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 gcwq 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 gcwq behaves as an unbound one.
61 * Note that DISASSOCIATED can be flipped only while holding
62 * assoc_mutex of all pools on the gcwq to avoid changing binding
63 * state while create_worker() is in progress.
65 GCWQ_DISASSOCIATED = 1 << 0, /* cpu can't serve workers */
66 GCWQ_FREEZING = 1 << 1, /* freeze in progress */
69 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
70 POOL_MANAGING_WORKERS = 1 << 1, /* managing workers */
73 WORKER_STARTED = 1 << 0, /* started */
74 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
76 WORKER_PREP = 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
80 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_UNBOUND |
83 NR_WORKER_POOLS = 2, /* # worker pools per gcwq */
85 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
87 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
88 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
90 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
91 /* call for help after 10ms
93 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
94 CREATE_COOLDOWN = HZ, /* time to breath after fail */
97 * Rescue workers are used only on emergencies and shared by
100 RESCUER_NICE_LEVEL = -20,
101 HIGHPRI_NICE_LEVEL = -20,
105 * Structure fields follow one of the following exclusion rules.
107 * I: Modifiable by initialization/destruction paths and read-only for
110 * P: Preemption protected. Disabling preemption is enough and should
111 * only be modified and accessed from the local cpu.
113 * L: gcwq->lock protected. Access with gcwq->lock held.
115 * X: During normal operation, modification requires gcwq->lock and
116 * should be done only from local cpu. Either disabling preemption
117 * on local cpu or grabbing gcwq->lock is enough for read access.
118 * If GCWQ_DISASSOCIATED is set, it's identical to L.
120 * F: wq->flush_mutex protected.
122 * W: workqueue_lock protected.
125 /* struct worker is defined in workqueue_internal.h */
128 struct global_cwq *gcwq; /* I: the owning gcwq */
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 struct mutex assoc_mutex; /* protect GCWQ_DISASSOCIATED */
142 struct ida worker_ida; /* L: for worker IDs */
146 * Global per-cpu workqueue. There's one and only one for each cpu
147 * and all works are queued and processed here regardless of their
151 spinlock_t lock; /* the gcwq lock */
152 unsigned int cpu; /* I: the associated cpu */
153 unsigned int flags; /* L: GCWQ_* flags */
155 /* workers are chained either in busy_hash or pool idle_list */
156 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
157 /* L: hash of busy workers */
159 struct worker_pool pools[NR_WORKER_POOLS];
160 /* normal and highpri pools */
161 } ____cacheline_aligned_in_smp;
164 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
165 * work_struct->data are used for flags and thus cwqs need to be
166 * aligned at two's power of the number of flag bits.
168 struct cpu_workqueue_struct {
169 struct worker_pool *pool; /* I: the associated pool */
170 struct workqueue_struct *wq; /* I: the owning workqueue */
171 int work_color; /* L: current color */
172 int flush_color; /* L: flushing color */
173 int nr_in_flight[WORK_NR_COLORS];
174 /* L: nr of in_flight works */
175 int nr_active; /* L: nr of active works */
176 int max_active; /* L: max active works */
177 struct list_head delayed_works; /* L: delayed works */
181 * Structure used to wait for workqueue flush.
184 struct list_head list; /* F: list of flushers */
185 int flush_color; /* F: flush color waiting for */
186 struct completion done; /* flush completion */
190 * All cpumasks are assumed to be always set on UP and thus can't be
191 * used to determine whether there's something to be done.
194 typedef cpumask_var_t mayday_mask_t;
195 #define mayday_test_and_set_cpu(cpu, mask) \
196 cpumask_test_and_set_cpu((cpu), (mask))
197 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
198 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
199 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
200 #define free_mayday_mask(mask) free_cpumask_var((mask))
202 typedef unsigned long mayday_mask_t;
203 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
204 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
205 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
206 #define alloc_mayday_mask(maskp, gfp) true
207 #define free_mayday_mask(mask) do { } while (0)
211 * The externally visible workqueue abstraction is an array of
212 * per-CPU workqueues:
214 struct workqueue_struct {
215 unsigned int flags; /* W: WQ_* flags */
217 struct cpu_workqueue_struct __percpu *pcpu;
218 struct cpu_workqueue_struct *single;
220 } cpu_wq; /* I: cwq's */
221 struct list_head list; /* W: list of all workqueues */
223 struct mutex flush_mutex; /* protects wq flushing */
224 int work_color; /* F: current work color */
225 int flush_color; /* F: current flush color */
226 atomic_t nr_cwqs_to_flush; /* flush in progress */
227 struct wq_flusher *first_flusher; /* F: first flusher */
228 struct list_head flusher_queue; /* F: flush waiters */
229 struct list_head flusher_overflow; /* F: flush overflow list */
231 mayday_mask_t mayday_mask; /* cpus requesting rescue */
232 struct worker *rescuer; /* I: rescue worker */
234 int nr_drainers; /* W: drain in progress */
235 int saved_max_active; /* W: saved cwq max_active */
236 #ifdef CONFIG_LOCKDEP
237 struct lockdep_map lockdep_map;
239 char name[]; /* I: workqueue name */
242 struct workqueue_struct *system_wq __read_mostly;
243 EXPORT_SYMBOL_GPL(system_wq);
244 struct workqueue_struct *system_highpri_wq __read_mostly;
245 EXPORT_SYMBOL_GPL(system_highpri_wq);
246 struct workqueue_struct *system_long_wq __read_mostly;
247 EXPORT_SYMBOL_GPL(system_long_wq);
248 struct workqueue_struct *system_unbound_wq __read_mostly;
249 EXPORT_SYMBOL_GPL(system_unbound_wq);
250 struct workqueue_struct *system_freezable_wq __read_mostly;
251 EXPORT_SYMBOL_GPL(system_freezable_wq);
253 #define CREATE_TRACE_POINTS
254 #include <trace/events/workqueue.h>
256 #define for_each_worker_pool(pool, gcwq) \
257 for ((pool) = &(gcwq)->pools[0]; \
258 (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
260 #define for_each_busy_worker(worker, i, pos, gcwq) \
261 hash_for_each(gcwq->busy_hash, i, pos, worker, hentry)
263 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
266 if (cpu < nr_cpu_ids) {
268 cpu = cpumask_next(cpu, mask);
269 if (cpu < nr_cpu_ids)
273 return WORK_CPU_UNBOUND;
275 return WORK_CPU_NONE;
278 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
279 struct workqueue_struct *wq)
281 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
287 * An extra gcwq is defined for an invalid cpu number
288 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
289 * specific CPU. The following iterators are similar to
290 * for_each_*_cpu() iterators but also considers the unbound gcwq.
292 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
293 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
294 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
295 * WORK_CPU_UNBOUND for unbound workqueues
297 #define for_each_gcwq_cpu(cpu) \
298 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
299 (cpu) < WORK_CPU_NONE; \
300 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
302 #define for_each_online_gcwq_cpu(cpu) \
303 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
304 (cpu) < WORK_CPU_NONE; \
305 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
307 #define for_each_cwq_cpu(cpu, wq) \
308 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
309 (cpu) < WORK_CPU_NONE; \
310 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
312 #ifdef CONFIG_DEBUG_OBJECTS_WORK
314 static struct debug_obj_descr work_debug_descr;
316 static void *work_debug_hint(void *addr)
318 return ((struct work_struct *) addr)->func;
322 * fixup_init is called when:
323 * - an active object is initialized
325 static int work_fixup_init(void *addr, enum debug_obj_state state)
327 struct work_struct *work = addr;
330 case ODEBUG_STATE_ACTIVE:
331 cancel_work_sync(work);
332 debug_object_init(work, &work_debug_descr);
340 * fixup_activate is called when:
341 * - an active object is activated
342 * - an unknown object is activated (might be a statically initialized object)
344 static int work_fixup_activate(void *addr, enum debug_obj_state state)
346 struct work_struct *work = addr;
350 case ODEBUG_STATE_NOTAVAILABLE:
352 * This is not really a fixup. The work struct was
353 * statically initialized. We just make sure that it
354 * is tracked in the object tracker.
356 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
357 debug_object_init(work, &work_debug_descr);
358 debug_object_activate(work, &work_debug_descr);
364 case ODEBUG_STATE_ACTIVE:
373 * fixup_free is called when:
374 * - an active object is freed
376 static int work_fixup_free(void *addr, enum debug_obj_state state)
378 struct work_struct *work = addr;
381 case ODEBUG_STATE_ACTIVE:
382 cancel_work_sync(work);
383 debug_object_free(work, &work_debug_descr);
390 static struct debug_obj_descr work_debug_descr = {
391 .name = "work_struct",
392 .debug_hint = work_debug_hint,
393 .fixup_init = work_fixup_init,
394 .fixup_activate = work_fixup_activate,
395 .fixup_free = work_fixup_free,
398 static inline void debug_work_activate(struct work_struct *work)
400 debug_object_activate(work, &work_debug_descr);
403 static inline void debug_work_deactivate(struct work_struct *work)
405 debug_object_deactivate(work, &work_debug_descr);
408 void __init_work(struct work_struct *work, int onstack)
411 debug_object_init_on_stack(work, &work_debug_descr);
413 debug_object_init(work, &work_debug_descr);
415 EXPORT_SYMBOL_GPL(__init_work);
417 void destroy_work_on_stack(struct work_struct *work)
419 debug_object_free(work, &work_debug_descr);
421 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
424 static inline void debug_work_activate(struct work_struct *work) { }
425 static inline void debug_work_deactivate(struct work_struct *work) { }
428 /* Serializes the accesses to the list of workqueues. */
429 static DEFINE_SPINLOCK(workqueue_lock);
430 static LIST_HEAD(workqueues);
431 static bool workqueue_freezing; /* W: have wqs started freezing? */
434 * The almighty global cpu workqueues. nr_running is the only field
435 * which is expected to be used frequently by other cpus via
436 * try_to_wake_up(). Put it in a separate cacheline.
438 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
439 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]);
442 * Global cpu workqueue and nr_running counter for unbound gcwq. The
443 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
444 * workers have WORKER_UNBOUND set.
446 static struct global_cwq unbound_global_cwq;
447 static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = {
448 [0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */
451 static int worker_thread(void *__worker);
452 static unsigned int work_cpu(struct work_struct *work);
454 static int worker_pool_pri(struct worker_pool *pool)
456 return pool - pool->gcwq->pools;
459 static struct global_cwq *get_gcwq(unsigned int cpu)
461 if (cpu != WORK_CPU_UNBOUND)
462 return &per_cpu(global_cwq, cpu);
464 return &unbound_global_cwq;
467 static atomic_t *get_pool_nr_running(struct worker_pool *pool)
469 int cpu = pool->gcwq->cpu;
470 int idx = worker_pool_pri(pool);
472 if (cpu != WORK_CPU_UNBOUND)
473 return &per_cpu(pool_nr_running, cpu)[idx];
475 return &unbound_pool_nr_running[idx];
478 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
479 struct workqueue_struct *wq)
481 if (!(wq->flags & WQ_UNBOUND)) {
482 if (likely(cpu < nr_cpu_ids))
483 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
484 } else if (likely(cpu == WORK_CPU_UNBOUND))
485 return wq->cpu_wq.single;
489 static unsigned int work_color_to_flags(int color)
491 return color << WORK_STRUCT_COLOR_SHIFT;
494 static int get_work_color(struct work_struct *work)
496 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
497 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
500 static int work_next_color(int color)
502 return (color + 1) % WORK_NR_COLORS;
506 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
507 * contain the pointer to the queued cwq. Once execution starts, the flag
508 * is cleared and the high bits contain OFFQ flags and CPU number.
510 * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling()
511 * and clear_work_data() can be used to set the cwq, cpu or clear
512 * work->data. These functions should only be called while the work is
513 * owned - ie. while the PENDING bit is set.
515 * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to
516 * a work. gcwq is available once the work has been queued anywhere after
517 * initialization until it is sync canceled. cwq is available only while
518 * the work item is queued.
520 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
521 * canceled. While being canceled, a work item may have its PENDING set
522 * but stay off timer and worklist for arbitrarily long and nobody should
523 * try to steal the PENDING bit.
525 static inline void set_work_data(struct work_struct *work, unsigned long data,
528 BUG_ON(!work_pending(work));
529 atomic_long_set(&work->data, data | flags | work_static(work));
532 static void set_work_cwq(struct work_struct *work,
533 struct cpu_workqueue_struct *cwq,
534 unsigned long extra_flags)
536 set_work_data(work, (unsigned long)cwq,
537 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
540 static void set_work_cpu_and_clear_pending(struct work_struct *work,
544 * The following wmb is paired with the implied mb in
545 * test_and_set_bit(PENDING) and ensures all updates to @work made
546 * here are visible to and precede any updates by the next PENDING
550 set_work_data(work, (unsigned long)cpu << WORK_OFFQ_CPU_SHIFT, 0);
553 static void clear_work_data(struct work_struct *work)
555 smp_wmb(); /* see set_work_cpu_and_clear_pending() */
556 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
559 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
561 unsigned long data = atomic_long_read(&work->data);
563 if (data & WORK_STRUCT_CWQ)
564 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
569 static struct global_cwq *get_work_gcwq(struct work_struct *work)
571 unsigned long data = atomic_long_read(&work->data);
574 if (data & WORK_STRUCT_CWQ)
575 return ((struct cpu_workqueue_struct *)
576 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq;
578 cpu = data >> WORK_OFFQ_CPU_SHIFT;
579 if (cpu == WORK_CPU_NONE)
582 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
583 return get_gcwq(cpu);
586 static void mark_work_canceling(struct work_struct *work)
588 struct global_cwq *gcwq = get_work_gcwq(work);
589 unsigned long cpu = gcwq ? gcwq->cpu : WORK_CPU_NONE;
591 set_work_data(work, (cpu << WORK_OFFQ_CPU_SHIFT) | WORK_OFFQ_CANCELING,
592 WORK_STRUCT_PENDING);
595 static bool work_is_canceling(struct work_struct *work)
597 unsigned long data = atomic_long_read(&work->data);
599 return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING);
603 * Policy functions. These define the policies on how the global worker
604 * pools are managed. Unless noted otherwise, these functions assume that
605 * they're being called with gcwq->lock held.
608 static bool __need_more_worker(struct worker_pool *pool)
610 return !atomic_read(get_pool_nr_running(pool));
614 * Need to wake up a worker? Called from anything but currently
617 * Note that, because unbound workers never contribute to nr_running, this
618 * function will always return %true for unbound gcwq as long as the
619 * worklist isn't empty.
621 static bool need_more_worker(struct worker_pool *pool)
623 return !list_empty(&pool->worklist) && __need_more_worker(pool);
626 /* Can I start working? Called from busy but !running workers. */
627 static bool may_start_working(struct worker_pool *pool)
629 return pool->nr_idle;
632 /* Do I need to keep working? Called from currently running workers. */
633 static bool keep_working(struct worker_pool *pool)
635 atomic_t *nr_running = get_pool_nr_running(pool);
637 return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
640 /* Do we need a new worker? Called from manager. */
641 static bool need_to_create_worker(struct worker_pool *pool)
643 return need_more_worker(pool) && !may_start_working(pool);
646 /* Do I need to be the manager? */
647 static bool need_to_manage_workers(struct worker_pool *pool)
649 return need_to_create_worker(pool) ||
650 (pool->flags & POOL_MANAGE_WORKERS);
653 /* Do we have too many workers and should some go away? */
654 static bool too_many_workers(struct worker_pool *pool)
656 bool managing = pool->flags & POOL_MANAGING_WORKERS;
657 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
658 int nr_busy = pool->nr_workers - nr_idle;
661 * nr_idle and idle_list may disagree if idle rebinding is in
662 * progress. Never return %true if idle_list is empty.
664 if (list_empty(&pool->idle_list))
667 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
674 /* Return the first worker. Safe with preemption disabled */
675 static struct worker *first_worker(struct worker_pool *pool)
677 if (unlikely(list_empty(&pool->idle_list)))
680 return list_first_entry(&pool->idle_list, struct worker, entry);
684 * wake_up_worker - wake up an idle worker
685 * @pool: worker pool to wake worker from
687 * Wake up the first idle worker of @pool.
690 * spin_lock_irq(gcwq->lock).
692 static void wake_up_worker(struct worker_pool *pool)
694 struct worker *worker = first_worker(pool);
697 wake_up_process(worker->task);
701 * wq_worker_waking_up - a worker is waking up
702 * @task: task waking up
703 * @cpu: CPU @task is waking up to
705 * This function is called during try_to_wake_up() when a worker is
709 * spin_lock_irq(rq->lock)
711 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
713 struct worker *worker = kthread_data(task);
715 if (!(worker->flags & WORKER_NOT_RUNNING)) {
716 WARN_ON_ONCE(worker->pool->gcwq->cpu != cpu);
717 atomic_inc(get_pool_nr_running(worker->pool));
722 * wq_worker_sleeping - a worker is going to sleep
723 * @task: task going to sleep
724 * @cpu: CPU in question, must be the current CPU number
726 * This function is called during schedule() when a busy worker is
727 * going to sleep. Worker on the same cpu can be woken up by
728 * returning pointer to its task.
731 * spin_lock_irq(rq->lock)
734 * Worker task on @cpu to wake up, %NULL if none.
736 struct task_struct *wq_worker_sleeping(struct task_struct *task,
739 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
740 struct worker_pool *pool;
741 atomic_t *nr_running;
744 * Rescuers, which may not have all the fields set up like normal
745 * workers, also reach here, let's not access anything before
746 * checking NOT_RUNNING.
748 if (worker->flags & WORKER_NOT_RUNNING)
752 nr_running = get_pool_nr_running(pool);
754 /* this can only happen on the local cpu */
755 BUG_ON(cpu != raw_smp_processor_id());
758 * The counterpart of the following dec_and_test, implied mb,
759 * worklist not empty test sequence is in insert_work().
760 * Please read comment there.
762 * NOT_RUNNING is clear. This means that we're bound to and
763 * running on the local cpu w/ rq lock held and preemption
764 * disabled, which in turn means that none else could be
765 * manipulating idle_list, so dereferencing idle_list without gcwq
768 if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
769 to_wakeup = first_worker(pool);
770 return to_wakeup ? to_wakeup->task : NULL;
774 * worker_set_flags - set worker flags and adjust nr_running accordingly
776 * @flags: flags to set
777 * @wakeup: wakeup an idle worker if necessary
779 * Set @flags in @worker->flags and adjust nr_running accordingly. If
780 * nr_running becomes zero and @wakeup is %true, an idle worker is
784 * spin_lock_irq(gcwq->lock)
786 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
789 struct worker_pool *pool = worker->pool;
791 WARN_ON_ONCE(worker->task != current);
794 * If transitioning into NOT_RUNNING, adjust nr_running and
795 * wake up an idle worker as necessary if requested by
798 if ((flags & WORKER_NOT_RUNNING) &&
799 !(worker->flags & WORKER_NOT_RUNNING)) {
800 atomic_t *nr_running = get_pool_nr_running(pool);
803 if (atomic_dec_and_test(nr_running) &&
804 !list_empty(&pool->worklist))
805 wake_up_worker(pool);
807 atomic_dec(nr_running);
810 worker->flags |= flags;
814 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
816 * @flags: flags to clear
818 * Clear @flags in @worker->flags and adjust nr_running accordingly.
821 * spin_lock_irq(gcwq->lock)
823 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
825 struct worker_pool *pool = worker->pool;
826 unsigned int oflags = worker->flags;
828 WARN_ON_ONCE(worker->task != current);
830 worker->flags &= ~flags;
833 * If transitioning out of NOT_RUNNING, increment nr_running. Note
834 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
835 * of multiple flags, not a single flag.
837 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
838 if (!(worker->flags & WORKER_NOT_RUNNING))
839 atomic_inc(get_pool_nr_running(pool));
843 * find_worker_executing_work - find worker which is executing a work
844 * @gcwq: gcwq of interest
845 * @work: work to find worker for
847 * Find a worker which is executing @work on @gcwq by searching
848 * @gcwq->busy_hash which is keyed by the address of @work. For a worker
849 * to match, its current execution should match the address of @work and
850 * its work function. This is to avoid unwanted dependency between
851 * unrelated work executions through a work item being recycled while still
854 * This is a bit tricky. A work item may be freed once its execution
855 * starts and nothing prevents the freed area from being recycled for
856 * another work item. If the same work item address ends up being reused
857 * before the original execution finishes, workqueue will identify the
858 * recycled work item as currently executing and make it wait until the
859 * current execution finishes, introducing an unwanted dependency.
861 * This function checks the work item address, work function and workqueue
862 * to avoid false positives. Note that this isn't complete as one may
863 * construct a work function which can introduce dependency onto itself
864 * through a recycled work item. Well, if somebody wants to shoot oneself
865 * in the foot that badly, there's only so much we can do, and if such
866 * deadlock actually occurs, it should be easy to locate the culprit work
870 * spin_lock_irq(gcwq->lock).
873 * Pointer to worker which is executing @work if found, NULL
876 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
877 struct work_struct *work)
879 struct worker *worker;
880 struct hlist_node *tmp;
882 hash_for_each_possible(gcwq->busy_hash, worker, tmp, hentry,
884 if (worker->current_work == work &&
885 worker->current_func == work->func)
892 * move_linked_works - move linked works to a list
893 * @work: start of series of works to be scheduled
894 * @head: target list to append @work to
895 * @nextp: out paramter for nested worklist walking
897 * Schedule linked works starting from @work to @head. Work series to
898 * be scheduled starts at @work and includes any consecutive work with
899 * WORK_STRUCT_LINKED set in its predecessor.
901 * If @nextp is not NULL, it's updated to point to the next work of
902 * the last scheduled work. This allows move_linked_works() to be
903 * nested inside outer list_for_each_entry_safe().
906 * spin_lock_irq(gcwq->lock).
908 static void move_linked_works(struct work_struct *work, struct list_head *head,
909 struct work_struct **nextp)
911 struct work_struct *n;
914 * Linked worklist will always end before the end of the list,
915 * use NULL for list head.
917 list_for_each_entry_safe_from(work, n, NULL, entry) {
918 list_move_tail(&work->entry, head);
919 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
924 * If we're already inside safe list traversal and have moved
925 * multiple works to the scheduled queue, the next position
926 * needs to be updated.
932 static void cwq_activate_delayed_work(struct work_struct *work)
934 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
936 trace_workqueue_activate_work(work);
937 move_linked_works(work, &cwq->pool->worklist, NULL);
938 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
942 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
944 struct work_struct *work = list_first_entry(&cwq->delayed_works,
945 struct work_struct, entry);
947 cwq_activate_delayed_work(work);
951 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
952 * @cwq: cwq of interest
953 * @color: color of work which left the queue
955 * A work either has completed or is removed from pending queue,
956 * decrement nr_in_flight of its cwq and handle workqueue flushing.
959 * spin_lock_irq(gcwq->lock).
961 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
963 /* ignore uncolored works */
964 if (color == WORK_NO_COLOR)
967 cwq->nr_in_flight[color]--;
970 if (!list_empty(&cwq->delayed_works)) {
971 /* one down, submit a delayed one */
972 if (cwq->nr_active < cwq->max_active)
973 cwq_activate_first_delayed(cwq);
976 /* is flush in progress and are we at the flushing tip? */
977 if (likely(cwq->flush_color != color))
980 /* are there still in-flight works? */
981 if (cwq->nr_in_flight[color])
984 /* this cwq is done, clear flush_color */
985 cwq->flush_color = -1;
988 * If this was the last cwq, wake up the first flusher. It
989 * will handle the rest.
991 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
992 complete(&cwq->wq->first_flusher->done);
996 * try_to_grab_pending - steal work item from worklist and disable irq
997 * @work: work item to steal
998 * @is_dwork: @work is a delayed_work
999 * @flags: place to store irq state
1001 * Try to grab PENDING bit of @work. This function can handle @work in any
1002 * stable state - idle, on timer or on worklist. Return values are
1004 * 1 if @work was pending and we successfully stole PENDING
1005 * 0 if @work was idle and we claimed PENDING
1006 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1007 * -ENOENT if someone else is canceling @work, this state may persist
1008 * for arbitrarily long
1010 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1011 * interrupted while holding PENDING and @work off queue, irq must be
1012 * disabled on entry. This, combined with delayed_work->timer being
1013 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1015 * On successful return, >= 0, irq is disabled and the caller is
1016 * responsible for releasing it using local_irq_restore(*@flags).
1018 * This function is safe to call from any context including IRQ handler.
1020 static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1021 unsigned long *flags)
1023 struct global_cwq *gcwq;
1025 local_irq_save(*flags);
1027 /* try to steal the timer if it exists */
1029 struct delayed_work *dwork = to_delayed_work(work);
1032 * dwork->timer is irqsafe. If del_timer() fails, it's
1033 * guaranteed that the timer is not queued anywhere and not
1034 * running on the local CPU.
1036 if (likely(del_timer(&dwork->timer)))
1040 /* try to claim PENDING the normal way */
1041 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1045 * The queueing is in progress, or it is already queued. Try to
1046 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1048 gcwq = get_work_gcwq(work);
1052 spin_lock(&gcwq->lock);
1053 if (!list_empty(&work->entry)) {
1055 * This work is queued, but perhaps we locked the wrong gcwq.
1056 * In that case we must see the new value after rmb(), see
1057 * insert_work()->wmb().
1060 if (gcwq == get_work_gcwq(work)) {
1061 debug_work_deactivate(work);
1064 * A delayed work item cannot be grabbed directly
1065 * because it might have linked NO_COLOR work items
1066 * which, if left on the delayed_list, will confuse
1067 * cwq->nr_active management later on and cause
1068 * stall. Make sure the work item is activated
1071 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1072 cwq_activate_delayed_work(work);
1074 list_del_init(&work->entry);
1075 cwq_dec_nr_in_flight(get_work_cwq(work),
1076 get_work_color(work));
1078 spin_unlock(&gcwq->lock);
1082 spin_unlock(&gcwq->lock);
1084 local_irq_restore(*flags);
1085 if (work_is_canceling(work))
1092 * insert_work - insert a work into gcwq
1093 * @cwq: cwq @work belongs to
1094 * @work: work to insert
1095 * @head: insertion point
1096 * @extra_flags: extra WORK_STRUCT_* flags to set
1098 * Insert @work which belongs to @cwq into @gcwq after @head.
1099 * @extra_flags is or'd to work_struct flags.
1102 * spin_lock_irq(gcwq->lock).
1104 static void insert_work(struct cpu_workqueue_struct *cwq,
1105 struct work_struct *work, struct list_head *head,
1106 unsigned int extra_flags)
1108 struct worker_pool *pool = cwq->pool;
1110 /* we own @work, set data and link */
1111 set_work_cwq(work, cwq, extra_flags);
1114 * Ensure that we get the right work->data if we see the
1115 * result of list_add() below, see try_to_grab_pending().
1119 list_add_tail(&work->entry, head);
1122 * Ensure either worker_sched_deactivated() sees the above
1123 * list_add_tail() or we see zero nr_running to avoid workers
1124 * lying around lazily while there are works to be processed.
1128 if (__need_more_worker(pool))
1129 wake_up_worker(pool);
1133 * Test whether @work is being queued from another work executing on the
1134 * same workqueue. This is rather expensive and should only be used from
1137 static bool is_chained_work(struct workqueue_struct *wq)
1139 unsigned long flags;
1142 for_each_gcwq_cpu(cpu) {
1143 struct global_cwq *gcwq = get_gcwq(cpu);
1144 struct worker *worker;
1145 struct hlist_node *pos;
1148 spin_lock_irqsave(&gcwq->lock, flags);
1149 for_each_busy_worker(worker, i, pos, gcwq) {
1150 if (worker->task != current)
1152 spin_unlock_irqrestore(&gcwq->lock, flags);
1154 * I'm @worker, no locking necessary. See if @work
1155 * is headed to the same workqueue.
1157 return worker->current_cwq->wq == wq;
1159 spin_unlock_irqrestore(&gcwq->lock, flags);
1164 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1165 struct work_struct *work)
1167 struct global_cwq *gcwq;
1168 struct cpu_workqueue_struct *cwq;
1169 struct list_head *worklist;
1170 unsigned int work_flags;
1171 unsigned int req_cpu = cpu;
1174 * While a work item is PENDING && off queue, a task trying to
1175 * steal the PENDING will busy-loop waiting for it to either get
1176 * queued or lose PENDING. Grabbing PENDING and queueing should
1177 * happen with IRQ disabled.
1179 WARN_ON_ONCE(!irqs_disabled());
1181 debug_work_activate(work);
1183 /* if dying, only works from the same workqueue are allowed */
1184 if (unlikely(wq->flags & WQ_DRAINING) &&
1185 WARN_ON_ONCE(!is_chained_work(wq)))
1188 /* determine gcwq to use */
1189 if (!(wq->flags & WQ_UNBOUND)) {
1190 struct global_cwq *last_gcwq;
1192 if (cpu == WORK_CPU_UNBOUND)
1193 cpu = raw_smp_processor_id();
1196 * It's multi cpu. If @work was previously on a different
1197 * cpu, it might still be running there, in which case the
1198 * work needs to be queued on that cpu to guarantee
1201 gcwq = get_gcwq(cpu);
1202 last_gcwq = get_work_gcwq(work);
1204 if (last_gcwq && last_gcwq != gcwq) {
1205 struct worker *worker;
1207 spin_lock(&last_gcwq->lock);
1209 worker = find_worker_executing_work(last_gcwq, work);
1211 if (worker && worker->current_cwq->wq == wq)
1214 /* meh... not running there, queue here */
1215 spin_unlock(&last_gcwq->lock);
1216 spin_lock(&gcwq->lock);
1219 spin_lock(&gcwq->lock);
1222 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1223 spin_lock(&gcwq->lock);
1226 /* gcwq determined, get cwq and queue */
1227 cwq = get_cwq(gcwq->cpu, wq);
1228 trace_workqueue_queue_work(req_cpu, cwq, work);
1230 if (WARN_ON(!list_empty(&work->entry))) {
1231 spin_unlock(&gcwq->lock);
1235 cwq->nr_in_flight[cwq->work_color]++;
1236 work_flags = work_color_to_flags(cwq->work_color);
1238 if (likely(cwq->nr_active < cwq->max_active)) {
1239 trace_workqueue_activate_work(work);
1241 worklist = &cwq->pool->worklist;
1243 work_flags |= WORK_STRUCT_DELAYED;
1244 worklist = &cwq->delayed_works;
1247 insert_work(cwq, work, worklist, work_flags);
1249 spin_unlock(&gcwq->lock);
1253 * queue_work_on - queue work on specific cpu
1254 * @cpu: CPU number to execute work on
1255 * @wq: workqueue to use
1256 * @work: work to queue
1258 * Returns %false if @work was already on a queue, %true otherwise.
1260 * We queue the work to a specific CPU, the caller must ensure it
1263 bool queue_work_on(int cpu, struct workqueue_struct *wq,
1264 struct work_struct *work)
1267 unsigned long flags;
1269 local_irq_save(flags);
1271 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1272 __queue_work(cpu, wq, work);
1276 local_irq_restore(flags);
1279 EXPORT_SYMBOL_GPL(queue_work_on);
1282 * queue_work - queue work on a workqueue
1283 * @wq: workqueue to use
1284 * @work: work to queue
1286 * Returns %false if @work was already on a queue, %true otherwise.
1288 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1289 * it can be processed by another CPU.
1291 bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1293 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1295 EXPORT_SYMBOL_GPL(queue_work);
1297 void delayed_work_timer_fn(unsigned long __data)
1299 struct delayed_work *dwork = (struct delayed_work *)__data;
1300 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1302 /* should have been called from irqsafe timer with irq already off */
1303 __queue_work(dwork->cpu, cwq->wq, &dwork->work);
1305 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1307 static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1308 struct delayed_work *dwork, unsigned long delay)
1310 struct timer_list *timer = &dwork->timer;
1311 struct work_struct *work = &dwork->work;
1314 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1315 timer->data != (unsigned long)dwork);
1316 WARN_ON_ONCE(timer_pending(timer));
1317 WARN_ON_ONCE(!list_empty(&work->entry));
1320 * If @delay is 0, queue @dwork->work immediately. This is for
1321 * both optimization and correctness. The earliest @timer can
1322 * expire is on the closest next tick and delayed_work users depend
1323 * on that there's no such delay when @delay is 0.
1326 __queue_work(cpu, wq, &dwork->work);
1330 timer_stats_timer_set_start_info(&dwork->timer);
1333 * This stores cwq for the moment, for the timer_fn. Note that the
1334 * work's gcwq is preserved to allow reentrance detection for
1337 if (!(wq->flags & WQ_UNBOUND)) {
1338 struct global_cwq *gcwq = get_work_gcwq(work);
1341 * If we cannot get the last gcwq from @work directly,
1342 * select the last CPU such that it avoids unnecessarily
1343 * triggering non-reentrancy check in __queue_work().
1348 if (lcpu == WORK_CPU_UNBOUND)
1349 lcpu = raw_smp_processor_id();
1351 lcpu = WORK_CPU_UNBOUND;
1354 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1357 timer->expires = jiffies + delay;
1359 if (unlikely(cpu != WORK_CPU_UNBOUND))
1360 add_timer_on(timer, cpu);
1366 * queue_delayed_work_on - queue work on specific CPU after delay
1367 * @cpu: CPU number to execute work on
1368 * @wq: workqueue to use
1369 * @dwork: work to queue
1370 * @delay: number of jiffies to wait before queueing
1372 * Returns %false if @work was already on a queue, %true otherwise. If
1373 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1376 bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1377 struct delayed_work *dwork, unsigned long delay)
1379 struct work_struct *work = &dwork->work;
1381 unsigned long flags;
1383 /* read the comment in __queue_work() */
1384 local_irq_save(flags);
1386 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1387 __queue_delayed_work(cpu, wq, dwork, delay);
1391 local_irq_restore(flags);
1394 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1397 * queue_delayed_work - queue work on a workqueue after delay
1398 * @wq: workqueue to use
1399 * @dwork: delayable work to queue
1400 * @delay: number of jiffies to wait before queueing
1402 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1404 bool queue_delayed_work(struct workqueue_struct *wq,
1405 struct delayed_work *dwork, unsigned long delay)
1407 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1409 EXPORT_SYMBOL_GPL(queue_delayed_work);
1412 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1413 * @cpu: CPU number to execute work on
1414 * @wq: workqueue to use
1415 * @dwork: work to queue
1416 * @delay: number of jiffies to wait before queueing
1418 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1419 * modify @dwork's timer so that it expires after @delay. If @delay is
1420 * zero, @work is guaranteed to be scheduled immediately regardless of its
1423 * Returns %false if @dwork was idle and queued, %true if @dwork was
1424 * pending and its timer was modified.
1426 * This function is safe to call from any context including IRQ handler.
1427 * See try_to_grab_pending() for details.
1429 bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1430 struct delayed_work *dwork, unsigned long delay)
1432 unsigned long flags;
1436 ret = try_to_grab_pending(&dwork->work, true, &flags);
1437 } while (unlikely(ret == -EAGAIN));
1439 if (likely(ret >= 0)) {
1440 __queue_delayed_work(cpu, wq, dwork, delay);
1441 local_irq_restore(flags);
1444 /* -ENOENT from try_to_grab_pending() becomes %true */
1447 EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1450 * mod_delayed_work - modify delay of or queue a delayed work
1451 * @wq: workqueue to use
1452 * @dwork: work to queue
1453 * @delay: number of jiffies to wait before queueing
1455 * mod_delayed_work_on() on local CPU.
1457 bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1458 unsigned long delay)
1460 return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1462 EXPORT_SYMBOL_GPL(mod_delayed_work);
1465 * worker_enter_idle - enter idle state
1466 * @worker: worker which is entering idle state
1468 * @worker is entering idle state. Update stats and idle timer if
1472 * spin_lock_irq(gcwq->lock).
1474 static void worker_enter_idle(struct worker *worker)
1476 struct worker_pool *pool = worker->pool;
1477 struct global_cwq *gcwq = pool->gcwq;
1479 BUG_ON(worker->flags & WORKER_IDLE);
1480 BUG_ON(!list_empty(&worker->entry) &&
1481 (worker->hentry.next || worker->hentry.pprev));
1483 /* can't use worker_set_flags(), also called from start_worker() */
1484 worker->flags |= WORKER_IDLE;
1486 worker->last_active = jiffies;
1488 /* idle_list is LIFO */
1489 list_add(&worker->entry, &pool->idle_list);
1491 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1492 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1495 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1496 * gcwq->lock between setting %WORKER_UNBOUND and zapping
1497 * nr_running, the warning may trigger spuriously. Check iff
1498 * unbind is not in progress.
1500 WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
1501 pool->nr_workers == pool->nr_idle &&
1502 atomic_read(get_pool_nr_running(pool)));
1506 * worker_leave_idle - leave idle state
1507 * @worker: worker which is leaving idle state
1509 * @worker is leaving idle state. Update stats.
1512 * spin_lock_irq(gcwq->lock).
1514 static void worker_leave_idle(struct worker *worker)
1516 struct worker_pool *pool = worker->pool;
1518 BUG_ON(!(worker->flags & WORKER_IDLE));
1519 worker_clr_flags(worker, WORKER_IDLE);
1521 list_del_init(&worker->entry);
1525 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1528 * Works which are scheduled while the cpu is online must at least be
1529 * scheduled to a worker which is bound to the cpu so that if they are
1530 * flushed from cpu callbacks while cpu is going down, they are
1531 * guaranteed to execute on the cpu.
1533 * This function is to be used by rogue workers and rescuers to bind
1534 * themselves to the target cpu and may race with cpu going down or
1535 * coming online. kthread_bind() can't be used because it may put the
1536 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1537 * verbatim as it's best effort and blocking and gcwq may be
1538 * [dis]associated in the meantime.
1540 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1541 * binding against %GCWQ_DISASSOCIATED which is set during
1542 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1543 * enters idle state or fetches works without dropping lock, it can
1544 * guarantee the scheduling requirement described in the first paragraph.
1547 * Might sleep. Called without any lock but returns with gcwq->lock
1551 * %true if the associated gcwq is online (@worker is successfully
1552 * bound), %false if offline.
1554 static bool worker_maybe_bind_and_lock(struct worker *worker)
1555 __acquires(&gcwq->lock)
1557 struct global_cwq *gcwq = worker->pool->gcwq;
1558 struct task_struct *task = worker->task;
1562 * The following call may fail, succeed or succeed
1563 * without actually migrating the task to the cpu if
1564 * it races with cpu hotunplug operation. Verify
1565 * against GCWQ_DISASSOCIATED.
1567 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1568 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1570 spin_lock_irq(&gcwq->lock);
1571 if (gcwq->flags & GCWQ_DISASSOCIATED)
1573 if (task_cpu(task) == gcwq->cpu &&
1574 cpumask_equal(¤t->cpus_allowed,
1575 get_cpu_mask(gcwq->cpu)))
1577 spin_unlock_irq(&gcwq->lock);
1580 * We've raced with CPU hot[un]plug. Give it a breather
1581 * and retry migration. cond_resched() is required here;
1582 * otherwise, we might deadlock against cpu_stop trying to
1583 * bring down the CPU on non-preemptive kernel.
1591 * Rebind an idle @worker to its CPU. worker_thread() will test
1592 * list_empty(@worker->entry) before leaving idle and call this function.
1594 static void idle_worker_rebind(struct worker *worker)
1596 struct global_cwq *gcwq = worker->pool->gcwq;
1598 /* CPU may go down again inbetween, clear UNBOUND only on success */
1599 if (worker_maybe_bind_and_lock(worker))
1600 worker_clr_flags(worker, WORKER_UNBOUND);
1602 /* rebind complete, become available again */
1603 list_add(&worker->entry, &worker->pool->idle_list);
1604 spin_unlock_irq(&gcwq->lock);
1608 * Function for @worker->rebind.work used to rebind unbound busy workers to
1609 * the associated cpu which is coming back online. This is scheduled by
1610 * cpu up but can race with other cpu hotplug operations and may be
1611 * executed twice without intervening cpu down.
1613 static void busy_worker_rebind_fn(struct work_struct *work)
1615 struct worker *worker = container_of(work, struct worker, rebind_work);
1616 struct global_cwq *gcwq = worker->pool->gcwq;
1618 if (worker_maybe_bind_and_lock(worker))
1619 worker_clr_flags(worker, WORKER_UNBOUND);
1621 spin_unlock_irq(&gcwq->lock);
1625 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1626 * @gcwq: gcwq of interest
1628 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1629 * is different for idle and busy ones.
1631 * Idle ones will be removed from the idle_list and woken up. They will
1632 * add themselves back after completing rebind. This ensures that the
1633 * idle_list doesn't contain any unbound workers when re-bound busy workers
1634 * try to perform local wake-ups for concurrency management.
1636 * Busy workers can rebind after they finish their current work items.
1637 * Queueing the rebind work item at the head of the scheduled list is
1638 * enough. Note that nr_running will be properly bumped as busy workers
1641 * On return, all non-manager workers are scheduled for rebind - see
1642 * manage_workers() for the manager special case. Any idle worker
1643 * including the manager will not appear on @idle_list until rebind is
1644 * complete, making local wake-ups safe.
1646 static void rebind_workers(struct global_cwq *gcwq)
1648 struct worker_pool *pool;
1649 struct worker *worker, *n;
1650 struct hlist_node *pos;
1653 lockdep_assert_held(&gcwq->lock);
1655 for_each_worker_pool(pool, gcwq)
1656 lockdep_assert_held(&pool->assoc_mutex);
1658 /* dequeue and kick idle ones */
1659 for_each_worker_pool(pool, gcwq) {
1660 list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1662 * idle workers should be off @pool->idle_list
1663 * until rebind is complete to avoid receiving
1664 * premature local wake-ups.
1666 list_del_init(&worker->entry);
1669 * worker_thread() will see the above dequeuing
1670 * and call idle_worker_rebind().
1672 wake_up_process(worker->task);
1676 /* rebind busy workers */
1677 for_each_busy_worker(worker, i, pos, gcwq) {
1678 struct work_struct *rebind_work = &worker->rebind_work;
1679 struct workqueue_struct *wq;
1681 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1682 work_data_bits(rebind_work)))
1685 debug_work_activate(rebind_work);
1688 * wq doesn't really matter but let's keep @worker->pool
1689 * and @cwq->pool consistent for sanity.
1691 if (worker_pool_pri(worker->pool))
1692 wq = system_highpri_wq;
1696 insert_work(get_cwq(gcwq->cpu, wq), rebind_work,
1697 worker->scheduled.next,
1698 work_color_to_flags(WORK_NO_COLOR));
1702 static struct worker *alloc_worker(void)
1704 struct worker *worker;
1706 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1708 INIT_LIST_HEAD(&worker->entry);
1709 INIT_LIST_HEAD(&worker->scheduled);
1710 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1711 /* on creation a worker is in !idle && prep state */
1712 worker->flags = WORKER_PREP;
1718 * create_worker - create a new workqueue worker
1719 * @pool: pool the new worker will belong to
1721 * Create a new worker which is bound to @pool. The returned worker
1722 * can be started by calling start_worker() or destroyed using
1726 * Might sleep. Does GFP_KERNEL allocations.
1729 * Pointer to the newly created worker.
1731 static struct worker *create_worker(struct worker_pool *pool)
1733 struct global_cwq *gcwq = pool->gcwq;
1734 const char *pri = worker_pool_pri(pool) ? "H" : "";
1735 struct worker *worker = NULL;
1738 spin_lock_irq(&gcwq->lock);
1739 while (ida_get_new(&pool->worker_ida, &id)) {
1740 spin_unlock_irq(&gcwq->lock);
1741 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1743 spin_lock_irq(&gcwq->lock);
1745 spin_unlock_irq(&gcwq->lock);
1747 worker = alloc_worker();
1751 worker->pool = pool;
1754 if (gcwq->cpu != WORK_CPU_UNBOUND)
1755 worker->task = kthread_create_on_node(worker_thread,
1756 worker, cpu_to_node(gcwq->cpu),
1757 "kworker/%u:%d%s", gcwq->cpu, id, pri);
1759 worker->task = kthread_create(worker_thread, worker,
1760 "kworker/u:%d%s", id, pri);
1761 if (IS_ERR(worker->task))
1764 if (worker_pool_pri(pool))
1765 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1768 * Determine CPU binding of the new worker depending on
1769 * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the
1770 * flag remains stable across this function. See the comments
1771 * above the flag definition for details.
1773 * As an unbound worker may later become a regular one if CPU comes
1774 * online, make sure every worker has %PF_THREAD_BOUND set.
1776 if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
1777 kthread_bind(worker->task, gcwq->cpu);
1779 worker->task->flags |= PF_THREAD_BOUND;
1780 worker->flags |= WORKER_UNBOUND;
1786 spin_lock_irq(&gcwq->lock);
1787 ida_remove(&pool->worker_ida, id);
1788 spin_unlock_irq(&gcwq->lock);
1795 * start_worker - start a newly created worker
1796 * @worker: worker to start
1798 * Make the gcwq aware of @worker and start it.
1801 * spin_lock_irq(gcwq->lock).
1803 static void start_worker(struct worker *worker)
1805 worker->flags |= WORKER_STARTED;
1806 worker->pool->nr_workers++;
1807 worker_enter_idle(worker);
1808 wake_up_process(worker->task);
1812 * destroy_worker - destroy a workqueue worker
1813 * @worker: worker to be destroyed
1815 * Destroy @worker and adjust @gcwq stats accordingly.
1818 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1820 static void destroy_worker(struct worker *worker)
1822 struct worker_pool *pool = worker->pool;
1823 struct global_cwq *gcwq = pool->gcwq;
1824 int id = worker->id;
1826 /* sanity check frenzy */
1827 BUG_ON(worker->current_work);
1828 BUG_ON(!list_empty(&worker->scheduled));
1830 if (worker->flags & WORKER_STARTED)
1832 if (worker->flags & WORKER_IDLE)
1835 list_del_init(&worker->entry);
1836 worker->flags |= WORKER_DIE;
1838 spin_unlock_irq(&gcwq->lock);
1840 kthread_stop(worker->task);
1843 spin_lock_irq(&gcwq->lock);
1844 ida_remove(&pool->worker_ida, id);
1847 static void idle_worker_timeout(unsigned long __pool)
1849 struct worker_pool *pool = (void *)__pool;
1850 struct global_cwq *gcwq = pool->gcwq;
1852 spin_lock_irq(&gcwq->lock);
1854 if (too_many_workers(pool)) {
1855 struct worker *worker;
1856 unsigned long expires;
1858 /* idle_list is kept in LIFO order, check the last one */
1859 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1860 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1862 if (time_before(jiffies, expires))
1863 mod_timer(&pool->idle_timer, expires);
1865 /* it's been idle for too long, wake up manager */
1866 pool->flags |= POOL_MANAGE_WORKERS;
1867 wake_up_worker(pool);
1871 spin_unlock_irq(&gcwq->lock);
1874 static bool send_mayday(struct work_struct *work)
1876 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1877 struct workqueue_struct *wq = cwq->wq;
1880 if (!(wq->flags & WQ_RESCUER))
1883 /* mayday mayday mayday */
1884 cpu = cwq->pool->gcwq->cpu;
1885 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1886 if (cpu == WORK_CPU_UNBOUND)
1888 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1889 wake_up_process(wq->rescuer->task);
1893 static void gcwq_mayday_timeout(unsigned long __pool)
1895 struct worker_pool *pool = (void *)__pool;
1896 struct global_cwq *gcwq = pool->gcwq;
1897 struct work_struct *work;
1899 spin_lock_irq(&gcwq->lock);
1901 if (need_to_create_worker(pool)) {
1903 * We've been trying to create a new worker but
1904 * haven't been successful. We might be hitting an
1905 * allocation deadlock. Send distress signals to
1908 list_for_each_entry(work, &pool->worklist, entry)
1912 spin_unlock_irq(&gcwq->lock);
1914 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1918 * maybe_create_worker - create a new worker if necessary
1919 * @pool: pool to create a new worker for
1921 * Create a new worker for @pool if necessary. @pool is guaranteed to
1922 * have at least one idle worker on return from this function. If
1923 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1924 * sent to all rescuers with works scheduled on @pool to resolve
1925 * possible allocation deadlock.
1927 * On return, need_to_create_worker() is guaranteed to be false and
1928 * may_start_working() true.
1931 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1932 * multiple times. Does GFP_KERNEL allocations. Called only from
1936 * false if no action was taken and gcwq->lock stayed locked, true
1939 static bool maybe_create_worker(struct worker_pool *pool)
1940 __releases(&gcwq->lock)
1941 __acquires(&gcwq->lock)
1943 struct global_cwq *gcwq = pool->gcwq;
1945 if (!need_to_create_worker(pool))
1948 spin_unlock_irq(&gcwq->lock);
1950 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1951 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1954 struct worker *worker;
1956 worker = create_worker(pool);
1958 del_timer_sync(&pool->mayday_timer);
1959 spin_lock_irq(&gcwq->lock);
1960 start_worker(worker);
1961 BUG_ON(need_to_create_worker(pool));
1965 if (!need_to_create_worker(pool))
1968 __set_current_state(TASK_INTERRUPTIBLE);
1969 schedule_timeout(CREATE_COOLDOWN);
1971 if (!need_to_create_worker(pool))
1975 del_timer_sync(&pool->mayday_timer);
1976 spin_lock_irq(&gcwq->lock);
1977 if (need_to_create_worker(pool))
1983 * maybe_destroy_worker - destroy workers which have been idle for a while
1984 * @pool: pool to destroy workers for
1986 * Destroy @pool workers which have been idle for longer than
1987 * IDLE_WORKER_TIMEOUT.
1990 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1991 * multiple times. Called only from manager.
1994 * false if no action was taken and gcwq->lock stayed locked, true
1997 static bool maybe_destroy_workers(struct worker_pool *pool)
2001 while (too_many_workers(pool)) {
2002 struct worker *worker;
2003 unsigned long expires;
2005 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2006 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2008 if (time_before(jiffies, expires)) {
2009 mod_timer(&pool->idle_timer, expires);
2013 destroy_worker(worker);
2021 * manage_workers - manage worker pool
2024 * Assume the manager role and manage gcwq worker pool @worker belongs
2025 * to. At any given time, there can be only zero or one manager per
2026 * gcwq. The exclusion is handled automatically by this function.
2028 * The caller can safely start processing works on false return. On
2029 * true return, it's guaranteed that need_to_create_worker() is false
2030 * and may_start_working() is true.
2033 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2034 * multiple times. Does GFP_KERNEL allocations.
2037 * false if no action was taken and gcwq->lock stayed locked, true if
2038 * some action was taken.
2040 static bool manage_workers(struct worker *worker)
2042 struct worker_pool *pool = worker->pool;
2045 if (pool->flags & POOL_MANAGING_WORKERS)
2048 pool->flags |= POOL_MANAGING_WORKERS;
2051 * To simplify both worker management and CPU hotplug, hold off
2052 * management while hotplug is in progress. CPU hotplug path can't
2053 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2054 * lead to idle worker depletion (all become busy thinking someone
2055 * else is managing) which in turn can result in deadlock under
2056 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2057 * manager against CPU hotplug.
2059 * assoc_mutex would always be free unless CPU hotplug is in
2060 * progress. trylock first without dropping @gcwq->lock.
2062 if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2063 spin_unlock_irq(&pool->gcwq->lock);
2064 mutex_lock(&pool->assoc_mutex);
2066 * CPU hotplug could have happened while we were waiting
2067 * for assoc_mutex. Hotplug itself can't handle us
2068 * because manager isn't either on idle or busy list, and
2069 * @gcwq's state and ours could have deviated.
2071 * As hotplug is now excluded via assoc_mutex, we can
2072 * simply try to bind. It will succeed or fail depending
2073 * on @gcwq's current state. Try it and adjust
2074 * %WORKER_UNBOUND accordingly.
2076 if (worker_maybe_bind_and_lock(worker))
2077 worker->flags &= ~WORKER_UNBOUND;
2079 worker->flags |= WORKER_UNBOUND;
2084 pool->flags &= ~POOL_MANAGE_WORKERS;
2087 * Destroy and then create so that may_start_working() is true
2090 ret |= maybe_destroy_workers(pool);
2091 ret |= maybe_create_worker(pool);
2093 pool->flags &= ~POOL_MANAGING_WORKERS;
2094 mutex_unlock(&pool->assoc_mutex);
2099 * process_one_work - process single work
2101 * @work: work to process
2103 * Process @work. This function contains all the logics necessary to
2104 * process a single work including synchronization against and
2105 * interaction with other workers on the same cpu, queueing and
2106 * flushing. As long as context requirement is met, any worker can
2107 * call this function to process a work.
2110 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2112 static void process_one_work(struct worker *worker, struct work_struct *work)
2113 __releases(&gcwq->lock)
2114 __acquires(&gcwq->lock)
2116 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2117 struct worker_pool *pool = worker->pool;
2118 struct global_cwq *gcwq = pool->gcwq;
2119 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2121 struct worker *collision;
2122 #ifdef CONFIG_LOCKDEP
2124 * It is permissible to free the struct work_struct from
2125 * inside the function that is called from it, this we need to
2126 * take into account for lockdep too. To avoid bogus "held
2127 * lock freed" warnings as well as problems when looking into
2128 * work->lockdep_map, make a copy and use that here.
2130 struct lockdep_map lockdep_map;
2132 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2135 * Ensure we're on the correct CPU. DISASSOCIATED test is
2136 * necessary to avoid spurious warnings from rescuers servicing the
2137 * unbound or a disassociated gcwq.
2139 WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2140 !(gcwq->flags & GCWQ_DISASSOCIATED) &&
2141 raw_smp_processor_id() != gcwq->cpu);
2144 * A single work shouldn't be executed concurrently by
2145 * multiple workers on a single cpu. Check whether anyone is
2146 * already processing the work. If so, defer the work to the
2147 * currently executing one.
2149 collision = find_worker_executing_work(gcwq, work);
2150 if (unlikely(collision)) {
2151 move_linked_works(work, &collision->scheduled, NULL);
2155 /* claim and dequeue */
2156 debug_work_deactivate(work);
2157 hash_add(gcwq->busy_hash, &worker->hentry, (unsigned long)work);
2158 worker->current_work = work;
2159 worker->current_func = work->func;
2160 worker->current_cwq = cwq;
2161 work_color = get_work_color(work);
2163 list_del_init(&work->entry);
2166 * CPU intensive works don't participate in concurrency
2167 * management. They're the scheduler's responsibility.
2169 if (unlikely(cpu_intensive))
2170 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2173 * Unbound gcwq isn't concurrency managed and work items should be
2174 * executed ASAP. Wake up another worker if necessary.
2176 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2177 wake_up_worker(pool);
2180 * Record the last CPU and clear PENDING which should be the last
2181 * update to @work. Also, do this inside @gcwq->lock so that
2182 * PENDING and queued state changes happen together while IRQ is
2185 set_work_cpu_and_clear_pending(work, gcwq->cpu);
2187 spin_unlock_irq(&gcwq->lock);
2189 lock_map_acquire_read(&cwq->wq->lockdep_map);
2190 lock_map_acquire(&lockdep_map);
2191 trace_workqueue_execute_start(work);
2192 worker->current_func(work);
2194 * While we must be careful to not use "work" after this, the trace
2195 * point will only record its address.
2197 trace_workqueue_execute_end(work);
2198 lock_map_release(&lockdep_map);
2199 lock_map_release(&cwq->wq->lockdep_map);
2201 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2202 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2203 " last function: %pf\n",
2204 current->comm, preempt_count(), task_pid_nr(current),
2205 worker->current_func);
2206 debug_show_held_locks(current);
2210 spin_lock_irq(&gcwq->lock);
2212 /* clear cpu intensive status */
2213 if (unlikely(cpu_intensive))
2214 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2216 /* we're done with it, release */
2217 hash_del(&worker->hentry);
2218 worker->current_work = NULL;
2219 worker->current_func = NULL;
2220 worker->current_cwq = NULL;
2221 cwq_dec_nr_in_flight(cwq, work_color);
2225 * process_scheduled_works - process scheduled works
2228 * Process all scheduled works. Please note that the scheduled list
2229 * may change while processing a work, so this function repeatedly
2230 * fetches a work from the top and executes it.
2233 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2236 static void process_scheduled_works(struct worker *worker)
2238 while (!list_empty(&worker->scheduled)) {
2239 struct work_struct *work = list_first_entry(&worker->scheduled,
2240 struct work_struct, entry);
2241 process_one_work(worker, work);
2246 * worker_thread - the worker thread function
2249 * The gcwq worker thread function. There's a single dynamic pool of
2250 * these per each cpu. These workers process all works regardless of
2251 * their specific target workqueue. The only exception is works which
2252 * belong to workqueues with a rescuer which will be explained in
2255 static int worker_thread(void *__worker)
2257 struct worker *worker = __worker;
2258 struct worker_pool *pool = worker->pool;
2259 struct global_cwq *gcwq = pool->gcwq;
2261 /* tell the scheduler that this is a workqueue worker */
2262 worker->task->flags |= PF_WQ_WORKER;
2264 spin_lock_irq(&gcwq->lock);
2266 /* we are off idle list if destruction or rebind is requested */
2267 if (unlikely(list_empty(&worker->entry))) {
2268 spin_unlock_irq(&gcwq->lock);
2270 /* if DIE is set, destruction is requested */
2271 if (worker->flags & WORKER_DIE) {
2272 worker->task->flags &= ~PF_WQ_WORKER;
2276 /* otherwise, rebind */
2277 idle_worker_rebind(worker);
2281 worker_leave_idle(worker);
2283 /* no more worker necessary? */
2284 if (!need_more_worker(pool))
2287 /* do we need to manage? */
2288 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2292 * ->scheduled list can only be filled while a worker is
2293 * preparing to process a work or actually processing it.
2294 * Make sure nobody diddled with it while I was sleeping.
2296 BUG_ON(!list_empty(&worker->scheduled));
2299 * When control reaches this point, we're guaranteed to have
2300 * at least one idle worker or that someone else has already
2301 * assumed the manager role.
2303 worker_clr_flags(worker, WORKER_PREP);
2306 struct work_struct *work =
2307 list_first_entry(&pool->worklist,
2308 struct work_struct, entry);
2310 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2311 /* optimization path, not strictly necessary */
2312 process_one_work(worker, work);
2313 if (unlikely(!list_empty(&worker->scheduled)))
2314 process_scheduled_works(worker);
2316 move_linked_works(work, &worker->scheduled, NULL);
2317 process_scheduled_works(worker);
2319 } while (keep_working(pool));
2321 worker_set_flags(worker, WORKER_PREP, false);
2323 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2327 * gcwq->lock is held and there's no work to process and no
2328 * need to manage, sleep. Workers are woken up only while
2329 * holding gcwq->lock or from local cpu, so setting the
2330 * current state before releasing gcwq->lock is enough to
2331 * prevent losing any event.
2333 worker_enter_idle(worker);
2334 __set_current_state(TASK_INTERRUPTIBLE);
2335 spin_unlock_irq(&gcwq->lock);
2341 * rescuer_thread - the rescuer thread function
2344 * Workqueue rescuer thread function. There's one rescuer for each
2345 * workqueue which has WQ_RESCUER set.
2347 * Regular work processing on a gcwq may block trying to create a new
2348 * worker which uses GFP_KERNEL allocation which has slight chance of
2349 * developing into deadlock if some works currently on the same queue
2350 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2351 * the problem rescuer solves.
2353 * When such condition is possible, the gcwq summons rescuers of all
2354 * workqueues which have works queued on the gcwq and let them process
2355 * those works so that forward progress can be guaranteed.
2357 * This should happen rarely.
2359 static int rescuer_thread(void *__rescuer)
2361 struct worker *rescuer = __rescuer;
2362 struct workqueue_struct *wq = rescuer->rescue_wq;
2363 struct list_head *scheduled = &rescuer->scheduled;
2364 bool is_unbound = wq->flags & WQ_UNBOUND;
2367 set_user_nice(current, RESCUER_NICE_LEVEL);
2370 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2371 * doesn't participate in concurrency management.
2373 rescuer->task->flags |= PF_WQ_WORKER;
2375 set_current_state(TASK_INTERRUPTIBLE);
2377 if (kthread_should_stop()) {
2378 __set_current_state(TASK_RUNNING);
2379 rescuer->task->flags &= ~PF_WQ_WORKER;
2384 * See whether any cpu is asking for help. Unbounded
2385 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2387 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2388 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2389 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2390 struct worker_pool *pool = cwq->pool;
2391 struct global_cwq *gcwq = pool->gcwq;
2392 struct work_struct *work, *n;
2394 __set_current_state(TASK_RUNNING);
2395 mayday_clear_cpu(cpu, wq->mayday_mask);
2397 /* migrate to the target cpu if possible */
2398 rescuer->pool = pool;
2399 worker_maybe_bind_and_lock(rescuer);
2402 * Slurp in all works issued via this workqueue and
2405 BUG_ON(!list_empty(&rescuer->scheduled));
2406 list_for_each_entry_safe(work, n, &pool->worklist, entry)
2407 if (get_work_cwq(work) == cwq)
2408 move_linked_works(work, scheduled, &n);
2410 process_scheduled_works(rescuer);
2413 * Leave this gcwq. If keep_working() is %true, notify a
2414 * regular worker; otherwise, we end up with 0 concurrency
2415 * and stalling the execution.
2417 if (keep_working(pool))
2418 wake_up_worker(pool);
2420 spin_unlock_irq(&gcwq->lock);
2423 /* rescuers should never participate in concurrency management */
2424 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2430 struct work_struct work;
2431 struct completion done;
2434 static void wq_barrier_func(struct work_struct *work)
2436 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2437 complete(&barr->done);
2441 * insert_wq_barrier - insert a barrier work
2442 * @cwq: cwq to insert barrier into
2443 * @barr: wq_barrier to insert
2444 * @target: target work to attach @barr to
2445 * @worker: worker currently executing @target, NULL if @target is not executing
2447 * @barr is linked to @target such that @barr is completed only after
2448 * @target finishes execution. Please note that the ordering
2449 * guarantee is observed only with respect to @target and on the local
2452 * Currently, a queued barrier can't be canceled. This is because
2453 * try_to_grab_pending() can't determine whether the work to be
2454 * grabbed is at the head of the queue and thus can't clear LINKED
2455 * flag of the previous work while there must be a valid next work
2456 * after a work with LINKED flag set.
2458 * Note that when @worker is non-NULL, @target may be modified
2459 * underneath us, so we can't reliably determine cwq from @target.
2462 * spin_lock_irq(gcwq->lock).
2464 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2465 struct wq_barrier *barr,
2466 struct work_struct *target, struct worker *worker)
2468 struct list_head *head;
2469 unsigned int linked = 0;
2472 * debugobject calls are safe here even with gcwq->lock locked
2473 * as we know for sure that this will not trigger any of the
2474 * checks and call back into the fixup functions where we
2477 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2478 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2479 init_completion(&barr->done);
2482 * If @target is currently being executed, schedule the
2483 * barrier to the worker; otherwise, put it after @target.
2486 head = worker->scheduled.next;
2488 unsigned long *bits = work_data_bits(target);
2490 head = target->entry.next;
2491 /* there can already be other linked works, inherit and set */
2492 linked = *bits & WORK_STRUCT_LINKED;
2493 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2496 debug_work_activate(&barr->work);
2497 insert_work(cwq, &barr->work, head,
2498 work_color_to_flags(WORK_NO_COLOR) | linked);
2502 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2503 * @wq: workqueue being flushed
2504 * @flush_color: new flush color, < 0 for no-op
2505 * @work_color: new work color, < 0 for no-op
2507 * Prepare cwqs for workqueue flushing.
2509 * If @flush_color is non-negative, flush_color on all cwqs should be
2510 * -1. If no cwq has in-flight commands at the specified color, all
2511 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2512 * has in flight commands, its cwq->flush_color is set to
2513 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2514 * wakeup logic is armed and %true is returned.
2516 * The caller should have initialized @wq->first_flusher prior to
2517 * calling this function with non-negative @flush_color. If
2518 * @flush_color is negative, no flush color update is done and %false
2521 * If @work_color is non-negative, all cwqs should have the same
2522 * work_color which is previous to @work_color and all will be
2523 * advanced to @work_color.
2526 * mutex_lock(wq->flush_mutex).
2529 * %true if @flush_color >= 0 and there's something to flush. %false
2532 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2533 int flush_color, int work_color)
2538 if (flush_color >= 0) {
2539 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2540 atomic_set(&wq->nr_cwqs_to_flush, 1);
2543 for_each_cwq_cpu(cpu, wq) {
2544 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2545 struct global_cwq *gcwq = cwq->pool->gcwq;
2547 spin_lock_irq(&gcwq->lock);
2549 if (flush_color >= 0) {
2550 BUG_ON(cwq->flush_color != -1);
2552 if (cwq->nr_in_flight[flush_color]) {
2553 cwq->flush_color = flush_color;
2554 atomic_inc(&wq->nr_cwqs_to_flush);
2559 if (work_color >= 0) {
2560 BUG_ON(work_color != work_next_color(cwq->work_color));
2561 cwq->work_color = work_color;
2564 spin_unlock_irq(&gcwq->lock);
2567 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2568 complete(&wq->first_flusher->done);
2574 * flush_workqueue - ensure that any scheduled work has run to completion.
2575 * @wq: workqueue to flush
2577 * Forces execution of the workqueue and blocks until its completion.
2578 * This is typically used in driver shutdown handlers.
2580 * We sleep until all works which were queued on entry have been handled,
2581 * but we are not livelocked by new incoming ones.
2583 void flush_workqueue(struct workqueue_struct *wq)
2585 struct wq_flusher this_flusher = {
2586 .list = LIST_HEAD_INIT(this_flusher.list),
2588 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2592 lock_map_acquire(&wq->lockdep_map);
2593 lock_map_release(&wq->lockdep_map);
2595 mutex_lock(&wq->flush_mutex);
2598 * Start-to-wait phase
2600 next_color = work_next_color(wq->work_color);
2602 if (next_color != wq->flush_color) {
2604 * Color space is not full. The current work_color
2605 * becomes our flush_color and work_color is advanced
2608 BUG_ON(!list_empty(&wq->flusher_overflow));
2609 this_flusher.flush_color = wq->work_color;
2610 wq->work_color = next_color;
2612 if (!wq->first_flusher) {
2613 /* no flush in progress, become the first flusher */
2614 BUG_ON(wq->flush_color != this_flusher.flush_color);
2616 wq->first_flusher = &this_flusher;
2618 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2620 /* nothing to flush, done */
2621 wq->flush_color = next_color;
2622 wq->first_flusher = NULL;
2627 BUG_ON(wq->flush_color == this_flusher.flush_color);
2628 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2629 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2633 * Oops, color space is full, wait on overflow queue.
2634 * The next flush completion will assign us
2635 * flush_color and transfer to flusher_queue.
2637 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2640 mutex_unlock(&wq->flush_mutex);
2642 wait_for_completion(&this_flusher.done);
2645 * Wake-up-and-cascade phase
2647 * First flushers are responsible for cascading flushes and
2648 * handling overflow. Non-first flushers can simply return.
2650 if (wq->first_flusher != &this_flusher)
2653 mutex_lock(&wq->flush_mutex);
2655 /* we might have raced, check again with mutex held */
2656 if (wq->first_flusher != &this_flusher)
2659 wq->first_flusher = NULL;
2661 BUG_ON(!list_empty(&this_flusher.list));
2662 BUG_ON(wq->flush_color != this_flusher.flush_color);
2665 struct wq_flusher *next, *tmp;
2667 /* complete all the flushers sharing the current flush color */
2668 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2669 if (next->flush_color != wq->flush_color)
2671 list_del_init(&next->list);
2672 complete(&next->done);
2675 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2676 wq->flush_color != work_next_color(wq->work_color));
2678 /* this flush_color is finished, advance by one */
2679 wq->flush_color = work_next_color(wq->flush_color);
2681 /* one color has been freed, handle overflow queue */
2682 if (!list_empty(&wq->flusher_overflow)) {
2684 * Assign the same color to all overflowed
2685 * flushers, advance work_color and append to
2686 * flusher_queue. This is the start-to-wait
2687 * phase for these overflowed flushers.
2689 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2690 tmp->flush_color = wq->work_color;
2692 wq->work_color = work_next_color(wq->work_color);
2694 list_splice_tail_init(&wq->flusher_overflow,
2695 &wq->flusher_queue);
2696 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2699 if (list_empty(&wq->flusher_queue)) {
2700 BUG_ON(wq->flush_color != wq->work_color);
2705 * Need to flush more colors. Make the next flusher
2706 * the new first flusher and arm cwqs.
2708 BUG_ON(wq->flush_color == wq->work_color);
2709 BUG_ON(wq->flush_color != next->flush_color);
2711 list_del_init(&next->list);
2712 wq->first_flusher = next;
2714 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2718 * Meh... this color is already done, clear first
2719 * flusher and repeat cascading.
2721 wq->first_flusher = NULL;
2725 mutex_unlock(&wq->flush_mutex);
2727 EXPORT_SYMBOL_GPL(flush_workqueue);
2730 * drain_workqueue - drain a workqueue
2731 * @wq: workqueue to drain
2733 * Wait until the workqueue becomes empty. While draining is in progress,
2734 * only chain queueing is allowed. IOW, only currently pending or running
2735 * work items on @wq can queue further work items on it. @wq is flushed
2736 * repeatedly until it becomes empty. The number of flushing is detemined
2737 * by the depth of chaining and should be relatively short. Whine if it
2740 void drain_workqueue(struct workqueue_struct *wq)
2742 unsigned int flush_cnt = 0;
2746 * __queue_work() needs to test whether there are drainers, is much
2747 * hotter than drain_workqueue() and already looks at @wq->flags.
2748 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2750 spin_lock(&workqueue_lock);
2751 if (!wq->nr_drainers++)
2752 wq->flags |= WQ_DRAINING;
2753 spin_unlock(&workqueue_lock);
2755 flush_workqueue(wq);
2757 for_each_cwq_cpu(cpu, wq) {
2758 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2761 spin_lock_irq(&cwq->pool->gcwq->lock);
2762 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2763 spin_unlock_irq(&cwq->pool->gcwq->lock);
2768 if (++flush_cnt == 10 ||
2769 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2770 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2771 wq->name, flush_cnt);
2775 spin_lock(&workqueue_lock);
2776 if (!--wq->nr_drainers)
2777 wq->flags &= ~WQ_DRAINING;
2778 spin_unlock(&workqueue_lock);
2780 EXPORT_SYMBOL_GPL(drain_workqueue);
2782 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2784 struct worker *worker = NULL;
2785 struct global_cwq *gcwq;
2786 struct cpu_workqueue_struct *cwq;
2789 gcwq = get_work_gcwq(work);
2793 spin_lock_irq(&gcwq->lock);
2794 if (!list_empty(&work->entry)) {
2796 * See the comment near try_to_grab_pending()->smp_rmb().
2797 * If it was re-queued to a different gcwq under us, we
2798 * are not going to wait.
2801 cwq = get_work_cwq(work);
2802 if (unlikely(!cwq || gcwq != cwq->pool->gcwq))
2805 worker = find_worker_executing_work(gcwq, work);
2808 cwq = worker->current_cwq;
2811 insert_wq_barrier(cwq, barr, work, worker);
2812 spin_unlock_irq(&gcwq->lock);
2815 * If @max_active is 1 or rescuer is in use, flushing another work
2816 * item on the same workqueue may lead to deadlock. Make sure the
2817 * flusher is not running on the same workqueue by verifying write
2820 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2821 lock_map_acquire(&cwq->wq->lockdep_map);
2823 lock_map_acquire_read(&cwq->wq->lockdep_map);
2824 lock_map_release(&cwq->wq->lockdep_map);
2828 spin_unlock_irq(&gcwq->lock);
2833 * flush_work - wait for a work to finish executing the last queueing instance
2834 * @work: the work to flush
2836 * Wait until @work has finished execution. @work is guaranteed to be idle
2837 * on return if it hasn't been requeued since flush started.
2840 * %true if flush_work() waited for the work to finish execution,
2841 * %false if it was already idle.
2843 bool flush_work(struct work_struct *work)
2845 struct wq_barrier barr;
2847 lock_map_acquire(&work->lockdep_map);
2848 lock_map_release(&work->lockdep_map);
2850 if (start_flush_work(work, &barr)) {
2851 wait_for_completion(&barr.done);
2852 destroy_work_on_stack(&barr.work);
2858 EXPORT_SYMBOL_GPL(flush_work);
2860 static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2862 unsigned long flags;
2866 ret = try_to_grab_pending(work, is_dwork, &flags);
2868 * If someone else is canceling, wait for the same event it
2869 * would be waiting for before retrying.
2871 if (unlikely(ret == -ENOENT))
2873 } while (unlikely(ret < 0));
2875 /* tell other tasks trying to grab @work to back off */
2876 mark_work_canceling(work);
2877 local_irq_restore(flags);
2880 clear_work_data(work);
2885 * cancel_work_sync - cancel a work and wait for it to finish
2886 * @work: the work to cancel
2888 * Cancel @work and wait for its execution to finish. This function
2889 * can be used even if the work re-queues itself or migrates to
2890 * another workqueue. On return from this function, @work is
2891 * guaranteed to be not pending or executing on any CPU.
2893 * cancel_work_sync(&delayed_work->work) must not be used for
2894 * delayed_work's. Use cancel_delayed_work_sync() instead.
2896 * The caller must ensure that the workqueue on which @work was last
2897 * queued can't be destroyed before this function returns.
2900 * %true if @work was pending, %false otherwise.
2902 bool cancel_work_sync(struct work_struct *work)
2904 return __cancel_work_timer(work, false);
2906 EXPORT_SYMBOL_GPL(cancel_work_sync);
2909 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2910 * @dwork: the delayed work to flush
2912 * Delayed timer is cancelled and the pending work is queued for
2913 * immediate execution. Like flush_work(), this function only
2914 * considers the last queueing instance of @dwork.
2917 * %true if flush_work() waited for the work to finish execution,
2918 * %false if it was already idle.
2920 bool flush_delayed_work(struct delayed_work *dwork)
2922 local_irq_disable();
2923 if (del_timer_sync(&dwork->timer))
2924 __queue_work(dwork->cpu,
2925 get_work_cwq(&dwork->work)->wq, &dwork->work);
2927 return flush_work(&dwork->work);
2929 EXPORT_SYMBOL(flush_delayed_work);
2932 * cancel_delayed_work - cancel a delayed work
2933 * @dwork: delayed_work to cancel
2935 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2936 * and canceled; %false if wasn't pending. Note that the work callback
2937 * function may still be running on return, unless it returns %true and the
2938 * work doesn't re-arm itself. Explicitly flush or use
2939 * cancel_delayed_work_sync() to wait on it.
2941 * This function is safe to call from any context including IRQ handler.
2943 bool cancel_delayed_work(struct delayed_work *dwork)
2945 unsigned long flags;
2949 ret = try_to_grab_pending(&dwork->work, true, &flags);
2950 } while (unlikely(ret == -EAGAIN));
2952 if (unlikely(ret < 0))
2955 set_work_cpu_and_clear_pending(&dwork->work, work_cpu(&dwork->work));
2956 local_irq_restore(flags);
2959 EXPORT_SYMBOL(cancel_delayed_work);
2962 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2963 * @dwork: the delayed work cancel
2965 * This is cancel_work_sync() for delayed works.
2968 * %true if @dwork was pending, %false otherwise.
2970 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2972 return __cancel_work_timer(&dwork->work, true);
2974 EXPORT_SYMBOL(cancel_delayed_work_sync);
2977 * schedule_work_on - put work task on a specific cpu
2978 * @cpu: cpu to put the work task on
2979 * @work: job to be done
2981 * This puts a job on a specific cpu
2983 bool schedule_work_on(int cpu, struct work_struct *work)
2985 return queue_work_on(cpu, system_wq, work);
2987 EXPORT_SYMBOL(schedule_work_on);
2990 * schedule_work - put work task in global workqueue
2991 * @work: job to be done
2993 * Returns %false if @work was already on the kernel-global workqueue and
2996 * This puts a job in the kernel-global workqueue if it was not already
2997 * queued and leaves it in the same position on the kernel-global
2998 * workqueue otherwise.
3000 bool schedule_work(struct work_struct *work)
3002 return queue_work(system_wq, work);
3004 EXPORT_SYMBOL(schedule_work);
3007 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3009 * @dwork: job to be done
3010 * @delay: number of jiffies to wait
3012 * After waiting for a given time this puts a job in the kernel-global
3013 * workqueue on the specified CPU.
3015 bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3016 unsigned long delay)
3018 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3020 EXPORT_SYMBOL(schedule_delayed_work_on);
3023 * schedule_delayed_work - put work task in global workqueue after delay
3024 * @dwork: job to be done
3025 * @delay: number of jiffies to wait or 0 for immediate execution
3027 * After waiting for a given time this puts a job in the kernel-global
3030 bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3032 return queue_delayed_work(system_wq, dwork, delay);
3034 EXPORT_SYMBOL(schedule_delayed_work);
3037 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3038 * @func: the function to call
3040 * schedule_on_each_cpu() executes @func on each online CPU using the
3041 * system workqueue and blocks until all CPUs have completed.
3042 * schedule_on_each_cpu() is very slow.
3045 * 0 on success, -errno on failure.
3047 int schedule_on_each_cpu(work_func_t func)
3050 struct work_struct __percpu *works;
3052 works = alloc_percpu(struct work_struct);
3058 for_each_online_cpu(cpu) {
3059 struct work_struct *work = per_cpu_ptr(works, cpu);
3061 INIT_WORK(work, func);
3062 schedule_work_on(cpu, work);
3065 for_each_online_cpu(cpu)
3066 flush_work(per_cpu_ptr(works, cpu));
3074 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3076 * Forces execution of the kernel-global workqueue and blocks until its
3079 * Think twice before calling this function! It's very easy to get into
3080 * trouble if you don't take great care. Either of the following situations
3081 * will lead to deadlock:
3083 * One of the work items currently on the workqueue needs to acquire
3084 * a lock held by your code or its caller.
3086 * Your code is running in the context of a work routine.
3088 * They will be detected by lockdep when they occur, but the first might not
3089 * occur very often. It depends on what work items are on the workqueue and
3090 * what locks they need, which you have no control over.
3092 * In most situations flushing the entire workqueue is overkill; you merely
3093 * need to know that a particular work item isn't queued and isn't running.
3094 * In such cases you should use cancel_delayed_work_sync() or
3095 * cancel_work_sync() instead.
3097 void flush_scheduled_work(void)
3099 flush_workqueue(system_wq);
3101 EXPORT_SYMBOL(flush_scheduled_work);
3104 * execute_in_process_context - reliably execute the routine with user context
3105 * @fn: the function to execute
3106 * @ew: guaranteed storage for the execute work structure (must
3107 * be available when the work executes)
3109 * Executes the function immediately if process context is available,
3110 * otherwise schedules the function for delayed execution.
3112 * Returns: 0 - function was executed
3113 * 1 - function was scheduled for execution
3115 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3117 if (!in_interrupt()) {
3122 INIT_WORK(&ew->work, fn);
3123 schedule_work(&ew->work);
3127 EXPORT_SYMBOL_GPL(execute_in_process_context);
3129 int keventd_up(void)
3131 return system_wq != NULL;
3134 static int alloc_cwqs(struct workqueue_struct *wq)
3137 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3138 * Make sure that the alignment isn't lower than that of
3139 * unsigned long long.
3141 const size_t size = sizeof(struct cpu_workqueue_struct);
3142 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3143 __alignof__(unsigned long long));
3145 if (!(wq->flags & WQ_UNBOUND))
3146 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3151 * Allocate enough room to align cwq and put an extra
3152 * pointer at the end pointing back to the originally
3153 * allocated pointer which will be used for free.
3155 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3157 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3158 *(void **)(wq->cpu_wq.single + 1) = ptr;
3162 /* just in case, make sure it's actually aligned */
3163 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3164 return wq->cpu_wq.v ? 0 : -ENOMEM;
3167 static void free_cwqs(struct workqueue_struct *wq)
3169 if (!(wq->flags & WQ_UNBOUND))
3170 free_percpu(wq->cpu_wq.pcpu);
3171 else if (wq->cpu_wq.single) {
3172 /* the pointer to free is stored right after the cwq */
3173 kfree(*(void **)(wq->cpu_wq.single + 1));
3177 static int wq_clamp_max_active(int max_active, unsigned int flags,
3180 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3182 if (max_active < 1 || max_active > lim)
3183 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3184 max_active, name, 1, lim);
3186 return clamp_val(max_active, 1, lim);
3189 struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3192 struct lock_class_key *key,
3193 const char *lock_name, ...)
3195 va_list args, args1;
3196 struct workqueue_struct *wq;
3200 /* determine namelen, allocate wq and format name */
3201 va_start(args, lock_name);
3202 va_copy(args1, args);
3203 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3205 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3209 vsnprintf(wq->name, namelen, fmt, args1);
3214 * Workqueues which may be used during memory reclaim should
3215 * have a rescuer to guarantee forward progress.
3217 if (flags & WQ_MEM_RECLAIM)
3218 flags |= WQ_RESCUER;
3220 max_active = max_active ?: WQ_DFL_ACTIVE;
3221 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3225 wq->saved_max_active = max_active;
3226 mutex_init(&wq->flush_mutex);
3227 atomic_set(&wq->nr_cwqs_to_flush, 0);
3228 INIT_LIST_HEAD(&wq->flusher_queue);
3229 INIT_LIST_HEAD(&wq->flusher_overflow);
3231 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3232 INIT_LIST_HEAD(&wq->list);
3234 if (alloc_cwqs(wq) < 0)
3237 for_each_cwq_cpu(cpu, wq) {
3238 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3239 struct global_cwq *gcwq = get_gcwq(cpu);
3240 int pool_idx = (bool)(flags & WQ_HIGHPRI);
3242 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3243 cwq->pool = &gcwq->pools[pool_idx];
3245 cwq->flush_color = -1;
3246 cwq->max_active = max_active;
3247 INIT_LIST_HEAD(&cwq->delayed_works);
3250 if (flags & WQ_RESCUER) {
3251 struct worker *rescuer;
3253 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3256 wq->rescuer = rescuer = alloc_worker();
3260 rescuer->rescue_wq = wq;
3261 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3263 if (IS_ERR(rescuer->task))
3266 rescuer->task->flags |= PF_THREAD_BOUND;
3267 wake_up_process(rescuer->task);
3271 * workqueue_lock protects global freeze state and workqueues
3272 * list. Grab it, set max_active accordingly and add the new
3273 * workqueue to workqueues list.
3275 spin_lock(&workqueue_lock);
3277 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3278 for_each_cwq_cpu(cpu, wq)
3279 get_cwq(cpu, wq)->max_active = 0;
3281 list_add(&wq->list, &workqueues);
3283 spin_unlock(&workqueue_lock);
3289 free_mayday_mask(wq->mayday_mask);
3295 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3298 * destroy_workqueue - safely terminate a workqueue
3299 * @wq: target workqueue
3301 * Safely destroy a workqueue. All work currently pending will be done first.
3303 void destroy_workqueue(struct workqueue_struct *wq)
3307 /* drain it before proceeding with destruction */
3308 drain_workqueue(wq);
3311 * wq list is used to freeze wq, remove from list after
3312 * flushing is complete in case freeze races us.
3314 spin_lock(&workqueue_lock);
3315 list_del(&wq->list);
3316 spin_unlock(&workqueue_lock);
3319 for_each_cwq_cpu(cpu, wq) {
3320 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3323 for (i = 0; i < WORK_NR_COLORS; i++)
3324 BUG_ON(cwq->nr_in_flight[i]);
3325 BUG_ON(cwq->nr_active);
3326 BUG_ON(!list_empty(&cwq->delayed_works));
3329 if (wq->flags & WQ_RESCUER) {
3330 kthread_stop(wq->rescuer->task);
3331 free_mayday_mask(wq->mayday_mask);
3338 EXPORT_SYMBOL_GPL(destroy_workqueue);
3341 * cwq_set_max_active - adjust max_active of a cwq
3342 * @cwq: target cpu_workqueue_struct
3343 * @max_active: new max_active value.
3345 * Set @cwq->max_active to @max_active and activate delayed works if
3349 * spin_lock_irq(gcwq->lock).
3351 static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active)
3353 cwq->max_active = max_active;
3355 while (!list_empty(&cwq->delayed_works) &&
3356 cwq->nr_active < cwq->max_active)
3357 cwq_activate_first_delayed(cwq);
3361 * workqueue_set_max_active - adjust max_active of a workqueue
3362 * @wq: target workqueue
3363 * @max_active: new max_active value.
3365 * Set max_active of @wq to @max_active.
3368 * Don't call from IRQ context.
3370 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3374 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3376 spin_lock(&workqueue_lock);
3378 wq->saved_max_active = max_active;
3380 for_each_cwq_cpu(cpu, wq) {
3381 struct global_cwq *gcwq = get_gcwq(cpu);
3383 spin_lock_irq(&gcwq->lock);
3385 if (!(wq->flags & WQ_FREEZABLE) ||
3386 !(gcwq->flags & GCWQ_FREEZING))
3387 cwq_set_max_active(get_cwq(gcwq->cpu, wq), max_active);
3389 spin_unlock_irq(&gcwq->lock);
3392 spin_unlock(&workqueue_lock);
3394 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3397 * workqueue_congested - test whether a workqueue is congested
3398 * @cpu: CPU in question
3399 * @wq: target workqueue
3401 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3402 * no synchronization around this function and the test result is
3403 * unreliable and only useful as advisory hints or for debugging.
3406 * %true if congested, %false otherwise.
3408 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3410 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3412 return !list_empty(&cwq->delayed_works);
3414 EXPORT_SYMBOL_GPL(workqueue_congested);
3417 * work_cpu - return the last known associated cpu for @work
3418 * @work: the work of interest
3421 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3423 static unsigned int work_cpu(struct work_struct *work)
3425 struct global_cwq *gcwq = get_work_gcwq(work);
3427 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3431 * work_busy - test whether a work is currently pending or running
3432 * @work: the work to be tested
3434 * Test whether @work is currently pending or running. There is no
3435 * synchronization around this function and the test result is
3436 * unreliable and only useful as advisory hints or for debugging.
3437 * Especially for reentrant wqs, the pending state might hide the
3441 * OR'd bitmask of WORK_BUSY_* bits.
3443 unsigned int work_busy(struct work_struct *work)
3445 struct global_cwq *gcwq = get_work_gcwq(work);
3446 unsigned long flags;
3447 unsigned int ret = 0;
3452 spin_lock_irqsave(&gcwq->lock, flags);
3454 if (work_pending(work))
3455 ret |= WORK_BUSY_PENDING;
3456 if (find_worker_executing_work(gcwq, work))
3457 ret |= WORK_BUSY_RUNNING;
3459 spin_unlock_irqrestore(&gcwq->lock, flags);
3463 EXPORT_SYMBOL_GPL(work_busy);
3468 * There are two challenges in supporting CPU hotplug. Firstly, there
3469 * are a lot of assumptions on strong associations among work, cwq and
3470 * gcwq which make migrating pending and scheduled works very
3471 * difficult to implement without impacting hot paths. Secondly,
3472 * gcwqs serve mix of short, long and very long running works making
3473 * blocked draining impractical.
3475 * This is solved by allowing a gcwq to be disassociated from the CPU
3476 * running as an unbound one and allowing it to be reattached later if the
3477 * cpu comes back online.
3480 /* claim manager positions of all pools */
3481 static void gcwq_claim_assoc_and_lock(struct global_cwq *gcwq)
3483 struct worker_pool *pool;
3485 for_each_worker_pool(pool, gcwq)
3486 mutex_lock_nested(&pool->assoc_mutex, pool - gcwq->pools);
3487 spin_lock_irq(&gcwq->lock);
3490 /* release manager positions */
3491 static void gcwq_release_assoc_and_unlock(struct global_cwq *gcwq)
3493 struct worker_pool *pool;
3495 spin_unlock_irq(&gcwq->lock);
3496 for_each_worker_pool(pool, gcwq)
3497 mutex_unlock(&pool->assoc_mutex);
3500 static void gcwq_unbind_fn(struct work_struct *work)
3502 struct global_cwq *gcwq = get_gcwq(smp_processor_id());
3503 struct worker_pool *pool;
3504 struct worker *worker;
3505 struct hlist_node *pos;
3508 BUG_ON(gcwq->cpu != smp_processor_id());
3510 gcwq_claim_assoc_and_lock(gcwq);
3513 * We've claimed all manager positions. Make all workers unbound
3514 * and set DISASSOCIATED. Before this, all workers except for the
3515 * ones which are still executing works from before the last CPU
3516 * down must be on the cpu. After this, they may become diasporas.
3518 for_each_worker_pool(pool, gcwq)
3519 list_for_each_entry(worker, &pool->idle_list, entry)
3520 worker->flags |= WORKER_UNBOUND;
3522 for_each_busy_worker(worker, i, pos, gcwq)
3523 worker->flags |= WORKER_UNBOUND;
3525 gcwq->flags |= GCWQ_DISASSOCIATED;
3527 gcwq_release_assoc_and_unlock(gcwq);
3530 * Call schedule() so that we cross rq->lock and thus can guarantee
3531 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3532 * as scheduler callbacks may be invoked from other cpus.
3537 * Sched callbacks are disabled now. Zap nr_running. After this,
3538 * nr_running stays zero and need_more_worker() and keep_working()
3539 * are always true as long as the worklist is not empty. @gcwq now
3540 * behaves as unbound (in terms of concurrency management) gcwq
3541 * which is served by workers tied to the CPU.
3543 * On return from this function, the current worker would trigger
3544 * unbound chain execution of pending work items if other workers
3547 for_each_worker_pool(pool, gcwq)
3548 atomic_set(get_pool_nr_running(pool), 0);
3552 * Workqueues should be brought up before normal priority CPU notifiers.
3553 * This will be registered high priority CPU notifier.
3555 static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3556 unsigned long action,
3559 unsigned int cpu = (unsigned long)hcpu;
3560 struct global_cwq *gcwq = get_gcwq(cpu);
3561 struct worker_pool *pool;
3563 switch (action & ~CPU_TASKS_FROZEN) {
3564 case CPU_UP_PREPARE:
3565 for_each_worker_pool(pool, gcwq) {
3566 struct worker *worker;
3568 if (pool->nr_workers)
3571 worker = create_worker(pool);
3575 spin_lock_irq(&gcwq->lock);
3576 start_worker(worker);
3577 spin_unlock_irq(&gcwq->lock);
3581 case CPU_DOWN_FAILED:
3583 gcwq_claim_assoc_and_lock(gcwq);
3584 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3585 rebind_workers(gcwq);
3586 gcwq_release_assoc_and_unlock(gcwq);
3593 * Workqueues should be brought down after normal priority CPU notifiers.
3594 * This will be registered as low priority CPU notifier.
3596 static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3597 unsigned long action,
3600 unsigned int cpu = (unsigned long)hcpu;
3601 struct work_struct unbind_work;
3603 switch (action & ~CPU_TASKS_FROZEN) {
3604 case CPU_DOWN_PREPARE:
3605 /* unbinding should happen on the local CPU */
3606 INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
3607 queue_work_on(cpu, system_highpri_wq, &unbind_work);
3608 flush_work(&unbind_work);
3616 struct work_for_cpu {
3617 struct work_struct work;
3623 static void work_for_cpu_fn(struct work_struct *work)
3625 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3627 wfc->ret = wfc->fn(wfc->arg);
3631 * work_on_cpu - run a function in user context on a particular cpu
3632 * @cpu: the cpu to run on
3633 * @fn: the function to run
3634 * @arg: the function arg
3636 * This will return the value @fn returns.
3637 * It is up to the caller to ensure that the cpu doesn't go offline.
3638 * The caller must not hold any locks which would prevent @fn from completing.
3640 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3642 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3644 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3645 schedule_work_on(cpu, &wfc.work);
3646 flush_work(&wfc.work);
3649 EXPORT_SYMBOL_GPL(work_on_cpu);
3650 #endif /* CONFIG_SMP */
3652 #ifdef CONFIG_FREEZER
3655 * freeze_workqueues_begin - begin freezing workqueues
3657 * Start freezing workqueues. After this function returns, all freezable
3658 * workqueues will queue new works to their frozen_works list instead of
3662 * Grabs and releases workqueue_lock and gcwq->lock's.
3664 void freeze_workqueues_begin(void)
3668 spin_lock(&workqueue_lock);
3670 BUG_ON(workqueue_freezing);
3671 workqueue_freezing = true;
3673 for_each_gcwq_cpu(cpu) {
3674 struct global_cwq *gcwq = get_gcwq(cpu);
3675 struct workqueue_struct *wq;
3677 spin_lock_irq(&gcwq->lock);
3679 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3680 gcwq->flags |= GCWQ_FREEZING;
3682 list_for_each_entry(wq, &workqueues, list) {
3683 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3685 if (cwq && wq->flags & WQ_FREEZABLE)
3686 cwq->max_active = 0;
3689 spin_unlock_irq(&gcwq->lock);
3692 spin_unlock(&workqueue_lock);
3696 * freeze_workqueues_busy - are freezable workqueues still busy?
3698 * Check whether freezing is complete. This function must be called
3699 * between freeze_workqueues_begin() and thaw_workqueues().
3702 * Grabs and releases workqueue_lock.
3705 * %true if some freezable workqueues are still busy. %false if freezing
3708 bool freeze_workqueues_busy(void)
3713 spin_lock(&workqueue_lock);
3715 BUG_ON(!workqueue_freezing);
3717 for_each_gcwq_cpu(cpu) {
3718 struct workqueue_struct *wq;
3720 * nr_active is monotonically decreasing. It's safe
3721 * to peek without lock.
3723 list_for_each_entry(wq, &workqueues, list) {
3724 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3726 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3729 BUG_ON(cwq->nr_active < 0);
3730 if (cwq->nr_active) {
3737 spin_unlock(&workqueue_lock);
3742 * thaw_workqueues - thaw workqueues
3744 * Thaw workqueues. Normal queueing is restored and all collected
3745 * frozen works are transferred to their respective gcwq worklists.
3748 * Grabs and releases workqueue_lock and gcwq->lock's.
3750 void thaw_workqueues(void)
3754 spin_lock(&workqueue_lock);
3756 if (!workqueue_freezing)
3759 for_each_gcwq_cpu(cpu) {
3760 struct global_cwq *gcwq = get_gcwq(cpu);
3761 struct worker_pool *pool;
3762 struct workqueue_struct *wq;
3764 spin_lock_irq(&gcwq->lock);
3766 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3767 gcwq->flags &= ~GCWQ_FREEZING;
3769 list_for_each_entry(wq, &workqueues, list) {
3770 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3772 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3775 /* restore max_active and repopulate worklist */
3776 cwq_set_max_active(cwq, wq->saved_max_active);
3779 for_each_worker_pool(pool, gcwq)
3780 wake_up_worker(pool);
3782 spin_unlock_irq(&gcwq->lock);
3785 workqueue_freezing = false;
3787 spin_unlock(&workqueue_lock);
3789 #endif /* CONFIG_FREEZER */
3791 static int __init init_workqueues(void)
3795 /* make sure we have enough bits for OFFQ CPU number */
3796 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) <
3799 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3800 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3802 /* initialize gcwqs */
3803 for_each_gcwq_cpu(cpu) {
3804 struct global_cwq *gcwq = get_gcwq(cpu);
3805 struct worker_pool *pool;
3807 spin_lock_init(&gcwq->lock);
3809 gcwq->flags |= GCWQ_DISASSOCIATED;
3811 hash_init(gcwq->busy_hash);
3813 for_each_worker_pool(pool, gcwq) {
3815 INIT_LIST_HEAD(&pool->worklist);
3816 INIT_LIST_HEAD(&pool->idle_list);
3818 init_timer_deferrable(&pool->idle_timer);
3819 pool->idle_timer.function = idle_worker_timeout;
3820 pool->idle_timer.data = (unsigned long)pool;
3822 setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
3823 (unsigned long)pool);
3825 mutex_init(&pool->assoc_mutex);
3826 ida_init(&pool->worker_ida);
3830 /* create the initial worker */
3831 for_each_online_gcwq_cpu(cpu) {
3832 struct global_cwq *gcwq = get_gcwq(cpu);
3833 struct worker_pool *pool;
3835 if (cpu != WORK_CPU_UNBOUND)
3836 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3838 for_each_worker_pool(pool, gcwq) {
3839 struct worker *worker;
3841 worker = create_worker(pool);
3843 spin_lock_irq(&gcwq->lock);
3844 start_worker(worker);
3845 spin_unlock_irq(&gcwq->lock);
3849 system_wq = alloc_workqueue("events", 0, 0);
3850 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3851 system_long_wq = alloc_workqueue("events_long", 0, 0);
3852 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3853 WQ_UNBOUND_MAX_ACTIVE);
3854 system_freezable_wq = alloc_workqueue("events_freezable",
3856 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3857 !system_unbound_wq || !system_freezable_wq);
3860 early_initcall(init_workqueues);