2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/hashtable.h>
46 #include "workqueue_internal.h"
52 * A bound pool is either associated or disassociated with its CPU.
53 * While associated (!DISASSOCIATED), all workers are bound to the
54 * CPU and none has %WORKER_UNBOUND set and concurrency management
57 * While DISASSOCIATED, the cpu may be offline and all workers have
58 * %WORKER_UNBOUND set and concurrency management disabled, and may
59 * be executing on any CPU. The pool behaves as an unbound one.
61 * Note that DISASSOCIATED can be flipped only while holding
62 * assoc_mutex to avoid changing binding state while
63 * create_worker() is in progress.
65 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
66 POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */
67 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
68 POOL_FREEZING = 1 << 3, /* freeze in progress */
71 WORKER_STARTED = 1 << 0, /* started */
72 WORKER_DIE = 1 << 1, /* die die die */
73 WORKER_IDLE = 1 << 2, /* is idle */
74 WORKER_PREP = 1 << 3, /* preparing to run works */
75 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
76 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
78 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND |
81 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
83 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
85 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
86 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
88 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
89 /* call for help after 10ms
91 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
92 CREATE_COOLDOWN = HZ, /* time to breath after fail */
95 * Rescue workers are used only on emergencies and shared by
98 RESCUER_NICE_LEVEL = -20,
99 HIGHPRI_NICE_LEVEL = -20,
103 * Structure fields follow one of the following exclusion rules.
105 * I: Modifiable by initialization/destruction paths and read-only for
108 * P: Preemption protected. Disabling preemption is enough and should
109 * only be modified and accessed from the local cpu.
111 * L: pool->lock protected. Access with pool->lock held.
113 * X: During normal operation, modification requires pool->lock and should
114 * be done only from local cpu. Either disabling preemption on local
115 * cpu or grabbing pool->lock is enough for read access. If
116 * POOL_DISASSOCIATED is set, it's identical to L.
118 * F: wq->flush_mutex protected.
120 * W: workqueue_lock protected.
123 /* struct worker is defined in workqueue_internal.h */
126 spinlock_t lock; /* the pool lock */
127 unsigned int cpu; /* I: the associated cpu */
128 int id; /* I: pool ID */
129 unsigned int flags; /* X: flags */
131 struct list_head worklist; /* L: list of pending works */
132 int nr_workers; /* L: total number of workers */
134 /* nr_idle includes the ones off idle_list for rebinding */
135 int nr_idle; /* L: currently idle ones */
137 struct list_head idle_list; /* X: list of idle workers */
138 struct timer_list idle_timer; /* L: worker idle timeout */
139 struct timer_list mayday_timer; /* L: SOS timer for workers */
141 /* workers are chained either in busy_hash or idle_list */
142 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
143 /* L: hash of busy workers */
145 struct mutex assoc_mutex; /* protect POOL_DISASSOCIATED */
146 struct ida worker_ida; /* L: for worker IDs */
149 * The current concurrency level. As it's likely to be accessed
150 * from other CPUs during try_to_wake_up(), put it in a separate
153 atomic_t nr_running ____cacheline_aligned_in_smp;
154 } ____cacheline_aligned_in_smp;
157 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
158 * work_struct->data are used for flags and thus cwqs need to be
159 * aligned at two's power of the number of flag bits.
161 struct cpu_workqueue_struct {
162 struct worker_pool *pool; /* I: the associated pool */
163 struct workqueue_struct *wq; /* I: the owning workqueue */
164 int work_color; /* L: current color */
165 int flush_color; /* L: flushing color */
166 int nr_in_flight[WORK_NR_COLORS];
167 /* L: nr of in_flight works */
168 int nr_active; /* L: nr of active works */
169 int max_active; /* L: max active works */
170 struct list_head delayed_works; /* L: delayed works */
174 * Structure used to wait for workqueue flush.
177 struct list_head list; /* F: list of flushers */
178 int flush_color; /* F: flush color waiting for */
179 struct completion done; /* flush completion */
183 * All cpumasks are assumed to be always set on UP and thus can't be
184 * used to determine whether there's something to be done.
187 typedef cpumask_var_t mayday_mask_t;
188 #define mayday_test_and_set_cpu(cpu, mask) \
189 cpumask_test_and_set_cpu((cpu), (mask))
190 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
191 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
192 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
193 #define free_mayday_mask(mask) free_cpumask_var((mask))
195 typedef unsigned long mayday_mask_t;
196 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
197 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
198 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
199 #define alloc_mayday_mask(maskp, gfp) true
200 #define free_mayday_mask(mask) do { } while (0)
204 * The externally visible workqueue abstraction is an array of
205 * per-CPU workqueues:
207 struct workqueue_struct {
208 unsigned int flags; /* W: WQ_* flags */
210 struct cpu_workqueue_struct __percpu *pcpu;
211 struct cpu_workqueue_struct *single;
213 } cpu_wq; /* I: cwq's */
214 struct list_head list; /* W: list of all workqueues */
216 struct mutex flush_mutex; /* protects wq flushing */
217 int work_color; /* F: current work color */
218 int flush_color; /* F: current flush color */
219 atomic_t nr_cwqs_to_flush; /* flush in progress */
220 struct wq_flusher *first_flusher; /* F: first flusher */
221 struct list_head flusher_queue; /* F: flush waiters */
222 struct list_head flusher_overflow; /* F: flush overflow list */
224 mayday_mask_t mayday_mask; /* cpus requesting rescue */
225 struct worker *rescuer; /* I: rescue worker */
227 int nr_drainers; /* W: drain in progress */
228 int saved_max_active; /* W: saved cwq max_active */
229 #ifdef CONFIG_LOCKDEP
230 struct lockdep_map lockdep_map;
232 char name[]; /* I: workqueue name */
235 struct workqueue_struct *system_wq __read_mostly;
236 EXPORT_SYMBOL_GPL(system_wq);
237 struct workqueue_struct *system_highpri_wq __read_mostly;
238 EXPORT_SYMBOL_GPL(system_highpri_wq);
239 struct workqueue_struct *system_long_wq __read_mostly;
240 EXPORT_SYMBOL_GPL(system_long_wq);
241 struct workqueue_struct *system_unbound_wq __read_mostly;
242 EXPORT_SYMBOL_GPL(system_unbound_wq);
243 struct workqueue_struct *system_freezable_wq __read_mostly;
244 EXPORT_SYMBOL_GPL(system_freezable_wq);
246 #define CREATE_TRACE_POINTS
247 #include <trace/events/workqueue.h>
249 #define for_each_std_worker_pool(pool, cpu) \
250 for ((pool) = &std_worker_pools(cpu)[0]; \
251 (pool) < &std_worker_pools(cpu)[NR_STD_WORKER_POOLS]; (pool)++)
253 #define for_each_busy_worker(worker, i, pos, pool) \
254 hash_for_each(pool->busy_hash, i, pos, worker, hentry)
256 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
259 if (cpu < nr_cpu_ids) {
261 cpu = cpumask_next(cpu, mask);
262 if (cpu < nr_cpu_ids)
266 return WORK_CPU_UNBOUND;
271 static inline int __next_cwq_cpu(int cpu, const struct cpumask *mask,
272 struct workqueue_struct *wq)
274 return __next_wq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
280 * An extra cpu number is defined using an invalid cpu number
281 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
282 * specific CPU. The following iterators are similar to for_each_*_cpu()
283 * iterators but also considers the unbound CPU.
285 * for_each_wq_cpu() : possible CPUs + WORK_CPU_UNBOUND
286 * for_each_online_wq_cpu() : online CPUs + WORK_CPU_UNBOUND
287 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
288 * WORK_CPU_UNBOUND for unbound workqueues
290 #define for_each_wq_cpu(cpu) \
291 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, 3); \
292 (cpu) < WORK_CPU_END; \
293 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, 3))
295 #define for_each_online_wq_cpu(cpu) \
296 for ((cpu) = __next_wq_cpu(-1, cpu_online_mask, 3); \
297 (cpu) < WORK_CPU_END; \
298 (cpu) = __next_wq_cpu((cpu), cpu_online_mask, 3))
300 #define for_each_cwq_cpu(cpu, wq) \
301 for ((cpu) = __next_cwq_cpu(-1, cpu_possible_mask, (wq)); \
302 (cpu) < WORK_CPU_END; \
303 (cpu) = __next_cwq_cpu((cpu), cpu_possible_mask, (wq)))
305 #ifdef CONFIG_DEBUG_OBJECTS_WORK
307 static struct debug_obj_descr work_debug_descr;
309 static void *work_debug_hint(void *addr)
311 return ((struct work_struct *) addr)->func;
315 * fixup_init is called when:
316 * - an active object is initialized
318 static int work_fixup_init(void *addr, enum debug_obj_state state)
320 struct work_struct *work = addr;
323 case ODEBUG_STATE_ACTIVE:
324 cancel_work_sync(work);
325 debug_object_init(work, &work_debug_descr);
333 * fixup_activate is called when:
334 * - an active object is activated
335 * - an unknown object is activated (might be a statically initialized object)
337 static int work_fixup_activate(void *addr, enum debug_obj_state state)
339 struct work_struct *work = addr;
343 case ODEBUG_STATE_NOTAVAILABLE:
345 * This is not really a fixup. The work struct was
346 * statically initialized. We just make sure that it
347 * is tracked in the object tracker.
349 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
350 debug_object_init(work, &work_debug_descr);
351 debug_object_activate(work, &work_debug_descr);
357 case ODEBUG_STATE_ACTIVE:
366 * fixup_free is called when:
367 * - an active object is freed
369 static int work_fixup_free(void *addr, enum debug_obj_state state)
371 struct work_struct *work = addr;
374 case ODEBUG_STATE_ACTIVE:
375 cancel_work_sync(work);
376 debug_object_free(work, &work_debug_descr);
383 static struct debug_obj_descr work_debug_descr = {
384 .name = "work_struct",
385 .debug_hint = work_debug_hint,
386 .fixup_init = work_fixup_init,
387 .fixup_activate = work_fixup_activate,
388 .fixup_free = work_fixup_free,
391 static inline void debug_work_activate(struct work_struct *work)
393 debug_object_activate(work, &work_debug_descr);
396 static inline void debug_work_deactivate(struct work_struct *work)
398 debug_object_deactivate(work, &work_debug_descr);
401 void __init_work(struct work_struct *work, int onstack)
404 debug_object_init_on_stack(work, &work_debug_descr);
406 debug_object_init(work, &work_debug_descr);
408 EXPORT_SYMBOL_GPL(__init_work);
410 void destroy_work_on_stack(struct work_struct *work)
412 debug_object_free(work, &work_debug_descr);
414 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
417 static inline void debug_work_activate(struct work_struct *work) { }
418 static inline void debug_work_deactivate(struct work_struct *work) { }
421 /* Serializes the accesses to the list of workqueues. */
422 static DEFINE_SPINLOCK(workqueue_lock);
423 static LIST_HEAD(workqueues);
424 static bool workqueue_freezing; /* W: have wqs started freezing? */
427 * The CPU and unbound standard worker pools. The unbound ones have
428 * POOL_DISASSOCIATED set, and their workers have WORKER_UNBOUND set.
430 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
431 cpu_std_worker_pools);
432 static struct worker_pool unbound_std_worker_pools[NR_STD_WORKER_POOLS];
434 /* idr of all pools */
435 static DEFINE_MUTEX(worker_pool_idr_mutex);
436 static DEFINE_IDR(worker_pool_idr);
438 static int worker_thread(void *__worker);
440 static struct worker_pool *std_worker_pools(int cpu)
442 if (cpu != WORK_CPU_UNBOUND)
443 return per_cpu(cpu_std_worker_pools, cpu);
445 return unbound_std_worker_pools;
448 static int std_worker_pool_pri(struct worker_pool *pool)
450 return pool - std_worker_pools(pool->cpu);
453 /* allocate ID and assign it to @pool */
454 static int worker_pool_assign_id(struct worker_pool *pool)
458 mutex_lock(&worker_pool_idr_mutex);
459 idr_pre_get(&worker_pool_idr, GFP_KERNEL);
460 ret = idr_get_new(&worker_pool_idr, pool, &pool->id);
461 mutex_unlock(&worker_pool_idr_mutex);
467 * Lookup worker_pool by id. The idr currently is built during boot and
468 * never modified. Don't worry about locking for now.
470 static struct worker_pool *worker_pool_by_id(int pool_id)
472 return idr_find(&worker_pool_idr, pool_id);
475 static struct worker_pool *get_std_worker_pool(int cpu, bool highpri)
477 struct worker_pool *pools = std_worker_pools(cpu);
479 return &pools[highpri];
482 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
483 struct workqueue_struct *wq)
485 if (!(wq->flags & WQ_UNBOUND)) {
486 if (likely(cpu < nr_cpu_ids))
487 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
488 } else if (likely(cpu == WORK_CPU_UNBOUND))
489 return wq->cpu_wq.single;
493 static unsigned int work_color_to_flags(int color)
495 return color << WORK_STRUCT_COLOR_SHIFT;
498 static int get_work_color(struct work_struct *work)
500 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
501 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
504 static int work_next_color(int color)
506 return (color + 1) % WORK_NR_COLORS;
510 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
511 * contain the pointer to the queued cwq. Once execution starts, the flag
512 * is cleared and the high bits contain OFFQ flags and pool ID.
514 * set_work_cwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
515 * and clear_work_data() can be used to set the cwq, pool or clear
516 * work->data. These functions should only be called while the work is
517 * owned - ie. while the PENDING bit is set.
519 * get_work_pool() and get_work_cwq() can be used to obtain the pool or cwq
520 * corresponding to a work. Pool is available once the work has been
521 * queued anywhere after initialization until it is sync canceled. cwq is
522 * available only while the work item is queued.
524 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
525 * canceled. While being canceled, a work item may have its PENDING set
526 * but stay off timer and worklist for arbitrarily long and nobody should
527 * try to steal the PENDING bit.
529 static inline void set_work_data(struct work_struct *work, unsigned long data,
532 BUG_ON(!work_pending(work));
533 atomic_long_set(&work->data, data | flags | work_static(work));
536 static void set_work_cwq(struct work_struct *work,
537 struct cpu_workqueue_struct *cwq,
538 unsigned long extra_flags)
540 set_work_data(work, (unsigned long)cwq,
541 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
544 static void set_work_pool_and_keep_pending(struct work_struct *work,
547 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
548 WORK_STRUCT_PENDING);
551 static void set_work_pool_and_clear_pending(struct work_struct *work,
555 * The following wmb is paired with the implied mb in
556 * test_and_set_bit(PENDING) and ensures all updates to @work made
557 * here are visible to and precede any updates by the next PENDING
561 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
564 static void clear_work_data(struct work_struct *work)
566 smp_wmb(); /* see set_work_pool_and_clear_pending() */
567 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
570 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
572 unsigned long data = atomic_long_read(&work->data);
574 if (data & WORK_STRUCT_CWQ)
575 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
581 * get_work_pool - return the worker_pool a given work was associated with
582 * @work: the work item of interest
584 * Return the worker_pool @work was last associated with. %NULL if none.
586 static struct worker_pool *get_work_pool(struct work_struct *work)
588 unsigned long data = atomic_long_read(&work->data);
589 struct worker_pool *pool;
592 if (data & WORK_STRUCT_CWQ)
593 return ((struct cpu_workqueue_struct *)
594 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
596 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
597 if (pool_id == WORK_OFFQ_POOL_NONE)
600 pool = worker_pool_by_id(pool_id);
606 * get_work_pool_id - return the worker pool ID a given work is associated with
607 * @work: the work item of interest
609 * Return the worker_pool ID @work was last associated with.
610 * %WORK_OFFQ_POOL_NONE if none.
612 static int get_work_pool_id(struct work_struct *work)
614 unsigned long data = atomic_long_read(&work->data);
616 if (data & WORK_STRUCT_CWQ)
617 return ((struct cpu_workqueue_struct *)
618 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
620 return data >> WORK_OFFQ_POOL_SHIFT;
623 static void mark_work_canceling(struct work_struct *work)
625 unsigned long pool_id = get_work_pool_id(work);
627 pool_id <<= WORK_OFFQ_POOL_SHIFT;
628 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
631 static bool work_is_canceling(struct work_struct *work)
633 unsigned long data = atomic_long_read(&work->data);
635 return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING);
639 * Policy functions. These define the policies on how the global worker
640 * pools are managed. Unless noted otherwise, these functions assume that
641 * they're being called with pool->lock held.
644 static bool __need_more_worker(struct worker_pool *pool)
646 return !atomic_read(&pool->nr_running);
650 * Need to wake up a worker? Called from anything but currently
653 * Note that, because unbound workers never contribute to nr_running, this
654 * function will always return %true for unbound pools as long as the
655 * worklist isn't empty.
657 static bool need_more_worker(struct worker_pool *pool)
659 return !list_empty(&pool->worklist) && __need_more_worker(pool);
662 /* Can I start working? Called from busy but !running workers. */
663 static bool may_start_working(struct worker_pool *pool)
665 return pool->nr_idle;
668 /* Do I need to keep working? Called from currently running workers. */
669 static bool keep_working(struct worker_pool *pool)
671 return !list_empty(&pool->worklist) &&
672 atomic_read(&pool->nr_running) <= 1;
675 /* Do we need a new worker? Called from manager. */
676 static bool need_to_create_worker(struct worker_pool *pool)
678 return need_more_worker(pool) && !may_start_working(pool);
681 /* Do I need to be the manager? */
682 static bool need_to_manage_workers(struct worker_pool *pool)
684 return need_to_create_worker(pool) ||
685 (pool->flags & POOL_MANAGE_WORKERS);
688 /* Do we have too many workers and should some go away? */
689 static bool too_many_workers(struct worker_pool *pool)
691 bool managing = pool->flags & POOL_MANAGING_WORKERS;
692 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
693 int nr_busy = pool->nr_workers - nr_idle;
696 * nr_idle and idle_list may disagree if idle rebinding is in
697 * progress. Never return %true if idle_list is empty.
699 if (list_empty(&pool->idle_list))
702 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
709 /* Return the first worker. Safe with preemption disabled */
710 static struct worker *first_worker(struct worker_pool *pool)
712 if (unlikely(list_empty(&pool->idle_list)))
715 return list_first_entry(&pool->idle_list, struct worker, entry);
719 * wake_up_worker - wake up an idle worker
720 * @pool: worker pool to wake worker from
722 * Wake up the first idle worker of @pool.
725 * spin_lock_irq(pool->lock).
727 static void wake_up_worker(struct worker_pool *pool)
729 struct worker *worker = first_worker(pool);
732 wake_up_process(worker->task);
736 * wq_worker_waking_up - a worker is waking up
737 * @task: task waking up
738 * @cpu: CPU @task is waking up to
740 * This function is called during try_to_wake_up() when a worker is
744 * spin_lock_irq(rq->lock)
746 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
748 struct worker *worker = kthread_data(task);
750 if (!(worker->flags & WORKER_NOT_RUNNING)) {
751 WARN_ON_ONCE(worker->pool->cpu != cpu);
752 atomic_inc(&worker->pool->nr_running);
757 * wq_worker_sleeping - a worker is going to sleep
758 * @task: task going to sleep
759 * @cpu: CPU in question, must be the current CPU number
761 * This function is called during schedule() when a busy worker is
762 * going to sleep. Worker on the same cpu can be woken up by
763 * returning pointer to its task.
766 * spin_lock_irq(rq->lock)
769 * Worker task on @cpu to wake up, %NULL if none.
771 struct task_struct *wq_worker_sleeping(struct task_struct *task,
774 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
775 struct worker_pool *pool;
778 * Rescuers, which may not have all the fields set up like normal
779 * workers, also reach here, let's not access anything before
780 * checking NOT_RUNNING.
782 if (worker->flags & WORKER_NOT_RUNNING)
787 /* this can only happen on the local cpu */
788 BUG_ON(cpu != raw_smp_processor_id());
791 * The counterpart of the following dec_and_test, implied mb,
792 * worklist not empty test sequence is in insert_work().
793 * Please read comment there.
795 * NOT_RUNNING is clear. This means that we're bound to and
796 * running on the local cpu w/ rq lock held and preemption
797 * disabled, which in turn means that none else could be
798 * manipulating idle_list, so dereferencing idle_list without pool
801 if (atomic_dec_and_test(&pool->nr_running) &&
802 !list_empty(&pool->worklist))
803 to_wakeup = first_worker(pool);
804 return to_wakeup ? to_wakeup->task : NULL;
808 * worker_set_flags - set worker flags and adjust nr_running accordingly
810 * @flags: flags to set
811 * @wakeup: wakeup an idle worker if necessary
813 * Set @flags in @worker->flags and adjust nr_running accordingly. If
814 * nr_running becomes zero and @wakeup is %true, an idle worker is
818 * spin_lock_irq(pool->lock)
820 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
823 struct worker_pool *pool = worker->pool;
825 WARN_ON_ONCE(worker->task != current);
828 * If transitioning into NOT_RUNNING, adjust nr_running and
829 * wake up an idle worker as necessary if requested by
832 if ((flags & WORKER_NOT_RUNNING) &&
833 !(worker->flags & WORKER_NOT_RUNNING)) {
835 if (atomic_dec_and_test(&pool->nr_running) &&
836 !list_empty(&pool->worklist))
837 wake_up_worker(pool);
839 atomic_dec(&pool->nr_running);
842 worker->flags |= flags;
846 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
848 * @flags: flags to clear
850 * Clear @flags in @worker->flags and adjust nr_running accordingly.
853 * spin_lock_irq(pool->lock)
855 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
857 struct worker_pool *pool = worker->pool;
858 unsigned int oflags = worker->flags;
860 WARN_ON_ONCE(worker->task != current);
862 worker->flags &= ~flags;
865 * If transitioning out of NOT_RUNNING, increment nr_running. Note
866 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
867 * of multiple flags, not a single flag.
869 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
870 if (!(worker->flags & WORKER_NOT_RUNNING))
871 atomic_inc(&pool->nr_running);
875 * find_worker_executing_work - find worker which is executing a work
876 * @pool: pool of interest
877 * @work: work to find worker for
879 * Find a worker which is executing @work on @pool by searching
880 * @pool->busy_hash which is keyed by the address of @work. For a worker
881 * to match, its current execution should match the address of @work and
882 * its work function. This is to avoid unwanted dependency between
883 * unrelated work executions through a work item being recycled while still
886 * This is a bit tricky. A work item may be freed once its execution
887 * starts and nothing prevents the freed area from being recycled for
888 * another work item. If the same work item address ends up being reused
889 * before the original execution finishes, workqueue will identify the
890 * recycled work item as currently executing and make it wait until the
891 * current execution finishes, introducing an unwanted dependency.
893 * This function checks the work item address, work function and workqueue
894 * to avoid false positives. Note that this isn't complete as one may
895 * construct a work function which can introduce dependency onto itself
896 * through a recycled work item. Well, if somebody wants to shoot oneself
897 * in the foot that badly, there's only so much we can do, and if such
898 * deadlock actually occurs, it should be easy to locate the culprit work
902 * spin_lock_irq(pool->lock).
905 * Pointer to worker which is executing @work if found, NULL
908 static struct worker *find_worker_executing_work(struct worker_pool *pool,
909 struct work_struct *work)
911 struct worker *worker;
912 struct hlist_node *tmp;
914 hash_for_each_possible(pool->busy_hash, worker, tmp, hentry,
916 if (worker->current_work == work &&
917 worker->current_func == work->func)
924 * move_linked_works - move linked works to a list
925 * @work: start of series of works to be scheduled
926 * @head: target list to append @work to
927 * @nextp: out paramter for nested worklist walking
929 * Schedule linked works starting from @work to @head. Work series to
930 * be scheduled starts at @work and includes any consecutive work with
931 * WORK_STRUCT_LINKED set in its predecessor.
933 * If @nextp is not NULL, it's updated to point to the next work of
934 * the last scheduled work. This allows move_linked_works() to be
935 * nested inside outer list_for_each_entry_safe().
938 * spin_lock_irq(pool->lock).
940 static void move_linked_works(struct work_struct *work, struct list_head *head,
941 struct work_struct **nextp)
943 struct work_struct *n;
946 * Linked worklist will always end before the end of the list,
947 * use NULL for list head.
949 list_for_each_entry_safe_from(work, n, NULL, entry) {
950 list_move_tail(&work->entry, head);
951 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
956 * If we're already inside safe list traversal and have moved
957 * multiple works to the scheduled queue, the next position
958 * needs to be updated.
964 static void cwq_activate_delayed_work(struct work_struct *work)
966 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
968 trace_workqueue_activate_work(work);
969 move_linked_works(work, &cwq->pool->worklist, NULL);
970 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
974 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
976 struct work_struct *work = list_first_entry(&cwq->delayed_works,
977 struct work_struct, entry);
979 cwq_activate_delayed_work(work);
983 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
984 * @cwq: cwq of interest
985 * @color: color of work which left the queue
987 * A work either has completed or is removed from pending queue,
988 * decrement nr_in_flight of its cwq and handle workqueue flushing.
991 * spin_lock_irq(pool->lock).
993 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
995 /* ignore uncolored works */
996 if (color == WORK_NO_COLOR)
999 cwq->nr_in_flight[color]--;
1002 if (!list_empty(&cwq->delayed_works)) {
1003 /* one down, submit a delayed one */
1004 if (cwq->nr_active < cwq->max_active)
1005 cwq_activate_first_delayed(cwq);
1008 /* is flush in progress and are we at the flushing tip? */
1009 if (likely(cwq->flush_color != color))
1012 /* are there still in-flight works? */
1013 if (cwq->nr_in_flight[color])
1016 /* this cwq is done, clear flush_color */
1017 cwq->flush_color = -1;
1020 * If this was the last cwq, wake up the first flusher. It
1021 * will handle the rest.
1023 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1024 complete(&cwq->wq->first_flusher->done);
1028 * try_to_grab_pending - steal work item from worklist and disable irq
1029 * @work: work item to steal
1030 * @is_dwork: @work is a delayed_work
1031 * @flags: place to store irq state
1033 * Try to grab PENDING bit of @work. This function can handle @work in any
1034 * stable state - idle, on timer or on worklist. Return values are
1036 * 1 if @work was pending and we successfully stole PENDING
1037 * 0 if @work was idle and we claimed PENDING
1038 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1039 * -ENOENT if someone else is canceling @work, this state may persist
1040 * for arbitrarily long
1042 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1043 * interrupted while holding PENDING and @work off queue, irq must be
1044 * disabled on entry. This, combined with delayed_work->timer being
1045 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1047 * On successful return, >= 0, irq is disabled and the caller is
1048 * responsible for releasing it using local_irq_restore(*@flags).
1050 * This function is safe to call from any context including IRQ handler.
1052 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1053 unsigned long *flags)
1055 struct worker_pool *pool;
1056 struct cpu_workqueue_struct *cwq;
1058 local_irq_save(*flags);
1060 /* try to steal the timer if it exists */
1062 struct delayed_work *dwork = to_delayed_work(work);
1065 * dwork->timer is irqsafe. If del_timer() fails, it's
1066 * guaranteed that the timer is not queued anywhere and not
1067 * running on the local CPU.
1069 if (likely(del_timer(&dwork->timer)))
1073 /* try to claim PENDING the normal way */
1074 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1078 * The queueing is in progress, or it is already queued. Try to
1079 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1081 pool = get_work_pool(work);
1085 spin_lock(&pool->lock);
1087 * work->data is guaranteed to point to cwq only while the work
1088 * item is queued on cwq->wq, and both updating work->data to point
1089 * to cwq on queueing and to pool on dequeueing are done under
1090 * cwq->pool->lock. This in turn guarantees that, if work->data
1091 * points to cwq which is associated with a locked pool, the work
1092 * item is currently queued on that pool.
1094 cwq = get_work_cwq(work);
1095 if (cwq && cwq->pool == pool) {
1096 debug_work_deactivate(work);
1099 * A delayed work item cannot be grabbed directly because
1100 * it might have linked NO_COLOR work items which, if left
1101 * on the delayed_list, will confuse cwq->nr_active
1102 * management later on and cause stall. Make sure the work
1103 * item is activated before grabbing.
1105 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1106 cwq_activate_delayed_work(work);
1108 list_del_init(&work->entry);
1109 cwq_dec_nr_in_flight(get_work_cwq(work), get_work_color(work));
1111 /* work->data points to cwq iff queued, point to pool */
1112 set_work_pool_and_keep_pending(work, pool->id);
1114 spin_unlock(&pool->lock);
1117 spin_unlock(&pool->lock);
1119 local_irq_restore(*flags);
1120 if (work_is_canceling(work))
1127 * insert_work - insert a work into a pool
1128 * @cwq: cwq @work belongs to
1129 * @work: work to insert
1130 * @head: insertion point
1131 * @extra_flags: extra WORK_STRUCT_* flags to set
1133 * Insert @work which belongs to @cwq after @head. @extra_flags is or'd to
1134 * work_struct flags.
1137 * spin_lock_irq(pool->lock).
1139 static void insert_work(struct cpu_workqueue_struct *cwq,
1140 struct work_struct *work, struct list_head *head,
1141 unsigned int extra_flags)
1143 struct worker_pool *pool = cwq->pool;
1145 /* we own @work, set data and link */
1146 set_work_cwq(work, cwq, extra_flags);
1147 list_add_tail(&work->entry, head);
1150 * Ensure either worker_sched_deactivated() sees the above
1151 * list_add_tail() or we see zero nr_running to avoid workers
1152 * lying around lazily while there are works to be processed.
1156 if (__need_more_worker(pool))
1157 wake_up_worker(pool);
1161 * Test whether @work is being queued from another work executing on the
1164 static bool is_chained_work(struct workqueue_struct *wq)
1166 struct worker *worker;
1168 worker = current_wq_worker();
1170 * Return %true iff I'm a worker execuing a work item on @wq. If
1171 * I'm @worker, it's safe to dereference it without locking.
1173 return worker && worker->current_cwq->wq == wq;
1176 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1177 struct work_struct *work)
1179 struct cpu_workqueue_struct *cwq;
1180 struct list_head *worklist;
1181 unsigned int work_flags;
1182 unsigned int req_cpu = cpu;
1185 * While a work item is PENDING && off queue, a task trying to
1186 * steal the PENDING will busy-loop waiting for it to either get
1187 * queued or lose PENDING. Grabbing PENDING and queueing should
1188 * happen with IRQ disabled.
1190 WARN_ON_ONCE(!irqs_disabled());
1192 debug_work_activate(work);
1194 /* if dying, only works from the same workqueue are allowed */
1195 if (unlikely(wq->flags & WQ_DRAINING) &&
1196 WARN_ON_ONCE(!is_chained_work(wq)))
1199 /* determine the cwq to use */
1200 if (!(wq->flags & WQ_UNBOUND)) {
1201 struct worker_pool *last_pool;
1203 if (cpu == WORK_CPU_UNBOUND)
1204 cpu = raw_smp_processor_id();
1207 * It's multi cpu. If @work was previously on a different
1208 * cpu, it might still be running there, in which case the
1209 * work needs to be queued on that cpu to guarantee
1212 cwq = get_cwq(cpu, wq);
1213 last_pool = get_work_pool(work);
1215 if (last_pool && last_pool != cwq->pool) {
1216 struct worker *worker;
1218 spin_lock(&last_pool->lock);
1220 worker = find_worker_executing_work(last_pool, work);
1222 if (worker && worker->current_cwq->wq == wq) {
1223 cwq = get_cwq(last_pool->cpu, wq);
1225 /* meh... not running there, queue here */
1226 spin_unlock(&last_pool->lock);
1227 spin_lock(&cwq->pool->lock);
1230 spin_lock(&cwq->pool->lock);
1233 cwq = get_cwq(WORK_CPU_UNBOUND, wq);
1234 spin_lock(&cwq->pool->lock);
1237 /* cwq determined, queue */
1238 trace_workqueue_queue_work(req_cpu, cwq, work);
1240 if (WARN_ON(!list_empty(&work->entry))) {
1241 spin_unlock(&cwq->pool->lock);
1245 cwq->nr_in_flight[cwq->work_color]++;
1246 work_flags = work_color_to_flags(cwq->work_color);
1248 if (likely(cwq->nr_active < cwq->max_active)) {
1249 trace_workqueue_activate_work(work);
1251 worklist = &cwq->pool->worklist;
1253 work_flags |= WORK_STRUCT_DELAYED;
1254 worklist = &cwq->delayed_works;
1257 insert_work(cwq, work, worklist, work_flags);
1259 spin_unlock(&cwq->pool->lock);
1263 * queue_work_on - queue work on specific cpu
1264 * @cpu: CPU number to execute work on
1265 * @wq: workqueue to use
1266 * @work: work to queue
1268 * Returns %false if @work was already on a queue, %true otherwise.
1270 * We queue the work to a specific CPU, the caller must ensure it
1273 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1274 struct work_struct *work)
1277 unsigned long flags;
1279 local_irq_save(flags);
1281 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1282 __queue_work(cpu, wq, work);
1286 local_irq_restore(flags);
1289 EXPORT_SYMBOL_GPL(queue_work_on);
1292 * queue_work - queue work on a workqueue
1293 * @wq: workqueue to use
1294 * @work: work to queue
1296 * Returns %false if @work was already on a queue, %true otherwise.
1298 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1299 * it can be processed by another CPU.
1301 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1303 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1305 EXPORT_SYMBOL_GPL(queue_work);
1307 void delayed_work_timer_fn(unsigned long __data)
1309 struct delayed_work *dwork = (struct delayed_work *)__data;
1311 /* should have been called from irqsafe timer with irq already off */
1312 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1314 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1316 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1317 struct delayed_work *dwork, unsigned long delay)
1319 struct timer_list *timer = &dwork->timer;
1320 struct work_struct *work = &dwork->work;
1322 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1323 timer->data != (unsigned long)dwork);
1324 WARN_ON_ONCE(timer_pending(timer));
1325 WARN_ON_ONCE(!list_empty(&work->entry));
1328 * If @delay is 0, queue @dwork->work immediately. This is for
1329 * both optimization and correctness. The earliest @timer can
1330 * expire is on the closest next tick and delayed_work users depend
1331 * on that there's no such delay when @delay is 0.
1334 __queue_work(cpu, wq, &dwork->work);
1338 timer_stats_timer_set_start_info(&dwork->timer);
1342 timer->expires = jiffies + delay;
1344 if (unlikely(cpu != WORK_CPU_UNBOUND))
1345 add_timer_on(timer, cpu);
1351 * queue_delayed_work_on - queue work on specific CPU after delay
1352 * @cpu: CPU number to execute work on
1353 * @wq: workqueue to use
1354 * @dwork: work to queue
1355 * @delay: number of jiffies to wait before queueing
1357 * Returns %false if @work was already on a queue, %true otherwise. If
1358 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1361 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1362 struct delayed_work *dwork, unsigned long delay)
1364 struct work_struct *work = &dwork->work;
1366 unsigned long flags;
1368 /* read the comment in __queue_work() */
1369 local_irq_save(flags);
1371 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1372 __queue_delayed_work(cpu, wq, dwork, delay);
1376 local_irq_restore(flags);
1379 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1382 * queue_delayed_work - queue work on a workqueue after delay
1383 * @wq: workqueue to use
1384 * @dwork: delayable work to queue
1385 * @delay: number of jiffies to wait before queueing
1387 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1389 bool queue_delayed_work(struct workqueue_struct *wq,
1390 struct delayed_work *dwork, unsigned long delay)
1392 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1394 EXPORT_SYMBOL_GPL(queue_delayed_work);
1397 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1398 * @cpu: CPU number to execute work on
1399 * @wq: workqueue to use
1400 * @dwork: work to queue
1401 * @delay: number of jiffies to wait before queueing
1403 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1404 * modify @dwork's timer so that it expires after @delay. If @delay is
1405 * zero, @work is guaranteed to be scheduled immediately regardless of its
1408 * Returns %false if @dwork was idle and queued, %true if @dwork was
1409 * pending and its timer was modified.
1411 * This function is safe to call from any context including IRQ handler.
1412 * See try_to_grab_pending() for details.
1414 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1415 struct delayed_work *dwork, unsigned long delay)
1417 unsigned long flags;
1421 ret = try_to_grab_pending(&dwork->work, true, &flags);
1422 } while (unlikely(ret == -EAGAIN));
1424 if (likely(ret >= 0)) {
1425 __queue_delayed_work(cpu, wq, dwork, delay);
1426 local_irq_restore(flags);
1429 /* -ENOENT from try_to_grab_pending() becomes %true */
1432 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1435 * mod_delayed_work - modify delay of or queue a delayed work
1436 * @wq: workqueue to use
1437 * @dwork: work to queue
1438 * @delay: number of jiffies to wait before queueing
1440 * mod_delayed_work_on() on local CPU.
1442 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1443 unsigned long delay)
1445 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1447 EXPORT_SYMBOL_GPL(mod_delayed_work);
1450 * worker_enter_idle - enter idle state
1451 * @worker: worker which is entering idle state
1453 * @worker is entering idle state. Update stats and idle timer if
1457 * spin_lock_irq(pool->lock).
1459 static void worker_enter_idle(struct worker *worker)
1461 struct worker_pool *pool = worker->pool;
1463 BUG_ON(worker->flags & WORKER_IDLE);
1464 BUG_ON(!list_empty(&worker->entry) &&
1465 (worker->hentry.next || worker->hentry.pprev));
1467 /* can't use worker_set_flags(), also called from start_worker() */
1468 worker->flags |= WORKER_IDLE;
1470 worker->last_active = jiffies;
1472 /* idle_list is LIFO */
1473 list_add(&worker->entry, &pool->idle_list);
1475 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1476 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1479 * Sanity check nr_running. Because wq_unbind_fn() releases
1480 * pool->lock between setting %WORKER_UNBOUND and zapping
1481 * nr_running, the warning may trigger spuriously. Check iff
1482 * unbind is not in progress.
1484 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1485 pool->nr_workers == pool->nr_idle &&
1486 atomic_read(&pool->nr_running));
1490 * worker_leave_idle - leave idle state
1491 * @worker: worker which is leaving idle state
1493 * @worker is leaving idle state. Update stats.
1496 * spin_lock_irq(pool->lock).
1498 static void worker_leave_idle(struct worker *worker)
1500 struct worker_pool *pool = worker->pool;
1502 BUG_ON(!(worker->flags & WORKER_IDLE));
1503 worker_clr_flags(worker, WORKER_IDLE);
1505 list_del_init(&worker->entry);
1509 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock pool
1512 * Works which are scheduled while the cpu is online must at least be
1513 * scheduled to a worker which is bound to the cpu so that if they are
1514 * flushed from cpu callbacks while cpu is going down, they are
1515 * guaranteed to execute on the cpu.
1517 * This function is to be used by rogue workers and rescuers to bind
1518 * themselves to the target cpu and may race with cpu going down or
1519 * coming online. kthread_bind() can't be used because it may put the
1520 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1521 * verbatim as it's best effort and blocking and pool may be
1522 * [dis]associated in the meantime.
1524 * This function tries set_cpus_allowed() and locks pool and verifies the
1525 * binding against %POOL_DISASSOCIATED which is set during
1526 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1527 * enters idle state or fetches works without dropping lock, it can
1528 * guarantee the scheduling requirement described in the first paragraph.
1531 * Might sleep. Called without any lock but returns with pool->lock
1535 * %true if the associated pool is online (@worker is successfully
1536 * bound), %false if offline.
1538 static bool worker_maybe_bind_and_lock(struct worker *worker)
1539 __acquires(&pool->lock)
1541 struct worker_pool *pool = worker->pool;
1542 struct task_struct *task = worker->task;
1546 * The following call may fail, succeed or succeed
1547 * without actually migrating the task to the cpu if
1548 * it races with cpu hotunplug operation. Verify
1549 * against POOL_DISASSOCIATED.
1551 if (!(pool->flags & POOL_DISASSOCIATED))
1552 set_cpus_allowed_ptr(task, get_cpu_mask(pool->cpu));
1554 spin_lock_irq(&pool->lock);
1555 if (pool->flags & POOL_DISASSOCIATED)
1557 if (task_cpu(task) == pool->cpu &&
1558 cpumask_equal(¤t->cpus_allowed,
1559 get_cpu_mask(pool->cpu)))
1561 spin_unlock_irq(&pool->lock);
1564 * We've raced with CPU hot[un]plug. Give it a breather
1565 * and retry migration. cond_resched() is required here;
1566 * otherwise, we might deadlock against cpu_stop trying to
1567 * bring down the CPU on non-preemptive kernel.
1575 * Rebind an idle @worker to its CPU. worker_thread() will test
1576 * list_empty(@worker->entry) before leaving idle and call this function.
1578 static void idle_worker_rebind(struct worker *worker)
1580 /* CPU may go down again inbetween, clear UNBOUND only on success */
1581 if (worker_maybe_bind_and_lock(worker))
1582 worker_clr_flags(worker, WORKER_UNBOUND);
1584 /* rebind complete, become available again */
1585 list_add(&worker->entry, &worker->pool->idle_list);
1586 spin_unlock_irq(&worker->pool->lock);
1590 * Function for @worker->rebind.work used to rebind unbound busy workers to
1591 * the associated cpu which is coming back online. This is scheduled by
1592 * cpu up but can race with other cpu hotplug operations and may be
1593 * executed twice without intervening cpu down.
1595 static void busy_worker_rebind_fn(struct work_struct *work)
1597 struct worker *worker = container_of(work, struct worker, rebind_work);
1599 if (worker_maybe_bind_and_lock(worker))
1600 worker_clr_flags(worker, WORKER_UNBOUND);
1602 spin_unlock_irq(&worker->pool->lock);
1606 * rebind_workers - rebind all workers of a pool to the associated CPU
1607 * @pool: pool of interest
1609 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1610 * is different for idle and busy ones.
1612 * Idle ones will be removed from the idle_list and woken up. They will
1613 * add themselves back after completing rebind. This ensures that the
1614 * idle_list doesn't contain any unbound workers when re-bound busy workers
1615 * try to perform local wake-ups for concurrency management.
1617 * Busy workers can rebind after they finish their current work items.
1618 * Queueing the rebind work item at the head of the scheduled list is
1619 * enough. Note that nr_running will be properly bumped as busy workers
1622 * On return, all non-manager workers are scheduled for rebind - see
1623 * manage_workers() for the manager special case. Any idle worker
1624 * including the manager will not appear on @idle_list until rebind is
1625 * complete, making local wake-ups safe.
1627 static void rebind_workers(struct worker_pool *pool)
1629 struct worker *worker, *n;
1630 struct hlist_node *pos;
1633 lockdep_assert_held(&pool->assoc_mutex);
1634 lockdep_assert_held(&pool->lock);
1636 /* dequeue and kick idle ones */
1637 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1639 * idle workers should be off @pool->idle_list until rebind
1640 * is complete to avoid receiving premature local wake-ups.
1642 list_del_init(&worker->entry);
1645 * worker_thread() will see the above dequeuing and call
1646 * idle_worker_rebind().
1648 wake_up_process(worker->task);
1651 /* rebind busy workers */
1652 for_each_busy_worker(worker, i, pos, pool) {
1653 struct work_struct *rebind_work = &worker->rebind_work;
1654 struct workqueue_struct *wq;
1656 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1657 work_data_bits(rebind_work)))
1660 debug_work_activate(rebind_work);
1663 * wq doesn't really matter but let's keep @worker->pool
1664 * and @cwq->pool consistent for sanity.
1666 if (std_worker_pool_pri(worker->pool))
1667 wq = system_highpri_wq;
1671 insert_work(get_cwq(pool->cpu, wq), rebind_work,
1672 worker->scheduled.next,
1673 work_color_to_flags(WORK_NO_COLOR));
1677 static struct worker *alloc_worker(void)
1679 struct worker *worker;
1681 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1683 INIT_LIST_HEAD(&worker->entry);
1684 INIT_LIST_HEAD(&worker->scheduled);
1685 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1686 /* on creation a worker is in !idle && prep state */
1687 worker->flags = WORKER_PREP;
1693 * create_worker - create a new workqueue worker
1694 * @pool: pool the new worker will belong to
1696 * Create a new worker which is bound to @pool. The returned worker
1697 * can be started by calling start_worker() or destroyed using
1701 * Might sleep. Does GFP_KERNEL allocations.
1704 * Pointer to the newly created worker.
1706 static struct worker *create_worker(struct worker_pool *pool)
1708 const char *pri = std_worker_pool_pri(pool) ? "H" : "";
1709 struct worker *worker = NULL;
1712 spin_lock_irq(&pool->lock);
1713 while (ida_get_new(&pool->worker_ida, &id)) {
1714 spin_unlock_irq(&pool->lock);
1715 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1717 spin_lock_irq(&pool->lock);
1719 spin_unlock_irq(&pool->lock);
1721 worker = alloc_worker();
1725 worker->pool = pool;
1728 if (pool->cpu != WORK_CPU_UNBOUND)
1729 worker->task = kthread_create_on_node(worker_thread,
1730 worker, cpu_to_node(pool->cpu),
1731 "kworker/%u:%d%s", pool->cpu, id, pri);
1733 worker->task = kthread_create(worker_thread, worker,
1734 "kworker/u:%d%s", id, pri);
1735 if (IS_ERR(worker->task))
1738 if (std_worker_pool_pri(pool))
1739 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1742 * Determine CPU binding of the new worker depending on
1743 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1744 * flag remains stable across this function. See the comments
1745 * above the flag definition for details.
1747 * As an unbound worker may later become a regular one if CPU comes
1748 * online, make sure every worker has %PF_THREAD_BOUND set.
1750 if (!(pool->flags & POOL_DISASSOCIATED)) {
1751 kthread_bind(worker->task, pool->cpu);
1753 worker->task->flags |= PF_THREAD_BOUND;
1754 worker->flags |= WORKER_UNBOUND;
1760 spin_lock_irq(&pool->lock);
1761 ida_remove(&pool->worker_ida, id);
1762 spin_unlock_irq(&pool->lock);
1769 * start_worker - start a newly created worker
1770 * @worker: worker to start
1772 * Make the pool aware of @worker and start it.
1775 * spin_lock_irq(pool->lock).
1777 static void start_worker(struct worker *worker)
1779 worker->flags |= WORKER_STARTED;
1780 worker->pool->nr_workers++;
1781 worker_enter_idle(worker);
1782 wake_up_process(worker->task);
1786 * destroy_worker - destroy a workqueue worker
1787 * @worker: worker to be destroyed
1789 * Destroy @worker and adjust @pool stats accordingly.
1792 * spin_lock_irq(pool->lock) which is released and regrabbed.
1794 static void destroy_worker(struct worker *worker)
1796 struct worker_pool *pool = worker->pool;
1797 int id = worker->id;
1799 /* sanity check frenzy */
1800 BUG_ON(worker->current_work);
1801 BUG_ON(!list_empty(&worker->scheduled));
1803 if (worker->flags & WORKER_STARTED)
1805 if (worker->flags & WORKER_IDLE)
1808 list_del_init(&worker->entry);
1809 worker->flags |= WORKER_DIE;
1811 spin_unlock_irq(&pool->lock);
1813 kthread_stop(worker->task);
1816 spin_lock_irq(&pool->lock);
1817 ida_remove(&pool->worker_ida, id);
1820 static void idle_worker_timeout(unsigned long __pool)
1822 struct worker_pool *pool = (void *)__pool;
1824 spin_lock_irq(&pool->lock);
1826 if (too_many_workers(pool)) {
1827 struct worker *worker;
1828 unsigned long expires;
1830 /* idle_list is kept in LIFO order, check the last one */
1831 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1832 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1834 if (time_before(jiffies, expires))
1835 mod_timer(&pool->idle_timer, expires);
1837 /* it's been idle for too long, wake up manager */
1838 pool->flags |= POOL_MANAGE_WORKERS;
1839 wake_up_worker(pool);
1843 spin_unlock_irq(&pool->lock);
1846 static bool send_mayday(struct work_struct *work)
1848 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1849 struct workqueue_struct *wq = cwq->wq;
1852 if (!(wq->flags & WQ_RESCUER))
1855 /* mayday mayday mayday */
1856 cpu = cwq->pool->cpu;
1857 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1858 if (cpu == WORK_CPU_UNBOUND)
1860 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1861 wake_up_process(wq->rescuer->task);
1865 static void pool_mayday_timeout(unsigned long __pool)
1867 struct worker_pool *pool = (void *)__pool;
1868 struct work_struct *work;
1870 spin_lock_irq(&pool->lock);
1872 if (need_to_create_worker(pool)) {
1874 * We've been trying to create a new worker but
1875 * haven't been successful. We might be hitting an
1876 * allocation deadlock. Send distress signals to
1879 list_for_each_entry(work, &pool->worklist, entry)
1883 spin_unlock_irq(&pool->lock);
1885 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1889 * maybe_create_worker - create a new worker if necessary
1890 * @pool: pool to create a new worker for
1892 * Create a new worker for @pool if necessary. @pool is guaranteed to
1893 * have at least one idle worker on return from this function. If
1894 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1895 * sent to all rescuers with works scheduled on @pool to resolve
1896 * possible allocation deadlock.
1898 * On return, need_to_create_worker() is guaranteed to be false and
1899 * may_start_working() true.
1902 * spin_lock_irq(pool->lock) which may be released and regrabbed
1903 * multiple times. Does GFP_KERNEL allocations. Called only from
1907 * false if no action was taken and pool->lock stayed locked, true
1910 static bool maybe_create_worker(struct worker_pool *pool)
1911 __releases(&pool->lock)
1912 __acquires(&pool->lock)
1914 if (!need_to_create_worker(pool))
1917 spin_unlock_irq(&pool->lock);
1919 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1920 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1923 struct worker *worker;
1925 worker = create_worker(pool);
1927 del_timer_sync(&pool->mayday_timer);
1928 spin_lock_irq(&pool->lock);
1929 start_worker(worker);
1930 BUG_ON(need_to_create_worker(pool));
1934 if (!need_to_create_worker(pool))
1937 __set_current_state(TASK_INTERRUPTIBLE);
1938 schedule_timeout(CREATE_COOLDOWN);
1940 if (!need_to_create_worker(pool))
1944 del_timer_sync(&pool->mayday_timer);
1945 spin_lock_irq(&pool->lock);
1946 if (need_to_create_worker(pool))
1952 * maybe_destroy_worker - destroy workers which have been idle for a while
1953 * @pool: pool to destroy workers for
1955 * Destroy @pool workers which have been idle for longer than
1956 * IDLE_WORKER_TIMEOUT.
1959 * spin_lock_irq(pool->lock) which may be released and regrabbed
1960 * multiple times. Called only from manager.
1963 * false if no action was taken and pool->lock stayed locked, true
1966 static bool maybe_destroy_workers(struct worker_pool *pool)
1970 while (too_many_workers(pool)) {
1971 struct worker *worker;
1972 unsigned long expires;
1974 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1975 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1977 if (time_before(jiffies, expires)) {
1978 mod_timer(&pool->idle_timer, expires);
1982 destroy_worker(worker);
1990 * manage_workers - manage worker pool
1993 * Assume the manager role and manage the worker pool @worker belongs
1994 * to. At any given time, there can be only zero or one manager per
1995 * pool. The exclusion is handled automatically by this function.
1997 * The caller can safely start processing works on false return. On
1998 * true return, it's guaranteed that need_to_create_worker() is false
1999 * and may_start_working() is true.
2002 * spin_lock_irq(pool->lock) which may be released and regrabbed
2003 * multiple times. Does GFP_KERNEL allocations.
2006 * spin_lock_irq(pool->lock) which may be released and regrabbed
2007 * multiple times. Does GFP_KERNEL allocations.
2009 static bool manage_workers(struct worker *worker)
2011 struct worker_pool *pool = worker->pool;
2014 if (pool->flags & POOL_MANAGING_WORKERS)
2017 pool->flags |= POOL_MANAGING_WORKERS;
2020 * To simplify both worker management and CPU hotplug, hold off
2021 * management while hotplug is in progress. CPU hotplug path can't
2022 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2023 * lead to idle worker depletion (all become busy thinking someone
2024 * else is managing) which in turn can result in deadlock under
2025 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2026 * manager against CPU hotplug.
2028 * assoc_mutex would always be free unless CPU hotplug is in
2029 * progress. trylock first without dropping @pool->lock.
2031 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2032 spin_unlock_irq(&pool->lock);
2033 mutex_lock(&pool->assoc_mutex);
2035 * CPU hotplug could have happened while we were waiting
2036 * for assoc_mutex. Hotplug itself can't handle us
2037 * because manager isn't either on idle or busy list, and
2038 * @pool's state and ours could have deviated.
2040 * As hotplug is now excluded via assoc_mutex, we can
2041 * simply try to bind. It will succeed or fail depending
2042 * on @pool's current state. Try it and adjust
2043 * %WORKER_UNBOUND accordingly.
2045 if (worker_maybe_bind_and_lock(worker))
2046 worker->flags &= ~WORKER_UNBOUND;
2048 worker->flags |= WORKER_UNBOUND;
2053 pool->flags &= ~POOL_MANAGE_WORKERS;
2056 * Destroy and then create so that may_start_working() is true
2059 ret |= maybe_destroy_workers(pool);
2060 ret |= maybe_create_worker(pool);
2062 pool->flags &= ~POOL_MANAGING_WORKERS;
2063 mutex_unlock(&pool->assoc_mutex);
2068 * process_one_work - process single work
2070 * @work: work to process
2072 * Process @work. This function contains all the logics necessary to
2073 * process a single work including synchronization against and
2074 * interaction with other workers on the same cpu, queueing and
2075 * flushing. As long as context requirement is met, any worker can
2076 * call this function to process a work.
2079 * spin_lock_irq(pool->lock) which is released and regrabbed.
2081 static void process_one_work(struct worker *worker, struct work_struct *work)
2082 __releases(&pool->lock)
2083 __acquires(&pool->lock)
2085 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2086 struct worker_pool *pool = worker->pool;
2087 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2089 struct worker *collision;
2090 #ifdef CONFIG_LOCKDEP
2092 * It is permissible to free the struct work_struct from
2093 * inside the function that is called from it, this we need to
2094 * take into account for lockdep too. To avoid bogus "held
2095 * lock freed" warnings as well as problems when looking into
2096 * work->lockdep_map, make a copy and use that here.
2098 struct lockdep_map lockdep_map;
2100 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2103 * Ensure we're on the correct CPU. DISASSOCIATED test is
2104 * necessary to avoid spurious warnings from rescuers servicing the
2105 * unbound or a disassociated pool.
2107 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2108 !(pool->flags & POOL_DISASSOCIATED) &&
2109 raw_smp_processor_id() != pool->cpu);
2112 * A single work shouldn't be executed concurrently by
2113 * multiple workers on a single cpu. Check whether anyone is
2114 * already processing the work. If so, defer the work to the
2115 * currently executing one.
2117 collision = find_worker_executing_work(pool, work);
2118 if (unlikely(collision)) {
2119 move_linked_works(work, &collision->scheduled, NULL);
2123 /* claim and dequeue */
2124 debug_work_deactivate(work);
2125 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2126 worker->current_work = work;
2127 worker->current_func = work->func;
2128 worker->current_cwq = cwq;
2129 work_color = get_work_color(work);
2131 list_del_init(&work->entry);
2134 * CPU intensive works don't participate in concurrency
2135 * management. They're the scheduler's responsibility.
2137 if (unlikely(cpu_intensive))
2138 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2141 * Unbound pool isn't concurrency managed and work items should be
2142 * executed ASAP. Wake up another worker if necessary.
2144 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2145 wake_up_worker(pool);
2148 * Record the last pool and clear PENDING which should be the last
2149 * update to @work. Also, do this inside @pool->lock so that
2150 * PENDING and queued state changes happen together while IRQ is
2153 set_work_pool_and_clear_pending(work, pool->id);
2155 spin_unlock_irq(&pool->lock);
2157 lock_map_acquire_read(&cwq->wq->lockdep_map);
2158 lock_map_acquire(&lockdep_map);
2159 trace_workqueue_execute_start(work);
2160 worker->current_func(work);
2162 * While we must be careful to not use "work" after this, the trace
2163 * point will only record its address.
2165 trace_workqueue_execute_end(work);
2166 lock_map_release(&lockdep_map);
2167 lock_map_release(&cwq->wq->lockdep_map);
2169 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2170 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2171 " last function: %pf\n",
2172 current->comm, preempt_count(), task_pid_nr(current),
2173 worker->current_func);
2174 debug_show_held_locks(current);
2178 spin_lock_irq(&pool->lock);
2180 /* clear cpu intensive status */
2181 if (unlikely(cpu_intensive))
2182 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2184 /* we're done with it, release */
2185 hash_del(&worker->hentry);
2186 worker->current_work = NULL;
2187 worker->current_func = NULL;
2188 worker->current_cwq = NULL;
2189 cwq_dec_nr_in_flight(cwq, work_color);
2193 * process_scheduled_works - process scheduled works
2196 * Process all scheduled works. Please note that the scheduled list
2197 * may change while processing a work, so this function repeatedly
2198 * fetches a work from the top and executes it.
2201 * spin_lock_irq(pool->lock) which may be released and regrabbed
2204 static void process_scheduled_works(struct worker *worker)
2206 while (!list_empty(&worker->scheduled)) {
2207 struct work_struct *work = list_first_entry(&worker->scheduled,
2208 struct work_struct, entry);
2209 process_one_work(worker, work);
2214 * worker_thread - the worker thread function
2217 * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools
2218 * of these per each cpu. These workers process all works regardless of
2219 * their specific target workqueue. The only exception is works which
2220 * belong to workqueues with a rescuer which will be explained in
2223 static int worker_thread(void *__worker)
2225 struct worker *worker = __worker;
2226 struct worker_pool *pool = worker->pool;
2228 /* tell the scheduler that this is a workqueue worker */
2229 worker->task->flags |= PF_WQ_WORKER;
2231 spin_lock_irq(&pool->lock);
2233 /* we are off idle list if destruction or rebind is requested */
2234 if (unlikely(list_empty(&worker->entry))) {
2235 spin_unlock_irq(&pool->lock);
2237 /* if DIE is set, destruction is requested */
2238 if (worker->flags & WORKER_DIE) {
2239 worker->task->flags &= ~PF_WQ_WORKER;
2243 /* otherwise, rebind */
2244 idle_worker_rebind(worker);
2248 worker_leave_idle(worker);
2250 /* no more worker necessary? */
2251 if (!need_more_worker(pool))
2254 /* do we need to manage? */
2255 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2259 * ->scheduled list can only be filled while a worker is
2260 * preparing to process a work or actually processing it.
2261 * Make sure nobody diddled with it while I was sleeping.
2263 BUG_ON(!list_empty(&worker->scheduled));
2266 * When control reaches this point, we're guaranteed to have
2267 * at least one idle worker or that someone else has already
2268 * assumed the manager role.
2270 worker_clr_flags(worker, WORKER_PREP);
2273 struct work_struct *work =
2274 list_first_entry(&pool->worklist,
2275 struct work_struct, entry);
2277 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2278 /* optimization path, not strictly necessary */
2279 process_one_work(worker, work);
2280 if (unlikely(!list_empty(&worker->scheduled)))
2281 process_scheduled_works(worker);
2283 move_linked_works(work, &worker->scheduled, NULL);
2284 process_scheduled_works(worker);
2286 } while (keep_working(pool));
2288 worker_set_flags(worker, WORKER_PREP, false);
2290 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2294 * pool->lock is held and there's no work to process and no need to
2295 * manage, sleep. Workers are woken up only while holding
2296 * pool->lock or from local cpu, so setting the current state
2297 * before releasing pool->lock is enough to prevent losing any
2300 worker_enter_idle(worker);
2301 __set_current_state(TASK_INTERRUPTIBLE);
2302 spin_unlock_irq(&pool->lock);
2308 * rescuer_thread - the rescuer thread function
2311 * Workqueue rescuer thread function. There's one rescuer for each
2312 * workqueue which has WQ_RESCUER set.
2314 * Regular work processing on a pool may block trying to create a new
2315 * worker which uses GFP_KERNEL allocation which has slight chance of
2316 * developing into deadlock if some works currently on the same queue
2317 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2318 * the problem rescuer solves.
2320 * When such condition is possible, the pool summons rescuers of all
2321 * workqueues which have works queued on the pool and let them process
2322 * those works so that forward progress can be guaranteed.
2324 * This should happen rarely.
2326 static int rescuer_thread(void *__rescuer)
2328 struct worker *rescuer = __rescuer;
2329 struct workqueue_struct *wq = rescuer->rescue_wq;
2330 struct list_head *scheduled = &rescuer->scheduled;
2331 bool is_unbound = wq->flags & WQ_UNBOUND;
2334 set_user_nice(current, RESCUER_NICE_LEVEL);
2337 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2338 * doesn't participate in concurrency management.
2340 rescuer->task->flags |= PF_WQ_WORKER;
2342 set_current_state(TASK_INTERRUPTIBLE);
2344 if (kthread_should_stop()) {
2345 __set_current_state(TASK_RUNNING);
2346 rescuer->task->flags &= ~PF_WQ_WORKER;
2351 * See whether any cpu is asking for help. Unbounded
2352 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2354 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2355 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2356 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2357 struct worker_pool *pool = cwq->pool;
2358 struct work_struct *work, *n;
2360 __set_current_state(TASK_RUNNING);
2361 mayday_clear_cpu(cpu, wq->mayday_mask);
2363 /* migrate to the target cpu if possible */
2364 rescuer->pool = pool;
2365 worker_maybe_bind_and_lock(rescuer);
2368 * Slurp in all works issued via this workqueue and
2371 BUG_ON(!list_empty(&rescuer->scheduled));
2372 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2373 if (get_work_cwq(work) == cwq)
2374 move_linked_works(work, scheduled, &n);
2376 process_scheduled_works(rescuer);
2379 * Leave this pool. If keep_working() is %true, notify a
2380 * regular worker; otherwise, we end up with 0 concurrency
2381 * and stalling the execution.
2383 if (keep_working(pool))
2384 wake_up_worker(pool);
2386 spin_unlock_irq(&pool->lock);
2389 /* rescuers should never participate in concurrency management */
2390 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2396 struct work_struct work;
2397 struct completion done;
2400 static void wq_barrier_func(struct work_struct *work)
2402 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2403 complete(&barr->done);
2407 * insert_wq_barrier - insert a barrier work
2408 * @cwq: cwq to insert barrier into
2409 * @barr: wq_barrier to insert
2410 * @target: target work to attach @barr to
2411 * @worker: worker currently executing @target, NULL if @target is not executing
2413 * @barr is linked to @target such that @barr is completed only after
2414 * @target finishes execution. Please note that the ordering
2415 * guarantee is observed only with respect to @target and on the local
2418 * Currently, a queued barrier can't be canceled. This is because
2419 * try_to_grab_pending() can't determine whether the work to be
2420 * grabbed is at the head of the queue and thus can't clear LINKED
2421 * flag of the previous work while there must be a valid next work
2422 * after a work with LINKED flag set.
2424 * Note that when @worker is non-NULL, @target may be modified
2425 * underneath us, so we can't reliably determine cwq from @target.
2428 * spin_lock_irq(pool->lock).
2430 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2431 struct wq_barrier *barr,
2432 struct work_struct *target, struct worker *worker)
2434 struct list_head *head;
2435 unsigned int linked = 0;
2438 * debugobject calls are safe here even with pool->lock locked
2439 * as we know for sure that this will not trigger any of the
2440 * checks and call back into the fixup functions where we
2443 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2444 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2445 init_completion(&barr->done);
2448 * If @target is currently being executed, schedule the
2449 * barrier to the worker; otherwise, put it after @target.
2452 head = worker->scheduled.next;
2454 unsigned long *bits = work_data_bits(target);
2456 head = target->entry.next;
2457 /* there can already be other linked works, inherit and set */
2458 linked = *bits & WORK_STRUCT_LINKED;
2459 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2462 debug_work_activate(&barr->work);
2463 insert_work(cwq, &barr->work, head,
2464 work_color_to_flags(WORK_NO_COLOR) | linked);
2468 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2469 * @wq: workqueue being flushed
2470 * @flush_color: new flush color, < 0 for no-op
2471 * @work_color: new work color, < 0 for no-op
2473 * Prepare cwqs for workqueue flushing.
2475 * If @flush_color is non-negative, flush_color on all cwqs should be
2476 * -1. If no cwq has in-flight commands at the specified color, all
2477 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2478 * has in flight commands, its cwq->flush_color is set to
2479 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2480 * wakeup logic is armed and %true is returned.
2482 * The caller should have initialized @wq->first_flusher prior to
2483 * calling this function with non-negative @flush_color. If
2484 * @flush_color is negative, no flush color update is done and %false
2487 * If @work_color is non-negative, all cwqs should have the same
2488 * work_color which is previous to @work_color and all will be
2489 * advanced to @work_color.
2492 * mutex_lock(wq->flush_mutex).
2495 * %true if @flush_color >= 0 and there's something to flush. %false
2498 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2499 int flush_color, int work_color)
2504 if (flush_color >= 0) {
2505 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2506 atomic_set(&wq->nr_cwqs_to_flush, 1);
2509 for_each_cwq_cpu(cpu, wq) {
2510 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2511 struct worker_pool *pool = cwq->pool;
2513 spin_lock_irq(&pool->lock);
2515 if (flush_color >= 0) {
2516 BUG_ON(cwq->flush_color != -1);
2518 if (cwq->nr_in_flight[flush_color]) {
2519 cwq->flush_color = flush_color;
2520 atomic_inc(&wq->nr_cwqs_to_flush);
2525 if (work_color >= 0) {
2526 BUG_ON(work_color != work_next_color(cwq->work_color));
2527 cwq->work_color = work_color;
2530 spin_unlock_irq(&pool->lock);
2533 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2534 complete(&wq->first_flusher->done);
2540 * flush_workqueue - ensure that any scheduled work has run to completion.
2541 * @wq: workqueue to flush
2543 * Forces execution of the workqueue and blocks until its completion.
2544 * This is typically used in driver shutdown handlers.
2546 * We sleep until all works which were queued on entry have been handled,
2547 * but we are not livelocked by new incoming ones.
2549 void flush_workqueue(struct workqueue_struct *wq)
2551 struct wq_flusher this_flusher = {
2552 .list = LIST_HEAD_INIT(this_flusher.list),
2554 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2558 lock_map_acquire(&wq->lockdep_map);
2559 lock_map_release(&wq->lockdep_map);
2561 mutex_lock(&wq->flush_mutex);
2564 * Start-to-wait phase
2566 next_color = work_next_color(wq->work_color);
2568 if (next_color != wq->flush_color) {
2570 * Color space is not full. The current work_color
2571 * becomes our flush_color and work_color is advanced
2574 BUG_ON(!list_empty(&wq->flusher_overflow));
2575 this_flusher.flush_color = wq->work_color;
2576 wq->work_color = next_color;
2578 if (!wq->first_flusher) {
2579 /* no flush in progress, become the first flusher */
2580 BUG_ON(wq->flush_color != this_flusher.flush_color);
2582 wq->first_flusher = &this_flusher;
2584 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2586 /* nothing to flush, done */
2587 wq->flush_color = next_color;
2588 wq->first_flusher = NULL;
2593 BUG_ON(wq->flush_color == this_flusher.flush_color);
2594 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2595 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2599 * Oops, color space is full, wait on overflow queue.
2600 * The next flush completion will assign us
2601 * flush_color and transfer to flusher_queue.
2603 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2606 mutex_unlock(&wq->flush_mutex);
2608 wait_for_completion(&this_flusher.done);
2611 * Wake-up-and-cascade phase
2613 * First flushers are responsible for cascading flushes and
2614 * handling overflow. Non-first flushers can simply return.
2616 if (wq->first_flusher != &this_flusher)
2619 mutex_lock(&wq->flush_mutex);
2621 /* we might have raced, check again with mutex held */
2622 if (wq->first_flusher != &this_flusher)
2625 wq->first_flusher = NULL;
2627 BUG_ON(!list_empty(&this_flusher.list));
2628 BUG_ON(wq->flush_color != this_flusher.flush_color);
2631 struct wq_flusher *next, *tmp;
2633 /* complete all the flushers sharing the current flush color */
2634 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2635 if (next->flush_color != wq->flush_color)
2637 list_del_init(&next->list);
2638 complete(&next->done);
2641 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2642 wq->flush_color != work_next_color(wq->work_color));
2644 /* this flush_color is finished, advance by one */
2645 wq->flush_color = work_next_color(wq->flush_color);
2647 /* one color has been freed, handle overflow queue */
2648 if (!list_empty(&wq->flusher_overflow)) {
2650 * Assign the same color to all overflowed
2651 * flushers, advance work_color and append to
2652 * flusher_queue. This is the start-to-wait
2653 * phase for these overflowed flushers.
2655 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2656 tmp->flush_color = wq->work_color;
2658 wq->work_color = work_next_color(wq->work_color);
2660 list_splice_tail_init(&wq->flusher_overflow,
2661 &wq->flusher_queue);
2662 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2665 if (list_empty(&wq->flusher_queue)) {
2666 BUG_ON(wq->flush_color != wq->work_color);
2671 * Need to flush more colors. Make the next flusher
2672 * the new first flusher and arm cwqs.
2674 BUG_ON(wq->flush_color == wq->work_color);
2675 BUG_ON(wq->flush_color != next->flush_color);
2677 list_del_init(&next->list);
2678 wq->first_flusher = next;
2680 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2684 * Meh... this color is already done, clear first
2685 * flusher and repeat cascading.
2687 wq->first_flusher = NULL;
2691 mutex_unlock(&wq->flush_mutex);
2693 EXPORT_SYMBOL_GPL(flush_workqueue);
2696 * drain_workqueue - drain a workqueue
2697 * @wq: workqueue to drain
2699 * Wait until the workqueue becomes empty. While draining is in progress,
2700 * only chain queueing is allowed. IOW, only currently pending or running
2701 * work items on @wq can queue further work items on it. @wq is flushed
2702 * repeatedly until it becomes empty. The number of flushing is detemined
2703 * by the depth of chaining and should be relatively short. Whine if it
2706 void drain_workqueue(struct workqueue_struct *wq)
2708 unsigned int flush_cnt = 0;
2712 * __queue_work() needs to test whether there are drainers, is much
2713 * hotter than drain_workqueue() and already looks at @wq->flags.
2714 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2716 spin_lock(&workqueue_lock);
2717 if (!wq->nr_drainers++)
2718 wq->flags |= WQ_DRAINING;
2719 spin_unlock(&workqueue_lock);
2721 flush_workqueue(wq);
2723 for_each_cwq_cpu(cpu, wq) {
2724 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2727 spin_lock_irq(&cwq->pool->lock);
2728 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2729 spin_unlock_irq(&cwq->pool->lock);
2734 if (++flush_cnt == 10 ||
2735 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2736 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2737 wq->name, flush_cnt);
2741 spin_lock(&workqueue_lock);
2742 if (!--wq->nr_drainers)
2743 wq->flags &= ~WQ_DRAINING;
2744 spin_unlock(&workqueue_lock);
2746 EXPORT_SYMBOL_GPL(drain_workqueue);
2748 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2750 struct worker *worker = NULL;
2751 struct worker_pool *pool;
2752 struct cpu_workqueue_struct *cwq;
2755 pool = get_work_pool(work);
2759 spin_lock_irq(&pool->lock);
2760 /* see the comment in try_to_grab_pending() with the same code */
2761 cwq = get_work_cwq(work);
2763 if (unlikely(cwq->pool != pool))
2766 worker = find_worker_executing_work(pool, work);
2769 cwq = worker->current_cwq;
2772 insert_wq_barrier(cwq, barr, work, worker);
2773 spin_unlock_irq(&pool->lock);
2776 * If @max_active is 1 or rescuer is in use, flushing another work
2777 * item on the same workqueue may lead to deadlock. Make sure the
2778 * flusher is not running on the same workqueue by verifying write
2781 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2782 lock_map_acquire(&cwq->wq->lockdep_map);
2784 lock_map_acquire_read(&cwq->wq->lockdep_map);
2785 lock_map_release(&cwq->wq->lockdep_map);
2789 spin_unlock_irq(&pool->lock);
2794 * flush_work - wait for a work to finish executing the last queueing instance
2795 * @work: the work to flush
2797 * Wait until @work has finished execution. @work is guaranteed to be idle
2798 * on return if it hasn't been requeued since flush started.
2801 * %true if flush_work() waited for the work to finish execution,
2802 * %false if it was already idle.
2804 bool flush_work(struct work_struct *work)
2806 struct wq_barrier barr;
2808 lock_map_acquire(&work->lockdep_map);
2809 lock_map_release(&work->lockdep_map);
2811 if (start_flush_work(work, &barr)) {
2812 wait_for_completion(&barr.done);
2813 destroy_work_on_stack(&barr.work);
2819 EXPORT_SYMBOL_GPL(flush_work);
2821 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2823 unsigned long flags;
2827 ret = try_to_grab_pending(work, is_dwork, &flags);
2829 * If someone else is canceling, wait for the same event it
2830 * would be waiting for before retrying.
2832 if (unlikely(ret == -ENOENT))
2834 } while (unlikely(ret < 0));
2836 /* tell other tasks trying to grab @work to back off */
2837 mark_work_canceling(work);
2838 local_irq_restore(flags);
2841 clear_work_data(work);
2846 * cancel_work_sync - cancel a work and wait for it to finish
2847 * @work: the work to cancel
2849 * Cancel @work and wait for its execution to finish. This function
2850 * can be used even if the work re-queues itself or migrates to
2851 * another workqueue. On return from this function, @work is
2852 * guaranteed to be not pending or executing on any CPU.
2854 * cancel_work_sync(&delayed_work->work) must not be used for
2855 * delayed_work's. Use cancel_delayed_work_sync() instead.
2857 * The caller must ensure that the workqueue on which @work was last
2858 * queued can't be destroyed before this function returns.
2861 * %true if @work was pending, %false otherwise.
2863 bool cancel_work_sync(struct work_struct *work)
2865 return __cancel_work_timer(work, false);
2867 EXPORT_SYMBOL_GPL(cancel_work_sync);
2870 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2871 * @dwork: the delayed work to flush
2873 * Delayed timer is cancelled and the pending work is queued for
2874 * immediate execution. Like flush_work(), this function only
2875 * considers the last queueing instance of @dwork.
2878 * %true if flush_work() waited for the work to finish execution,
2879 * %false if it was already idle.
2881 bool flush_delayed_work(struct delayed_work *dwork)
2883 local_irq_disable();
2884 if (del_timer_sync(&dwork->timer))
2885 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2887 return flush_work(&dwork->work);
2889 EXPORT_SYMBOL(flush_delayed_work);
2892 * cancel_delayed_work - cancel a delayed work
2893 * @dwork: delayed_work to cancel
2895 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2896 * and canceled; %false if wasn't pending. Note that the work callback
2897 * function may still be running on return, unless it returns %true and the
2898 * work doesn't re-arm itself. Explicitly flush or use
2899 * cancel_delayed_work_sync() to wait on it.
2901 * This function is safe to call from any context including IRQ handler.
2903 bool cancel_delayed_work(struct delayed_work *dwork)
2905 unsigned long flags;
2909 ret = try_to_grab_pending(&dwork->work, true, &flags);
2910 } while (unlikely(ret == -EAGAIN));
2912 if (unlikely(ret < 0))
2915 set_work_pool_and_clear_pending(&dwork->work,
2916 get_work_pool_id(&dwork->work));
2917 local_irq_restore(flags);
2920 EXPORT_SYMBOL(cancel_delayed_work);
2923 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2924 * @dwork: the delayed work cancel
2926 * This is cancel_work_sync() for delayed works.
2929 * %true if @dwork was pending, %false otherwise.
2931 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2933 return __cancel_work_timer(&dwork->work, true);
2935 EXPORT_SYMBOL(cancel_delayed_work_sync);
2938 * schedule_work_on - put work task on a specific cpu
2939 * @cpu: cpu to put the work task on
2940 * @work: job to be done
2942 * This puts a job on a specific cpu
2944 bool schedule_work_on(int cpu, struct work_struct *work)
2946 return queue_work_on(cpu, system_wq, work);
2948 EXPORT_SYMBOL(schedule_work_on);
2951 * schedule_work - put work task in global workqueue
2952 * @work: job to be done
2954 * Returns %false if @work was already on the kernel-global workqueue and
2957 * This puts a job in the kernel-global workqueue if it was not already
2958 * queued and leaves it in the same position on the kernel-global
2959 * workqueue otherwise.
2961 bool schedule_work(struct work_struct *work)
2963 return queue_work(system_wq, work);
2965 EXPORT_SYMBOL(schedule_work);
2968 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2970 * @dwork: job to be done
2971 * @delay: number of jiffies to wait
2973 * After waiting for a given time this puts a job in the kernel-global
2974 * workqueue on the specified CPU.
2976 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
2977 unsigned long delay)
2979 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2981 EXPORT_SYMBOL(schedule_delayed_work_on);
2984 * schedule_delayed_work - put work task in global workqueue after delay
2985 * @dwork: job to be done
2986 * @delay: number of jiffies to wait or 0 for immediate execution
2988 * After waiting for a given time this puts a job in the kernel-global
2991 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
2993 return queue_delayed_work(system_wq, dwork, delay);
2995 EXPORT_SYMBOL(schedule_delayed_work);
2998 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2999 * @func: the function to call
3001 * schedule_on_each_cpu() executes @func on each online CPU using the
3002 * system workqueue and blocks until all CPUs have completed.
3003 * schedule_on_each_cpu() is very slow.
3006 * 0 on success, -errno on failure.
3008 int schedule_on_each_cpu(work_func_t func)
3011 struct work_struct __percpu *works;
3013 works = alloc_percpu(struct work_struct);
3019 for_each_online_cpu(cpu) {
3020 struct work_struct *work = per_cpu_ptr(works, cpu);
3022 INIT_WORK(work, func);
3023 schedule_work_on(cpu, work);
3026 for_each_online_cpu(cpu)
3027 flush_work(per_cpu_ptr(works, cpu));
3035 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3037 * Forces execution of the kernel-global workqueue and blocks until its
3040 * Think twice before calling this function! It's very easy to get into
3041 * trouble if you don't take great care. Either of the following situations
3042 * will lead to deadlock:
3044 * One of the work items currently on the workqueue needs to acquire
3045 * a lock held by your code or its caller.
3047 * Your code is running in the context of a work routine.
3049 * They will be detected by lockdep when they occur, but the first might not
3050 * occur very often. It depends on what work items are on the workqueue and
3051 * what locks they need, which you have no control over.
3053 * In most situations flushing the entire workqueue is overkill; you merely
3054 * need to know that a particular work item isn't queued and isn't running.
3055 * In such cases you should use cancel_delayed_work_sync() or
3056 * cancel_work_sync() instead.
3058 void flush_scheduled_work(void)
3060 flush_workqueue(system_wq);
3062 EXPORT_SYMBOL(flush_scheduled_work);
3065 * execute_in_process_context - reliably execute the routine with user context
3066 * @fn: the function to execute
3067 * @ew: guaranteed storage for the execute work structure (must
3068 * be available when the work executes)
3070 * Executes the function immediately if process context is available,
3071 * otherwise schedules the function for delayed execution.
3073 * Returns: 0 - function was executed
3074 * 1 - function was scheduled for execution
3076 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3078 if (!in_interrupt()) {
3083 INIT_WORK(&ew->work, fn);
3084 schedule_work(&ew->work);
3088 EXPORT_SYMBOL_GPL(execute_in_process_context);
3090 int keventd_up(void)
3092 return system_wq != NULL;
3095 static int alloc_cwqs(struct workqueue_struct *wq)
3098 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3099 * Make sure that the alignment isn't lower than that of
3100 * unsigned long long.
3102 const size_t size = sizeof(struct cpu_workqueue_struct);
3103 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3104 __alignof__(unsigned long long));
3106 if (!(wq->flags & WQ_UNBOUND))
3107 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3112 * Allocate enough room to align cwq and put an extra
3113 * pointer at the end pointing back to the originally
3114 * allocated pointer which will be used for free.
3116 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3118 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3119 *(void **)(wq->cpu_wq.single + 1) = ptr;
3123 /* just in case, make sure it's actually aligned */
3124 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3125 return wq->cpu_wq.v ? 0 : -ENOMEM;
3128 static void free_cwqs(struct workqueue_struct *wq)
3130 if (!(wq->flags & WQ_UNBOUND))
3131 free_percpu(wq->cpu_wq.pcpu);
3132 else if (wq->cpu_wq.single) {
3133 /* the pointer to free is stored right after the cwq */
3134 kfree(*(void **)(wq->cpu_wq.single + 1));
3138 static int wq_clamp_max_active(int max_active, unsigned int flags,
3141 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3143 if (max_active < 1 || max_active > lim)
3144 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3145 max_active, name, 1, lim);
3147 return clamp_val(max_active, 1, lim);
3150 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3153 struct lock_class_key *key,
3154 const char *lock_name, ...)
3156 va_list args, args1;
3157 struct workqueue_struct *wq;
3161 /* determine namelen, allocate wq and format name */
3162 va_start(args, lock_name);
3163 va_copy(args1, args);
3164 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3166 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3170 vsnprintf(wq->name, namelen, fmt, args1);
3175 * Workqueues which may be used during memory reclaim should
3176 * have a rescuer to guarantee forward progress.
3178 if (flags & WQ_MEM_RECLAIM)
3179 flags |= WQ_RESCUER;
3181 max_active = max_active ?: WQ_DFL_ACTIVE;
3182 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3186 wq->saved_max_active = max_active;
3187 mutex_init(&wq->flush_mutex);
3188 atomic_set(&wq->nr_cwqs_to_flush, 0);
3189 INIT_LIST_HEAD(&wq->flusher_queue);
3190 INIT_LIST_HEAD(&wq->flusher_overflow);
3192 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3193 INIT_LIST_HEAD(&wq->list);
3195 if (alloc_cwqs(wq) < 0)
3198 for_each_cwq_cpu(cpu, wq) {
3199 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3201 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3202 cwq->pool = get_std_worker_pool(cpu, flags & WQ_HIGHPRI);
3204 cwq->flush_color = -1;
3205 cwq->max_active = max_active;
3206 INIT_LIST_HEAD(&cwq->delayed_works);
3209 if (flags & WQ_RESCUER) {
3210 struct worker *rescuer;
3212 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3215 wq->rescuer = rescuer = alloc_worker();
3219 rescuer->rescue_wq = wq;
3220 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3222 if (IS_ERR(rescuer->task))
3225 rescuer->task->flags |= PF_THREAD_BOUND;
3226 wake_up_process(rescuer->task);
3230 * workqueue_lock protects global freeze state and workqueues
3231 * list. Grab it, set max_active accordingly and add the new
3232 * workqueue to workqueues list.
3234 spin_lock(&workqueue_lock);
3236 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3237 for_each_cwq_cpu(cpu, wq)
3238 get_cwq(cpu, wq)->max_active = 0;
3240 list_add(&wq->list, &workqueues);
3242 spin_unlock(&workqueue_lock);
3248 free_mayday_mask(wq->mayday_mask);
3254 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3257 * destroy_workqueue - safely terminate a workqueue
3258 * @wq: target workqueue
3260 * Safely destroy a workqueue. All work currently pending will be done first.
3262 void destroy_workqueue(struct workqueue_struct *wq)
3266 /* drain it before proceeding with destruction */
3267 drain_workqueue(wq);
3270 * wq list is used to freeze wq, remove from list after
3271 * flushing is complete in case freeze races us.
3273 spin_lock(&workqueue_lock);
3274 list_del(&wq->list);
3275 spin_unlock(&workqueue_lock);
3278 for_each_cwq_cpu(cpu, wq) {
3279 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3282 for (i = 0; i < WORK_NR_COLORS; i++)
3283 BUG_ON(cwq->nr_in_flight[i]);
3284 BUG_ON(cwq->nr_active);
3285 BUG_ON(!list_empty(&cwq->delayed_works));
3288 if (wq->flags & WQ_RESCUER) {
3289 kthread_stop(wq->rescuer->task);
3290 free_mayday_mask(wq->mayday_mask);
3297 EXPORT_SYMBOL_GPL(destroy_workqueue);
3300 * cwq_set_max_active - adjust max_active of a cwq
3301 * @cwq: target cpu_workqueue_struct
3302 * @max_active: new max_active value.
3304 * Set @cwq->max_active to @max_active and activate delayed works if
3308 * spin_lock_irq(pool->lock).
3310 static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active)
3312 cwq->max_active = max_active;
3314 while (!list_empty(&cwq->delayed_works) &&
3315 cwq->nr_active < cwq->max_active)
3316 cwq_activate_first_delayed(cwq);
3320 * workqueue_set_max_active - adjust max_active of a workqueue
3321 * @wq: target workqueue
3322 * @max_active: new max_active value.
3324 * Set max_active of @wq to @max_active.
3327 * Don't call from IRQ context.
3329 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3333 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3335 spin_lock(&workqueue_lock);
3337 wq->saved_max_active = max_active;
3339 for_each_cwq_cpu(cpu, wq) {
3340 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3341 struct worker_pool *pool = cwq->pool;
3343 spin_lock_irq(&pool->lock);
3345 if (!(wq->flags & WQ_FREEZABLE) ||
3346 !(pool->flags & POOL_FREEZING))
3347 cwq_set_max_active(cwq, max_active);
3349 spin_unlock_irq(&pool->lock);
3352 spin_unlock(&workqueue_lock);
3354 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3357 * workqueue_congested - test whether a workqueue is congested
3358 * @cpu: CPU in question
3359 * @wq: target workqueue
3361 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3362 * no synchronization around this function and the test result is
3363 * unreliable and only useful as advisory hints or for debugging.
3366 * %true if congested, %false otherwise.
3368 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3370 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3372 return !list_empty(&cwq->delayed_works);
3374 EXPORT_SYMBOL_GPL(workqueue_congested);
3377 * work_busy - test whether a work is currently pending or running
3378 * @work: the work to be tested
3380 * Test whether @work is currently pending or running. There is no
3381 * synchronization around this function and the test result is
3382 * unreliable and only useful as advisory hints or for debugging.
3385 * OR'd bitmask of WORK_BUSY_* bits.
3387 unsigned int work_busy(struct work_struct *work)
3389 struct worker_pool *pool = get_work_pool(work);
3390 unsigned long flags;
3391 unsigned int ret = 0;
3393 if (work_pending(work))
3394 ret |= WORK_BUSY_PENDING;
3397 spin_lock_irqsave(&pool->lock, flags);
3398 if (find_worker_executing_work(pool, work))
3399 ret |= WORK_BUSY_RUNNING;
3400 spin_unlock_irqrestore(&pool->lock, flags);
3405 EXPORT_SYMBOL_GPL(work_busy);
3410 * There are two challenges in supporting CPU hotplug. Firstly, there
3411 * are a lot of assumptions on strong associations among work, cwq and
3412 * pool which make migrating pending and scheduled works very
3413 * difficult to implement without impacting hot paths. Secondly,
3414 * worker pools serve mix of short, long and very long running works making
3415 * blocked draining impractical.
3417 * This is solved by allowing the pools to be disassociated from the CPU
3418 * running as an unbound one and allowing it to be reattached later if the
3419 * cpu comes back online.
3422 static void wq_unbind_fn(struct work_struct *work)
3424 int cpu = smp_processor_id();
3425 struct worker_pool *pool;
3426 struct worker *worker;
3427 struct hlist_node *pos;
3430 for_each_std_worker_pool(pool, cpu) {
3431 BUG_ON(cpu != smp_processor_id());
3433 mutex_lock(&pool->assoc_mutex);
3434 spin_lock_irq(&pool->lock);
3437 * We've claimed all manager positions. Make all workers
3438 * unbound and set DISASSOCIATED. Before this, all workers
3439 * except for the ones which are still executing works from
3440 * before the last CPU down must be on the cpu. After
3441 * this, they may become diasporas.
3443 list_for_each_entry(worker, &pool->idle_list, entry)
3444 worker->flags |= WORKER_UNBOUND;
3446 for_each_busy_worker(worker, i, pos, pool)
3447 worker->flags |= WORKER_UNBOUND;
3449 pool->flags |= POOL_DISASSOCIATED;
3451 spin_unlock_irq(&pool->lock);
3452 mutex_unlock(&pool->assoc_mutex);
3456 * Call schedule() so that we cross rq->lock and thus can guarantee
3457 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3458 * as scheduler callbacks may be invoked from other cpus.
3463 * Sched callbacks are disabled now. Zap nr_running. After this,
3464 * nr_running stays zero and need_more_worker() and keep_working()
3465 * are always true as long as the worklist is not empty. Pools on
3466 * @cpu now behave as unbound (in terms of concurrency management)
3467 * pools which are served by workers tied to the CPU.
3469 * On return from this function, the current worker would trigger
3470 * unbound chain execution of pending work items if other workers
3473 for_each_std_worker_pool(pool, cpu)
3474 atomic_set(&pool->nr_running, 0);
3478 * Workqueues should be brought up before normal priority CPU notifiers.
3479 * This will be registered high priority CPU notifier.
3481 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3482 unsigned long action,
3485 unsigned int cpu = (unsigned long)hcpu;
3486 struct worker_pool *pool;
3488 switch (action & ~CPU_TASKS_FROZEN) {
3489 case CPU_UP_PREPARE:
3490 for_each_std_worker_pool(pool, cpu) {
3491 struct worker *worker;
3493 if (pool->nr_workers)
3496 worker = create_worker(pool);
3500 spin_lock_irq(&pool->lock);
3501 start_worker(worker);
3502 spin_unlock_irq(&pool->lock);
3506 case CPU_DOWN_FAILED:
3508 for_each_std_worker_pool(pool, cpu) {
3509 mutex_lock(&pool->assoc_mutex);
3510 spin_lock_irq(&pool->lock);
3512 pool->flags &= ~POOL_DISASSOCIATED;
3513 rebind_workers(pool);
3515 spin_unlock_irq(&pool->lock);
3516 mutex_unlock(&pool->assoc_mutex);
3524 * Workqueues should be brought down after normal priority CPU notifiers.
3525 * This will be registered as low priority CPU notifier.
3527 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3528 unsigned long action,
3531 unsigned int cpu = (unsigned long)hcpu;
3532 struct work_struct unbind_work;
3534 switch (action & ~CPU_TASKS_FROZEN) {
3535 case CPU_DOWN_PREPARE:
3536 /* unbinding should happen on the local CPU */
3537 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
3538 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3539 flush_work(&unbind_work);
3547 struct work_for_cpu {
3548 struct work_struct work;
3554 static void work_for_cpu_fn(struct work_struct *work)
3556 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3558 wfc->ret = wfc->fn(wfc->arg);
3562 * work_on_cpu - run a function in user context on a particular cpu
3563 * @cpu: the cpu to run on
3564 * @fn: the function to run
3565 * @arg: the function arg
3567 * This will return the value @fn returns.
3568 * It is up to the caller to ensure that the cpu doesn't go offline.
3569 * The caller must not hold any locks which would prevent @fn from completing.
3571 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3573 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3575 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3576 schedule_work_on(cpu, &wfc.work);
3577 flush_work(&wfc.work);
3580 EXPORT_SYMBOL_GPL(work_on_cpu);
3581 #endif /* CONFIG_SMP */
3583 #ifdef CONFIG_FREEZER
3586 * freeze_workqueues_begin - begin freezing workqueues
3588 * Start freezing workqueues. After this function returns, all freezable
3589 * workqueues will queue new works to their frozen_works list instead of
3593 * Grabs and releases workqueue_lock and pool->lock's.
3595 void freeze_workqueues_begin(void)
3599 spin_lock(&workqueue_lock);
3601 BUG_ON(workqueue_freezing);
3602 workqueue_freezing = true;
3604 for_each_wq_cpu(cpu) {
3605 struct worker_pool *pool;
3606 struct workqueue_struct *wq;
3608 for_each_std_worker_pool(pool, cpu) {
3609 spin_lock_irq(&pool->lock);
3611 WARN_ON_ONCE(pool->flags & POOL_FREEZING);
3612 pool->flags |= POOL_FREEZING;
3614 list_for_each_entry(wq, &workqueues, list) {
3615 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3617 if (cwq && cwq->pool == pool &&
3618 (wq->flags & WQ_FREEZABLE))
3619 cwq->max_active = 0;
3622 spin_unlock_irq(&pool->lock);
3626 spin_unlock(&workqueue_lock);
3630 * freeze_workqueues_busy - are freezable workqueues still busy?
3632 * Check whether freezing is complete. This function must be called
3633 * between freeze_workqueues_begin() and thaw_workqueues().
3636 * Grabs and releases workqueue_lock.
3639 * %true if some freezable workqueues are still busy. %false if freezing
3642 bool freeze_workqueues_busy(void)
3647 spin_lock(&workqueue_lock);
3649 BUG_ON(!workqueue_freezing);
3651 for_each_wq_cpu(cpu) {
3652 struct workqueue_struct *wq;
3654 * nr_active is monotonically decreasing. It's safe
3655 * to peek without lock.
3657 list_for_each_entry(wq, &workqueues, list) {
3658 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3660 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3663 BUG_ON(cwq->nr_active < 0);
3664 if (cwq->nr_active) {
3671 spin_unlock(&workqueue_lock);
3676 * thaw_workqueues - thaw workqueues
3678 * Thaw workqueues. Normal queueing is restored and all collected
3679 * frozen works are transferred to their respective pool worklists.
3682 * Grabs and releases workqueue_lock and pool->lock's.
3684 void thaw_workqueues(void)
3688 spin_lock(&workqueue_lock);
3690 if (!workqueue_freezing)
3693 for_each_wq_cpu(cpu) {
3694 struct worker_pool *pool;
3695 struct workqueue_struct *wq;
3697 for_each_std_worker_pool(pool, cpu) {
3698 spin_lock_irq(&pool->lock);
3700 WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
3701 pool->flags &= ~POOL_FREEZING;
3703 list_for_each_entry(wq, &workqueues, list) {
3704 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3706 if (!cwq || cwq->pool != pool ||
3707 !(wq->flags & WQ_FREEZABLE))
3710 /* restore max_active and repopulate worklist */
3711 cwq_set_max_active(cwq, wq->saved_max_active);
3714 wake_up_worker(pool);
3716 spin_unlock_irq(&pool->lock);
3720 workqueue_freezing = false;
3722 spin_unlock(&workqueue_lock);
3724 #endif /* CONFIG_FREEZER */
3726 static int __init init_workqueues(void)
3730 /* make sure we have enough bits for OFFQ pool ID */
3731 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
3732 WORK_CPU_END * NR_STD_WORKER_POOLS);
3734 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3735 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3737 /* initialize CPU pools */
3738 for_each_wq_cpu(cpu) {
3739 struct worker_pool *pool;
3741 for_each_std_worker_pool(pool, cpu) {
3742 spin_lock_init(&pool->lock);
3744 pool->flags |= POOL_DISASSOCIATED;
3745 INIT_LIST_HEAD(&pool->worklist);
3746 INIT_LIST_HEAD(&pool->idle_list);
3747 hash_init(pool->busy_hash);
3749 init_timer_deferrable(&pool->idle_timer);
3750 pool->idle_timer.function = idle_worker_timeout;
3751 pool->idle_timer.data = (unsigned long)pool;
3753 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3754 (unsigned long)pool);
3756 mutex_init(&pool->assoc_mutex);
3757 ida_init(&pool->worker_ida);
3760 BUG_ON(worker_pool_assign_id(pool));
3764 /* create the initial worker */
3765 for_each_online_wq_cpu(cpu) {
3766 struct worker_pool *pool;
3768 for_each_std_worker_pool(pool, cpu) {
3769 struct worker *worker;
3771 if (cpu != WORK_CPU_UNBOUND)
3772 pool->flags &= ~POOL_DISASSOCIATED;
3774 worker = create_worker(pool);
3776 spin_lock_irq(&pool->lock);
3777 start_worker(worker);
3778 spin_unlock_irq(&pool->lock);
3782 system_wq = alloc_workqueue("events", 0, 0);
3783 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3784 system_long_wq = alloc_workqueue("events_long", 0, 0);
3785 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3786 WQ_UNBOUND_MAX_ACTIVE);
3787 system_freezable_wq = alloc_workqueue("events_freezable",
3789 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3790 !system_unbound_wq || !system_freezable_wq);
3793 early_initcall(init_workqueues);