 /* L: nr of in_flight works */
161 int nr_active; /* L: nr of active works */
162 int max_active; /* L: max active works */
163 struct list_head delayed_works; /* L: delayed works */
167 * Structure used to wait for workqueue flush.
170 struct list_head list; /* F: list of flushers */
171 int flush_color; /* F: flush color waiting for */
172 struct completion done; /* flush completion */
176 * All cpumasks are assumed to be always set on UP and thus can't be
177 * used to determine whether there's something to be done.
180 typedef cpumask_var_t mayday_mask_t;
181 #define mayday_test_and_set_cpu(cpu, mask) \
182 cpumask_test_and_set_cpu((cpu), (mask))
183 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
184 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
185 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
186 #define free_mayday_mask(mask) free_cpumask_var((mask))
188 typedef unsigned long mayday_mask_t;
189 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
190 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
191 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
192 #define alloc_mayday_mask(maskp, gfp) true
193 #define free_mayday_mask(mask) do { } while (0)
197 * The externally visible workqueue abstraction is an array of
198 * per-CPU workqueues:
200 struct workqueue_struct {
201 unsigned int flags; /* W: WQ_* flags */
203 struct cpu_workqueue_struct __percpu *pcpu;
204 struct cpu_workqueue_struct *single;
206 } cpu_wq; /* I: cwq's */
207 struct list_head list; /* W: list of all workqueues */
209 struct mutex flush_mutex; /* protects wq flushing */
210 int work_color; /* F: current work color */
211 int flush_color; /* F: current flush color */
212 atomic_t nr_cwqs_to_flush; /* flush in progress */
213 struct wq_flusher *first_flusher; /* F: first flusher */
214 struct list_head flusher_queue; /* F: flush waiters */
215 struct list_head flusher_overflow; /* F: flush overflow list */
217 mayday_mask_t mayday_mask; /* cpus requesting rescue */
218 struct worker *rescuer; /* I: rescue worker */
220 int nr_drainers; /* W: drain in progress */
221 int saved_max_active; /* W: saved cwq max_active */
222 #ifdef CONFIG_LOCKDEP
223 struct lockdep_map lockdep_map;
225 char name[]; /* I: workqueue name */
228 struct workqueue_struct *system_wq __read_mostly;
229 EXPORT_SYMBOL_GPL(system_wq);
230 struct workqueue_struct *system_highpri_wq __read_mostly;
231 EXPORT_SYMBOL_GPL(system_highpri_wq);
232 struct workqueue_struct *system_long_wq __read_mostly;
233 EXPORT_SYMBOL_GPL(system_long_wq);
234 struct workqueue_struct *system_unbound_wq __read_mostly;
235 EXPORT_SYMBOL_GPL(system_unbound_wq);
236 struct workqueue_struct *system_freezable_wq __read_mostly;
237 EXPORT_SYMBOL_GPL(system_freezable_wq);
239 #define CREATE_TRACE_POINTS
240 #include <trace/events/workqueue.h>
242 #define for_each_std_worker_pool(pool, cpu) \
243 for ((pool) = &std_worker_pools(cpu)[0]; \
244 (pool) < &std_worker_pools(cpu)[NR_STD_WORKER_POOLS]; (pool)++)
246 #define for_each_busy_worker(worker, i, pos, pool) \
247 hash_for_each(pool->busy_hash, i, pos, worker, hentry)
249 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
252 if (cpu < nr_cpu_ids) {
254 cpu = cpumask_next(cpu, mask);
255 if (cpu < nr_cpu_ids)
259 return WORK_CPU_UNBOUND;
264 static inline int __next_cwq_cpu(int cpu, const struct cpumask *mask,
265 struct workqueue_struct *wq)
267 return __next_wq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
273 * An extra cpu number is defined using an invalid cpu number
274 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
275 * specific CPU. The following iterators are similar to for_each_*_cpu()
276 * iterators but also considers the unbound CPU.
278 * for_each_wq_cpu() : possible CPUs + WORK_CPU_UNBOUND
279 * for_each_online_wq_cpu() : online CPUs + WORK_CPU_UNBOUND
280 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
281 * WORK_CPU_UNBOUND for unbound workqueues
283 #define for_each_wq_cpu(cpu) \
284 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, 3); \
285 (cpu) < WORK_CPU_END; \
286 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, 3))
288 #define for_each_online_wq_cpu(cpu) \
289 for ((cpu) = __next_wq_cpu(-1, cpu_online_mask, 3); \
290 (cpu) < WORK_CPU_END; \
291 (cpu) = __next_wq_cpu((cpu), cpu_online_mask, 3))
293 #define for_each_cwq_cpu(cpu, wq) \
294 for ((cpu) = __next_cwq_cpu(-1, cpu_possible_mask, (wq)); \
295 (cpu) < WORK_CPU_END; \
296 (cpu) = __next_cwq_cpu((cpu), cpu_possible_mask, (wq)))
298 #ifdef CONFIG_DEBUG_OBJECTS_WORK
300 static struct debug_obj_descr work_debug_descr;
302 static void *work_debug_hint(void *addr)
304 return ((struct work_struct *) addr)->func;
308 * fixup_init is called when:
309 * - an active object is initialized
311 static int work_fixup_init(void *addr, enum debug_obj_state state)
313 struct work_struct *work = addr;
316 case ODEBUG_STATE_ACTIVE:
317 cancel_work_sync(work);
318 debug_object_init(work, &work_debug_descr);
326 * fixup_activate is called when:
327 * - an active object is activated
328 * - an unknown object is activated (might be a statically initialized object)
330 static int work_fixup_activate(void *addr, enum debug_obj_state state)
332 struct work_struct *work = addr;
336 case ODEBUG_STATE_NOTAVAILABLE:
338 * This is not really a fixup. The work struct was
339 * statically initialized. We just make sure that it
340 * is tracked in the object tracker.
342 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
343 debug_object_init(work, &work_debug_descr);
344 debug_object_activate(work, &work_debug_descr);
350 case ODEBUG_STATE_ACTIVE:
359 * fixup_free is called when:
360 * - an active object is freed
362 static int work_fixup_free(void *addr, enum debug_obj_state state)
364 struct work_struct *work = addr;
367 case ODEBUG_STATE_ACTIVE:
368 cancel_work_sync(work);
369 debug_object_free(work, &work_debug_descr);
376 static struct debug_obj_descr work_debug_descr = {
377 .name = "work_struct",
378 .debug_hint = work_debug_hint,
379 .fixup_init = work_fixup_init,
380 .fixup_activate = work_fixup_activate,
381 .fixup_free = work_fixup_free,
384 static inline void debug_work_activate(struct work_struct *work)
386 debug_object_activate(work, &work_debug_descr);
389 static inline void debug_work_deactivate(struct work_struct *work)
391 debug_object_deactivate(work, &work_debug_descr);
394 void __init_work(struct work_struct *work, int onstack)
397 debug_object_init_on_stack(work, &work_debug_descr);
399 debug_object_init(work, &work_debug_descr);
401 EXPORT_SYMBOL_GPL(__init_work);
403 void destroy_work_on_stack(struct work_struct *work)
405 debug_object_free(work, &work_debug_descr);
407 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
410 static inline void debug_work_activate(struct work_struct *work) { }
411 static inline void debug_work_deactivate(struct work_struct *work) { }
414 /* Serializes the accesses to the list of workqueues. */
415 static DEFINE_SPINLOCK(workqueue_lock);
416 static LIST_HEAD(workqueues);
417 static bool workqueue_freezing; /* W: have wqs started freezing? */
420 * The CPU standard worker pools. nr_running is the only field which is
421 * expected to be used frequently by other cpus via try_to_wake_up(). Put
422 * it in a separate cacheline.
424 static DEFINE_PER_CPU(struct worker_pool [NR_STD_WORKER_POOLS],
425 cpu_std_worker_pools);
426 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t [NR_STD_WORKER_POOLS],
427 cpu_std_pool_nr_running);
430 * Standard worker pools and nr_running counter for unbound CPU. The pools
431 * have POOL_DISASSOCIATED set, and all workers have WORKER_UNBOUND set.
433 static struct worker_pool unbound_std_worker_pools[NR_STD_WORKER_POOLS];
434 static atomic_t unbound_std_pool_nr_running[NR_STD_WORKER_POOLS] = {
435 [0 ... NR_STD_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */
438 /* idr of all pools */
439 static DEFINE_MUTEX(worker_pool_idr_mutex);
440 static DEFINE_IDR(worker_pool_idr);
442 static int worker_thread(void *__worker);
444 static struct worker_pool *std_worker_pools(int cpu)
446 if (cpu != WORK_CPU_UNBOUND)
447 return per_cpu(cpu_std_worker_pools, cpu);
449 return unbound_std_worker_pools;
452 static int std_worker_pool_pri(struct worker_pool *pool)
454 return pool - std_worker_pools(pool->cpu);
457 /* allocate ID and assign it to @pool */
458 static int worker_pool_assign_id(struct worker_pool *pool)
462 mutex_lock(&worker_pool_idr_mutex);
463 idr_pre_get(&worker_pool_idr, GFP_KERNEL);
464 ret = idr_get_new(&worker_pool_idr, pool, &pool->id);
465 mutex_unlock(&worker_pool_idr_mutex);
471 * Lookup worker_pool by id. The idr currently is built during boot and
472 * never modified. Don't worry about locking for now.
474 static struct worker_pool *worker_pool_by_id(int pool_id)
476 return idr_find(&worker_pool_idr, pool_id);
479 static struct worker_pool *get_std_worker_pool(int cpu, bool highpri)
481 struct worker_pool *pools = std_worker_pools(cpu);
483 return &pools[highpri];
486 static atomic_t *get_pool_nr_running(struct worker_pool *pool)
489 int idx = std_worker_pool_pri(pool);
491 if (cpu != WORK_CPU_UNBOUND)
492 return &per_cpu(cpu_std_pool_nr_running, cpu)[idx];
494 return &unbound_std_pool_nr_running[idx];
497 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
498 struct workqueue_struct *wq)
500 if (!(wq->flags & WQ_UNBOUND)) {
501 if (likely(cpu < nr_cpu_ids))
502 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
503 } else if (likely(cpu == WORK_CPU_UNBOUND))
504 return wq->cpu_wq.single;
508 static unsigned int work_color_to_flags(int color)
510 return color << WORK_STRUCT_COLOR_SHIFT;
513 static int get_work_color(struct work_struct *work)
515 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
516 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
519 static int work_next_color(int color)
521 return (color + 1) % WORK_NR_COLORS;
525 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
526 * contain the pointer to the queued cwq. Once execution starts, the flag
527 * is cleared and the high bits contain OFFQ flags and pool ID.
529 * set_work_cwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
530 * and clear_work_data() can be used to set the cwq, pool or clear
531 * work->data. These functions should only be called while the work is
532 * owned - ie. while the PENDING bit is set.
534 * get_work_pool() and get_work_cwq() can be used to obtain the pool or cwq
535 * corresponding to a work. Pool is available once the work has been
536 * queued anywhere after initialization until it is sync canceled. cwq is
537 * available only while the work item is queued.
539 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
540 * canceled. While being canceled, a work item may have its PENDING set
541 * but stay off timer and worklist for arbitrarily long and nobody should
542 * try to steal the PENDING bit.
544 static inline void set_work_data(struct work_struct *work, unsigned long data,
547 BUG_ON(!work_pending(work));
548 atomic_long_set(&work->data, data | flags | work_static(work));
551 static void set_work_cwq(struct work_struct *work,
552 struct cpu_workqueue_struct *cwq,
553 unsigned long extra_flags)
555 set_work_data(work, (unsigned long)cwq,
556 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
559 static void set_work_pool_and_keep_pending(struct work_struct *work,
562 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
563 WORK_STRUCT_PENDING);
566 static void set_work_pool_and_clear_pending(struct work_struct *work,
570 * The following wmb is paired with the implied mb in
571 * test_and_set_bit(PENDING) and ensures all updates to @work made
572 * here are visible to and precede any updates by the next PENDING
576 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
579 static void clear_work_data(struct work_struct *work)
581 smp_wmb(); /* see set_work_pool_and_clear_pending() */
582 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
585 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
587 unsigned long data = atomic_long_read(&work->data);
589 if (data & WORK_STRUCT_CWQ)
590 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
596 * get_work_pool - return the worker_pool a given work was associated with
597 * @work: the work item of interest
599 * Return the worker_pool @work was last associated with. %NULL if none.
601 static struct worker_pool *get_work_pool(struct work_struct *work)
603 unsigned long data = atomic_long_read(&work->data);
604 struct worker_pool *pool;
607 if (data & WORK_STRUCT_CWQ)
608 return ((struct cpu_workqueue_struct *)
609 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
611 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
612 if (pool_id == WORK_OFFQ_POOL_NONE)
615 pool = worker_pool_by_id(pool_id);
621 * get_work_pool_id - return the worker pool ID a given work is associated with
622 * @work: the work item of interest
624 * Return the worker_pool ID @work was last associated with.
625 * %WORK_OFFQ_POOL_NONE if none.
627 static int get_work_pool_id(struct work_struct *work)
629 struct worker_pool *pool = get_work_pool(work);
631 return pool ? pool->id : WORK_OFFQ_POOL_NONE;
634 static void mark_work_canceling(struct work_struct *work)
636 unsigned long pool_id = get_work_pool_id(work);
638 pool_id <<= WORK_OFFQ_POOL_SHIFT;
639 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
642 static bool work_is_canceling(struct work_struct *work)
644 unsigned long data = atomic_long_read(&work->data);
646 return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING);
650 * Policy functions. These define the policies on how the global worker
651 * pools are managed. Unless noted otherwise, these functions assume that
652 * they're being called with pool->lock held.
655 static bool __need_more_worker(struct worker_pool *pool)
657 return !atomic_read(get_pool_nr_running(pool));
661 * Need to wake up a worker? Called from anything but currently
664 * Note that, because unbound workers never contribute to nr_running, this
665 * function will always return %true for unbound pools as long as the
666 * worklist isn't empty.
668 static bool need_more_worker(struct worker_pool *pool)
670 return !list_empty(&pool->worklist) && __need_more_worker(pool);
673 /* Can I start working? Called from busy but !running workers. */
674 static bool may_start_working(struct worker_pool *pool)
676 return pool->nr_idle;
679 /* Do I need to keep working? Called from currently running workers. */
680 static bool keep_working(struct worker_pool *pool)
682 atomic_t *nr_running = get_pool_nr_running(pool);
684 return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
687 /* Do we need a new worker? Called from manager. */
688 static bool need_to_create_worker(struct worker_pool *pool)
690 return need_more_worker(pool) && !may_start_working(pool);
693 /* Do I need to be the manager? */
694 static bool need_to_manage_workers(struct worker_pool *pool)
696 return need_to_create_worker(pool) ||
697 (pool->flags & POOL_MANAGE_WORKERS);
700 /* Do we have too many workers and should some go away? */
701 static bool too_many_workers(struct worker_pool *pool)
703 bool managing = pool->flags & POOL_MANAGING_WORKERS;
704 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
705 int nr_busy = pool->nr_workers - nr_idle;
708 * nr_idle and idle_list may disagree if idle rebinding is in
709 * progress. Never return %true if idle_list is empty.
711 if (list_empty(&pool->idle_list))
714 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
721 /* Return the first worker. Safe with preemption disabled */
722 static struct worker *first_worker(struct worker_pool *pool)
724 if (unlikely(list_empty(&pool->idle_list)))
727 return list_first_entry(&pool->idle_list, struct worker, entry);
731 * wake_up_worker - wake up an idle worker
732 * @pool: worker pool to wake worker from
734 * Wake up the first idle worker of @pool.
737 * spin_lock_irq(pool->lock).
739 static void wake_up_worker(struct worker_pool *pool)
741 struct worker *worker = first_worker(pool);
744 wake_up_process(worker->task);
748 * wq_worker_waking_up - a worker is waking up
749 * @task: task waking up
750 * @cpu: CPU @task is waking up to
752 * This function is called during try_to_wake_up() when a worker is
756 * spin_lock_irq(rq->lock)
758 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
760 struct worker *worker = kthread_data(task);
762 if (!(worker->flags & WORKER_NOT_RUNNING)) {
763 WARN_ON_ONCE(worker->pool->cpu != cpu);
764 atomic_inc(get_pool_nr_running(worker->pool));
769 * wq_worker_sleeping - a worker is going to sleep
770 * @task: task going to sleep
771 * @cpu: CPU in question, must be the current CPU number
773 * This function is called during schedule() when a busy worker is
774 * going to sleep. Worker on the same cpu can be woken up by
775 * returning pointer to its task.
778 * spin_lock_irq(rq->lock)
781 * Worker task on @cpu to wake up, %NULL if none.
783 struct task_struct *wq_worker_sleeping(struct task_struct *task,
786 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
787 struct worker_pool *pool;
788 atomic_t *nr_running;
791 * Rescuers, which may not have all the fields set up like normal
792 * workers, also reach here, let's not access anything before
793 * checking NOT_RUNNING.
795 if (worker->flags & WORKER_NOT_RUNNING)
799 nr_running = get_pool_nr_running(pool);
801 /* this can only happen on the local cpu */
802 BUG_ON(cpu != raw_smp_processor_id());
805 * The counterpart of the following dec_and_test, implied mb,
806 * worklist not empty test sequence is in insert_work().
807 * Please read comment there.
809 * NOT_RUNNING is clear. This means that we're bound to and
810 * running on the local cpu w/ rq lock held and preemption
811 * disabled, which in turn means that none else could be
812 * manipulating idle_list, so dereferencing idle_list without pool
815 if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
816 to_wakeup = first_worker(pool);
817 return to_wakeup ? to_wakeup->task : NULL;
821 * worker_set_flags - set worker flags and adjust nr_running accordingly
823 * @flags: flags to set
824 * @wakeup: wakeup an idle worker if necessary
826 * Set @flags in @worker->flags and adjust nr_running accordingly. If
827 * nr_running becomes zero and @wakeup is %true, an idle worker is
831 * spin_lock_irq(pool->lock)
833 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
836 struct worker_pool *pool = worker->pool;
838 WARN_ON_ONCE(worker->task != current);
841 * If transitioning into NOT_RUNNING, adjust nr_running and
842 * wake up an idle worker as necessary if requested by
845 if ((flags & WORKER_NOT_RUNNING) &&
846 !(worker->flags & WORKER_NOT_RUNNING)) {
847 atomic_t *nr_running = get_pool_nr_running(pool);
850 if (atomic_dec_and_test(nr_running) &&
851 !list_empty(&pool->worklist))
852 wake_up_worker(pool);
854 atomic_dec(nr_running);
857 worker->flags |= flags;
861 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
863 * @flags: flags to clear
865 * Clear @flags in @worker->flags and adjust nr_running accordingly.
868 * spin_lock_irq(pool->lock)
870 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
872 struct worker_pool *pool = worker->pool;
873 unsigned int oflags = worker->flags;
875 WARN_ON_ONCE(worker->task != current);
877 worker->flags &= ~flags;
880 * If transitioning out of NOT_RUNNING, increment nr_running. Note
881 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
882 * of multiple flags, not a single flag.
884 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
885 if (!(worker->flags & WORKER_NOT_RUNNING))
886 atomic_inc(get_pool_nr_running(pool));
890 * find_worker_executing_work - find worker which is executing a work
891 * @pool: pool of interest
892 * @work: work to find worker for
894 * Find a worker which is executing @work on @pool by searching
895 * @pool->busy_hash which is keyed by the address of @work. For a worker
896 * to match, its current execution should match the address of @work and
897 * its work function. This is to avoid unwanted dependency between
898 * unrelated work executions through a work item being recycled while still
901 * This is a bit tricky. A work item may be freed once its execution
902 * starts and nothing prevents the freed area from being recycled for
903 * another work item. If the same work item address ends up being reused
904 * before the original execution finishes, workqueue will identify the
905 * recycled work item as currently executing and make it wait until the
906 * current execution finishes, introducing an unwanted dependency.
908 * This function checks the work item address, work function and workqueue
909 * to avoid false positives. Note that this isn't complete as one may
910 * construct a work function which can introduce dependency onto itself
911 * through a recycled work item. Well, if somebody wants to shoot oneself
912 * in the foot that badly, there's only so much we can do, and if such
913 * deadlock actually occurs, it should be easy to locate the culprit work
917 * spin_lock_irq(pool->lock).
920 * Pointer to worker which is executing @work if found, NULL
923 static struct worker *find_worker_executing_work(struct worker_pool *pool,
924 struct work_struct *work)
926 struct worker *worker;
927 struct hlist_node *tmp;
929 hash_for_each_possible(pool->busy_hash, worker, tmp, hentry,
931 if (worker->current_work == work &&
932 worker->current_func == work->func)
939 * move_linked_works - move linked works to a list
940 * @work: start of series of works to be scheduled
941 * @head: target list to append @work to
942 * @nextp: out paramter for nested worklist walking
944 * Schedule linked works starting from @work to @head. Work series to
945 * be scheduled starts at @work and includes any consecutive work with
946 * WORK_STRUCT_LINKED set in its predecessor.
948 * If @nextp is not NULL, it's updated to point to the next work of
949 * the last scheduled work. This allows move_linked_works() to be
950 * nested inside outer list_for_each_entry_safe().
953 * spin_lock_irq(pool->lock).
955 static void move_linked_works(struct work_struct *work, struct list_head *head,
956 struct work_struct **nextp)
958 struct work_struct *n;
961 * Linked worklist will always end before the end of the list,
962 * use NULL for list head.
964 list_for_each_entry_safe_from(work, n, NULL, entry) {
965 list_move_tail(&work->entry, head);
966 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
971 * If we're already inside safe list traversal and have moved
972 * multiple works to the scheduled queue, the next position
973 * needs to be updated.
979 static void cwq_activate_delayed_work(struct work_struct *work)
981 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
983 trace_workqueue_activate_work(work);
984 move_linked_works(work, &cwq->pool->worklist, NULL);
985 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
989 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
991 struct work_struct *work = list_first_entry(&cwq->delayed_works,
992 struct work_struct, entry);
994 cwq_activate_delayed_work(work);
998 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
999 * @cwq: cwq of interest
1000 * @color: color of work which left the queue
1002 * A work either has completed or is removed from pending queue,
1003 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1006 * spin_lock_irq(pool->lock).
1008 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
1010 /* ignore uncolored works */
1011 if (color == WORK_NO_COLOR)
1014 cwq->nr_in_flight[color]--;
1017 if (!list_empty(&cwq->delayed_works)) {
1018 /* one down, submit a delayed one */
1019 if (cwq->nr_active < cwq->max_active)
1020 cwq_activate_first_delayed(cwq);
1023 /* is flush in progress and are we at the flushing tip? */
1024 if (likely(cwq->flush_color != color))
1027 /* are there still in-flight works? */
1028 if (cwq->nr_in_flight[color])
1031 /* this cwq is done, clear flush_color */
1032 cwq->flush_color = -1;
1035 * If this was the last cwq, wake up the first flusher. It
1036 * will handle the rest.
1038 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1039 complete(&cwq->wq->first_flusher->done);
1043 * try_to_grab_pending - steal work item from worklist and disable irq
1044 * @work: work item to steal
1045 * @is_dwork: @work is a delayed_work
1046 * @flags: place to store irq state
1048 * Try to grab PENDING bit of @work. This function can handle @work in any
1049 * stable state - idle, on timer or on worklist. Return values are
1051 * 1 if @work was pending and we successfully stole PENDING
1052 * 0 if @work was idle and we claimed PENDING
1053 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1054 * -ENOENT if someone else is canceling @work, this state may persist
1055 * for arbitrarily long
1057 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1058 * interrupted while holding PENDING and @work off queue, irq must be
1059 * disabled on entry. This, combined with delayed_work->timer being
1060 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1062 * On successful return, >= 0, irq is disabled and the caller is
1063 * responsible for releasing it using local_irq_restore(*@flags).
1065 * This function is safe to call from any context including IRQ handler.
1067 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1068 unsigned long *flags)
1070 struct worker_pool *pool;
1071 struct cpu_workqueue_struct *cwq;
1073 local_irq_save(*flags);
1075 /* try to steal the timer if it exists */
1077 struct delayed_work *dwork = to_delayed_work(work);
1080 * dwork->timer is irqsafe. If del_timer() fails, it's
1081 * guaranteed that the timer is not queued anywhere and not
1082 * running on the local CPU.
1084 if (likely(del_timer(&dwork->timer)))
1088 /* try to claim PENDING the normal way */
1089 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1093 * The queueing is in progress, or it is already queued. Try to
1094 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1096 pool = get_work_pool(work);
1100 spin_lock(&pool->lock);
1102 * work->data is guaranteed to point to cwq only while the work
1103 * item is queued on cwq->wq, and both updating work->data to point
1104 * to cwq on queueing and to pool on dequeueing are done under
1105 * cwq->pool->lock. This in turn guarantees that, if work->data
1106 * points to cwq which is associated with a locked pool, the work
1107 * item is currently queued on that pool.
1109 cwq = get_work_cwq(work);
1110 if (cwq && cwq->pool == pool) {
1111 debug_work_deactivate(work);
1114 * A delayed work item cannot be grabbed directly because
1115 * it might have linked NO_COLOR work items which, if left
1116 * on the delayed_list, will confuse cwq->nr_active
1117 * management later on and cause stall. Make sure the work
1118 * item is activated before grabbing.
1120 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1121 cwq_activate_delayed_work(work);
1123 list_del_init(&work->entry);
1124 cwq_dec_nr_in_flight(get_work_cwq(work), get_work_color(work));
1126 /* work->data points to cwq iff queued, point to pool */
1127 set_work_pool_and_keep_pending(work, pool->id);
1129 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);
1162 list_add_tail(&work->entry, head);
1165 * Ensure either worker_sched_deactivated() sees the above
1166 * list_add_tail() or we see zero nr_running to avoid workers
1167 * lying around lazily while there are works to be processed.
1171 if (__need_more_worker(pool))
1172 wake_up_worker(pool);
1176 * Test whether @work is being queued from another work executing on the
1177 * same workqueue. This is rather expensive and should only be used from
1180 static bool is_chained_work(struct workqueue_struct *wq)
1182 unsigned long flags;
1185 for_each_wq_cpu(cpu) {
1186 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
1187 struct worker_pool *pool = cwq->pool;
1188 struct worker *worker;
1189 struct hlist_node *pos;
1192 spin_lock_irqsave(&pool->lock, flags);
1193 for_each_busy_worker(worker, i, pos, pool) {
1194 if (worker->task != current)
1196 spin_unlock_irqrestore(&pool->lock, flags);
1198 * I'm @worker, no locking necessary. See if @work
1199 * is headed to the same workqueue.
1201 return worker->current_cwq->wq == wq;
1203 spin_unlock_irqrestore(&pool->lock, flags);
1208 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1209 struct work_struct *work)
1211 bool highpri = wq->flags & WQ_HIGHPRI;
1212 struct worker_pool *pool;
1213 struct cpu_workqueue_struct *cwq;
1214 struct list_head *worklist;
1215 unsigned int work_flags;
1216 unsigned int req_cpu = cpu;
1219 * While a work item is PENDING && off queue, a task trying to
1220 * steal the PENDING will busy-loop waiting for it to either get
1221 * queued or lose PENDING. Grabbing PENDING and queueing should
1222 * happen with IRQ disabled.
1224 WARN_ON_ONCE(!irqs_disabled());
1226 debug_work_activate(work);
1228 /* if dying, only works from the same workqueue are allowed */
1229 if (unlikely(wq->flags & WQ_DRAINING) &&
1230 WARN_ON_ONCE(!is_chained_work(wq)))
1233 /* determine pool to use */
1234 if (!(wq->flags & WQ_UNBOUND)) {
1235 struct worker_pool *last_pool;
1237 if (cpu == WORK_CPU_UNBOUND)
1238 cpu = raw_smp_processor_id();
1241 * It's multi cpu. If @work was previously on a different
1242 * cpu, it might still be running there, in which case the
1243 * work needs to be queued on that cpu to guarantee
1246 pool = get_std_worker_pool(cpu, highpri);
1247 last_pool = get_work_pool(work);
1249 if (last_pool && last_pool != pool) {
1250 struct worker *worker;
1252 spin_lock(&last_pool->lock);
1254 worker = find_worker_executing_work(last_pool, work);
1256 if (worker && worker->current_cwq->wq == wq)
1259 /* meh... not running there, queue here */
1260 spin_unlock(&last_pool->lock);
1261 spin_lock(&pool->lock);
1264 spin_lock(&pool->lock);
1267 pool = get_std_worker_pool(WORK_CPU_UNBOUND, highpri);
1268 spin_lock(&pool->lock);
1271 /* pool determined, get cwq and queue */
1272 cwq = get_cwq(pool->cpu, wq);
1273 trace_workqueue_queue_work(req_cpu, cwq, work);
1275 if (WARN_ON(!list_empty(&work->entry))) {
1276 spin_unlock(&pool->lock);
1280 cwq->nr_in_flight[cwq->work_color]++;
1281 work_flags = work_color_to_flags(cwq->work_color);
1283 if (likely(cwq->nr_active < cwq->max_active)) {
1284 trace_workqueue_activate_work(work);
1286 worklist = &cwq->pool->worklist;
1288 work_flags |= WORK_STRUCT_DELAYED;
1289 worklist = &cwq->delayed_works;
1292 insert_work(cwq, work, worklist, work_flags);
1294 spin_unlock(&pool->lock);
1298 * queue_work_on - queue work on specific cpu
1299 * @cpu: CPU number to execute work on
1300 * @wq: workqueue to use
1301 * @work: work to queue
1303 * Returns %false if @work was already on a queue, %true otherwise.
1305 * We queue the work to a specific CPU, the caller must ensure it
1308 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1309 struct work_struct *work)
1312 unsigned long flags;
1314 local_irq_save(flags);
1316 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1317 __queue_work(cpu, wq, work);
1321 local_irq_restore(flags);
1324 EXPORT_SYMBOL_GPL(queue_work_on);
1327 * queue_work - queue work on a workqueue
1328 * @wq: workqueue to use
1329 * @work: work to queue
1331 * Returns %false if @work was already on a queue, %true otherwise.
1333 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1334 * it can be processed by another CPU.
1336 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1338 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1340 EXPORT_SYMBOL_GPL(queue_work);
1342 void delayed_work_timer_fn(unsigned long __data)
1344 struct delayed_work *dwork = (struct delayed_work *)__data;
1346 /* should have been called from irqsafe timer with irq already off */
1347 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1349 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1351 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1352 struct delayed_work *dwork, unsigned long delay)
1354 struct timer_list *timer = &dwork->timer;
1355 struct work_struct *work = &dwork->work;
1357 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1358 timer->data != (unsigned long)dwork);
1359 WARN_ON_ONCE(timer_pending(timer));
1360 WARN_ON_ONCE(!list_empty(&work->entry));
1363 * If @delay is 0, queue @dwork->work immediately. This is for
1364 * both optimization and correctness. The earliest @timer can
1365 * expire is on the closest next tick and delayed_work users depend
1366 * on that there's no such delay when @delay is 0.
1369 __queue_work(cpu, wq, &dwork->work);
1373 timer_stats_timer_set_start_info(&dwork->timer);
1377 timer->expires = jiffies + delay;
1379 if (unlikely(cpu != WORK_CPU_UNBOUND))
1380 add_timer_on(timer, cpu);
1386 * queue_delayed_work_on - queue work on specific CPU after delay
1387 * @cpu: CPU number to execute work on
1388 * @wq: workqueue to use
1389 * @dwork: work to queue
1390 * @delay: number of jiffies to wait before queueing
1392 * Returns %false if @work was already on a queue, %true otherwise. If
1393 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1396 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1397 struct delayed_work *dwork, unsigned long delay)
1399 struct work_struct *work = &dwork->work;
1401 unsigned long flags;
1403 /* read the comment in __queue_work() */
1404 local_irq_save(flags);
1406 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1407 __queue_delayed_work(cpu, wq, dwork, delay);
1411 local_irq_restore(flags);
1414 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1417 * queue_delayed_work - queue work on a workqueue after delay
1418 * @wq: workqueue to use
1419 * @dwork: delayable work to queue
1420 * @delay: number of jiffies to wait before queueing
1422 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1424 bool queue_delayed_work(struct workqueue_struct *wq,
1425 struct delayed_work *dwork, unsigned long delay)
1427 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1429 EXPORT_SYMBOL_GPL(queue_delayed_work);
1432 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1433 * @cpu: CPU number to execute work on
1434 * @wq: workqueue to use
1435 * @dwork: work to queue
1436 * @delay: number of jiffies to wait before queueing
1438 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1439 * modify @dwork's timer so that it expires after @delay. If @delay is
1440 * zero, @work is guaranteed to be scheduled immediately regardless of its
1443 * Returns %false if @dwork was idle and queued, %true if @dwork was
1444 * pending and its timer was modified.
1446 * This function is safe to call from any context including IRQ handler.
1447 * See try_to_grab_pending() for details.
1449 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1450 struct delayed_work *dwork, unsigned long delay)
1452 unsigned long flags;
1456 ret = try_to_grab_pending(&dwork->work, true, &flags);
1457 } while (unlikely(ret == -EAGAIN));
1459 if (likely(ret >= 0)) {
1460 __queue_delayed_work(cpu, wq, dwork, delay);
1461 local_irq_restore(flags);
1464 /* -ENOENT from try_to_grab_pending() becomes %true */
1467 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1470 * mod_delayed_work - modify delay of or queue a delayed work
1471 * @wq: workqueue to use
1472 * @dwork: work to queue
1473 * @delay: number of jiffies to wait before queueing
1475 * mod_delayed_work_on() on local CPU.
1477 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1478 unsigned long delay)
1480 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1482 EXPORT_SYMBOL_GPL(mod_delayed_work);
1485 * worker_enter_idle - enter idle state
1486 * @worker: worker which is entering idle state
1488 * @worker is entering idle state. Update stats and idle timer if
1492 * spin_lock_irq(pool->lock).
1494 static void worker_enter_idle(struct worker *worker)
1496 struct worker_pool *pool = worker->pool;
1498 BUG_ON(worker->flags & WORKER_IDLE);
1499 BUG_ON(!list_empty(&worker->entry) &&
1500 (worker->hentry.next || worker->hentry.pprev));
1502 /* can't use worker_set_flags(), also called from start_worker() */
1503 worker->flags |= WORKER_IDLE;
1505 worker->last_active = jiffies;
1507 /* idle_list is LIFO */
1508 list_add(&worker->entry, &pool->idle_list);
1510 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1511 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1514 * Sanity check nr_running. Because wq_unbind_fn() releases
1515 * pool->lock between setting %WORKER_UNBOUND and zapping
1516 * nr_running, the warning may trigger spuriously. Check iff
1517 * unbind is not in progress.
1519 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1520 pool->nr_workers == pool->nr_idle &&
1521 atomic_read(get_pool_nr_running(pool)));
1525 * worker_leave_idle - leave idle state
1526 * @worker: worker which is leaving idle state
1528 * @worker is leaving idle state. Update stats.
1531 * spin_lock_irq(pool->lock).
1533 static void worker_leave_idle(struct worker *worker)
1535 struct worker_pool *pool = worker->pool;
1537 BUG_ON(!(worker->flags & WORKER_IDLE));
1538 worker_clr_flags(worker, WORKER_IDLE);
1540 list_del_init(&worker->entry);
1544 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock pool
1547 * Works which are scheduled while the cpu is online must at least be
1548 * scheduled to a worker which is bound to the cpu so that if they are
1549 * flushed from cpu callbacks while cpu is going down, they are
1550 * guaranteed to execute on the cpu.
1552 * This function is to be used by rogue workers and rescuers to bind
1553 * themselves to the target cpu and may race with cpu going down or
1554 * coming online. kthread_bind() can't be used because it may put the
1555 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1556 * verbatim as it's best effort and blocking and pool may be
1557 * [dis]associated in the meantime.
1559 * This function tries set_cpus_allowed() and locks pool and verifies the
1560 * binding against %POOL_DISASSOCIATED which is set during
1561 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1562 * enters idle state or fetches works without dropping lock, it can
1563 * guarantee the scheduling requirement described in the first paragraph.
1566 * Might sleep. Called without any lock but returns with pool->lock
1570 * %true if the associated pool is online (@worker is successfully
1571 * bound), %false if offline.
1573 static bool worker_maybe_bind_and_lock(struct worker *worker)
1574 __acquires(&pool->lock)
1576 struct worker_pool *pool = worker->pool;
1577 struct task_struct *task = worker->task;
1581 * The following call may fail, succeed or succeed
1582 * without actually migrating the task to the cpu if
1583 * it races with cpu hotunplug operation. Verify
1584 * against POOL_DISASSOCIATED.
1586 if (!(pool->flags & POOL_DISASSOCIATED))
1587 set_cpus_allowed_ptr(task, get_cpu_mask(pool->cpu));
1589 spin_lock_irq(&pool->lock);
1590 if (pool->flags & POOL_DISASSOCIATED)
1592 if (task_cpu(task) == pool->cpu &&
1593 cpumask_equal(¤t->cpus_allowed,
1594 get_cpu_mask(pool->cpu)))
1596 spin_unlock_irq(&pool->lock);
1599 * We've raced with CPU hot[un]plug. Give it a breather
1600 * and retry migration. cond_resched() is required here;
1601 * otherwise, we might deadlock against cpu_stop trying to
1602 * bring down the CPU on non-preemptive kernel.
1610 * Rebind an idle @worker to its CPU. worker_thread() will test
1611 * list_empty(@worker->entry) before leaving idle and call this function.
1613 static void idle_worker_rebind(struct worker *worker)
1615 /* CPU may go down again inbetween, clear UNBOUND only on success */
1616 if (worker_maybe_bind_and_lock(worker))
1617 worker_clr_flags(worker, WORKER_UNBOUND);
1619 /* rebind complete, become available again */
1620 list_add(&worker->entry, &worker->pool->idle_list);
1621 spin_unlock_irq(&worker->pool->lock);
1625 * Function for @worker->rebind.work used to rebind unbound busy workers to
1626 * the associated cpu which is coming back online. This is scheduled by
1627 * cpu up but can race with other cpu hotplug operations and may be
1628 * executed twice without intervening cpu down.
1630 static void busy_worker_rebind_fn(struct work_struct *work)
1632 struct worker *worker = container_of(work, struct worker, rebind_work);
1634 if (worker_maybe_bind_and_lock(worker))
1635 worker_clr_flags(worker, WORKER_UNBOUND);
1637 spin_unlock_irq(&worker->pool->lock);
1641 * rebind_workers - rebind all workers of a pool to the associated CPU
1642 * @pool: pool of interest
1644 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1645 * is different for idle and busy ones.
1647 * Idle ones will be removed from the idle_list and woken up. They will
1648 * add themselves back after completing rebind. This ensures that the
1649 * idle_list doesn't contain any unbound workers when re-bound busy workers
1650 * try to perform local wake-ups for concurrency management.
1652 * Busy workers can rebind after they finish their current work items.
1653 * Queueing the rebind work item at the head of the scheduled list is
1654 * enough. Note that nr_running will be properly bumped as busy workers
1657 * On return, all non-manager workers are scheduled for rebind - see
1658 * manage_workers() for the manager special case. Any idle worker
1659 * including the manager will not appear on @idle_list until rebind is
1660 * complete, making local wake-ups safe.
1662 static void rebind_workers(struct worker_pool *pool)
1664 struct worker *worker, *n;
1665 struct hlist_node *pos;
1668 lockdep_assert_held(&pool->assoc_mutex);
1669 lockdep_assert_held(&pool->lock);
1671 /* dequeue and kick idle ones */
1672 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1674 * idle workers should be off @pool->idle_list until rebind
1675 * is complete to avoid receiving premature local wake-ups.
1677 list_del_init(&worker->entry);
1680 * worker_thread() will see the above dequeuing and call
1681 * idle_worker_rebind().
1683 wake_up_process(worker->task);
1686 /* rebind busy workers */
1687 for_each_busy_worker(worker, i, pos, pool) {
1688 struct work_struct *rebind_work = &worker->rebind_work;
1689 struct workqueue_struct *wq;
1691 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1692 work_data_bits(rebind_work)))
1695 debug_work_activate(rebind_work);
1698 * wq doesn't really matter but let's keep @worker->pool
1699 * and @cwq->pool consistent for sanity.
1701 if (std_worker_pool_pri(worker->pool))
1702 wq = system_highpri_wq;
1706 insert_work(get_cwq(pool->cpu, wq), rebind_work,
1707 worker->scheduled.next,
1708 work_color_to_flags(WORK_NO_COLOR));
1712 static struct worker *alloc_worker(void)
1714 struct worker *worker;
1716 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1718 INIT_LIST_HEAD(&worker->entry);
1719 INIT_LIST_HEAD(&worker->scheduled);
1720 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1721 /* on creation a worker is in !idle && prep state */
1722 worker->flags = WORKER_PREP;
1728 * create_worker - create a new workqueue worker
1729 * @pool: pool the new worker will belong to
1731 * Create a new worker which is bound to @pool. The returned worker
1732 * can be started by calling start_worker() or destroyed using
1736 * Might sleep. Does GFP_KERNEL allocations.
1739 * Pointer to the newly created worker.
1741 static struct worker *create_worker(struct worker_pool *pool)
1743 const char *pri = std_worker_pool_pri(pool) ? "H" : "";
1744 struct worker *worker = NULL;
1747 spin_lock_irq(&pool->lock);
1748 while (ida_get_new(&pool->worker_ida, &id)) {
1749 spin_unlock_irq(&pool->lock);
1750 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1752 spin_lock_irq(&pool->lock);
1754 spin_unlock_irq(&pool->lock);
1756 worker = alloc_worker();
1760 worker->pool = pool;
1763 if (pool->cpu != WORK_CPU_UNBOUND)
1764 worker->task = kthread_create_on_node(worker_thread,
1765 worker, cpu_to_node(pool->cpu),
1766 "kworker/%u:%d%s", pool->cpu, id, pri);
1768 worker->task = kthread_create(worker_thread, worker,
1769 "kworker/u:%d%s", id, pri);
1770 if (IS_ERR(worker->task))
1773 if (std_worker_pool_pri(pool))
1774 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1777 * Determine CPU binding of the new worker depending on
1778 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1779 * flag remains stable across this function. See the comments
1780 * above the flag definition for details.
1782 * As an unbound worker may later become a regular one if CPU comes
1783 * online, make sure every worker has %PF_THREAD_BOUND set.
1785 if (!(pool->flags & POOL_DISASSOCIATED)) {
1786 kthread_bind(worker->task, pool->cpu);
1788 worker->task->flags |= PF_THREAD_BOUND;
1789 worker->flags |= WORKER_UNBOUND;
1795 spin_lock_irq(&pool->lock);
1796 ida_remove(&pool->worker_ida, id);
1797 spin_unlock_irq(&pool->lock);
1804 * start_worker - start a newly created worker
1805 * @worker: worker to start
1807 * Make the pool aware of @worker and start it.
1810 * spin_lock_irq(pool->lock).
1812 static void start_worker(struct worker *worker)
1814 worker->flags |= WORKER_STARTED;
1815 worker->pool->nr_workers++;
1816 worker_enter_idle(worker);
1817 wake_up_process(worker->task);
1821 * destroy_worker - destroy a workqueue worker
1822 * @worker: worker to be destroyed
1824 * Destroy @worker and adjust @pool stats accordingly.
1827 * spin_lock_irq(pool->lock) which is released and regrabbed.
1829 static void destroy_worker(struct worker *worker)
1831 struct worker_pool *pool = worker->pool;
1832 int id = worker->id;
1834 /* sanity check frenzy */
1835 BUG_ON(worker->current_work);
1836 BUG_ON(!list_empty(&worker->scheduled));
1838 if (worker->flags & WORKER_STARTED)
1840 if (worker->flags & WORKER_IDLE)
1843 list_del_init(&worker->entry);
1844 worker->flags |= WORKER_DIE;
1846 spin_unlock_irq(&pool->lock);
1848 kthread_stop(worker->task);
1851 spin_lock_irq(&pool->lock);
1852 ida_remove(&pool->worker_ida, id);
1855 static void idle_worker_timeout(unsigned long __pool)
1857 struct worker_pool *pool = (void *)__pool;
1859 spin_lock_irq(&pool->lock);
1861 if (too_many_workers(pool)) {
1862 struct worker *worker;
1863 unsigned long expires;
1865 /* idle_list is kept in LIFO order, check the last one */
1866 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1867 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1869 if (time_before(jiffies, expires))
1870 mod_timer(&pool->idle_timer, expires);
1872 /* it's been idle for too long, wake up manager */
1873 pool->flags |= POOL_MANAGE_WORKERS;
1874 wake_up_worker(pool);
1878 spin_unlock_irq(&pool->lock);
1881 static bool send_mayday(struct work_struct *work)
1883 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1884 struct workqueue_struct *wq = cwq->wq;
1887 if (!(wq->flags & WQ_RESCUER))
1890 /* mayday mayday mayday */
1891 cpu = cwq->pool->cpu;
1892 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1893 if (cpu == WORK_CPU_UNBOUND)
1895 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1896 wake_up_process(wq->rescuer->task);
1900 static void pool_mayday_timeout(unsigned long __pool)
1902 struct worker_pool *pool = (void *)__pool;
1903 struct work_struct *work;
1905 spin_lock_irq(&pool->lock);
1907 if (need_to_create_worker(pool)) {
1909 * We've been trying to create a new worker but
1910 * haven't been successful. We might be hitting an
1911 * allocation deadlock. Send distress signals to
1914 list_for_each_entry(work, &pool->worklist, entry)
1918 spin_unlock_irq(&pool->lock);
1920 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1924 * maybe_create_worker - create a new worker if necessary
1925 * @pool: pool to create a new worker for
1927 * Create a new worker for @pool if necessary. @pool is guaranteed to
1928 * have at least one idle worker on return from this function. If
1929 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1930 * sent to all rescuers with works scheduled on @pool to resolve
1931 * possible allocation deadlock.
1933 * On return, need_to_create_worker() is guaranteed to be false and
1934 * may_start_working() true.
1937 * spin_lock_irq(pool->lock) which may be released and regrabbed
1938 * multiple times. Does GFP_KERNEL allocations. Called only from
1942 * false if no action was taken and pool->lock stayed locked, true
1945 static bool maybe_create_worker(struct worker_pool *pool)
1946 __releases(&pool->lock)
1947 __acquires(&pool->lock)
1949 if (!need_to_create_worker(pool))
1952 spin_unlock_irq(&pool->lock);
1954 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1955 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1958 struct worker *worker;
1960 worker = create_worker(pool);
1962 del_timer_sync(&pool->mayday_timer);
1963 spin_lock_irq(&pool->lock);
1964 start_worker(worker);
1965 BUG_ON(need_to_create_worker(pool));
1969 if (!need_to_create_worker(pool))
1972 __set_current_state(TASK_INTERRUPTIBLE);
1973 schedule_timeout(CREATE_COOLDOWN);
1975 if (!need_to_create_worker(pool))
1979 del_timer_sync(&pool->mayday_timer);
1980 spin_lock_irq(&pool->lock);
1981 if (need_to_create_worker(pool))
1987 * maybe_destroy_worker - destroy workers which have been idle for a while
1988 * @pool: pool to destroy workers for
1990 * Destroy @pool workers which have been idle for longer than
1991 * IDLE_WORKER_TIMEOUT.
1994 * spin_lock_irq(pool->lock) which may be released and regrabbed
1995 * multiple times. Called only from manager.
1998 * false if no action was taken and pool->lock stayed locked, true
2001 static bool maybe_destroy_workers(struct worker_pool *pool)
2005 while (too_many_workers(pool)) {
2006 struct worker *worker;
2007 unsigned long expires;
2009 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2010 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2012 if (time_before(jiffies, expires)) {
2013 mod_timer(&pool->idle_timer, expires);
2017 destroy_worker(worker);
2025 * manage_workers - manage worker pool
2028 * Assume the manager role and manage the worker pool @worker belongs
2029 * to. At any given time, there can be only zero or one manager per
2030 * pool. The exclusion is handled automatically by this function.
2032 * The caller can safely start processing works on false return. On
2033 * true return, it's guaranteed that need_to_create_worker() is false
2034 * and may_start_working() is true.
2037 * spin_lock_irq(pool->lock) which may be released and regrabbed
2038 * multiple times. Does GFP_KERNEL allocations.
2041 * spin_lock_irq(pool->lock) which may be released and regrabbed
2042 * multiple times. Does GFP_KERNEL allocations.
2044 static bool manage_workers(struct worker *worker)
2046 struct worker_pool *pool = worker->pool;
2049 if (pool->flags & POOL_MANAGING_WORKERS)
2052 pool->flags |= POOL_MANAGING_WORKERS;
2055 * To simplify both worker management and CPU hotplug, hold off
2056 * management while hotplug is in progress. CPU hotplug path can't
2057 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2058 * lead to idle worker depletion (all become busy thinking someone
2059 * else is managing) which in turn can result in deadlock under
2060 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2061 * manager against CPU hotplug.
2063 * assoc_mutex would always be free unless CPU hotplug is in
2064 * progress. trylock first without dropping @pool->lock.
2066 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2067 spin_unlock_irq(&pool->lock);
2068 mutex_lock(&pool->assoc_mutex);
2070 * CPU hotplug could have happened while we were waiting
2071 * for assoc_mutex. Hotplug itself can't handle us
2072 * because manager isn't either on idle or busy list, and
2073 * @pool's state and ours could have deviated.
2075 * As hotplug is now excluded via assoc_mutex, we can
2076 * simply try to bind. It will succeed or fail depending
2077 * on @pool's current state. Try it and adjust
2078 * %WORKER_UNBOUND accordingly.
2080 if (worker_maybe_bind_and_lock(worker))
2081 worker->flags &= ~WORKER_UNBOUND;
2083 worker->flags |= WORKER_UNBOUND;
2088 pool->flags &= ~POOL_MANAGE_WORKERS;
2091 * Destroy and then create so that may_start_working() is true
2094 ret |= maybe_destroy_workers(pool);
2095 ret |= maybe_create_worker(pool);
2097 pool->flags &= ~POOL_MANAGING_WORKERS;
2098 mutex_unlock(&pool->assoc_mutex);
2103 * process_one_work - process single work
2105 * @work: work to process
2107 * Process @work. This function contains all the logics necessary to
2108 * process a single work including synchronization against and
2109 * interaction with other workers on the same cpu, queueing and
2110 * flushing. As long as context requirement is met, any worker can
2111 * call this function to process a work.
2114 * spin_lock_irq(pool->lock) which is released and regrabbed.
2116 static void process_one_work(struct worker *worker, struct work_struct *work)
2117 __releases(&pool->lock)
2118 __acquires(&pool->lock)
2120 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2121 struct worker_pool *pool = worker->pool;
2122 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2124 struct worker *collision;
2125 #ifdef CONFIG_LOCKDEP
2127 * It is permissible to free the struct work_struct from
2128 * inside the function that is called from it, this we need to
2129 * take into account for lockdep too. To avoid bogus "held
2130 * lock freed" warnings as well as problems when looking into
2131 * work->lockdep_map, make a copy and use that here.
2133 struct lockdep_map lockdep_map;
2135 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2138 * Ensure we're on the correct CPU. DISASSOCIATED test is
2139 * necessary to avoid spurious warnings from rescuers servicing the
2140 * unbound or a disassociated pool.
2142 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2143 !(pool->flags & POOL_DISASSOCIATED) &&
2144 raw_smp_processor_id() != pool->cpu);
2147 * A single work shouldn't be executed concurrently by
2148 * multiple workers on a single cpu. Check whether anyone is
2149 * already processing the work. If so, defer the work to the
2150 * currently executing one.
2152 collision = find_worker_executing_work(pool, work);
2153 if (unlikely(collision)) {
2154 move_linked_works(work, &collision->scheduled, NULL);
2158 /* claim and dequeue */
2159 debug_work_deactivate(work);
2160 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2161 worker->current_work = work;
2162 worker->current_func = work->func;
2163 worker->current_cwq = cwq;
2164 work_color = get_work_color(work);
2166 list_del_init(&work->entry);
2169 * CPU intensive works don't participate in concurrency
2170 * management. They're the scheduler's responsibility.
2172 if (unlikely(cpu_intensive))
2173 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2176 * Unbound pool isn't concurrency managed and work items should be
2177 * executed ASAP. Wake up another worker if necessary.
2179 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2180 wake_up_worker(pool);
2183 * Record the last pool and clear PENDING which should be the last
2184 * update to @work. Also, do this inside @pool->lock so that
2185 * PENDING and queued state changes happen together while IRQ is
2188 set_work_pool_and_clear_pending(work, pool->id);
2190 spin_unlock_irq(&pool->lock);
2192 lock_map_acquire_read(&cwq->wq->lockdep_map);
2193 lock_map_acquire(&lockdep_map);
2194 trace_workqueue_execute_start(work);
2195 worker->current_func(work);
2197 * While we must be careful to not use "work" after this, the trace
2198 * point will only record its address.
2200 trace_workqueue_execute_end(work);
2201 lock_map_release(&lockdep_map);
2202 lock_map_release(&cwq->wq->lockdep_map);
2204 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2205 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2206 " last function: %pf\n",
2207 current->comm, preempt_count(), task_pid_nr(current),
2208 worker->current_func);
2209 debug_show_held_locks(current);
2213 spin_lock_irq(&pool->lock);
2215 /* clear cpu intensive status */
2216 if (unlikely(cpu_intensive))
2217 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2219 /* we're done with it, release */
2220 hash_del(&worker->hentry);
2221 worker->current_work = NULL;
2222 worker->current_func = NULL;
2223 worker->current_cwq = NULL;
2224 cwq_dec_nr_in_flight(cwq, work_color);
2228 * process_scheduled_works - process scheduled works
2231 * Process all scheduled works. Please note that the scheduled list
2232 * may change while processing a work, so this function repeatedly
2233 * fetches a work from the top and executes it.
2236 * spin_lock_irq(pool->lock) which may be released and regrabbed
2239 static void process_scheduled_works(struct worker *worker)
2241 while (!list_empty(&worker->scheduled)) {
2242 struct work_struct *work = list_first_entry(&worker->scheduled,
2243 struct work_struct, entry);
2244 process_one_work(worker, work);
2249 * worker_thread - the worker thread function
2252 * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools
2253 * of these per each cpu. These workers process all works regardless of
2254 * their specific target workqueue. The only exception is works which
2255 * belong to workqueues with a rescuer which will be explained in
2258 static int worker_thread(void *__worker)
2260 struct worker *worker = __worker;
2261 struct worker_pool *pool = worker->pool;
2263 /* tell the scheduler that this is a workqueue worker */
2264 worker->task->flags |= PF_WQ_WORKER;
2266 spin_lock_irq(&pool->lock);
2268 /* we are off idle list if destruction or rebind is requested */
2269 if (unlikely(list_empty(&worker->entry))) {
2270 spin_unlock_irq(&pool->lock);
2272 /* if DIE is set, destruction is requested */
2273 if (worker->flags & WORKER_DIE) {
2274 worker->task->flags &= ~PF_WQ_WORKER;
2278 /* otherwise, rebind */
2279 idle_worker_rebind(worker);
2283 worker_leave_idle(worker);
2285 /* no more worker necessary? */
2286 if (!need_more_worker(pool))
2289 /* do we need to manage? */
2290 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2294 * ->scheduled list can only be filled while a worker is
2295 * preparing to process a work or actually processing it.
2296 * Make sure nobody diddled with it while I was sleeping.
2298 BUG_ON(!list_empty(&worker->scheduled));
2301 * When control reaches this point, we're guaranteed to have
2302 * at least one idle worker or that someone else has already
2303 * assumed the manager role.
2305 worker_clr_flags(worker, WORKER_PREP);
2308 struct work_struct *work =
2309 list_first_entry(&pool->worklist,
2310 struct work_struct, entry);
2312 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2313 /* optimization path, not strictly necessary */
2314 process_one_work(worker, work);
2315 if (unlikely(!list_empty(&worker->scheduled)))
2316 process_scheduled_works(worker);
2318 move_linked_works(work, &worker->scheduled, NULL);
2319 process_scheduled_works(worker);
2321 } while (keep_working(pool));
2323 worker_set_flags(worker, WORKER_PREP, false);
2325 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2329 * pool->lock is held and there's no work to process and no need to
2330 * manage, sleep. Workers are woken up only while holding
2331 * pool->lock or from local cpu, so setting the current state
2332 * before releasing pool->lock is enough to prevent losing any
2335 worker_enter_idle(worker);
2336 __set_current_state(TASK_INTERRUPTIBLE);
2337 spin_unlock_irq(&pool->lock);
2343 * rescuer_thread - the rescuer thread function
2346 * Workqueue rescuer thread function. There's one rescuer for each
2347 * workqueue which has WQ_RESCUER set.
2349 * Regular work processing on a pool may block trying to create a new
2350 * worker which uses GFP_KERNEL allocation which has slight chance of
2351 * developing into deadlock if some works currently on the same queue
2352 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2353 * the problem rescuer solves.
2355 * When such condition is possible, the pool summons rescuers of all
2356 * workqueues which have works queued on the pool and let them process
2357 * those works so that forward progress can be guaranteed.
2359 * This should happen rarely.
2361 static int rescuer_thread(void *__rescuer)
2363 struct worker *rescuer = __rescuer;
2364 struct workqueue_struct *wq = rescuer->rescue_wq;
2365 struct list_head *scheduled = &rescuer->scheduled;
2366 bool is_unbound = wq->flags & WQ_UNBOUND;
2369 set_user_nice(current, RESCUER_NICE_LEVEL);
2372 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2373 * doesn't participate in concurrency management.
2375 rescuer->task->flags |= PF_WQ_WORKER;
2377 set_current_state(TASK_INTERRUPTIBLE);
2379 if (kthread_should_stop()) {
2380 __set_current_state(TASK_RUNNING);
2381 rescuer->task->flags &= ~PF_WQ_WORKER;
2386 * See whether any cpu is asking for help. Unbounded
2387 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2389 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2390 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2391 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2392 struct worker_pool *pool = cwq->pool;
2393 struct work_struct *work, *n;
2395 __set_current_state(TASK_RUNNING);
2396 mayday_clear_cpu(cpu, wq->mayday_mask);
2398 /* migrate to the target cpu if possible */
2399 rescuer->pool = pool;
2400 worker_maybe_bind_and_lock(rescuer);
2403 * Slurp in all works issued via this workqueue and
2406 BUG_ON(!list_empty(&rescuer->scheduled));
2407 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2408 if (get_work_cwq(work) == cwq)
2409 move_linked_works(work, scheduled, &n);
2411 process_scheduled_works(rescuer);
2414 * Leave this pool. If keep_working() is %true, notify a
2415 * regular worker; otherwise, we end up with 0 concurrency
2416 * and stalling the execution.
2418 if (keep_working(pool))
2419 wake_up_worker(pool);
2421 spin_unlock_irq(&pool->lock);
2424 /* rescuers should never participate in concurrency management */
2425 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2431 struct work_struct work;
2432 struct completion done;
2435 static void wq_barrier_func(struct work_struct *work)
2437 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2438 complete(&barr->done);
2442 * insert_wq_barrier - insert a barrier work
2443 * @cwq: cwq to insert barrier into
2444 * @barr: wq_barrier to insert
2445 * @target: target work to attach @barr to
2446 * @worker: worker currently executing @target, NULL if @target is not executing
2448 * @barr is linked to @target such that @barr is completed only after
2449 * @target finishes execution. Please note that the ordering
2450 * guarantee is observed only with respect to @target and on the local
2453 * Currently, a queued barrier can't be canceled. This is because
2454 * try_to_grab_pending() can't determine whether the work to be
2455 * grabbed is at the head of the queue and thus can't clear LINKED
2456 * flag of the previous work while there must be a valid next work
2457 * after a work with LINKED flag set.
2459 * Note that when @worker is non-NULL, @target may be modified
2460 * underneath us, so we can't reliably determine cwq from @target.
2463 * spin_lock_irq(pool->lock).
2465 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2466 struct wq_barrier *barr,
2467 struct work_struct *target, struct worker *worker)
2469 struct list_head *head;
2470 unsigned int linked = 0;
2473 * debugobject calls are safe here even with pool->lock locked
2474 * as we know for sure that this will not trigger any of the
2475 * checks and call back into the fixup functions where we
2478 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2479 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2480 init_completion(&barr->done);
2483 * If @target is currently being executed, schedule the
2484 * barrier to the worker; otherwise, put it after @target.
2487 head = worker->scheduled.next;
2489 unsigned long *bits = work_data_bits(target);
2491 head = target->entry.next;
2492 /* there can already be other linked works, inherit and set */
2493 linked = *bits & WORK_STRUCT_LINKED;
2494 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2497 debug_work_activate(&barr->work);
2498 insert_work(cwq, &barr->work, head,
2499 work_color_to_flags(WORK_NO_COLOR) | linked);
2503 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2504 * @wq: workqueue being flushed
2505 * @flush_color: new flush color, < 0 for no-op
2506 * @work_color: new work color, < 0 for no-op
2508 * Prepare cwqs for workqueue flushing.
2510 * If @flush_color is non-negative, flush_color on all cwqs should be
2511 * -1. If no cwq has in-flight commands at the specified color, all
2512 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2513 * has in flight commands, its cwq->flush_color is set to
2514 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2515 * wakeup logic is armed and %true is returned.
2517 * The caller should have initialized @wq->first_flusher prior to
2518 * calling this function with non-negative @flush_color. If
2519 * @flush_color is negative, no flush color update is done and %false
2522 * If @work_color is non-negative, all cwqs should have the same
2523 * work_color which is previous to @work_color and all will be
2524 * advanced to @work_color.
2527 * mutex_lock(wq->flush_mutex).
2530 * %true if @flush_color >= 0 and there's something to flush. %false
2533 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2534 int flush_color, int work_color)
2539 if (flush_color >= 0) {
2540 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2541 atomic_set(&wq->nr_cwqs_to_flush, 1);
2544 for_each_cwq_cpu(cpu, wq) {
2545 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2546 struct worker_pool *pool = cwq->pool;
2548 spin_lock_irq(&pool->lock);
2550 if (flush_color >= 0) {
2551 BUG_ON(cwq->flush_color != -1);
2553 if (cwq->nr_in_flight[flush_color]) {
2554 cwq->flush_color = flush_color;
2555 atomic_inc(&wq->nr_cwqs_to_flush);
2560 if (work_color >= 0) {
2561 BUG_ON(work_color != work_next_color(cwq->work_color));
2562 cwq->work_color = work_color;
2565 spin_unlock_irq(&pool->lock);
2568 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2569 complete(&wq->first_flusher->done);
2575 * flush_workqueue - ensure that any scheduled work has run to completion.
2576 * @wq: workqueue to flush
2578 * Forces execution of the workqueue and blocks until its completion.
2579 * This is typically used in driver shutdown handlers.
2581 * We sleep until all works which were queued on entry have been handled,
2582 * but we are not livelocked by new incoming ones.
2584 void flush_workqueue(struct workqueue_struct *wq)
2586 struct wq_flusher this_flusher = {
2587 .list = LIST_HEAD_INIT(this_flusher.list),
2589 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2593 lock_map_acquire(&wq->lockdep_map);
2594 lock_map_release(&wq->lockdep_map);
2596 mutex_lock(&wq->flush_mutex);
2599 * Start-to-wait phase
2601 next_color = work_next_color(wq->work_color);
2603 if (next_color != wq->flush_color) {
2605 * Color space is not full. The current work_color
2606 * becomes our flush_color and work_color is advanced
2609 BUG_ON(!list_empty(&wq->flusher_overflow));
2610 this_flusher.flush_color = wq->work_color;
2611 wq->work_color = next_color;
2613 if (!wq->first_flusher) {
2614 /* no flush in progress, become the first flusher */
2615 BUG_ON(wq->flush_color != this_flusher.flush_color);
2617 wq->first_flusher = &this_flusher;
2619 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2621 /* nothing to flush, done */
2622 wq->flush_color = next_color;
2623 wq->first_flusher = NULL;
2628 BUG_ON(wq->flush_color == this_flusher.flush_color);
2629 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2630 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2634 * Oops, color space is full, wait on overflow queue.
2635 * The next flush completion will assign us
2636 * flush_color and transfer to flusher_queue.
2638 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2641 mutex_unlock(&wq->flush_mutex);
2643 wait_for_completion(&this_flusher.done);
2646 * Wake-up-and-cascade phase
2648 * First flushers are responsible for cascading flushes and
2649 * handling overflow. Non-first flushers can simply return.
2651 if (wq->first_flusher != &this_flusher)
2654 mutex_lock(&wq->flush_mutex);
2656 /* we might have raced, check again with mutex held */
2657 if (wq->first_flusher != &this_flusher)
2660 wq->first_flusher = NULL;
2662 BUG_ON(!list_empty(&this_flusher.list));
2663 BUG_ON(wq->flush_color != this_flusher.flush_color);
2666 struct wq_flusher *next, *tmp;
2668 /* complete all the flushers sharing the current flush color */
2669 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2670 if (next->flush_color != wq->flush_color)
2672 list_del_init(&next->list);
2673 complete(&next->done);
2676 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2677 wq->flush_color != work_next_color(wq->work_color));
2679 /* this flush_color is finished, advance by one */
2680 wq->flush_color = work_next_color(wq->flush_color);
2682 /* one color has been freed, handle overflow queue */
2683 if (!list_empty(&wq->flusher_overflow)) {
2685 * Assign the same color to all overflowed
2686 * flushers, advance work_color and append to
2687 * flusher_queue. This is the start-to-wait
2688 * phase for these overflowed flushers.
2690 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2691 tmp->flush_color = wq->work_color;
2693 wq->work_color = work_next_color(wq->work_color);
2695 list_splice_tail_init(&wq->flusher_overflow,
2696 &wq->flusher_queue);
2697 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2700 if (list_empty(&wq->flusher_queue)) {
2701 BUG_ON(wq->flush_color != wq->work_color);
2706 * Need to flush more colors. Make the next flusher
2707 * the new first flusher and arm cwqs.
2709 BUG_ON(wq->flush_color == wq->work_color);
2710 BUG_ON(wq->flush_color != next->flush_color);
2712 list_del_init(&next->list);
2713 wq->first_flusher = next;
2715 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2719 * Meh... this color is already done, clear first
2720 * flusher and repeat cascading.
2722 wq->first_flusher = NULL;
2726 mutex_unlock(&wq->flush_mutex);
2728 EXPORT_SYMBOL_GPL(flush_workqueue);
2731 * drain_workqueue - drain a workqueue
2732 * @wq: workqueue to drain
2734 * Wait until the workqueue becomes empty. While draining is in progress,
2735 * only chain queueing is allowed. IOW, only currently pending or running
2736 * work items on @wq can queue further work items on it. @wq is flushed
2737 * repeatedly until it becomes empty. The number of flushing is detemined
2738 * by the depth of chaining and should be relatively short. Whine if it
2741 void drain_workqueue(struct workqueue_struct *wq)
2743 unsigned int flush_cnt = 0;
2747 * __queue_work() needs to test whether there are drainers, is much
2748 * hotter than drain_workqueue() and already looks at @wq->flags.
2749 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2751 spin_lock(&workqueue_lock);
2752 if (!wq->nr_drainers++)
2753 wq->flags |= WQ_DRAINING;
2754 spin_unlock(&workqueue_lock);
2756 flush_workqueue(wq);
2758 for_each_cwq_cpu(cpu, wq) {
2759 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2762 spin_lock_irq(&cwq->pool->lock);
2763 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2764 spin_unlock_irq(&cwq->pool->lock);
2769 if (++flush_cnt == 10 ||
2770 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2771 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2772 wq->name, flush_cnt);
2776 spin_lock(&workqueue_lock);
2777 if (!--wq->nr_drainers)
2778 wq->flags &= ~WQ_DRAINING;
2779 spin_unlock(&workqueue_lock);
2781 EXPORT_SYMBOL_GPL(drain_workqueue);
2783 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2785 struct worker *worker = NULL;
2786 struct worker_pool *pool;
2787 struct cpu_workqueue_struct *cwq;
2790 pool = get_work_pool(work);
2794 spin_lock_irq(&pool->lock);
2795 /* see the comment in try_to_grab_pending() with the same code */
2796 cwq = get_work_cwq(work);
2798 if (unlikely(cwq->pool != pool))
2801 worker = find_worker_executing_work(pool, work);
2804 cwq = worker->current_cwq;
2807 insert_wq_barrier(cwq, barr, work, worker);
2808 spin_unlock_irq(&pool->lock);
2811 * If @max_active is 1 or rescuer is in use, flushing another work
2812 * item on the same workqueue may lead to deadlock. Make sure the
2813 * flusher is not running on the same workqueue by verifying write
2816 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2817 lock_map_acquire(&cwq->wq->lockdep_map);
2819 lock_map_acquire_read(&cwq->wq->lockdep_map);
2820 lock_map_release(&cwq->wq->lockdep_map);
2824 spin_unlock_irq(&pool->lock);
2829 * flush_work - wait for a work to finish executing the last queueing instance
2830 * @work: the work to flush
2832 * Wait until @work has finished execution. @work is guaranteed to be idle
2833 * on return if it hasn't been requeued since flush started.
2836 * %true if flush_work() waited for the work to finish execution,
2837 * %false if it was already idle.
2839 bool flush_work(struct work_struct *work)
2841 struct wq_barrier barr;
2843 lock_map_acquire(&work->lockdep_map);
2844 lock_map_release(&work->lockdep_map);
2846 if (start_flush_work(work, &barr)) {
2847 wait_for_completion(&barr.done);
2848 destroy_work_on_stack(&barr.work);
2854 EXPORT_SYMBOL_GPL(flush_work);
2856 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2858 unsigned long flags;
2862 ret = try_to_grab_pending(work, is_dwork, &flags);
2864 * If someone else is canceling, wait for the same event it
2865 * would be waiting for before retrying.
2867 if (unlikely(ret == -ENOENT))
2869 } while (unlikely(ret < 0));
2871 /* tell other tasks trying to grab @work to back off */
2872 mark_work_canceling(work);
2873 local_irq_restore(flags);
2876 clear_work_data(work);
2881 * cancel_work_sync - cancel a work and wait for it to finish
2882 * @work: the work to cancel
2884 * Cancel @work and wait for its execution to finish. This function
2885 * can be used even if the work re-queues itself or migrates to
2886 * another workqueue. On return from this function, @work is
2887 * guaranteed to be not pending or executing on any CPU.
2889 * cancel_work_sync(&delayed_work->work) must not be used for
2890 * delayed_work's. Use cancel_delayed_work_sync() instead.
2892 * The caller must ensure that the workqueue on which @work was last
2893 * queued can't be destroyed before this function returns.
2896 * %true if @work was pending, %false otherwise.
2898 bool cancel_work_sync(struct work_struct *work)
2900 return __cancel_work_timer(work, false);
2902 EXPORT_SYMBOL_GPL(cancel_work_sync);
2905 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2906 * @dwork: the delayed work to flush
2908 * Delayed timer is cancelled and the pending work is queued for
2909 * immediate execution. Like flush_work(), this function only
2910 * considers the last queueing instance of @dwork.
2913 * %true if flush_work() waited for the work to finish execution,
2914 * %false if it was already idle.
2916 bool flush_delayed_work(struct delayed_work *dwork)
2918 local_irq_disable();
2919 if (del_timer_sync(&dwork->timer))
2920 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2922 return flush_work(&dwork->work);
2924 EXPORT_SYMBOL(flush_delayed_work);
2927 * cancel_delayed_work - cancel a delayed work
2928 * @dwork: delayed_work to cancel
2930 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2931 * and canceled; %false if wasn't pending. Note that the work callback
2932 * function may still be running on return, unless it returns %true and the
2933 * work doesn't re-arm itself. Explicitly flush or use
2934 * cancel_delayed_work_sync() to wait on it.
2936 * This function is safe to call from any context including IRQ handler.
2938 bool cancel_delayed_work(struct delayed_work *dwork)
2940 unsigned long flags;
2944 ret = try_to_grab_pending(&dwork->work, true, &flags);
2945 } while (unlikely(ret == -EAGAIN));
2947 if (unlikely(ret < 0))
2950 set_work_pool_and_clear_pending(&dwork->work,
2951 get_work_pool_id(&dwork->work));
2952 local_irq_restore(flags);
2955 EXPORT_SYMBOL(cancel_delayed_work);
2958 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2959 * @dwork: the delayed work cancel
2961 * This is cancel_work_sync() for delayed works.
2964 * %true if @dwork was pending, %false otherwise.
2966 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2968 return __cancel_work_timer(&dwork->work, true);
2970 EXPORT_SYMBOL(cancel_delayed_work_sync);
2973 * schedule_work_on - put work task on a specific cpu
2974 * @cpu: cpu to put the work task on
2975 * @work: job to be done
2977 * This puts a job on a specific cpu
2979 bool schedule_work_on(int cpu, struct work_struct *work)
2981 return queue_work_on(cpu, system_wq, work);
2983 EXPORT_SYMBOL(schedule_work_on);
2986 * schedule_work - put work task in global workqueue
2987 * @work: job to be done
2989 * Returns %false if @work was already on the kernel-global workqueue and
2992 * This puts a job in the kernel-global workqueue if it was not already
2993 * queued and leaves it in the same position on the kernel-global
2994 * workqueue otherwise.
2996 bool schedule_work(struct work_struct *work)
2998 return queue_work(system_wq, work);
3000 EXPORT_SYMBOL(schedule_work);
3003 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3005 * @dwork: job to be done
3006 * @delay: number of jiffies to wait
3008 * After waiting for a given time this puts a job in the kernel-global
3009 * workqueue on the specified CPU.
3011 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3012 unsigned long delay)
3014 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3016 EXPORT_SYMBOL(schedule_delayed_work_on);
3019 * schedule_delayed_work - put work task in global workqueue after delay
3020 * @dwork: job to be done
3021 * @delay: number of jiffies to wait or 0 for immediate execution
3023 * After waiting for a given time this puts a job in the kernel-global
3026 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3028 return queue_delayed_work(system_wq, dwork, delay);
3030 EXPORT_SYMBOL(schedule_delayed_work);
3033 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3034 * @func: the function to call
3036 * schedule_on_each_cpu() executes @func on each online CPU using the
3037 * system workqueue and blocks until all CPUs have completed.
3038 * schedule_on_each_cpu() is very slow.
3041 * 0 on success, -errno on failure.
3043 int schedule_on_each_cpu(work_func_t func)
3046 struct work_struct __percpu *works;
3048 works = alloc_percpu(struct work_struct);
3054 for_each_online_cpu(cpu) {
3055 struct work_struct *work = per_cpu_ptr(works, cpu);
3057 INIT_WORK(work, func);
3058 schedule_work_on(cpu, work);
3061 for_each_online_cpu(cpu)
3062 flush_work(per_cpu_ptr(works, cpu));
3070 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3072 * Forces execution of the kernel-global workqueue and blocks until its
3075 * Think twice before calling this function! It's very easy to get into
3076 * trouble if you don't take great care. Either of the following situations
3077 * will lead to deadlock:
3079 * One of the work items currently on the workqueue needs to acquire
3080 * a lock held by your code or its caller.
3082 * Your code is running in the context of a work routine.
3084 * They will be detected by lockdep when they occur, but the first might not
3085 * occur very often. It depends on what work items are on the workqueue and
3086 * what locks they need, which you have no control over.
3088 * In most situations flushing the entire workqueue is overkill; you merely
3089 * need to know that a particular work item isn't queued and isn't running.
3090 * In such cases you should use cancel_delayed_work_sync() or
3091 * cancel_work_sync() instead.
3093 void flush_scheduled_work(void)
3095 flush_workqueue(system_wq);
3097 EXPORT_SYMBOL(flush_scheduled_work);
3100 * execute_in_process_context - reliably execute the routine with user context
3101 * @fn: the function to execute
3102 * @ew: guaranteed storage for the execute work structure (must
3103 * be available when the work executes)
3105 * Executes the function immediately if process context is available,
3106 * otherwise schedules the function for delayed execution.
3108 * Returns: 0 - function was executed
3109 * 1 - function was scheduled for execution
3111 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3113 if (!in_interrupt()) {
3118 INIT_WORK(&ew->work, fn);
3119 schedule_work(&ew->work);
3123 EXPORT_SYMBOL_GPL(execute_in_process_context);
3125 int keventd_up(void)
3127 return system_wq != NULL;
3130 static int alloc_cwqs(struct workqueue_struct *wq)
3133 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3134 * Make sure that the alignment isn't lower than that of
3135 * unsigned long long.
3137 const size_t size = sizeof(struct cpu_workqueue_struct);
3138 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3139 __alignof__(unsigned long long));
3141 if (!(wq->flags & WQ_UNBOUND))
3142 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3147 * Allocate enough room to align cwq and put an extra
3148 * pointer at the end pointing back to the originally
3149 * allocated pointer which will be used for free.
3151 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3153 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3154 *(void **)(wq->cpu_wq.single + 1) = ptr;
3158 /* just in case, make sure it's actually aligned */
3159 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3160 return wq->cpu_wq.v ? 0 : -ENOMEM;
3163 static void free_cwqs(struct workqueue_struct *wq)
3165 if (!(wq->flags & WQ_UNBOUND))
3166 free_percpu(wq->cpu_wq.pcpu);
3167 else if (wq->cpu_wq.single) {
3168 /* the pointer to free is stored right after the cwq */
3169 kfree(*(void **)(wq->cpu_wq.single + 1));
3173 static int wq_clamp_max_active(int max_active, unsigned int flags,
3176 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3178 if (max_active < 1 || max_active > lim)
3179 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3180 max_active, name, 1, lim);
3182 return clamp_val(max_active, 1, lim);
3185 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3188 struct lock_class_key *key,
3189 const char *lock_name, ...)
3191 va_list args, args1;
3192 struct workqueue_struct *wq;
3196 /* determine namelen, allocate wq and format name */
3197 va_start(args, lock_name);
3198 va_copy(args1, args);
3199 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3201 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3205 vsnprintf(wq->name, namelen, fmt, args1);
3210 * Workqueues which may be used during memory reclaim should
3211 * have a rescuer to guarantee forward progress.
3213 if (flags & WQ_MEM_RECLAIM)
3214 flags |= WQ_RESCUER;
3216 max_active = max_active ?: WQ_DFL_ACTIVE;
3217 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3221 wq->saved_max_active = max_active;
3222 mutex_init(&wq->flush_mutex);
3223 atomic_set(&wq->nr_cwqs_to_flush, 0);
3224 INIT_LIST_HEAD(&wq->flusher_queue);
3225 INIT_LIST_HEAD(&wq->flusher_overflow);
3227 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3228 INIT_LIST_HEAD(&wq->list);
3230 if (alloc_cwqs(wq) < 0)
3233 for_each_cwq_cpu(cpu, wq) {
3234 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3236 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3237 cwq->pool = get_std_worker_pool(cpu, flags & WQ_HIGHPRI);
3239 cwq->flush_color = -1;
3240 cwq->max_active = max_active;
3241 INIT_LIST_HEAD(&cwq->delayed_works);
3244 if (flags & WQ_RESCUER) {
3245 struct worker *rescuer;
3247 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3250 wq->rescuer = rescuer = alloc_worker();
3254 rescuer->rescue_wq = wq;
3255 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3257 if (IS_ERR(rescuer->task))
3260 rescuer->task->flags |= PF_THREAD_BOUND;
3261 wake_up_process(rescuer->task);
3265 * workqueue_lock protects global freeze state and workqueues
3266 * list. Grab it, set max_active accordingly and add the new
3267 * workqueue to workqueues list.
3269 spin_lock(&workqueue_lock);
3271 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3272 for_each_cwq_cpu(cpu, wq)
3273 get_cwq(cpu, wq)->max_active = 0;
3275 list_add(&wq->list, &workqueues);
3277 spin_unlock(&workqueue_lock);
3283 free_mayday_mask(wq->mayday_mask);
3289 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3292 * destroy_workqueue - safely terminate a workqueue
3293 * @wq: target workqueue
3295 * Safely destroy a workqueue. All work currently pending will be done first.
3297 void destroy_workqueue(struct workqueue_struct *wq)
3301 /* drain it before proceeding with destruction */
3302 drain_workqueue(wq);
3305 * wq list is used to freeze wq, remove from list after
3306 * flushing is complete in case freeze races us.
3308 spin_lock(&workqueue_lock);
3309 list_del(&wq->list);
3310 spin_unlock(&workqueue_lock);
3313 for_each_cwq_cpu(cpu, wq) {
3314 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3317 for (i = 0; i < WORK_NR_COLORS; i++)
3318 BUG_ON(cwq->nr_in_flight[i]);
3319 BUG_ON(cwq->nr_active);
3320 BUG_ON(!list_empty(&cwq->delayed_works));
3323 if (wq->flags & WQ_RESCUER) {
3324 kthread_stop(wq->rescuer->task);
3325 free_mayday_mask(wq->mayday_mask);
3332 EXPORT_SYMBOL_GPL(destroy_workqueue);
3335 * cwq_set_max_active - adjust max_active of a cwq
3336 * @cwq: target cpu_workqueue_struct
3337 * @max_active: new max_active value.
3339 * Set @cwq->max_active to @max_active and activate delayed works if
3343 * spin_lock_irq(pool->lock).
3345 static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active)
3347 cwq->max_active = max_active;
3349 while (!list_empty(&cwq->delayed_works) &&
3350 cwq->nr_active < cwq->max_active)
3351 cwq_activate_first_delayed(cwq);
3355 * workqueue_set_max_active - adjust max_active of a workqueue
3356 * @wq: target workqueue
3357 * @max_active: new max_active value.
3359 * Set max_active of @wq to @max_active.
3362 * Don't call from IRQ context.
3364 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3368 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3370 spin_lock(&workqueue_lock);
3372 wq->saved_max_active = max_active;
3374 for_each_cwq_cpu(cpu, wq) {
3375 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3376 struct worker_pool *pool = cwq->pool;
3378 spin_lock_irq(&pool->lock);
3380 if (!(wq->flags & WQ_FREEZABLE) ||
3381 !(pool->flags & POOL_FREEZING))
3382 cwq_set_max_active(cwq, max_active);
3384 spin_unlock_irq(&pool->lock);
3387 spin_unlock(&workqueue_lock);
3389 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3392 * workqueue_congested - test whether a workqueue is congested
3393 * @cpu: CPU in question
3394 * @wq: target workqueue
3396 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3397 * no synchronization around this function and the test result is
3398 * unreliable and only useful as advisory hints or for debugging.
3401 * %true if congested, %false otherwise.
3403 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3405 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3407 return !list_empty(&cwq->delayed_works);
3409 EXPORT_SYMBOL_GPL(workqueue_congested);
3412 * work_busy - test whether a work is currently pending or running
3413 * @work: the work to be tested
3415 * Test whether @work is currently pending or running. There is no
3416 * synchronization around this function and the test result is
3417 * unreliable and only useful as advisory hints or for debugging.
3420 * OR'd bitmask of WORK_BUSY_* bits.
3422 unsigned int work_busy(struct work_struct *work)
3424 struct worker_pool *pool = get_work_pool(work);
3425 unsigned long flags;
3426 unsigned int ret = 0;
3428 if (work_pending(work))
3429 ret |= WORK_BUSY_PENDING;
3432 spin_lock_irqsave(&pool->lock, flags);
3433 if (find_worker_executing_work(pool, work))
3434 ret |= WORK_BUSY_RUNNING;
3435 spin_unlock_irqrestore(&pool->lock, flags);
3440 EXPORT_SYMBOL_GPL(work_busy);
3445 * There are two challenges in supporting CPU hotplug. Firstly, there
3446 * are a lot of assumptions on strong associations among work, cwq and
3447 * pool which make migrating pending and scheduled works very
3448 * difficult to implement without impacting hot paths. Secondly,
3449 * worker pools serve mix of short, long and very long running works making
3450 * blocked draining impractical.
3452 * This is solved by allowing the pools to be disassociated from the CPU
3453 * running as an unbound one and allowing it to be reattached later if the
3454 * cpu comes back online.
3457 static void wq_unbind_fn(struct work_struct *work)
3459 int cpu = smp_processor_id();
3460 struct worker_pool *pool;
3461 struct worker *worker;
3462 struct hlist_node *pos;
3465 for_each_std_worker_pool(pool, cpu) {
3466 BUG_ON(cpu != smp_processor_id());
3468 mutex_lock(&pool->assoc_mutex);
3469 spin_lock_irq(&pool->lock);
3472 * We've claimed all manager positions. Make all workers
3473 * unbound and set DISASSOCIATED. Before this, all workers
3474 * except for the ones which are still executing works from
3475 * before the last CPU down must be on the cpu. After
3476 * this, they may become diasporas.
3478 list_for_each_entry(worker, &pool->idle_list, entry)
3479 worker->flags |= WORKER_UNBOUND;
3481 for_each_busy_worker(worker, i, pos, pool)
3482 worker->flags |= WORKER_UNBOUND;
3484 pool->flags |= POOL_DISASSOCIATED;
3486 spin_unlock_irq(&pool->lock);
3487 mutex_unlock(&pool->assoc_mutex);
3491 * Call schedule() so that we cross rq->lock and thus can guarantee
3492 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3493 * as scheduler callbacks may be invoked from other cpus.
3498 * Sched callbacks are disabled now. Zap nr_running. After this,
3499 * nr_running stays zero and need_more_worker() and keep_working()
3500 * are always true as long as the worklist is not empty. Pools on
3501 * @cpu now behave as unbound (in terms of concurrency management)
3502 * pools which are served by workers tied to the CPU.
3504 * On return from this function, the current worker would trigger
3505 * unbound chain execution of pending work items if other workers
3508 for_each_std_worker_pool(pool, cpu)
3509 atomic_set(get_pool_nr_running(pool), 0);
3513 * Workqueues should be brought up before normal priority CPU notifiers.
3514 * This will be registered high priority CPU notifier.
3516 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3517 unsigned long action,
3520 unsigned int cpu = (unsigned long)hcpu;
3521 struct worker_pool *pool;
3523 switch (action & ~CPU_TASKS_FROZEN) {
3524 case CPU_UP_PREPARE:
3525 for_each_std_worker_pool(pool, cpu) {
3526 struct worker *worker;
3528 if (pool->nr_workers)
3531 worker = create_worker(pool);
3535 spin_lock_irq(&pool->lock);
3536 start_worker(worker);
3537 spin_unlock_irq(&pool->lock);
3541 case CPU_DOWN_FAILED:
3543 for_each_std_worker_pool(pool, cpu) {
3544 mutex_lock(&pool->assoc_mutex);
3545 spin_lock_irq(&pool->lock);
3547 pool->flags &= ~POOL_DISASSOCIATED;
3548 rebind_workers(pool);
3550 spin_unlock_irq(&pool->lock);
3551 mutex_unlock(&pool->assoc_mutex);
3559 * Workqueues should be brought down after normal priority CPU notifiers.
3560 * This will be registered as low priority CPU notifier.
3562 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3563 unsigned long action,
3566 unsigned int cpu = (unsigned long)hcpu;
3567 struct work_struct unbind_work;
3569 switch (action & ~CPU_TASKS_FROZEN) {
3570 case CPU_DOWN_PREPARE:
3571 /* unbinding should happen on the local CPU */
3572 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
3573 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3574 flush_work(&unbind_work);
3582 struct work_for_cpu {
3583 struct work_struct work;
3589 static void work_for_cpu_fn(struct work_struct *work)
3591 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3593 wfc->ret = wfc->fn(wfc->arg);
3597 * work_on_cpu - run a function in user context on a particular cpu
3598 * @cpu: the cpu to run on
3599 * @fn: the function to run
3600 * @arg: the function arg
3602 * This will return the value @fn returns.
3603 * It is up to the caller to ensure that the cpu doesn't go offline.
3604 * The caller must not hold any locks which would prevent @fn from completing.
3606 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3608 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3610 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3611 schedule_work_on(cpu, &wfc.work);
3612 flush_work(&wfc.work);
3615 EXPORT_SYMBOL_GPL(work_on_cpu);
3616 #endif /* CONFIG_SMP */
3618 #ifdef CONFIG_FREEZER
3621 * freeze_workqueues_begin - begin freezing workqueues
3623 * Start freezing workqueues. After this function returns, all freezable
3624 * workqueues will queue new works to their frozen_works list instead of
3628 * Grabs and releases workqueue_lock and pool->lock's.
3630 void freeze_workqueues_begin(void)
3634 spin_lock(&workqueue_lock);
3636 BUG_ON(workqueue_freezing);
3637 workqueue_freezing = true;
3639 for_each_wq_cpu(cpu) {
3640 struct worker_pool *pool;
3641 struct workqueue_struct *wq;
3643 for_each_std_worker_pool(pool, cpu) {
3644 spin_lock_irq(&pool->lock);
3646 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
3647 pool->flags |= POOL_FREEZING;
3649 list_for_each_entry(wq, &workqueues, list) {
3650 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3652 if (cwq && cwq->pool == pool &&
3653 (wq->flags & WQ_FREEZABLE))
3654 cwq->max_active = 0;
3657 spin_unlock_irq(&pool->lock);
3661 spin_unlock(&workqueue_lock);
3665 * freeze_workqueues_busy - are freezable workqueues still busy?
3667 * Check whether freezing is complete. This function must be called
3668 * between freeze_workqueues_begin() and thaw_workqueues().
3671 * Grabs and releases workqueue_lock.
3674 * %true if some freezable workqueues are still busy. %false if freezing
3677 bool freeze_workqueues_busy(void)
3682 spin_lock(&workqueue_lock);
3684 BUG_ON(!workqueue_freezing);
3686 for_each_wq_cpu(cpu) {
3687 struct workqueue_struct *wq;
3689 * nr_active is monotonically decreasing. It's safe
3690 * to peek without lock.
3692 list_for_each_entry(wq, &workqueues, list) {
3693 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3695 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3698 BUG_ON(cwq->nr_active < 0);
3699 if (cwq->nr_active) {
3706 spin_unlock(&workqueue_lock);
3711 * thaw_workqueues - thaw workqueues
3713 * Thaw workqueues. Normal queueing is restored and all collected
3714 * frozen works are transferred to their respective pool worklists.
3717 * Grabs and releases workqueue_lock and pool->lock's.
3719 void thaw_workqueues(void)
3723 spin_lock(&workqueue_lock);
3725 if (!workqueue_freezing)
3728 for_each_wq_cpu(cpu) {
3729 struct worker_pool *pool;
3730 struct workqueue_struct *wq;
3732 for_each_std_worker_pool(pool, cpu) {
3733 spin_lock_irq(&pool->lock);
3735 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
3736 pool->flags &= ~POOL_FREEZING;
3738 list_for_each_entry(wq, &workqueues, list) {
3739 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3741 if (!cwq || cwq->pool != pool ||
3742 !(wq->flags & WQ_FREEZABLE))
3745 /* restore max_active and repopulate worklist */
3746 cwq_set_max_active(cwq, wq->saved_max_active);
3749 wake_up_worker(pool);
3751 spin_unlock_irq(&pool->lock);
3755 workqueue_freezing = false;
3757 spin_unlock(&workqueue_lock);
3759 #endif /* CONFIG_FREEZER */
3761 static int __init init_workqueues(void)
3765 /* make sure we have enough bits for OFFQ pool ID */
3766 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
3767 WORK_CPU_END * NR_STD_WORKER_POOLS);
3769 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3770 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3772 /* initialize CPU pools */
3773 for_each_wq_cpu(cpu) {
3774 struct worker_pool *pool;
3776 for_each_std_worker_pool(pool, cpu) {
3777 spin_lock_init(&pool->lock);
3779 pool->flags |= POOL_DISASSOCIATED;
3780 INIT_LIST_HEAD(&pool->worklist);
3781 INIT_LIST_HEAD(&pool->idle_list);
3782 hash_init(pool->busy_hash);
3784 init_timer_deferrable(&pool->idle_timer);
3785 pool->idle_timer.function = idle_worker_timeout;
3786 pool->idle_timer.data = (unsigned long)pool;
3788 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3789 (unsigned long)pool);
3791 mutex_init(&pool->assoc_mutex);
3792 ida_init(&pool->worker_ida);
3795 BUG_ON(worker_pool_assign_id(pool));
3799 /* create the initial worker */
3800 for_each_online_wq_cpu(cpu) {
3801 struct worker_pool *pool;
3803 for_each_std_worker_pool(pool, cpu) {
3804 struct worker *worker;
3806 if (cpu != WORK_CPU_UNBOUND)
3807 pool->flags &= ~POOL_DISASSOCIATED;
3809 worker = create_worker(pool);
3811 spin_lock_irq(&pool->lock);
3812 start_worker(worker);
3813 spin_unlock_irq(&pool->lock);
3817 system_wq = alloc_workqueue("events", 0, 0);
3818 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3819 system_long_wq = alloc_workqueue("events_long", 0, 0);
3820 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3821 WQ_UNBOUND_MAX_ACTIVE);
3822 system_freezable_wq = alloc_workqueue("events_freezable",
3824 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3825 !system_unbound_wq || !system_freezable_wq);
3828 early_initcall(init_workqueues);