2 * linux/kernel/workqueue.c
4 * Generic mechanism for defining kernel helper threads for running
5 * arbitrary tasks in process context.
7 * Started by Ingo Molnar, Copyright (C) 2002
9 * Derived from the taskqueue/keventd code by:
11 * David Woodhouse <dwmw2@infradead.org>
13 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
14 * Theodore Ts'o <tytso@mit.edu>
16 * Made to use alloc_percpu by Christoph Lameter.
19 #include <linux/module.h>
20 #include <linux/kernel.h>
21 #include <linux/sched.h>
22 #include <linux/init.h>
23 #include <linux/signal.h>
24 #include <linux/completion.h>
25 #include <linux/workqueue.h>
26 #include <linux/slab.h>
27 #include <linux/cpu.h>
28 #include <linux/notifier.h>
29 #include <linux/kthread.h>
30 #include <linux/hardirq.h>
31 #include <linux/mempolicy.h>
32 #include <linux/freezer.h>
33 #include <linux/kallsyms.h>
34 #include <linux/debug_locks.h>
35 #include <linux/lockdep.h>
36 #include <linux/idr.h>
38 #include "workqueue_sched.h"
41 /* global_cwq flags */
42 GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
43 GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */
44 GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
45 GCWQ_FREEZING = 1 << 3, /* freeze in progress */
46 GCWQ_HIGHPRI_PENDING = 1 << 4, /* highpri works on queue */
49 WORKER_STARTED = 1 << 0, /* started */
50 WORKER_DIE = 1 << 1, /* die die die */
51 WORKER_IDLE = 1 << 2, /* is idle */
52 WORKER_PREP = 1 << 3, /* preparing to run works */
53 WORKER_ROGUE = 1 << 4, /* not bound to any cpu */
54 WORKER_REBIND = 1 << 5, /* mom is home, come back */
55 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
56 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
58 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND |
59 WORKER_CPU_INTENSIVE | WORKER_UNBOUND,
61 /* gcwq->trustee_state */
62 TRUSTEE_START = 0, /* start */
63 TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */
64 TRUSTEE_BUTCHER = 2, /* butcher workers */
65 TRUSTEE_RELEASE = 3, /* release workers */
66 TRUSTEE_DONE = 4, /* trustee is done */
68 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
69 BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
70 BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
72 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
73 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
75 MAYDAY_INITIAL_TIMEOUT = HZ / 100, /* call for help after 10ms */
76 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
77 CREATE_COOLDOWN = HZ, /* time to breath after fail */
78 TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */
81 * Rescue workers are used only on emergencies and shared by
84 RESCUER_NICE_LEVEL = -20,
88 * Structure fields follow one of the following exclusion rules.
90 * I: Modifiable by initialization/destruction paths and read-only for
93 * P: Preemption protected. Disabling preemption is enough and should
94 * only be modified and accessed from the local cpu.
96 * L: gcwq->lock protected. Access with gcwq->lock held.
98 * X: During normal operation, modification requires gcwq->lock and
99 * should be done only from local cpu. Either disabling preemption
100 * on local cpu or grabbing gcwq->lock is enough for read access.
101 * If GCWQ_DISASSOCIATED is set, it's identical to L.
103 * F: wq->flush_mutex protected.
105 * W: workqueue_lock protected.
111 * The poor guys doing the actual heavy lifting. All on-duty workers
112 * are either serving the manager role, on idle list or on busy hash.
115 /* on idle list while idle, on busy hash table while busy */
117 struct list_head entry; /* L: while idle */
118 struct hlist_node hentry; /* L: while busy */
121 struct work_struct *current_work; /* L: work being processed */
122 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
123 struct list_head scheduled; /* L: scheduled works */
124 struct task_struct *task; /* I: worker task */
125 struct global_cwq *gcwq; /* I: the associated gcwq */
126 /* 64 bytes boundary on 64bit, 32 on 32bit */
127 unsigned long last_active; /* L: last active timestamp */
128 unsigned int flags; /* X: flags */
129 int id; /* I: worker id */
130 struct work_struct rebind_work; /* L: rebind worker to cpu */
134 * Global per-cpu workqueue. There's one and only one for each cpu
135 * and all works are queued and processed here regardless of their
139 spinlock_t lock; /* the gcwq lock */
140 struct list_head worklist; /* L: list of pending works */
141 unsigned int cpu; /* I: the associated cpu */
142 unsigned int flags; /* L: GCWQ_* flags */
144 int nr_workers; /* L: total number of workers */
145 int nr_idle; /* L: currently idle ones */
147 /* workers are chained either in the idle_list or busy_hash */
148 struct list_head idle_list; /* X: list of idle workers */
149 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
150 /* L: hash of busy workers */
152 struct timer_list idle_timer; /* L: worker idle timeout */
153 struct timer_list mayday_timer; /* L: SOS timer for dworkers */
155 struct ida worker_ida; /* L: for worker IDs */
157 struct task_struct *trustee; /* L: for gcwq shutdown */
158 unsigned int trustee_state; /* L: trustee state */
159 wait_queue_head_t trustee_wait; /* trustee wait */
160 struct worker *first_idle; /* L: first idle worker */
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 global_cwq *gcwq; /* I: the associated gcwq */
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) alloc_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; /* I: 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 saved_max_active; /* W: saved cwq max_active */
235 const char *name; /* I: workqueue name */
236 #ifdef CONFIG_LOCKDEP
237 struct lockdep_map lockdep_map;
241 struct workqueue_struct *system_wq __read_mostly;
242 struct workqueue_struct *system_long_wq __read_mostly;
243 struct workqueue_struct *system_nrt_wq __read_mostly;
244 struct workqueue_struct *system_unbound_wq __read_mostly;
245 EXPORT_SYMBOL_GPL(system_wq);
246 EXPORT_SYMBOL_GPL(system_long_wq);
247 EXPORT_SYMBOL_GPL(system_nrt_wq);
248 EXPORT_SYMBOL_GPL(system_unbound_wq);
250 #define for_each_busy_worker(worker, i, pos, gcwq) \
251 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
252 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
254 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
257 if (cpu < nr_cpu_ids) {
259 cpu = cpumask_next(cpu, mask);
260 if (cpu < nr_cpu_ids)
264 return WORK_CPU_UNBOUND;
266 return WORK_CPU_NONE;
269 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
270 struct workqueue_struct *wq)
272 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
278 * An extra gcwq is defined for an invalid cpu number
279 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
280 * specific CPU. The following iterators are similar to
281 * for_each_*_cpu() iterators but also considers the unbound gcwq.
283 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
284 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
285 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
286 * WORK_CPU_UNBOUND for unbound workqueues
288 #define for_each_gcwq_cpu(cpu) \
289 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
290 (cpu) < WORK_CPU_NONE; \
291 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
293 #define for_each_online_gcwq_cpu(cpu) \
294 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
295 (cpu) < WORK_CPU_NONE; \
296 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
298 #define for_each_cwq_cpu(cpu, wq) \
299 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
300 (cpu) < WORK_CPU_NONE; \
301 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
303 #ifdef CONFIG_LOCKDEP
305 * in_workqueue_context() - in context of specified workqueue?
306 * @wq: the workqueue of interest
308 * Checks lockdep state to see if the current task is executing from
309 * within a workqueue item. This function exists only if lockdep is
312 int in_workqueue_context(struct workqueue_struct *wq)
314 return lock_is_held(&wq->lockdep_map);
318 #ifdef CONFIG_DEBUG_OBJECTS_WORK
320 static struct debug_obj_descr work_debug_descr;
323 * fixup_init is called when:
324 * - an active object is initialized
326 static int work_fixup_init(void *addr, enum debug_obj_state state)
328 struct work_struct *work = addr;
331 case ODEBUG_STATE_ACTIVE:
332 cancel_work_sync(work);
333 debug_object_init(work, &work_debug_descr);
341 * fixup_activate is called when:
342 * - an active object is activated
343 * - an unknown object is activated (might be a statically initialized object)
345 static int work_fixup_activate(void *addr, enum debug_obj_state state)
347 struct work_struct *work = addr;
351 case ODEBUG_STATE_NOTAVAILABLE:
353 * This is not really a fixup. The work struct was
354 * statically initialized. We just make sure that it
355 * is tracked in the object tracker.
357 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
358 debug_object_init(work, &work_debug_descr);
359 debug_object_activate(work, &work_debug_descr);
365 case ODEBUG_STATE_ACTIVE:
374 * fixup_free is called when:
375 * - an active object is freed
377 static int work_fixup_free(void *addr, enum debug_obj_state state)
379 struct work_struct *work = addr;
382 case ODEBUG_STATE_ACTIVE:
383 cancel_work_sync(work);
384 debug_object_free(work, &work_debug_descr);
391 static struct debug_obj_descr work_debug_descr = {
392 .name = "work_struct",
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, gcwq_nr_running);
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_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */
449 static int worker_thread(void *__worker);
451 static struct global_cwq *get_gcwq(unsigned int cpu)
453 if (cpu != WORK_CPU_UNBOUND)
454 return &per_cpu(global_cwq, cpu);
456 return &unbound_global_cwq;
459 static atomic_t *get_gcwq_nr_running(unsigned int cpu)
461 if (cpu != WORK_CPU_UNBOUND)
462 return &per_cpu(gcwq_nr_running, cpu);
464 return &unbound_gcwq_nr_running;
467 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
468 struct workqueue_struct *wq)
470 if (!(wq->flags & WQ_UNBOUND)) {
471 if (likely(cpu < nr_cpu_ids)) {
473 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
475 return wq->cpu_wq.single;
478 } else if (likely(cpu == WORK_CPU_UNBOUND))
479 return wq->cpu_wq.single;
483 static unsigned int work_color_to_flags(int color)
485 return color << WORK_STRUCT_COLOR_SHIFT;
488 static int get_work_color(struct work_struct *work)
490 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
491 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
494 static int work_next_color(int color)
496 return (color + 1) % WORK_NR_COLORS;
500 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
501 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
502 * cleared and the work data contains the cpu number it was last on.
504 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
505 * cwq, cpu or clear work->data. These functions should only be
506 * called while the work is owned - ie. while the PENDING bit is set.
508 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
509 * corresponding to a work. gcwq is available once the work has been
510 * queued anywhere after initialization. cwq is available only from
511 * queueing until execution starts.
513 static inline void set_work_data(struct work_struct *work, unsigned long data,
516 BUG_ON(!work_pending(work));
517 atomic_long_set(&work->data, data | flags | work_static(work));
520 static void set_work_cwq(struct work_struct *work,
521 struct cpu_workqueue_struct *cwq,
522 unsigned long extra_flags)
524 set_work_data(work, (unsigned long)cwq,
525 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
528 static void set_work_cpu(struct work_struct *work, unsigned int cpu)
530 set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
533 static void clear_work_data(struct work_struct *work)
535 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
538 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
540 unsigned long data = atomic_long_read(&work->data);
542 if (data & WORK_STRUCT_CWQ)
543 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
548 static struct global_cwq *get_work_gcwq(struct work_struct *work)
550 unsigned long data = atomic_long_read(&work->data);
553 if (data & WORK_STRUCT_CWQ)
554 return ((struct cpu_workqueue_struct *)
555 (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
557 cpu = data >> WORK_STRUCT_FLAG_BITS;
558 if (cpu == WORK_CPU_NONE)
561 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
562 return get_gcwq(cpu);
566 * Policy functions. These define the policies on how the global
567 * worker pool is managed. Unless noted otherwise, these functions
568 * assume that they're being called with gcwq->lock held.
571 static bool __need_more_worker(struct global_cwq *gcwq)
573 return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) ||
574 gcwq->flags & GCWQ_HIGHPRI_PENDING;
578 * Need to wake up a worker? Called from anything but currently
581 static bool need_more_worker(struct global_cwq *gcwq)
583 return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
586 /* Can I start working? Called from busy but !running workers. */
587 static bool may_start_working(struct global_cwq *gcwq)
589 return gcwq->nr_idle;
592 /* Do I need to keep working? Called from currently running workers. */
593 static bool keep_working(struct global_cwq *gcwq)
595 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
597 return !list_empty(&gcwq->worklist) && atomic_read(nr_running) <= 1;
600 /* Do we need a new worker? Called from manager. */
601 static bool need_to_create_worker(struct global_cwq *gcwq)
603 return need_more_worker(gcwq) && !may_start_working(gcwq);
606 /* Do I need to be the manager? */
607 static bool need_to_manage_workers(struct global_cwq *gcwq)
609 return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS;
612 /* Do we have too many workers and should some go away? */
613 static bool too_many_workers(struct global_cwq *gcwq)
615 bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
616 int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
617 int nr_busy = gcwq->nr_workers - nr_idle;
619 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
626 /* Return the first worker. Safe with preemption disabled */
627 static struct worker *first_worker(struct global_cwq *gcwq)
629 if (unlikely(list_empty(&gcwq->idle_list)))
632 return list_first_entry(&gcwq->idle_list, struct worker, entry);
636 * wake_up_worker - wake up an idle worker
637 * @gcwq: gcwq to wake worker for
639 * Wake up the first idle worker of @gcwq.
642 * spin_lock_irq(gcwq->lock).
644 static void wake_up_worker(struct global_cwq *gcwq)
646 struct worker *worker = first_worker(gcwq);
649 wake_up_process(worker->task);
653 * wq_worker_waking_up - a worker is waking up
654 * @task: task waking up
655 * @cpu: CPU @task is waking up to
657 * This function is called during try_to_wake_up() when a worker is
661 * spin_lock_irq(rq->lock)
663 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
665 struct worker *worker = kthread_data(task);
667 if (likely(!(worker->flags & WORKER_NOT_RUNNING)))
668 atomic_inc(get_gcwq_nr_running(cpu));
672 * wq_worker_sleeping - a worker is going to sleep
673 * @task: task going to sleep
674 * @cpu: CPU in question, must be the current CPU number
676 * This function is called during schedule() when a busy worker is
677 * going to sleep. Worker on the same cpu can be woken up by
678 * returning pointer to its task.
681 * spin_lock_irq(rq->lock)
684 * Worker task on @cpu to wake up, %NULL if none.
686 struct task_struct *wq_worker_sleeping(struct task_struct *task,
689 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
690 struct global_cwq *gcwq = get_gcwq(cpu);
691 atomic_t *nr_running = get_gcwq_nr_running(cpu);
693 if (unlikely(worker->flags & WORKER_NOT_RUNNING))
696 /* this can only happen on the local cpu */
697 BUG_ON(cpu != raw_smp_processor_id());
700 * The counterpart of the following dec_and_test, implied mb,
701 * worklist not empty test sequence is in insert_work().
702 * Please read comment there.
704 * NOT_RUNNING is clear. This means that trustee is not in
705 * charge and we're running on the local cpu w/ rq lock held
706 * and preemption disabled, which in turn means that none else
707 * could be manipulating idle_list, so dereferencing idle_list
708 * without gcwq lock is safe.
710 if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
711 to_wakeup = first_worker(gcwq);
712 return to_wakeup ? to_wakeup->task : NULL;
716 * worker_set_flags - set worker flags and adjust nr_running accordingly
718 * @flags: flags to set
719 * @wakeup: wakeup an idle worker if necessary
721 * Set @flags in @worker->flags and adjust nr_running accordingly. If
722 * nr_running becomes zero and @wakeup is %true, an idle worker is
726 * spin_lock_irq(gcwq->lock)
728 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
731 struct global_cwq *gcwq = worker->gcwq;
733 WARN_ON_ONCE(worker->task != current);
736 * If transitioning into NOT_RUNNING, adjust nr_running and
737 * wake up an idle worker as necessary if requested by
740 if ((flags & WORKER_NOT_RUNNING) &&
741 !(worker->flags & WORKER_NOT_RUNNING)) {
742 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
745 if (atomic_dec_and_test(nr_running) &&
746 !list_empty(&gcwq->worklist))
747 wake_up_worker(gcwq);
749 atomic_dec(nr_running);
752 worker->flags |= flags;
756 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
758 * @flags: flags to clear
760 * Clear @flags in @worker->flags and adjust nr_running accordingly.
763 * spin_lock_irq(gcwq->lock)
765 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
767 struct global_cwq *gcwq = worker->gcwq;
768 unsigned int oflags = worker->flags;
770 WARN_ON_ONCE(worker->task != current);
772 worker->flags &= ~flags;
774 /* if transitioning out of NOT_RUNNING, increment nr_running */
775 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
776 if (!(worker->flags & WORKER_NOT_RUNNING))
777 atomic_inc(get_gcwq_nr_running(gcwq->cpu));
781 * busy_worker_head - return the busy hash head for a work
782 * @gcwq: gcwq of interest
783 * @work: work to be hashed
785 * Return hash head of @gcwq for @work.
788 * spin_lock_irq(gcwq->lock).
791 * Pointer to the hash head.
793 static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
794 struct work_struct *work)
796 const int base_shift = ilog2(sizeof(struct work_struct));
797 unsigned long v = (unsigned long)work;
799 /* simple shift and fold hash, do we need something better? */
801 v += v >> BUSY_WORKER_HASH_ORDER;
802 v &= BUSY_WORKER_HASH_MASK;
804 return &gcwq->busy_hash[v];
808 * __find_worker_executing_work - find worker which is executing a work
809 * @gcwq: gcwq of interest
810 * @bwh: hash head as returned by busy_worker_head()
811 * @work: work to find worker for
813 * Find a worker which is executing @work on @gcwq. @bwh should be
814 * the hash head obtained by calling busy_worker_head() with the same
818 * spin_lock_irq(gcwq->lock).
821 * Pointer to worker which is executing @work if found, NULL
824 static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
825 struct hlist_head *bwh,
826 struct work_struct *work)
828 struct worker *worker;
829 struct hlist_node *tmp;
831 hlist_for_each_entry(worker, tmp, bwh, hentry)
832 if (worker->current_work == work)
838 * find_worker_executing_work - find worker which is executing a work
839 * @gcwq: gcwq of interest
840 * @work: work to find worker for
842 * Find a worker which is executing @work on @gcwq. This function is
843 * identical to __find_worker_executing_work() except that this
844 * function calculates @bwh itself.
847 * spin_lock_irq(gcwq->lock).
850 * Pointer to worker which is executing @work if found, NULL
853 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
854 struct work_struct *work)
856 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
861 * gcwq_determine_ins_pos - find insertion position
862 * @gcwq: gcwq of interest
863 * @cwq: cwq a work is being queued for
865 * A work for @cwq is about to be queued on @gcwq, determine insertion
866 * position for the work. If @cwq is for HIGHPRI wq, the work is
867 * queued at the head of the queue but in FIFO order with respect to
868 * other HIGHPRI works; otherwise, at the end of the queue. This
869 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
870 * there are HIGHPRI works pending.
873 * spin_lock_irq(gcwq->lock).
876 * Pointer to inserstion position.
878 static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq,
879 struct cpu_workqueue_struct *cwq)
881 struct work_struct *twork;
883 if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
884 return &gcwq->worklist;
886 list_for_each_entry(twork, &gcwq->worklist, entry) {
887 struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
889 if (!(tcwq->wq->flags & WQ_HIGHPRI))
893 gcwq->flags |= GCWQ_HIGHPRI_PENDING;
894 return &twork->entry;
898 * insert_work - insert a work into gcwq
899 * @cwq: cwq @work belongs to
900 * @work: work to insert
901 * @head: insertion point
902 * @extra_flags: extra WORK_STRUCT_* flags to set
904 * Insert @work which belongs to @cwq into @gcwq after @head.
905 * @extra_flags is or'd to work_struct flags.
908 * spin_lock_irq(gcwq->lock).
910 static void insert_work(struct cpu_workqueue_struct *cwq,
911 struct work_struct *work, struct list_head *head,
912 unsigned int extra_flags)
914 struct global_cwq *gcwq = cwq->gcwq;
916 /* we own @work, set data and link */
917 set_work_cwq(work, cwq, extra_flags);
920 * Ensure that we get the right work->data if we see the
921 * result of list_add() below, see try_to_grab_pending().
925 list_add_tail(&work->entry, head);
928 * Ensure either worker_sched_deactivated() sees the above
929 * list_add_tail() or we see zero nr_running to avoid workers
930 * lying around lazily while there are works to be processed.
934 if (__need_more_worker(gcwq))
935 wake_up_worker(gcwq);
938 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
939 struct work_struct *work)
941 struct global_cwq *gcwq;
942 struct cpu_workqueue_struct *cwq;
943 struct list_head *worklist;
944 unsigned int work_flags;
947 debug_work_activate(work);
949 if (WARN_ON_ONCE(wq->flags & WQ_DYING))
952 /* determine gcwq to use */
953 if (!(wq->flags & WQ_UNBOUND)) {
954 struct global_cwq *last_gcwq;
956 if (unlikely(cpu == WORK_CPU_UNBOUND))
957 cpu = raw_smp_processor_id();
960 * It's multi cpu. If @wq is non-reentrant and @work
961 * was previously on a different cpu, it might still
962 * be running there, in which case the work needs to
963 * be queued on that cpu to guarantee non-reentrance.
965 gcwq = get_gcwq(cpu);
966 if (wq->flags & WQ_NON_REENTRANT &&
967 (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
968 struct worker *worker;
970 spin_lock_irqsave(&last_gcwq->lock, flags);
972 worker = find_worker_executing_work(last_gcwq, work);
974 if (worker && worker->current_cwq->wq == wq)
977 /* meh... not running there, queue here */
978 spin_unlock_irqrestore(&last_gcwq->lock, flags);
979 spin_lock_irqsave(&gcwq->lock, flags);
982 spin_lock_irqsave(&gcwq->lock, flags);
984 gcwq = get_gcwq(WORK_CPU_UNBOUND);
985 spin_lock_irqsave(&gcwq->lock, flags);
988 /* gcwq determined, get cwq and queue */
989 cwq = get_cwq(gcwq->cpu, wq);
991 BUG_ON(!list_empty(&work->entry));
993 cwq->nr_in_flight[cwq->work_color]++;
994 work_flags = work_color_to_flags(cwq->work_color);
996 if (likely(cwq->nr_active < cwq->max_active)) {
998 worklist = gcwq_determine_ins_pos(gcwq, cwq);
1000 work_flags |= WORK_STRUCT_DELAYED;
1001 worklist = &cwq->delayed_works;
1004 insert_work(cwq, work, worklist, work_flags);
1006 spin_unlock_irqrestore(&gcwq->lock, flags);
1010 * queue_work - queue work on a workqueue
1011 * @wq: workqueue to use
1012 * @work: work to queue
1014 * Returns 0 if @work was already on a queue, non-zero otherwise.
1016 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1017 * it can be processed by another CPU.
1019 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
1023 ret = queue_work_on(get_cpu(), wq, work);
1028 EXPORT_SYMBOL_GPL(queue_work);
1031 * queue_work_on - queue work on specific cpu
1032 * @cpu: CPU number to execute work on
1033 * @wq: workqueue to use
1034 * @work: work to queue
1036 * Returns 0 if @work was already on a queue, non-zero otherwise.
1038 * We queue the work to a specific CPU, the caller must ensure it
1042 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
1046 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1047 __queue_work(cpu, wq, work);
1052 EXPORT_SYMBOL_GPL(queue_work_on);
1054 static void delayed_work_timer_fn(unsigned long __data)
1056 struct delayed_work *dwork = (struct delayed_work *)__data;
1057 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1059 __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
1063 * queue_delayed_work - queue work on a workqueue after delay
1064 * @wq: workqueue to use
1065 * @dwork: delayable work to queue
1066 * @delay: number of jiffies to wait before queueing
1068 * Returns 0 if @work was already on a queue, non-zero otherwise.
1070 int queue_delayed_work(struct workqueue_struct *wq,
1071 struct delayed_work *dwork, unsigned long delay)
1074 return queue_work(wq, &dwork->work);
1076 return queue_delayed_work_on(-1, wq, dwork, delay);
1078 EXPORT_SYMBOL_GPL(queue_delayed_work);
1081 * queue_delayed_work_on - queue work on specific CPU after delay
1082 * @cpu: CPU number to execute work on
1083 * @wq: workqueue to use
1084 * @dwork: work to queue
1085 * @delay: number of jiffies to wait before queueing
1087 * Returns 0 if @work was already on a queue, non-zero otherwise.
1089 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1090 struct delayed_work *dwork, unsigned long delay)
1093 struct timer_list *timer = &dwork->timer;
1094 struct work_struct *work = &dwork->work;
1096 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1099 BUG_ON(timer_pending(timer));
1100 BUG_ON(!list_empty(&work->entry));
1102 timer_stats_timer_set_start_info(&dwork->timer);
1105 * This stores cwq for the moment, for the timer_fn.
1106 * Note that the work's gcwq is preserved to allow
1107 * reentrance detection for delayed works.
1109 if (!(wq->flags & WQ_UNBOUND)) {
1110 struct global_cwq *gcwq = get_work_gcwq(work);
1112 if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1115 lcpu = raw_smp_processor_id();
1117 lcpu = WORK_CPU_UNBOUND;
1119 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1121 timer->expires = jiffies + delay;
1122 timer->data = (unsigned long)dwork;
1123 timer->function = delayed_work_timer_fn;
1125 if (unlikely(cpu >= 0))
1126 add_timer_on(timer, cpu);
1133 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1136 * worker_enter_idle - enter idle state
1137 * @worker: worker which is entering idle state
1139 * @worker is entering idle state. Update stats and idle timer if
1143 * spin_lock_irq(gcwq->lock).
1145 static void worker_enter_idle(struct worker *worker)
1147 struct global_cwq *gcwq = worker->gcwq;
1149 BUG_ON(worker->flags & WORKER_IDLE);
1150 BUG_ON(!list_empty(&worker->entry) &&
1151 (worker->hentry.next || worker->hentry.pprev));
1153 /* can't use worker_set_flags(), also called from start_worker() */
1154 worker->flags |= WORKER_IDLE;
1156 worker->last_active = jiffies;
1158 /* idle_list is LIFO */
1159 list_add(&worker->entry, &gcwq->idle_list);
1161 if (likely(!(worker->flags & WORKER_ROGUE))) {
1162 if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1163 mod_timer(&gcwq->idle_timer,
1164 jiffies + IDLE_WORKER_TIMEOUT);
1166 wake_up_all(&gcwq->trustee_wait);
1168 /* sanity check nr_running */
1169 WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle &&
1170 atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1174 * worker_leave_idle - leave idle state
1175 * @worker: worker which is leaving idle state
1177 * @worker is leaving idle state. Update stats.
1180 * spin_lock_irq(gcwq->lock).
1182 static void worker_leave_idle(struct worker *worker)
1184 struct global_cwq *gcwq = worker->gcwq;
1186 BUG_ON(!(worker->flags & WORKER_IDLE));
1187 worker_clr_flags(worker, WORKER_IDLE);
1189 list_del_init(&worker->entry);
1193 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1196 * Works which are scheduled while the cpu is online must at least be
1197 * scheduled to a worker which is bound to the cpu so that if they are
1198 * flushed from cpu callbacks while cpu is going down, they are
1199 * guaranteed to execute on the cpu.
1201 * This function is to be used by rogue workers and rescuers to bind
1202 * themselves to the target cpu and may race with cpu going down or
1203 * coming online. kthread_bind() can't be used because it may put the
1204 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1205 * verbatim as it's best effort and blocking and gcwq may be
1206 * [dis]associated in the meantime.
1208 * This function tries set_cpus_allowed() and locks gcwq and verifies
1209 * the binding against GCWQ_DISASSOCIATED which is set during
1210 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1211 * idle state or fetches works without dropping lock, it can guarantee
1212 * the scheduling requirement described in the first paragraph.
1215 * Might sleep. Called without any lock but returns with gcwq->lock
1219 * %true if the associated gcwq is online (@worker is successfully
1220 * bound), %false if offline.
1222 static bool worker_maybe_bind_and_lock(struct worker *worker)
1223 __acquires(&gcwq->lock)
1225 struct global_cwq *gcwq = worker->gcwq;
1226 struct task_struct *task = worker->task;
1230 * The following call may fail, succeed or succeed
1231 * without actually migrating the task to the cpu if
1232 * it races with cpu hotunplug operation. Verify
1233 * against GCWQ_DISASSOCIATED.
1235 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1236 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1238 spin_lock_irq(&gcwq->lock);
1239 if (gcwq->flags & GCWQ_DISASSOCIATED)
1241 if (task_cpu(task) == gcwq->cpu &&
1242 cpumask_equal(¤t->cpus_allowed,
1243 get_cpu_mask(gcwq->cpu)))
1245 spin_unlock_irq(&gcwq->lock);
1247 /* CPU has come up inbetween, retry migration */
1253 * Function for worker->rebind_work used to rebind rogue busy workers
1254 * to the associated cpu which is coming back online. This is
1255 * scheduled by cpu up but can race with other cpu hotplug operations
1256 * and may be executed twice without intervening cpu down.
1258 static void worker_rebind_fn(struct work_struct *work)
1260 struct worker *worker = container_of(work, struct worker, rebind_work);
1261 struct global_cwq *gcwq = worker->gcwq;
1263 if (worker_maybe_bind_and_lock(worker))
1264 worker_clr_flags(worker, WORKER_REBIND);
1266 spin_unlock_irq(&gcwq->lock);
1269 static struct worker *alloc_worker(void)
1271 struct worker *worker;
1273 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1275 INIT_LIST_HEAD(&worker->entry);
1276 INIT_LIST_HEAD(&worker->scheduled);
1277 INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1278 /* on creation a worker is in !idle && prep state */
1279 worker->flags = WORKER_PREP;
1285 * create_worker - create a new workqueue worker
1286 * @gcwq: gcwq the new worker will belong to
1287 * @bind: whether to set affinity to @cpu or not
1289 * Create a new worker which is bound to @gcwq. The returned worker
1290 * can be started by calling start_worker() or destroyed using
1294 * Might sleep. Does GFP_KERNEL allocations.
1297 * Pointer to the newly created worker.
1299 static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
1301 bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1302 struct worker *worker = NULL;
1305 spin_lock_irq(&gcwq->lock);
1306 while (ida_get_new(&gcwq->worker_ida, &id)) {
1307 spin_unlock_irq(&gcwq->lock);
1308 if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1310 spin_lock_irq(&gcwq->lock);
1312 spin_unlock_irq(&gcwq->lock);
1314 worker = alloc_worker();
1318 worker->gcwq = gcwq;
1321 if (!on_unbound_cpu)
1322 worker->task = kthread_create(worker_thread, worker,
1323 "kworker/%u:%d", gcwq->cpu, id);
1325 worker->task = kthread_create(worker_thread, worker,
1326 "kworker/u:%d", id);
1327 if (IS_ERR(worker->task))
1331 * A rogue worker will become a regular one if CPU comes
1332 * online later on. Make sure every worker has
1333 * PF_THREAD_BOUND set.
1335 if (bind && !on_unbound_cpu)
1336 kthread_bind(worker->task, gcwq->cpu);
1338 worker->task->flags |= PF_THREAD_BOUND;
1340 worker->flags |= WORKER_UNBOUND;
1346 spin_lock_irq(&gcwq->lock);
1347 ida_remove(&gcwq->worker_ida, id);
1348 spin_unlock_irq(&gcwq->lock);
1355 * start_worker - start a newly created worker
1356 * @worker: worker to start
1358 * Make the gcwq aware of @worker and start it.
1361 * spin_lock_irq(gcwq->lock).
1363 static void start_worker(struct worker *worker)
1365 worker->flags |= WORKER_STARTED;
1366 worker->gcwq->nr_workers++;
1367 worker_enter_idle(worker);
1368 wake_up_process(worker->task);
1372 * destroy_worker - destroy a workqueue worker
1373 * @worker: worker to be destroyed
1375 * Destroy @worker and adjust @gcwq stats accordingly.
1378 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1380 static void destroy_worker(struct worker *worker)
1382 struct global_cwq *gcwq = worker->gcwq;
1383 int id = worker->id;
1385 /* sanity check frenzy */
1386 BUG_ON(worker->current_work);
1387 BUG_ON(!list_empty(&worker->scheduled));
1389 if (worker->flags & WORKER_STARTED)
1391 if (worker->flags & WORKER_IDLE)
1394 list_del_init(&worker->entry);
1395 worker->flags |= WORKER_DIE;
1397 spin_unlock_irq(&gcwq->lock);
1399 kthread_stop(worker->task);
1402 spin_lock_irq(&gcwq->lock);
1403 ida_remove(&gcwq->worker_ida, id);
1406 static void idle_worker_timeout(unsigned long __gcwq)
1408 struct global_cwq *gcwq = (void *)__gcwq;
1410 spin_lock_irq(&gcwq->lock);
1412 if (too_many_workers(gcwq)) {
1413 struct worker *worker;
1414 unsigned long expires;
1416 /* idle_list is kept in LIFO order, check the last one */
1417 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1418 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1420 if (time_before(jiffies, expires))
1421 mod_timer(&gcwq->idle_timer, expires);
1423 /* it's been idle for too long, wake up manager */
1424 gcwq->flags |= GCWQ_MANAGE_WORKERS;
1425 wake_up_worker(gcwq);
1429 spin_unlock_irq(&gcwq->lock);
1432 static bool send_mayday(struct work_struct *work)
1434 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1435 struct workqueue_struct *wq = cwq->wq;
1438 if (!(wq->flags & WQ_RESCUER))
1441 /* mayday mayday mayday */
1442 cpu = cwq->gcwq->cpu;
1443 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1444 if (cpu == WORK_CPU_UNBOUND)
1446 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1447 wake_up_process(wq->rescuer->task);
1451 static void gcwq_mayday_timeout(unsigned long __gcwq)
1453 struct global_cwq *gcwq = (void *)__gcwq;
1454 struct work_struct *work;
1456 spin_lock_irq(&gcwq->lock);
1458 if (need_to_create_worker(gcwq)) {
1460 * We've been trying to create a new worker but
1461 * haven't been successful. We might be hitting an
1462 * allocation deadlock. Send distress signals to
1465 list_for_each_entry(work, &gcwq->worklist, entry)
1469 spin_unlock_irq(&gcwq->lock);
1471 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1475 * maybe_create_worker - create a new worker if necessary
1476 * @gcwq: gcwq to create a new worker for
1478 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1479 * have at least one idle worker on return from this function. If
1480 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1481 * sent to all rescuers with works scheduled on @gcwq to resolve
1482 * possible allocation deadlock.
1484 * On return, need_to_create_worker() is guaranteed to be false and
1485 * may_start_working() true.
1488 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1489 * multiple times. Does GFP_KERNEL allocations. Called only from
1493 * false if no action was taken and gcwq->lock stayed locked, true
1496 static bool maybe_create_worker(struct global_cwq *gcwq)
1497 __releases(&gcwq->lock)
1498 __acquires(&gcwq->lock)
1500 if (!need_to_create_worker(gcwq))
1503 spin_unlock_irq(&gcwq->lock);
1505 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1506 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1509 struct worker *worker;
1511 worker = create_worker(gcwq, true);
1513 del_timer_sync(&gcwq->mayday_timer);
1514 spin_lock_irq(&gcwq->lock);
1515 start_worker(worker);
1516 BUG_ON(need_to_create_worker(gcwq));
1520 if (!need_to_create_worker(gcwq))
1523 __set_current_state(TASK_INTERRUPTIBLE);
1524 schedule_timeout(CREATE_COOLDOWN);
1526 if (!need_to_create_worker(gcwq))
1530 del_timer_sync(&gcwq->mayday_timer);
1531 spin_lock_irq(&gcwq->lock);
1532 if (need_to_create_worker(gcwq))
1538 * maybe_destroy_worker - destroy workers which have been idle for a while
1539 * @gcwq: gcwq to destroy workers for
1541 * Destroy @gcwq workers which have been idle for longer than
1542 * IDLE_WORKER_TIMEOUT.
1545 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1546 * multiple times. Called only from manager.
1549 * false if no action was taken and gcwq->lock stayed locked, true
1552 static bool maybe_destroy_workers(struct global_cwq *gcwq)
1556 while (too_many_workers(gcwq)) {
1557 struct worker *worker;
1558 unsigned long expires;
1560 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1561 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1563 if (time_before(jiffies, expires)) {
1564 mod_timer(&gcwq->idle_timer, expires);
1568 destroy_worker(worker);
1576 * manage_workers - manage worker pool
1579 * Assume the manager role and manage gcwq worker pool @worker belongs
1580 * to. At any given time, there can be only zero or one manager per
1581 * gcwq. The exclusion is handled automatically by this function.
1583 * The caller can safely start processing works on false return. On
1584 * true return, it's guaranteed that need_to_create_worker() is false
1585 * and may_start_working() is true.
1588 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1589 * multiple times. Does GFP_KERNEL allocations.
1592 * false if no action was taken and gcwq->lock stayed locked, true if
1593 * some action was taken.
1595 static bool manage_workers(struct worker *worker)
1597 struct global_cwq *gcwq = worker->gcwq;
1600 if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1603 gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1604 gcwq->flags |= GCWQ_MANAGING_WORKERS;
1607 * Destroy and then create so that may_start_working() is true
1610 ret |= maybe_destroy_workers(gcwq);
1611 ret |= maybe_create_worker(gcwq);
1613 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1616 * The trustee might be waiting to take over the manager
1617 * position, tell it we're done.
1619 if (unlikely(gcwq->trustee))
1620 wake_up_all(&gcwq->trustee_wait);
1626 * move_linked_works - move linked works to a list
1627 * @work: start of series of works to be scheduled
1628 * @head: target list to append @work to
1629 * @nextp: out paramter for nested worklist walking
1631 * Schedule linked works starting from @work to @head. Work series to
1632 * be scheduled starts at @work and includes any consecutive work with
1633 * WORK_STRUCT_LINKED set in its predecessor.
1635 * If @nextp is not NULL, it's updated to point to the next work of
1636 * the last scheduled work. This allows move_linked_works() to be
1637 * nested inside outer list_for_each_entry_safe().
1640 * spin_lock_irq(gcwq->lock).
1642 static void move_linked_works(struct work_struct *work, struct list_head *head,
1643 struct work_struct **nextp)
1645 struct work_struct *n;
1648 * Linked worklist will always end before the end of the list,
1649 * use NULL for list head.
1651 list_for_each_entry_safe_from(work, n, NULL, entry) {
1652 list_move_tail(&work->entry, head);
1653 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1658 * If we're already inside safe list traversal and have moved
1659 * multiple works to the scheduled queue, the next position
1660 * needs to be updated.
1666 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1668 struct work_struct *work = list_first_entry(&cwq->delayed_works,
1669 struct work_struct, entry);
1670 struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1672 move_linked_works(work, pos, NULL);
1673 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1678 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1679 * @cwq: cwq of interest
1680 * @color: color of work which left the queue
1681 * @delayed: for a delayed work
1683 * A work either has completed or is removed from pending queue,
1684 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1687 * spin_lock_irq(gcwq->lock).
1689 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1692 /* ignore uncolored works */
1693 if (color == WORK_NO_COLOR)
1696 cwq->nr_in_flight[color]--;
1700 if (!list_empty(&cwq->delayed_works)) {
1701 /* one down, submit a delayed one */
1702 if (cwq->nr_active < cwq->max_active)
1703 cwq_activate_first_delayed(cwq);
1707 /* is flush in progress and are we at the flushing tip? */
1708 if (likely(cwq->flush_color != color))
1711 /* are there still in-flight works? */
1712 if (cwq->nr_in_flight[color])
1715 /* this cwq is done, clear flush_color */
1716 cwq->flush_color = -1;
1719 * If this was the last cwq, wake up the first flusher. It
1720 * will handle the rest.
1722 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1723 complete(&cwq->wq->first_flusher->done);
1727 * process_one_work - process single work
1729 * @work: work to process
1731 * Process @work. This function contains all the logics necessary to
1732 * process a single work including synchronization against and
1733 * interaction with other workers on the same cpu, queueing and
1734 * flushing. As long as context requirement is met, any worker can
1735 * call this function to process a work.
1738 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1740 static void process_one_work(struct worker *worker, struct work_struct *work)
1741 __releases(&gcwq->lock)
1742 __acquires(&gcwq->lock)
1744 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1745 struct global_cwq *gcwq = cwq->gcwq;
1746 struct hlist_head *bwh = busy_worker_head(gcwq, work);
1747 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1748 work_func_t f = work->func;
1750 struct worker *collision;
1751 #ifdef CONFIG_LOCKDEP
1753 * It is permissible to free the struct work_struct from
1754 * inside the function that is called from it, this we need to
1755 * take into account for lockdep too. To avoid bogus "held
1756 * lock freed" warnings as well as problems when looking into
1757 * work->lockdep_map, make a copy and use that here.
1759 struct lockdep_map lockdep_map = work->lockdep_map;
1762 * A single work shouldn't be executed concurrently by
1763 * multiple workers on a single cpu. Check whether anyone is
1764 * already processing the work. If so, defer the work to the
1765 * currently executing one.
1767 collision = __find_worker_executing_work(gcwq, bwh, work);
1768 if (unlikely(collision)) {
1769 move_linked_works(work, &collision->scheduled, NULL);
1773 /* claim and process */
1774 debug_work_deactivate(work);
1775 hlist_add_head(&worker->hentry, bwh);
1776 worker->current_work = work;
1777 worker->current_cwq = cwq;
1778 work_color = get_work_color(work);
1780 /* record the current cpu number in the work data and dequeue */
1781 set_work_cpu(work, gcwq->cpu);
1782 list_del_init(&work->entry);
1785 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1786 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1788 if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1789 struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1790 struct work_struct, entry);
1792 if (!list_empty(&gcwq->worklist) &&
1793 get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1794 wake_up_worker(gcwq);
1796 gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1800 * CPU intensive works don't participate in concurrency
1801 * management. They're the scheduler's responsibility.
1803 if (unlikely(cpu_intensive))
1804 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1806 spin_unlock_irq(&gcwq->lock);
1808 work_clear_pending(work);
1809 lock_map_acquire(&cwq->wq->lockdep_map);
1810 lock_map_acquire(&lockdep_map);
1812 lock_map_release(&lockdep_map);
1813 lock_map_release(&cwq->wq->lockdep_map);
1815 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1816 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1818 current->comm, preempt_count(), task_pid_nr(current));
1819 printk(KERN_ERR " last function: ");
1820 print_symbol("%s\n", (unsigned long)f);
1821 debug_show_held_locks(current);
1825 spin_lock_irq(&gcwq->lock);
1827 /* clear cpu intensive status */
1828 if (unlikely(cpu_intensive))
1829 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1831 /* we're done with it, release */
1832 hlist_del_init(&worker->hentry);
1833 worker->current_work = NULL;
1834 worker->current_cwq = NULL;
1835 cwq_dec_nr_in_flight(cwq, work_color, false);
1839 * process_scheduled_works - process scheduled works
1842 * Process all scheduled works. Please note that the scheduled list
1843 * may change while processing a work, so this function repeatedly
1844 * fetches a work from the top and executes it.
1847 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1850 static void process_scheduled_works(struct worker *worker)
1852 while (!list_empty(&worker->scheduled)) {
1853 struct work_struct *work = list_first_entry(&worker->scheduled,
1854 struct work_struct, entry);
1855 process_one_work(worker, work);
1860 * worker_thread - the worker thread function
1863 * The gcwq worker thread function. There's a single dynamic pool of
1864 * these per each cpu. These workers process all works regardless of
1865 * their specific target workqueue. The only exception is works which
1866 * belong to workqueues with a rescuer which will be explained in
1869 static int worker_thread(void *__worker)
1871 struct worker *worker = __worker;
1872 struct global_cwq *gcwq = worker->gcwq;
1874 /* tell the scheduler that this is a workqueue worker */
1875 worker->task->flags |= PF_WQ_WORKER;
1877 spin_lock_irq(&gcwq->lock);
1879 /* DIE can be set only while we're idle, checking here is enough */
1880 if (worker->flags & WORKER_DIE) {
1881 spin_unlock_irq(&gcwq->lock);
1882 worker->task->flags &= ~PF_WQ_WORKER;
1886 worker_leave_idle(worker);
1888 /* no more worker necessary? */
1889 if (!need_more_worker(gcwq))
1892 /* do we need to manage? */
1893 if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1897 * ->scheduled list can only be filled while a worker is
1898 * preparing to process a work or actually processing it.
1899 * Make sure nobody diddled with it while I was sleeping.
1901 BUG_ON(!list_empty(&worker->scheduled));
1904 * When control reaches this point, we're guaranteed to have
1905 * at least one idle worker or that someone else has already
1906 * assumed the manager role.
1908 worker_clr_flags(worker, WORKER_PREP);
1911 struct work_struct *work =
1912 list_first_entry(&gcwq->worklist,
1913 struct work_struct, entry);
1915 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1916 /* optimization path, not strictly necessary */
1917 process_one_work(worker, work);
1918 if (unlikely(!list_empty(&worker->scheduled)))
1919 process_scheduled_works(worker);
1921 move_linked_works(work, &worker->scheduled, NULL);
1922 process_scheduled_works(worker);
1924 } while (keep_working(gcwq));
1926 worker_set_flags(worker, WORKER_PREP, false);
1928 if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1932 * gcwq->lock is held and there's no work to process and no
1933 * need to manage, sleep. Workers are woken up only while
1934 * holding gcwq->lock or from local cpu, so setting the
1935 * current state before releasing gcwq->lock is enough to
1936 * prevent losing any event.
1938 worker_enter_idle(worker);
1939 __set_current_state(TASK_INTERRUPTIBLE);
1940 spin_unlock_irq(&gcwq->lock);
1946 * rescuer_thread - the rescuer thread function
1947 * @__wq: the associated workqueue
1949 * Workqueue rescuer thread function. There's one rescuer for each
1950 * workqueue which has WQ_RESCUER set.
1952 * Regular work processing on a gcwq may block trying to create a new
1953 * worker which uses GFP_KERNEL allocation which has slight chance of
1954 * developing into deadlock if some works currently on the same queue
1955 * need to be processed to satisfy the GFP_KERNEL allocation. This is
1956 * the problem rescuer solves.
1958 * When such condition is possible, the gcwq summons rescuers of all
1959 * workqueues which have works queued on the gcwq and let them process
1960 * those works so that forward progress can be guaranteed.
1962 * This should happen rarely.
1964 static int rescuer_thread(void *__wq)
1966 struct workqueue_struct *wq = __wq;
1967 struct worker *rescuer = wq->rescuer;
1968 struct list_head *scheduled = &rescuer->scheduled;
1969 bool is_unbound = wq->flags & WQ_UNBOUND;
1972 set_user_nice(current, RESCUER_NICE_LEVEL);
1974 set_current_state(TASK_INTERRUPTIBLE);
1976 if (kthread_should_stop())
1980 * See whether any cpu is asking for help. Unbounded
1981 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
1983 for_each_mayday_cpu(cpu, wq->mayday_mask) {
1984 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
1985 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
1986 struct global_cwq *gcwq = cwq->gcwq;
1987 struct work_struct *work, *n;
1989 __set_current_state(TASK_RUNNING);
1990 mayday_clear_cpu(cpu, wq->mayday_mask);
1992 /* migrate to the target cpu if possible */
1993 rescuer->gcwq = gcwq;
1994 worker_maybe_bind_and_lock(rescuer);
1997 * Slurp in all works issued via this workqueue and
2000 BUG_ON(!list_empty(&rescuer->scheduled));
2001 list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2002 if (get_work_cwq(work) == cwq)
2003 move_linked_works(work, scheduled, &n);
2005 process_scheduled_works(rescuer);
2006 spin_unlock_irq(&gcwq->lock);
2014 struct work_struct work;
2015 struct completion done;
2018 static void wq_barrier_func(struct work_struct *work)
2020 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2021 complete(&barr->done);
2025 * insert_wq_barrier - insert a barrier work
2026 * @cwq: cwq to insert barrier into
2027 * @barr: wq_barrier to insert
2028 * @target: target work to attach @barr to
2029 * @worker: worker currently executing @target, NULL if @target is not executing
2031 * @barr is linked to @target such that @barr is completed only after
2032 * @target finishes execution. Please note that the ordering
2033 * guarantee is observed only with respect to @target and on the local
2036 * Currently, a queued barrier can't be canceled. This is because
2037 * try_to_grab_pending() can't determine whether the work to be
2038 * grabbed is at the head of the queue and thus can't clear LINKED
2039 * flag of the previous work while there must be a valid next work
2040 * after a work with LINKED flag set.
2042 * Note that when @worker is non-NULL, @target may be modified
2043 * underneath us, so we can't reliably determine cwq from @target.
2046 * spin_lock_irq(gcwq->lock).
2048 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2049 struct wq_barrier *barr,
2050 struct work_struct *target, struct worker *worker)
2052 struct list_head *head;
2053 unsigned int linked = 0;
2056 * debugobject calls are safe here even with gcwq->lock locked
2057 * as we know for sure that this will not trigger any of the
2058 * checks and call back into the fixup functions where we
2061 INIT_WORK_ON_STACK(&barr->work, wq_barrier_func);
2062 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2063 init_completion(&barr->done);
2066 * If @target is currently being executed, schedule the
2067 * barrier to the worker; otherwise, put it after @target.
2070 head = worker->scheduled.next;
2072 unsigned long *bits = work_data_bits(target);
2074 head = target->entry.next;
2075 /* there can already be other linked works, inherit and set */
2076 linked = *bits & WORK_STRUCT_LINKED;
2077 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2080 debug_work_activate(&barr->work);
2081 insert_work(cwq, &barr->work, head,
2082 work_color_to_flags(WORK_NO_COLOR) | linked);
2086 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2087 * @wq: workqueue being flushed
2088 * @flush_color: new flush color, < 0 for no-op
2089 * @work_color: new work color, < 0 for no-op
2091 * Prepare cwqs for workqueue flushing.
2093 * If @flush_color is non-negative, flush_color on all cwqs should be
2094 * -1. If no cwq has in-flight commands at the specified color, all
2095 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2096 * has in flight commands, its cwq->flush_color is set to
2097 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2098 * wakeup logic is armed and %true is returned.
2100 * The caller should have initialized @wq->first_flusher prior to
2101 * calling this function with non-negative @flush_color. If
2102 * @flush_color is negative, no flush color update is done and %false
2105 * If @work_color is non-negative, all cwqs should have the same
2106 * work_color which is previous to @work_color and all will be
2107 * advanced to @work_color.
2110 * mutex_lock(wq->flush_mutex).
2113 * %true if @flush_color >= 0 and there's something to flush. %false
2116 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2117 int flush_color, int work_color)
2122 if (flush_color >= 0) {
2123 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2124 atomic_set(&wq->nr_cwqs_to_flush, 1);
2127 for_each_cwq_cpu(cpu, wq) {
2128 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2129 struct global_cwq *gcwq = cwq->gcwq;
2131 spin_lock_irq(&gcwq->lock);
2133 if (flush_color >= 0) {
2134 BUG_ON(cwq->flush_color != -1);
2136 if (cwq->nr_in_flight[flush_color]) {
2137 cwq->flush_color = flush_color;
2138 atomic_inc(&wq->nr_cwqs_to_flush);
2143 if (work_color >= 0) {
2144 BUG_ON(work_color != work_next_color(cwq->work_color));
2145 cwq->work_color = work_color;
2148 spin_unlock_irq(&gcwq->lock);
2151 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2152 complete(&wq->first_flusher->done);
2158 * flush_workqueue - ensure that any scheduled work has run to completion.
2159 * @wq: workqueue to flush
2161 * Forces execution of the workqueue and blocks until its completion.
2162 * This is typically used in driver shutdown handlers.
2164 * We sleep until all works which were queued on entry have been handled,
2165 * but we are not livelocked by new incoming ones.
2167 void flush_workqueue(struct workqueue_struct *wq)
2169 struct wq_flusher this_flusher = {
2170 .list = LIST_HEAD_INIT(this_flusher.list),
2172 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2176 lock_map_acquire(&wq->lockdep_map);
2177 lock_map_release(&wq->lockdep_map);
2179 mutex_lock(&wq->flush_mutex);
2182 * Start-to-wait phase
2184 next_color = work_next_color(wq->work_color);
2186 if (next_color != wq->flush_color) {
2188 * Color space is not full. The current work_color
2189 * becomes our flush_color and work_color is advanced
2192 BUG_ON(!list_empty(&wq->flusher_overflow));
2193 this_flusher.flush_color = wq->work_color;
2194 wq->work_color = next_color;
2196 if (!wq->first_flusher) {
2197 /* no flush in progress, become the first flusher */
2198 BUG_ON(wq->flush_color != this_flusher.flush_color);
2200 wq->first_flusher = &this_flusher;
2202 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2204 /* nothing to flush, done */
2205 wq->flush_color = next_color;
2206 wq->first_flusher = NULL;
2211 BUG_ON(wq->flush_color == this_flusher.flush_color);
2212 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2213 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2217 * Oops, color space is full, wait on overflow queue.
2218 * The next flush completion will assign us
2219 * flush_color and transfer to flusher_queue.
2221 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2224 mutex_unlock(&wq->flush_mutex);
2226 wait_for_completion(&this_flusher.done);
2229 * Wake-up-and-cascade phase
2231 * First flushers are responsible for cascading flushes and
2232 * handling overflow. Non-first flushers can simply return.
2234 if (wq->first_flusher != &this_flusher)
2237 mutex_lock(&wq->flush_mutex);
2239 /* we might have raced, check again with mutex held */
2240 if (wq->first_flusher != &this_flusher)
2243 wq->first_flusher = NULL;
2245 BUG_ON(!list_empty(&this_flusher.list));
2246 BUG_ON(wq->flush_color != this_flusher.flush_color);
2249 struct wq_flusher *next, *tmp;
2251 /* complete all the flushers sharing the current flush color */
2252 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2253 if (next->flush_color != wq->flush_color)
2255 list_del_init(&next->list);
2256 complete(&next->done);
2259 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2260 wq->flush_color != work_next_color(wq->work_color));
2262 /* this flush_color is finished, advance by one */
2263 wq->flush_color = work_next_color(wq->flush_color);
2265 /* one color has been freed, handle overflow queue */
2266 if (!list_empty(&wq->flusher_overflow)) {
2268 * Assign the same color to all overflowed
2269 * flushers, advance work_color and append to
2270 * flusher_queue. This is the start-to-wait
2271 * phase for these overflowed flushers.
2273 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2274 tmp->flush_color = wq->work_color;
2276 wq->work_color = work_next_color(wq->work_color);
2278 list_splice_tail_init(&wq->flusher_overflow,
2279 &wq->flusher_queue);
2280 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2283 if (list_empty(&wq->flusher_queue)) {
2284 BUG_ON(wq->flush_color != wq->work_color);
2289 * Need to flush more colors. Make the next flusher
2290 * the new first flusher and arm cwqs.
2292 BUG_ON(wq->flush_color == wq->work_color);
2293 BUG_ON(wq->flush_color != next->flush_color);
2295 list_del_init(&next->list);
2296 wq->first_flusher = next;
2298 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2302 * Meh... this color is already done, clear first
2303 * flusher and repeat cascading.
2305 wq->first_flusher = NULL;
2309 mutex_unlock(&wq->flush_mutex);
2311 EXPORT_SYMBOL_GPL(flush_workqueue);
2314 * flush_work - block until a work_struct's callback has terminated
2315 * @work: the work which is to be flushed
2317 * Returns false if @work has already terminated.
2319 * It is expected that, prior to calling flush_work(), the caller has
2320 * arranged for the work to not be requeued, otherwise it doesn't make
2321 * sense to use this function.
2323 int flush_work(struct work_struct *work)
2325 struct worker *worker = NULL;
2326 struct global_cwq *gcwq;
2327 struct cpu_workqueue_struct *cwq;
2328 struct wq_barrier barr;
2331 gcwq = get_work_gcwq(work);
2335 spin_lock_irq(&gcwq->lock);
2336 if (!list_empty(&work->entry)) {
2338 * See the comment near try_to_grab_pending()->smp_rmb().
2339 * If it was re-queued to a different gcwq under us, we
2340 * are not going to wait.
2343 cwq = get_work_cwq(work);
2344 if (unlikely(!cwq || gcwq != cwq->gcwq))
2347 worker = find_worker_executing_work(gcwq, work);
2350 cwq = worker->current_cwq;
2353 insert_wq_barrier(cwq, &barr, work, worker);
2354 spin_unlock_irq(&gcwq->lock);
2356 lock_map_acquire(&cwq->wq->lockdep_map);
2357 lock_map_release(&cwq->wq->lockdep_map);
2359 wait_for_completion(&barr.done);
2360 destroy_work_on_stack(&barr.work);
2363 spin_unlock_irq(&gcwq->lock);
2366 EXPORT_SYMBOL_GPL(flush_work);
2369 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2370 * so this work can't be re-armed in any way.
2372 static int try_to_grab_pending(struct work_struct *work)
2374 struct global_cwq *gcwq;
2377 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2381 * The queueing is in progress, or it is already queued. Try to
2382 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2384 gcwq = get_work_gcwq(work);
2388 spin_lock_irq(&gcwq->lock);
2389 if (!list_empty(&work->entry)) {
2391 * This work is queued, but perhaps we locked the wrong gcwq.
2392 * In that case we must see the new value after rmb(), see
2393 * insert_work()->wmb().
2396 if (gcwq == get_work_gcwq(work)) {
2397 debug_work_deactivate(work);
2398 list_del_init(&work->entry);
2399 cwq_dec_nr_in_flight(get_work_cwq(work),
2400 get_work_color(work),
2401 *work_data_bits(work) & WORK_STRUCT_DELAYED);
2405 spin_unlock_irq(&gcwq->lock);
2410 static void wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2412 struct wq_barrier barr;
2413 struct worker *worker;
2415 spin_lock_irq(&gcwq->lock);
2417 worker = find_worker_executing_work(gcwq, work);
2418 if (unlikely(worker))
2419 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2421 spin_unlock_irq(&gcwq->lock);
2423 if (unlikely(worker)) {
2424 wait_for_completion(&barr.done);
2425 destroy_work_on_stack(&barr.work);
2429 static void wait_on_work(struct work_struct *work)
2435 lock_map_acquire(&work->lockdep_map);
2436 lock_map_release(&work->lockdep_map);
2438 for_each_gcwq_cpu(cpu)
2439 wait_on_cpu_work(get_gcwq(cpu), work);
2442 static int __cancel_work_timer(struct work_struct *work,
2443 struct timer_list* timer)
2448 ret = (timer && likely(del_timer(timer)));
2450 ret = try_to_grab_pending(work);
2452 } while (unlikely(ret < 0));
2454 clear_work_data(work);
2459 * cancel_work_sync - block until a work_struct's callback has terminated
2460 * @work: the work which is to be flushed
2462 * Returns true if @work was pending.
2464 * cancel_work_sync() will cancel the work if it is queued. If the work's
2465 * callback appears to be running, cancel_work_sync() will block until it
2468 * It is possible to use this function if the work re-queues itself. It can
2469 * cancel the work even if it migrates to another workqueue, however in that
2470 * case it only guarantees that work->func() has completed on the last queued
2473 * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
2474 * pending, otherwise it goes into a busy-wait loop until the timer expires.
2476 * The caller must ensure that workqueue_struct on which this work was last
2477 * queued can't be destroyed before this function returns.
2479 int cancel_work_sync(struct work_struct *work)
2481 return __cancel_work_timer(work, NULL);
2483 EXPORT_SYMBOL_GPL(cancel_work_sync);
2486 * cancel_delayed_work_sync - reliably kill off a delayed work.
2487 * @dwork: the delayed work struct
2489 * Returns true if @dwork was pending.
2491 * It is possible to use this function if @dwork rearms itself via queue_work()
2492 * or queue_delayed_work(). See also the comment for cancel_work_sync().
2494 int cancel_delayed_work_sync(struct delayed_work *dwork)
2496 return __cancel_work_timer(&dwork->work, &dwork->timer);
2498 EXPORT_SYMBOL(cancel_delayed_work_sync);
2501 * schedule_work - put work task in global workqueue
2502 * @work: job to be done
2504 * Returns zero if @work was already on the kernel-global workqueue and
2505 * non-zero otherwise.
2507 * This puts a job in the kernel-global workqueue if it was not already
2508 * queued and leaves it in the same position on the kernel-global
2509 * workqueue otherwise.
2511 int schedule_work(struct work_struct *work)
2513 return queue_work(system_wq, work);
2515 EXPORT_SYMBOL(schedule_work);
2518 * schedule_work_on - put work task on a specific cpu
2519 * @cpu: cpu to put the work task on
2520 * @work: job to be done
2522 * This puts a job on a specific cpu
2524 int schedule_work_on(int cpu, struct work_struct *work)
2526 return queue_work_on(cpu, system_wq, work);
2528 EXPORT_SYMBOL(schedule_work_on);
2531 * schedule_delayed_work - put work task in global workqueue after delay
2532 * @dwork: job to be done
2533 * @delay: number of jiffies to wait or 0 for immediate execution
2535 * After waiting for a given time this puts a job in the kernel-global
2538 int schedule_delayed_work(struct delayed_work *dwork,
2539 unsigned long delay)
2541 return queue_delayed_work(system_wq, dwork, delay);
2543 EXPORT_SYMBOL(schedule_delayed_work);
2546 * flush_delayed_work - block until a dwork_struct's callback has terminated
2547 * @dwork: the delayed work which is to be flushed
2549 * Any timeout is cancelled, and any pending work is run immediately.
2551 void flush_delayed_work(struct delayed_work *dwork)
2553 if (del_timer_sync(&dwork->timer)) {
2554 __queue_work(get_cpu(), get_work_cwq(&dwork->work)->wq,
2558 flush_work(&dwork->work);
2560 EXPORT_SYMBOL(flush_delayed_work);
2563 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2565 * @dwork: job to be done
2566 * @delay: number of jiffies to wait
2568 * After waiting for a given time this puts a job in the kernel-global
2569 * workqueue on the specified CPU.
2571 int schedule_delayed_work_on(int cpu,
2572 struct delayed_work *dwork, unsigned long delay)
2574 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2576 EXPORT_SYMBOL(schedule_delayed_work_on);
2579 * schedule_on_each_cpu - call a function on each online CPU from keventd
2580 * @func: the function to call
2582 * Returns zero on success.
2583 * Returns -ve errno on failure.
2585 * schedule_on_each_cpu() is very slow.
2587 int schedule_on_each_cpu(work_func_t func)
2590 struct work_struct __percpu *works;
2592 works = alloc_percpu(struct work_struct);
2598 for_each_online_cpu(cpu) {
2599 struct work_struct *work = per_cpu_ptr(works, cpu);
2601 INIT_WORK(work, func);
2602 schedule_work_on(cpu, work);
2605 for_each_online_cpu(cpu)
2606 flush_work(per_cpu_ptr(works, cpu));
2614 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2616 * Forces execution of the kernel-global workqueue and blocks until its
2619 * Think twice before calling this function! It's very easy to get into
2620 * trouble if you don't take great care. Either of the following situations
2621 * will lead to deadlock:
2623 * One of the work items currently on the workqueue needs to acquire
2624 * a lock held by your code or its caller.
2626 * Your code is running in the context of a work routine.
2628 * They will be detected by lockdep when they occur, but the first might not
2629 * occur very often. It depends on what work items are on the workqueue and
2630 * what locks they need, which you have no control over.
2632 * In most situations flushing the entire workqueue is overkill; you merely
2633 * need to know that a particular work item isn't queued and isn't running.
2634 * In such cases you should use cancel_delayed_work_sync() or
2635 * cancel_work_sync() instead.
2637 void flush_scheduled_work(void)
2639 flush_workqueue(system_wq);
2641 EXPORT_SYMBOL(flush_scheduled_work);
2644 * execute_in_process_context - reliably execute the routine with user context
2645 * @fn: the function to execute
2646 * @ew: guaranteed storage for the execute work structure (must
2647 * be available when the work executes)
2649 * Executes the function immediately if process context is available,
2650 * otherwise schedules the function for delayed execution.
2652 * Returns: 0 - function was executed
2653 * 1 - function was scheduled for execution
2655 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2657 if (!in_interrupt()) {
2662 INIT_WORK(&ew->work, fn);
2663 schedule_work(&ew->work);
2667 EXPORT_SYMBOL_GPL(execute_in_process_context);
2669 int keventd_up(void)
2671 return system_wq != NULL;
2674 static int alloc_cwqs(struct workqueue_struct *wq)
2677 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2678 * Make sure that the alignment isn't lower than that of
2679 * unsigned long long.
2681 const size_t size = sizeof(struct cpu_workqueue_struct);
2682 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2683 __alignof__(unsigned long long));
2685 bool percpu = !(wq->flags & WQ_UNBOUND);
2687 bool percpu = false;
2691 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2696 * Allocate enough room to align cwq and put an extra
2697 * pointer at the end pointing back to the originally
2698 * allocated pointer which will be used for free.
2700 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2702 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2703 *(void **)(wq->cpu_wq.single + 1) = ptr;
2707 /* just in case, make sure it's actually aligned */
2708 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2709 return wq->cpu_wq.v ? 0 : -ENOMEM;
2712 static void free_cwqs(struct workqueue_struct *wq)
2715 bool percpu = !(wq->flags & WQ_UNBOUND);
2717 bool percpu = false;
2721 free_percpu(wq->cpu_wq.pcpu);
2722 else if (wq->cpu_wq.single) {
2723 /* the pointer to free is stored right after the cwq */
2724 kfree(*(void **)(wq->cpu_wq.single + 1));
2728 static int wq_clamp_max_active(int max_active, unsigned int flags,
2731 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2733 if (max_active < 1 || max_active > lim)
2734 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2735 "is out of range, clamping between %d and %d\n",
2736 max_active, name, 1, lim);
2738 return clamp_val(max_active, 1, lim);
2741 struct workqueue_struct *__alloc_workqueue_key(const char *name,
2744 struct lock_class_key *key,
2745 const char *lock_name)
2747 struct workqueue_struct *wq;
2751 * Unbound workqueues aren't concurrency managed and should be
2752 * dispatched to workers immediately.
2754 if (flags & WQ_UNBOUND)
2755 flags |= WQ_HIGHPRI;
2757 max_active = max_active ?: WQ_DFL_ACTIVE;
2758 max_active = wq_clamp_max_active(max_active, flags, name);
2760 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
2765 wq->saved_max_active = max_active;
2766 mutex_init(&wq->flush_mutex);
2767 atomic_set(&wq->nr_cwqs_to_flush, 0);
2768 INIT_LIST_HEAD(&wq->flusher_queue);
2769 INIT_LIST_HEAD(&wq->flusher_overflow);
2772 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
2773 INIT_LIST_HEAD(&wq->list);
2775 if (alloc_cwqs(wq) < 0)
2778 for_each_cwq_cpu(cpu, wq) {
2779 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2780 struct global_cwq *gcwq = get_gcwq(cpu);
2782 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
2785 cwq->flush_color = -1;
2786 cwq->max_active = max_active;
2787 INIT_LIST_HEAD(&cwq->delayed_works);
2790 if (flags & WQ_RESCUER) {
2791 struct worker *rescuer;
2793 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
2796 wq->rescuer = rescuer = alloc_worker();
2800 rescuer->task = kthread_create(rescuer_thread, wq, "%s", name);
2801 if (IS_ERR(rescuer->task))
2804 rescuer->task->flags |= PF_THREAD_BOUND;
2805 wake_up_process(rescuer->task);
2809 * workqueue_lock protects global freeze state and workqueues
2810 * list. Grab it, set max_active accordingly and add the new
2811 * workqueue to workqueues list.
2813 spin_lock(&workqueue_lock);
2815 if (workqueue_freezing && wq->flags & WQ_FREEZEABLE)
2816 for_each_cwq_cpu(cpu, wq)
2817 get_cwq(cpu, wq)->max_active = 0;
2819 list_add(&wq->list, &workqueues);
2821 spin_unlock(&workqueue_lock);
2827 free_mayday_mask(wq->mayday_mask);
2833 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
2836 * destroy_workqueue - safely terminate a workqueue
2837 * @wq: target workqueue
2839 * Safely destroy a workqueue. All work currently pending will be done first.
2841 void destroy_workqueue(struct workqueue_struct *wq)
2845 wq->flags |= WQ_DYING;
2846 flush_workqueue(wq);
2849 * wq list is used to freeze wq, remove from list after
2850 * flushing is complete in case freeze races us.
2852 spin_lock(&workqueue_lock);
2853 list_del(&wq->list);
2854 spin_unlock(&workqueue_lock);
2857 for_each_cwq_cpu(cpu, wq) {
2858 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2861 for (i = 0; i < WORK_NR_COLORS; i++)
2862 BUG_ON(cwq->nr_in_flight[i]);
2863 BUG_ON(cwq->nr_active);
2864 BUG_ON(!list_empty(&cwq->delayed_works));
2867 if (wq->flags & WQ_RESCUER) {
2868 kthread_stop(wq->rescuer->task);
2869 free_mayday_mask(wq->mayday_mask);
2876 EXPORT_SYMBOL_GPL(destroy_workqueue);
2879 * workqueue_set_max_active - adjust max_active of a workqueue
2880 * @wq: target workqueue
2881 * @max_active: new max_active value.
2883 * Set max_active of @wq to @max_active.
2886 * Don't call from IRQ context.
2888 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
2892 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
2894 spin_lock(&workqueue_lock);
2896 wq->saved_max_active = max_active;
2898 for_each_cwq_cpu(cpu, wq) {
2899 struct global_cwq *gcwq = get_gcwq(cpu);
2901 spin_lock_irq(&gcwq->lock);
2903 if (!(wq->flags & WQ_FREEZEABLE) ||
2904 !(gcwq->flags & GCWQ_FREEZING))
2905 get_cwq(gcwq->cpu, wq)->max_active = max_active;
2907 spin_unlock_irq(&gcwq->lock);
2910 spin_unlock(&workqueue_lock);
2912 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
2915 * workqueue_congested - test whether a workqueue is congested
2916 * @cpu: CPU in question
2917 * @wq: target workqueue
2919 * Test whether @wq's cpu workqueue for @cpu is congested. There is
2920 * no synchronization around this function and the test result is
2921 * unreliable and only useful as advisory hints or for debugging.
2924 * %true if congested, %false otherwise.
2926 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
2928 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2930 return !list_empty(&cwq->delayed_works);
2932 EXPORT_SYMBOL_GPL(workqueue_congested);
2935 * work_cpu - return the last known associated cpu for @work
2936 * @work: the work of interest
2939 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
2941 unsigned int work_cpu(struct work_struct *work)
2943 struct global_cwq *gcwq = get_work_gcwq(work);
2945 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
2947 EXPORT_SYMBOL_GPL(work_cpu);
2950 * work_busy - test whether a work is currently pending or running
2951 * @work: the work to be tested
2953 * Test whether @work is currently pending or running. There is no
2954 * synchronization around this function and the test result is
2955 * unreliable and only useful as advisory hints or for debugging.
2956 * Especially for reentrant wqs, the pending state might hide the
2960 * OR'd bitmask of WORK_BUSY_* bits.
2962 unsigned int work_busy(struct work_struct *work)
2964 struct global_cwq *gcwq = get_work_gcwq(work);
2965 unsigned long flags;
2966 unsigned int ret = 0;
2971 spin_lock_irqsave(&gcwq->lock, flags);
2973 if (work_pending(work))
2974 ret |= WORK_BUSY_PENDING;
2975 if (find_worker_executing_work(gcwq, work))
2976 ret |= WORK_BUSY_RUNNING;
2978 spin_unlock_irqrestore(&gcwq->lock, flags);
2982 EXPORT_SYMBOL_GPL(work_busy);
2987 * There are two challenges in supporting CPU hotplug. Firstly, there
2988 * are a lot of assumptions on strong associations among work, cwq and
2989 * gcwq which make migrating pending and scheduled works very
2990 * difficult to implement without impacting hot paths. Secondly,
2991 * gcwqs serve mix of short, long and very long running works making
2992 * blocked draining impractical.
2994 * This is solved by allowing a gcwq to be detached from CPU, running
2995 * it with unbound (rogue) workers and allowing it to be reattached
2996 * later if the cpu comes back online. A separate thread is created
2997 * to govern a gcwq in such state and is called the trustee of the
3000 * Trustee states and their descriptions.
3002 * START Command state used on startup. On CPU_DOWN_PREPARE, a
3003 * new trustee is started with this state.
3005 * IN_CHARGE Once started, trustee will enter this state after
3006 * assuming the manager role and making all existing
3007 * workers rogue. DOWN_PREPARE waits for trustee to
3008 * enter this state. After reaching IN_CHARGE, trustee
3009 * tries to execute the pending worklist until it's empty
3010 * and the state is set to BUTCHER, or the state is set
3013 * BUTCHER Command state which is set by the cpu callback after
3014 * the cpu has went down. Once this state is set trustee
3015 * knows that there will be no new works on the worklist
3016 * and once the worklist is empty it can proceed to
3017 * killing idle workers.
3019 * RELEASE Command state which is set by the cpu callback if the
3020 * cpu down has been canceled or it has come online
3021 * again. After recognizing this state, trustee stops
3022 * trying to drain or butcher and clears ROGUE, rebinds
3023 * all remaining workers back to the cpu and releases
3026 * DONE Trustee will enter this state after BUTCHER or RELEASE
3029 * trustee CPU draining
3030 * took over down complete
3031 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3033 * | CPU is back online v return workers |
3034 * ----------------> RELEASE --------------
3038 * trustee_wait_event_timeout - timed event wait for trustee
3039 * @cond: condition to wait for
3040 * @timeout: timeout in jiffies
3042 * wait_event_timeout() for trustee to use. Handles locking and
3043 * checks for RELEASE request.
3046 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3047 * multiple times. To be used by trustee.
3050 * Positive indicating left time if @cond is satisfied, 0 if timed
3051 * out, -1 if canceled.
3053 #define trustee_wait_event_timeout(cond, timeout) ({ \
3054 long __ret = (timeout); \
3055 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3057 spin_unlock_irq(&gcwq->lock); \
3058 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3059 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3061 spin_lock_irq(&gcwq->lock); \
3063 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3067 * trustee_wait_event - event wait for trustee
3068 * @cond: condition to wait for
3070 * wait_event() for trustee to use. Automatically handles locking and
3071 * checks for CANCEL request.
3074 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3075 * multiple times. To be used by trustee.
3078 * 0 if @cond is satisfied, -1 if canceled.
3080 #define trustee_wait_event(cond) ({ \
3082 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3083 __ret1 < 0 ? -1 : 0; \
3086 static int __cpuinit trustee_thread(void *__gcwq)
3088 struct global_cwq *gcwq = __gcwq;
3089 struct worker *worker;
3090 struct work_struct *work;
3091 struct hlist_node *pos;
3095 BUG_ON(gcwq->cpu != smp_processor_id());
3097 spin_lock_irq(&gcwq->lock);
3099 * Claim the manager position and make all workers rogue.
3100 * Trustee must be bound to the target cpu and can't be
3103 BUG_ON(gcwq->cpu != smp_processor_id());
3104 rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3107 gcwq->flags |= GCWQ_MANAGING_WORKERS;
3109 list_for_each_entry(worker, &gcwq->idle_list, entry)
3110 worker->flags |= WORKER_ROGUE;
3112 for_each_busy_worker(worker, i, pos, gcwq)
3113 worker->flags |= WORKER_ROGUE;
3116 * Call schedule() so that we cross rq->lock and thus can
3117 * guarantee sched callbacks see the rogue flag. This is
3118 * necessary as scheduler callbacks may be invoked from other
3121 spin_unlock_irq(&gcwq->lock);
3123 spin_lock_irq(&gcwq->lock);
3126 * Sched callbacks are disabled now. Zap nr_running. After
3127 * this, nr_running stays zero and need_more_worker() and
3128 * keep_working() are always true as long as the worklist is
3131 atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3133 spin_unlock_irq(&gcwq->lock);
3134 del_timer_sync(&gcwq->idle_timer);
3135 spin_lock_irq(&gcwq->lock);
3138 * We're now in charge. Notify and proceed to drain. We need
3139 * to keep the gcwq running during the whole CPU down
3140 * procedure as other cpu hotunplug callbacks may need to
3141 * flush currently running tasks.
3143 gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3144 wake_up_all(&gcwq->trustee_wait);
3147 * The original cpu is in the process of dying and may go away
3148 * anytime now. When that happens, we and all workers would
3149 * be migrated to other cpus. Try draining any left work. We
3150 * want to get it over with ASAP - spam rescuers, wake up as
3151 * many idlers as necessary and create new ones till the
3152 * worklist is empty. Note that if the gcwq is frozen, there
3153 * may be frozen works in freezeable cwqs. Don't declare
3154 * completion while frozen.
3156 while (gcwq->nr_workers != gcwq->nr_idle ||
3157 gcwq->flags & GCWQ_FREEZING ||
3158 gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3161 list_for_each_entry(work, &gcwq->worklist, entry) {
3166 list_for_each_entry(worker, &gcwq->idle_list, entry) {
3169 wake_up_process(worker->task);
3172 if (need_to_create_worker(gcwq)) {
3173 spin_unlock_irq(&gcwq->lock);
3174 worker = create_worker(gcwq, false);
3175 spin_lock_irq(&gcwq->lock);
3177 worker->flags |= WORKER_ROGUE;
3178 start_worker(worker);
3182 /* give a breather */
3183 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3188 * Either all works have been scheduled and cpu is down, or
3189 * cpu down has already been canceled. Wait for and butcher
3190 * all workers till we're canceled.
3193 rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3194 while (!list_empty(&gcwq->idle_list))
3195 destroy_worker(list_first_entry(&gcwq->idle_list,
3196 struct worker, entry));
3197 } while (gcwq->nr_workers && rc >= 0);
3200 * At this point, either draining has completed and no worker
3201 * is left, or cpu down has been canceled or the cpu is being
3202 * brought back up. There shouldn't be any idle one left.
3203 * Tell the remaining busy ones to rebind once it finishes the
3204 * currently scheduled works by scheduling the rebind_work.
3206 WARN_ON(!list_empty(&gcwq->idle_list));
3208 for_each_busy_worker(worker, i, pos, gcwq) {
3209 struct work_struct *rebind_work = &worker->rebind_work;
3212 * Rebind_work may race with future cpu hotplug
3213 * operations. Use a separate flag to mark that
3214 * rebinding is scheduled.
3216 worker->flags |= WORKER_REBIND;
3217 worker->flags &= ~WORKER_ROGUE;
3219 /* queue rebind_work, wq doesn't matter, use the default one */
3220 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3221 work_data_bits(rebind_work)))
3224 debug_work_activate(rebind_work);
3225 insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3226 worker->scheduled.next,
3227 work_color_to_flags(WORK_NO_COLOR));
3230 /* relinquish manager role */
3231 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3233 /* notify completion */
3234 gcwq->trustee = NULL;
3235 gcwq->trustee_state = TRUSTEE_DONE;
3236 wake_up_all(&gcwq->trustee_wait);
3237 spin_unlock_irq(&gcwq->lock);
3242 * wait_trustee_state - wait for trustee to enter the specified state
3243 * @gcwq: gcwq the trustee of interest belongs to
3244 * @state: target state to wait for
3246 * Wait for the trustee to reach @state. DONE is already matched.
3249 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3250 * multiple times. To be used by cpu_callback.
3252 static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3253 __releases(&gcwq->lock)
3254 __acquires(&gcwq->lock)
3256 if (!(gcwq->trustee_state == state ||
3257 gcwq->trustee_state == TRUSTEE_DONE)) {
3258 spin_unlock_irq(&gcwq->lock);
3259 __wait_event(gcwq->trustee_wait,
3260 gcwq->trustee_state == state ||
3261 gcwq->trustee_state == TRUSTEE_DONE);
3262 spin_lock_irq(&gcwq->lock);
3266 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3267 unsigned long action,
3270 unsigned int cpu = (unsigned long)hcpu;
3271 struct global_cwq *gcwq = get_gcwq(cpu);
3272 struct task_struct *new_trustee = NULL;
3273 struct worker *uninitialized_var(new_worker);
3274 unsigned long flags;
3276 action &= ~CPU_TASKS_FROZEN;
3279 case CPU_DOWN_PREPARE:
3280 new_trustee = kthread_create(trustee_thread, gcwq,
3281 "workqueue_trustee/%d\n", cpu);
3282 if (IS_ERR(new_trustee))
3283 return notifier_from_errno(PTR_ERR(new_trustee));
3284 kthread_bind(new_trustee, cpu);
3286 case CPU_UP_PREPARE:
3287 BUG_ON(gcwq->first_idle);
3288 new_worker = create_worker(gcwq, false);
3291 kthread_stop(new_trustee);
3296 /* some are called w/ irq disabled, don't disturb irq status */
3297 spin_lock_irqsave(&gcwq->lock, flags);
3300 case CPU_DOWN_PREPARE:
3301 /* initialize trustee and tell it to acquire the gcwq */
3302 BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
3303 gcwq->trustee = new_trustee;
3304 gcwq->trustee_state = TRUSTEE_START;
3305 wake_up_process(gcwq->trustee);
3306 wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3308 case CPU_UP_PREPARE:
3309 BUG_ON(gcwq->first_idle);
3310 gcwq->first_idle = new_worker;
3315 * Before this, the trustee and all workers except for
3316 * the ones which are still executing works from
3317 * before the last CPU down must be on the cpu. After
3318 * this, they'll all be diasporas.
3320 gcwq->flags |= GCWQ_DISASSOCIATED;
3324 gcwq->trustee_state = TRUSTEE_BUTCHER;
3326 case CPU_UP_CANCELED:
3327 destroy_worker(gcwq->first_idle);
3328 gcwq->first_idle = NULL;
3331 case CPU_DOWN_FAILED:
3333 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3334 if (gcwq->trustee_state != TRUSTEE_DONE) {
3335 gcwq->trustee_state = TRUSTEE_RELEASE;
3336 wake_up_process(gcwq->trustee);
3337 wait_trustee_state(gcwq, TRUSTEE_DONE);
3341 * Trustee is done and there might be no worker left.
3342 * Put the first_idle in and request a real manager to
3345 spin_unlock_irq(&gcwq->lock);
3346 kthread_bind(gcwq->first_idle->task, cpu);
3347 spin_lock_irq(&gcwq->lock);
3348 gcwq->flags |= GCWQ_MANAGE_WORKERS;
3349 start_worker(gcwq->first_idle);
3350 gcwq->first_idle = NULL;
3354 spin_unlock_irqrestore(&gcwq->lock, flags);
3356 return notifier_from_errno(0);
3361 struct work_for_cpu {
3362 struct completion completion;
3368 static int do_work_for_cpu(void *_wfc)
3370 struct work_for_cpu *wfc = _wfc;
3371 wfc->ret = wfc->fn(wfc->arg);
3372 complete(&wfc->completion);
3377 * work_on_cpu - run a function in user context on a particular cpu
3378 * @cpu: the cpu to run on
3379 * @fn: the function to run
3380 * @arg: the function arg
3382 * This will return the value @fn returns.
3383 * It is up to the caller to ensure that the cpu doesn't go offline.
3384 * The caller must not hold any locks which would prevent @fn from completing.
3386 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3388 struct task_struct *sub_thread;
3389 struct work_for_cpu wfc = {
3390 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3395 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3396 if (IS_ERR(sub_thread))
3397 return PTR_ERR(sub_thread);
3398 kthread_bind(sub_thread, cpu);
3399 wake_up_process(sub_thread);
3400 wait_for_completion(&wfc.completion);
3403 EXPORT_SYMBOL_GPL(work_on_cpu);
3404 #endif /* CONFIG_SMP */
3406 #ifdef CONFIG_FREEZER
3409 * freeze_workqueues_begin - begin freezing workqueues
3411 * Start freezing workqueues. After this function returns, all
3412 * freezeable workqueues will queue new works to their frozen_works
3413 * list instead of gcwq->worklist.
3416 * Grabs and releases workqueue_lock and gcwq->lock's.
3418 void freeze_workqueues_begin(void)
3422 spin_lock(&workqueue_lock);
3424 BUG_ON(workqueue_freezing);
3425 workqueue_freezing = true;
3427 for_each_gcwq_cpu(cpu) {
3428 struct global_cwq *gcwq = get_gcwq(cpu);
3429 struct workqueue_struct *wq;
3431 spin_lock_irq(&gcwq->lock);
3433 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3434 gcwq->flags |= GCWQ_FREEZING;
3436 list_for_each_entry(wq, &workqueues, list) {
3437 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3439 if (cwq && wq->flags & WQ_FREEZEABLE)
3440 cwq->max_active = 0;
3443 spin_unlock_irq(&gcwq->lock);
3446 spin_unlock(&workqueue_lock);
3450 * freeze_workqueues_busy - are freezeable workqueues still busy?
3452 * Check whether freezing is complete. This function must be called
3453 * between freeze_workqueues_begin() and thaw_workqueues().
3456 * Grabs and releases workqueue_lock.
3459 * %true if some freezeable workqueues are still busy. %false if
3460 * freezing is complete.
3462 bool freeze_workqueues_busy(void)
3467 spin_lock(&workqueue_lock);
3469 BUG_ON(!workqueue_freezing);
3471 for_each_gcwq_cpu(cpu) {
3472 struct workqueue_struct *wq;
3474 * nr_active is monotonically decreasing. It's safe
3475 * to peek without lock.
3477 list_for_each_entry(wq, &workqueues, list) {
3478 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3480 if (!cwq || !(wq->flags & WQ_FREEZEABLE))
3483 BUG_ON(cwq->nr_active < 0);
3484 if (cwq->nr_active) {
3491 spin_unlock(&workqueue_lock);
3496 * thaw_workqueues - thaw workqueues
3498 * Thaw workqueues. Normal queueing is restored and all collected
3499 * frozen works are transferred to their respective gcwq worklists.
3502 * Grabs and releases workqueue_lock and gcwq->lock's.
3504 void thaw_workqueues(void)
3508 spin_lock(&workqueue_lock);
3510 if (!workqueue_freezing)
3513 for_each_gcwq_cpu(cpu) {
3514 struct global_cwq *gcwq = get_gcwq(cpu);
3515 struct workqueue_struct *wq;
3517 spin_lock_irq(&gcwq->lock);
3519 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3520 gcwq->flags &= ~GCWQ_FREEZING;
3522 list_for_each_entry(wq, &workqueues, list) {
3523 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3525 if (!cwq || !(wq->flags & WQ_FREEZEABLE))
3528 /* restore max_active and repopulate worklist */
3529 cwq->max_active = wq->saved_max_active;
3531 while (!list_empty(&cwq->delayed_works) &&
3532 cwq->nr_active < cwq->max_active)
3533 cwq_activate_first_delayed(cwq);
3536 wake_up_worker(gcwq);
3538 spin_unlock_irq(&gcwq->lock);
3541 workqueue_freezing = false;
3543 spin_unlock(&workqueue_lock);
3545 #endif /* CONFIG_FREEZER */
3547 static int __init init_workqueues(void)
3552 cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
3554 /* initialize gcwqs */
3555 for_each_gcwq_cpu(cpu) {
3556 struct global_cwq *gcwq = get_gcwq(cpu);
3558 spin_lock_init(&gcwq->lock);
3559 INIT_LIST_HEAD(&gcwq->worklist);
3561 gcwq->flags |= GCWQ_DISASSOCIATED;
3563 INIT_LIST_HEAD(&gcwq->idle_list);
3564 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3565 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3567 init_timer_deferrable(&gcwq->idle_timer);
3568 gcwq->idle_timer.function = idle_worker_timeout;
3569 gcwq->idle_timer.data = (unsigned long)gcwq;
3571 setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3572 (unsigned long)gcwq);
3574 ida_init(&gcwq->worker_ida);
3576 gcwq->trustee_state = TRUSTEE_DONE;
3577 init_waitqueue_head(&gcwq->trustee_wait);
3580 /* create the initial worker */
3581 for_each_online_gcwq_cpu(cpu) {
3582 struct global_cwq *gcwq = get_gcwq(cpu);
3583 struct worker *worker;
3585 if (cpu != WORK_CPU_UNBOUND)
3586 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3587 worker = create_worker(gcwq, true);
3589 spin_lock_irq(&gcwq->lock);
3590 start_worker(worker);
3591 spin_unlock_irq(&gcwq->lock);
3594 system_wq = alloc_workqueue("events", 0, 0);
3595 system_long_wq = alloc_workqueue("events_long", 0, 0);
3596 system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3597 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3598 WQ_UNBOUND_MAX_ACTIVE);
3599 BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq);
3602 early_initcall(init_workqueues);