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: Set during initialization and read-only afterwards.
92 * P: Preemption protected. Disabling preemption is enough and should
93 * only be modified and accessed from the local cpu.
95 * L: gcwq->lock protected. Access with gcwq->lock held.
97 * X: During normal operation, modification requires gcwq->lock and
98 * should be done only from local cpu. Either disabling preemption
99 * on local cpu or grabbing gcwq->lock is enough for read access.
100 * If GCWQ_DISASSOCIATED is set, it's identical to L.
102 * F: wq->flush_mutex protected.
104 * W: workqueue_lock protected.
110 * The poor guys doing the actual heavy lifting. All on-duty workers
111 * are either serving the manager role, on idle list or on busy hash.
114 /* on idle list while idle, on busy hash table while busy */
116 struct list_head entry; /* L: while idle */
117 struct hlist_node hentry; /* L: while busy */
120 struct work_struct *current_work; /* L: work being processed */
121 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
122 struct list_head scheduled; /* L: scheduled works */
123 struct task_struct *task; /* I: worker task */
124 struct global_cwq *gcwq; /* I: the associated gcwq */
125 /* 64 bytes boundary on 64bit, 32 on 32bit */
126 unsigned long last_active; /* L: last active timestamp */
127 unsigned int flags; /* X: flags */
128 int id; /* I: worker id */
129 struct work_struct rebind_work; /* L: rebind worker to cpu */
133 * Global per-cpu workqueue. There's one and only one for each cpu
134 * and all works are queued and processed here regardless of their
138 spinlock_t lock; /* the gcwq lock */
139 struct list_head worklist; /* L: list of pending works */
140 unsigned int cpu; /* I: the associated cpu */
141 unsigned int flags; /* L: GCWQ_* flags */
143 int nr_workers; /* L: total number of workers */
144 int nr_idle; /* L: currently idle ones */
146 /* workers are chained either in the idle_list or busy_hash */
147 struct list_head idle_list; /* X: list of idle workers */
148 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
149 /* L: hash of busy workers */
151 struct timer_list idle_timer; /* L: worker idle timeout */
152 struct timer_list mayday_timer; /* L: SOS timer for dworkers */
154 struct ida worker_ida; /* L: for worker IDs */
156 struct task_struct *trustee; /* L: for gcwq shutdown */
157 unsigned int trustee_state; /* L: trustee state */
158 wait_queue_head_t trustee_wait; /* trustee wait */
159 struct worker *first_idle; /* L: first idle worker */
160 } ____cacheline_aligned_in_smp;
163 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
164 * work_struct->data are used for flags and thus cwqs need to be
165 * aligned at two's power of the number of flag bits.
167 struct cpu_workqueue_struct {
168 struct global_cwq *gcwq; /* I: the associated gcwq */
169 struct workqueue_struct *wq; /* I: the owning workqueue */
170 int work_color; /* L: current color */
171 int flush_color; /* L: flushing color */
172 int nr_in_flight[WORK_NR_COLORS];
173 /* L: nr of in_flight works */
174 int nr_active; /* L: nr of active works */
175 int max_active; /* L: max active works */
176 struct list_head delayed_works; /* L: delayed works */
180 * Structure used to wait for workqueue flush.
183 struct list_head list; /* F: list of flushers */
184 int flush_color; /* F: flush color waiting for */
185 struct completion done; /* flush completion */
189 * All cpumasks are assumed to be always set on UP and thus can't be
190 * used to determine whether there's something to be done.
193 typedef cpumask_var_t mayday_mask_t;
194 #define mayday_test_and_set_cpu(cpu, mask) \
195 cpumask_test_and_set_cpu((cpu), (mask))
196 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
197 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
198 #define alloc_mayday_mask(maskp, gfp) alloc_cpumask_var((maskp), (gfp))
199 #define free_mayday_mask(mask) free_cpumask_var((mask))
201 typedef unsigned long mayday_mask_t;
202 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
203 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
204 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
205 #define alloc_mayday_mask(maskp, gfp) true
206 #define free_mayday_mask(mask) do { } while (0)
210 * The externally visible workqueue abstraction is an array of
211 * per-CPU workqueues:
213 struct workqueue_struct {
214 unsigned int flags; /* I: WQ_* flags */
216 struct cpu_workqueue_struct __percpu *pcpu;
217 struct cpu_workqueue_struct *single;
219 } cpu_wq; /* I: cwq's */
220 struct list_head list; /* W: list of all workqueues */
222 struct mutex flush_mutex; /* protects wq flushing */
223 int work_color; /* F: current work color */
224 int flush_color; /* F: current flush color */
225 atomic_t nr_cwqs_to_flush; /* flush in progress */
226 struct wq_flusher *first_flusher; /* F: first flusher */
227 struct list_head flusher_queue; /* F: flush waiters */
228 struct list_head flusher_overflow; /* F: flush overflow list */
230 mayday_mask_t mayday_mask; /* cpus requesting rescue */
231 struct worker *rescuer; /* I: rescue worker */
233 int saved_max_active; /* W: saved cwq max_active */
234 const char *name; /* I: workqueue name */
235 #ifdef CONFIG_LOCKDEP
236 struct lockdep_map lockdep_map;
240 struct workqueue_struct *system_wq __read_mostly;
241 struct workqueue_struct *system_long_wq __read_mostly;
242 struct workqueue_struct *system_nrt_wq __read_mostly;
243 struct workqueue_struct *system_unbound_wq __read_mostly;
244 EXPORT_SYMBOL_GPL(system_wq);
245 EXPORT_SYMBOL_GPL(system_long_wq);
246 EXPORT_SYMBOL_GPL(system_nrt_wq);
247 EXPORT_SYMBOL_GPL(system_unbound_wq);
249 #define for_each_busy_worker(worker, i, pos, gcwq) \
250 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
251 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
253 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
256 if (cpu < nr_cpu_ids) {
258 cpu = cpumask_next(cpu, mask);
259 if (cpu < nr_cpu_ids)
263 return WORK_CPU_UNBOUND;
265 return WORK_CPU_NONE;
268 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
269 struct workqueue_struct *wq)
271 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
277 * An extra gcwq is defined for an invalid cpu number
278 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
279 * specific CPU. The following iterators are similar to
280 * for_each_*_cpu() iterators but also considers the unbound gcwq.
282 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
283 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
284 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
285 * WORK_CPU_UNBOUND for unbound workqueues
287 #define for_each_gcwq_cpu(cpu) \
288 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
289 (cpu) < WORK_CPU_NONE; \
290 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
292 #define for_each_online_gcwq_cpu(cpu) \
293 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
294 (cpu) < WORK_CPU_NONE; \
295 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
297 #define for_each_cwq_cpu(cpu, wq) \
298 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
299 (cpu) < WORK_CPU_NONE; \
300 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
302 #ifdef CONFIG_LOCKDEP
304 * in_workqueue_context() - in context of specified workqueue?
305 * @wq: the workqueue of interest
307 * Checks lockdep state to see if the current task is executing from
308 * within a workqueue item. This function exists only if lockdep is
311 int in_workqueue_context(struct workqueue_struct *wq)
313 return lock_is_held(&wq->lockdep_map);
317 #ifdef CONFIG_DEBUG_OBJECTS_WORK
319 static struct debug_obj_descr work_debug_descr;
322 * fixup_init is called when:
323 * - an active object is initialized
325 static int work_fixup_init(void *addr, enum debug_obj_state state)
327 struct work_struct *work = addr;
330 case ODEBUG_STATE_ACTIVE:
331 cancel_work_sync(work);
332 debug_object_init(work, &work_debug_descr);
340 * fixup_activate is called when:
341 * - an active object is activated
342 * - an unknown object is activated (might be a statically initialized object)
344 static int work_fixup_activate(void *addr, enum debug_obj_state state)
346 struct work_struct *work = addr;
350 case ODEBUG_STATE_NOTAVAILABLE:
352 * This is not really a fixup. The work struct was
353 * statically initialized. We just make sure that it
354 * is tracked in the object tracker.
356 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
357 debug_object_init(work, &work_debug_descr);
358 debug_object_activate(work, &work_debug_descr);
364 case ODEBUG_STATE_ACTIVE:
373 * fixup_free is called when:
374 * - an active object is freed
376 static int work_fixup_free(void *addr, enum debug_obj_state state)
378 struct work_struct *work = addr;
381 case ODEBUG_STATE_ACTIVE:
382 cancel_work_sync(work);
383 debug_object_free(work, &work_debug_descr);
390 static struct debug_obj_descr work_debug_descr = {
391 .name = "work_struct",
392 .fixup_init = work_fixup_init,
393 .fixup_activate = work_fixup_activate,
394 .fixup_free = work_fixup_free,
397 static inline void debug_work_activate(struct work_struct *work)
399 debug_object_activate(work, &work_debug_descr);
402 static inline void debug_work_deactivate(struct work_struct *work)
404 debug_object_deactivate(work, &work_debug_descr);
407 void __init_work(struct work_struct *work, int onstack)
410 debug_object_init_on_stack(work, &work_debug_descr);
412 debug_object_init(work, &work_debug_descr);
414 EXPORT_SYMBOL_GPL(__init_work);
416 void destroy_work_on_stack(struct work_struct *work)
418 debug_object_free(work, &work_debug_descr);
420 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
423 static inline void debug_work_activate(struct work_struct *work) { }
424 static inline void debug_work_deactivate(struct work_struct *work) { }
427 /* Serializes the accesses to the list of workqueues. */
428 static DEFINE_SPINLOCK(workqueue_lock);
429 static LIST_HEAD(workqueues);
430 static bool workqueue_freezing; /* W: have wqs started freezing? */
433 * The almighty global cpu workqueues. nr_running is the only field
434 * which is expected to be used frequently by other cpus via
435 * try_to_wake_up(). Put it in a separate cacheline.
437 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
438 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
441 * Global cpu workqueue and nr_running counter for unbound gcwq. The
442 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
443 * workers have WORKER_UNBOUND set.
445 static struct global_cwq unbound_global_cwq;
446 static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */
448 static int worker_thread(void *__worker);
450 static struct global_cwq *get_gcwq(unsigned int cpu)
452 if (cpu != WORK_CPU_UNBOUND)
453 return &per_cpu(global_cwq, cpu);
455 return &unbound_global_cwq;
458 static atomic_t *get_gcwq_nr_running(unsigned int cpu)
460 if (cpu != WORK_CPU_UNBOUND)
461 return &per_cpu(gcwq_nr_running, cpu);
463 return &unbound_gcwq_nr_running;
466 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
467 struct workqueue_struct *wq)
469 if (!(wq->flags & WQ_UNBOUND)) {
470 if (likely(cpu < nr_cpu_ids)) {
472 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
474 return wq->cpu_wq.single;
477 } else if (likely(cpu == WORK_CPU_UNBOUND))
478 return wq->cpu_wq.single;
482 static unsigned int work_color_to_flags(int color)
484 return color << WORK_STRUCT_COLOR_SHIFT;
487 static int get_work_color(struct work_struct *work)
489 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
490 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
493 static int work_next_color(int color)
495 return (color + 1) % WORK_NR_COLORS;
499 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
500 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
501 * cleared and the work data contains the cpu number it was last on.
503 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
504 * cwq, cpu or clear work->data. These functions should only be
505 * called while the work is owned - ie. while the PENDING bit is set.
507 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
508 * corresponding to a work. gcwq is available once the work has been
509 * queued anywhere after initialization. cwq is available only from
510 * queueing until execution starts.
512 static inline void set_work_data(struct work_struct *work, unsigned long data,
515 BUG_ON(!work_pending(work));
516 atomic_long_set(&work->data, data | flags | work_static(work));
519 static void set_work_cwq(struct work_struct *work,
520 struct cpu_workqueue_struct *cwq,
521 unsigned long extra_flags)
523 set_work_data(work, (unsigned long)cwq,
524 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
527 static void set_work_cpu(struct work_struct *work, unsigned int cpu)
529 set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
532 static void clear_work_data(struct work_struct *work)
534 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
537 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
539 unsigned long data = atomic_long_read(&work->data);
541 if (data & WORK_STRUCT_CWQ)
542 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
547 static struct global_cwq *get_work_gcwq(struct work_struct *work)
549 unsigned long data = atomic_long_read(&work->data);
552 if (data & WORK_STRUCT_CWQ)
553 return ((struct cpu_workqueue_struct *)
554 (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
556 cpu = data >> WORK_STRUCT_FLAG_BITS;
557 if (cpu == WORK_CPU_NONE)
560 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
561 return get_gcwq(cpu);
565 * Policy functions. These define the policies on how the global
566 * worker pool is managed. Unless noted otherwise, these functions
567 * assume that they're being called with gcwq->lock held.
570 static bool __need_more_worker(struct global_cwq *gcwq)
572 return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) ||
573 gcwq->flags & GCWQ_HIGHPRI_PENDING;
577 * Need to wake up a worker? Called from anything but currently
580 static bool need_more_worker(struct global_cwq *gcwq)
582 return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
585 /* Can I start working? Called from busy but !running workers. */
586 static bool may_start_working(struct global_cwq *gcwq)
588 return gcwq->nr_idle;
591 /* Do I need to keep working? Called from currently running workers. */
592 static bool keep_working(struct global_cwq *gcwq)
594 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
596 return !list_empty(&gcwq->worklist) && atomic_read(nr_running) <= 1;
599 /* Do we need a new worker? Called from manager. */
600 static bool need_to_create_worker(struct global_cwq *gcwq)
602 return need_more_worker(gcwq) && !may_start_working(gcwq);
605 /* Do I need to be the manager? */
606 static bool need_to_manage_workers(struct global_cwq *gcwq)
608 return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS;
611 /* Do we have too many workers and should some go away? */
612 static bool too_many_workers(struct global_cwq *gcwq)
614 bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
615 int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
616 int nr_busy = gcwq->nr_workers - nr_idle;
618 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
625 /* Return the first worker. Safe with preemption disabled */
626 static struct worker *first_worker(struct global_cwq *gcwq)
628 if (unlikely(list_empty(&gcwq->idle_list)))
631 return list_first_entry(&gcwq->idle_list, struct worker, entry);
635 * wake_up_worker - wake up an idle worker
636 * @gcwq: gcwq to wake worker for
638 * Wake up the first idle worker of @gcwq.
641 * spin_lock_irq(gcwq->lock).
643 static void wake_up_worker(struct global_cwq *gcwq)
645 struct worker *worker = first_worker(gcwq);
648 wake_up_process(worker->task);
652 * wq_worker_waking_up - a worker is waking up
653 * @task: task waking up
654 * @cpu: CPU @task is waking up to
656 * This function is called during try_to_wake_up() when a worker is
660 * spin_lock_irq(rq->lock)
662 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
664 struct worker *worker = kthread_data(task);
666 if (likely(!(worker->flags & WORKER_NOT_RUNNING)))
667 atomic_inc(get_gcwq_nr_running(cpu));
671 * wq_worker_sleeping - a worker is going to sleep
672 * @task: task going to sleep
673 * @cpu: CPU in question, must be the current CPU number
675 * This function is called during schedule() when a busy worker is
676 * going to sleep. Worker on the same cpu can be woken up by
677 * returning pointer to its task.
680 * spin_lock_irq(rq->lock)
683 * Worker task on @cpu to wake up, %NULL if none.
685 struct task_struct *wq_worker_sleeping(struct task_struct *task,
688 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
689 struct global_cwq *gcwq = get_gcwq(cpu);
690 atomic_t *nr_running = get_gcwq_nr_running(cpu);
692 if (unlikely(worker->flags & WORKER_NOT_RUNNING))
695 /* this can only happen on the local cpu */
696 BUG_ON(cpu != raw_smp_processor_id());
699 * The counterpart of the following dec_and_test, implied mb,
700 * worklist not empty test sequence is in insert_work().
701 * Please read comment there.
703 * NOT_RUNNING is clear. This means that trustee is not in
704 * charge and we're running on the local cpu w/ rq lock held
705 * and preemption disabled, which in turn means that none else
706 * could be manipulating idle_list, so dereferencing idle_list
707 * without gcwq lock is safe.
709 if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
710 to_wakeup = first_worker(gcwq);
711 return to_wakeup ? to_wakeup->task : NULL;
715 * worker_set_flags - set worker flags and adjust nr_running accordingly
717 * @flags: flags to set
718 * @wakeup: wakeup an idle worker if necessary
720 * Set @flags in @worker->flags and adjust nr_running accordingly. If
721 * nr_running becomes zero and @wakeup is %true, an idle worker is
725 * spin_lock_irq(gcwq->lock)
727 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
730 struct global_cwq *gcwq = worker->gcwq;
732 WARN_ON_ONCE(worker->task != current);
735 * If transitioning into NOT_RUNNING, adjust nr_running and
736 * wake up an idle worker as necessary if requested by
739 if ((flags & WORKER_NOT_RUNNING) &&
740 !(worker->flags & WORKER_NOT_RUNNING)) {
741 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
744 if (atomic_dec_and_test(nr_running) &&
745 !list_empty(&gcwq->worklist))
746 wake_up_worker(gcwq);
748 atomic_dec(nr_running);
751 worker->flags |= flags;
755 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
757 * @flags: flags to clear
759 * Clear @flags in @worker->flags and adjust nr_running accordingly.
762 * spin_lock_irq(gcwq->lock)
764 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
766 struct global_cwq *gcwq = worker->gcwq;
767 unsigned int oflags = worker->flags;
769 WARN_ON_ONCE(worker->task != current);
771 worker->flags &= ~flags;
773 /* if transitioning out of NOT_RUNNING, increment nr_running */
774 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
775 if (!(worker->flags & WORKER_NOT_RUNNING))
776 atomic_inc(get_gcwq_nr_running(gcwq->cpu));
780 * busy_worker_head - return the busy hash head for a work
781 * @gcwq: gcwq of interest
782 * @work: work to be hashed
784 * Return hash head of @gcwq for @work.
787 * spin_lock_irq(gcwq->lock).
790 * Pointer to the hash head.
792 static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
793 struct work_struct *work)
795 const int base_shift = ilog2(sizeof(struct work_struct));
796 unsigned long v = (unsigned long)work;
798 /* simple shift and fold hash, do we need something better? */
800 v += v >> BUSY_WORKER_HASH_ORDER;
801 v &= BUSY_WORKER_HASH_MASK;
803 return &gcwq->busy_hash[v];
807 * __find_worker_executing_work - find worker which is executing a work
808 * @gcwq: gcwq of interest
809 * @bwh: hash head as returned by busy_worker_head()
810 * @work: work to find worker for
812 * Find a worker which is executing @work on @gcwq. @bwh should be
813 * the hash head obtained by calling busy_worker_head() with the same
817 * spin_lock_irq(gcwq->lock).
820 * Pointer to worker which is executing @work if found, NULL
823 static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
824 struct hlist_head *bwh,
825 struct work_struct *work)
827 struct worker *worker;
828 struct hlist_node *tmp;
830 hlist_for_each_entry(worker, tmp, bwh, hentry)
831 if (worker->current_work == work)
837 * find_worker_executing_work - find worker which is executing a work
838 * @gcwq: gcwq of interest
839 * @work: work to find worker for
841 * Find a worker which is executing @work on @gcwq. This function is
842 * identical to __find_worker_executing_work() except that this
843 * function calculates @bwh itself.
846 * spin_lock_irq(gcwq->lock).
849 * Pointer to worker which is executing @work if found, NULL
852 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
853 struct work_struct *work)
855 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
860 * gcwq_determine_ins_pos - find insertion position
861 * @gcwq: gcwq of interest
862 * @cwq: cwq a work is being queued for
864 * A work for @cwq is about to be queued on @gcwq, determine insertion
865 * position for the work. If @cwq is for HIGHPRI wq, the work is
866 * queued at the head of the queue but in FIFO order with respect to
867 * other HIGHPRI works; otherwise, at the end of the queue. This
868 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
869 * there are HIGHPRI works pending.
872 * spin_lock_irq(gcwq->lock).
875 * Pointer to inserstion position.
877 static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq,
878 struct cpu_workqueue_struct *cwq)
880 struct work_struct *twork;
882 if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
883 return &gcwq->worklist;
885 list_for_each_entry(twork, &gcwq->worklist, entry) {
886 struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
888 if (!(tcwq->wq->flags & WQ_HIGHPRI))
892 gcwq->flags |= GCWQ_HIGHPRI_PENDING;
893 return &twork->entry;
897 * insert_work - insert a work into gcwq
898 * @cwq: cwq @work belongs to
899 * @work: work to insert
900 * @head: insertion point
901 * @extra_flags: extra WORK_STRUCT_* flags to set
903 * Insert @work which belongs to @cwq into @gcwq after @head.
904 * @extra_flags is or'd to work_struct flags.
907 * spin_lock_irq(gcwq->lock).
909 static void insert_work(struct cpu_workqueue_struct *cwq,
910 struct work_struct *work, struct list_head *head,
911 unsigned int extra_flags)
913 struct global_cwq *gcwq = cwq->gcwq;
915 /* we own @work, set data and link */
916 set_work_cwq(work, cwq, extra_flags);
919 * Ensure that we get the right work->data if we see the
920 * result of list_add() below, see try_to_grab_pending().
924 list_add_tail(&work->entry, head);
927 * Ensure either worker_sched_deactivated() sees the above
928 * list_add_tail() or we see zero nr_running to avoid workers
929 * lying around lazily while there are works to be processed.
933 if (__need_more_worker(gcwq))
934 wake_up_worker(gcwq);
937 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
938 struct work_struct *work)
940 struct global_cwq *gcwq;
941 struct cpu_workqueue_struct *cwq;
942 struct list_head *worklist;
945 debug_work_activate(work);
947 /* determine gcwq to use */
948 if (!(wq->flags & WQ_UNBOUND)) {
949 struct global_cwq *last_gcwq;
951 if (unlikely(cpu == WORK_CPU_UNBOUND))
952 cpu = raw_smp_processor_id();
955 * It's multi cpu. If @wq is non-reentrant and @work
956 * was previously on a different cpu, it might still
957 * be running there, in which case the work needs to
958 * be queued on that cpu to guarantee non-reentrance.
960 gcwq = get_gcwq(cpu);
961 if (wq->flags & WQ_NON_REENTRANT &&
962 (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
963 struct worker *worker;
965 spin_lock_irqsave(&last_gcwq->lock, flags);
967 worker = find_worker_executing_work(last_gcwq, work);
969 if (worker && worker->current_cwq->wq == wq)
972 /* meh... not running there, queue here */
973 spin_unlock_irqrestore(&last_gcwq->lock, flags);
974 spin_lock_irqsave(&gcwq->lock, flags);
977 spin_lock_irqsave(&gcwq->lock, flags);
979 gcwq = get_gcwq(WORK_CPU_UNBOUND);
980 spin_lock_irqsave(&gcwq->lock, flags);
983 /* gcwq determined, get cwq and queue */
984 cwq = get_cwq(gcwq->cpu, wq);
986 BUG_ON(!list_empty(&work->entry));
988 cwq->nr_in_flight[cwq->work_color]++;
990 if (likely(cwq->nr_active < cwq->max_active)) {
992 worklist = gcwq_determine_ins_pos(gcwq, cwq);
994 worklist = &cwq->delayed_works;
996 insert_work(cwq, work, worklist, work_color_to_flags(cwq->work_color));
998 spin_unlock_irqrestore(&gcwq->lock, flags);
1002 * queue_work - queue work on a workqueue
1003 * @wq: workqueue to use
1004 * @work: work to queue
1006 * Returns 0 if @work was already on a queue, non-zero otherwise.
1008 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1009 * it can be processed by another CPU.
1011 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
1015 ret = queue_work_on(get_cpu(), wq, work);
1020 EXPORT_SYMBOL_GPL(queue_work);
1023 * queue_work_on - queue work on specific cpu
1024 * @cpu: CPU number to execute work on
1025 * @wq: workqueue to use
1026 * @work: work to queue
1028 * Returns 0 if @work was already on a queue, non-zero otherwise.
1030 * We queue the work to a specific CPU, the caller must ensure it
1034 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
1038 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1039 __queue_work(cpu, wq, work);
1044 EXPORT_SYMBOL_GPL(queue_work_on);
1046 static void delayed_work_timer_fn(unsigned long __data)
1048 struct delayed_work *dwork = (struct delayed_work *)__data;
1049 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1051 __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
1055 * queue_delayed_work - queue work on a workqueue after delay
1056 * @wq: workqueue to use
1057 * @dwork: delayable work to queue
1058 * @delay: number of jiffies to wait before queueing
1060 * Returns 0 if @work was already on a queue, non-zero otherwise.
1062 int queue_delayed_work(struct workqueue_struct *wq,
1063 struct delayed_work *dwork, unsigned long delay)
1066 return queue_work(wq, &dwork->work);
1068 return queue_delayed_work_on(-1, wq, dwork, delay);
1070 EXPORT_SYMBOL_GPL(queue_delayed_work);
1073 * queue_delayed_work_on - queue work on specific CPU after delay
1074 * @cpu: CPU number to execute work on
1075 * @wq: workqueue to use
1076 * @dwork: work to queue
1077 * @delay: number of jiffies to wait before queueing
1079 * Returns 0 if @work was already on a queue, non-zero otherwise.
1081 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1082 struct delayed_work *dwork, unsigned long delay)
1085 struct timer_list *timer = &dwork->timer;
1086 struct work_struct *work = &dwork->work;
1088 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1091 BUG_ON(timer_pending(timer));
1092 BUG_ON(!list_empty(&work->entry));
1094 timer_stats_timer_set_start_info(&dwork->timer);
1097 * This stores cwq for the moment, for the timer_fn.
1098 * Note that the work's gcwq is preserved to allow
1099 * reentrance detection for delayed works.
1101 if (!(wq->flags & WQ_UNBOUND)) {
1102 struct global_cwq *gcwq = get_work_gcwq(work);
1104 if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1107 lcpu = raw_smp_processor_id();
1109 lcpu = WORK_CPU_UNBOUND;
1111 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1113 timer->expires = jiffies + delay;
1114 timer->data = (unsigned long)dwork;
1115 timer->function = delayed_work_timer_fn;
1117 if (unlikely(cpu >= 0))
1118 add_timer_on(timer, cpu);
1125 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1128 * worker_enter_idle - enter idle state
1129 * @worker: worker which is entering idle state
1131 * @worker is entering idle state. Update stats and idle timer if
1135 * spin_lock_irq(gcwq->lock).
1137 static void worker_enter_idle(struct worker *worker)
1139 struct global_cwq *gcwq = worker->gcwq;
1141 BUG_ON(worker->flags & WORKER_IDLE);
1142 BUG_ON(!list_empty(&worker->entry) &&
1143 (worker->hentry.next || worker->hentry.pprev));
1145 /* can't use worker_set_flags(), also called from start_worker() */
1146 worker->flags |= WORKER_IDLE;
1148 worker->last_active = jiffies;
1150 /* idle_list is LIFO */
1151 list_add(&worker->entry, &gcwq->idle_list);
1153 if (likely(!(worker->flags & WORKER_ROGUE))) {
1154 if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1155 mod_timer(&gcwq->idle_timer,
1156 jiffies + IDLE_WORKER_TIMEOUT);
1158 wake_up_all(&gcwq->trustee_wait);
1160 /* sanity check nr_running */
1161 WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle &&
1162 atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1166 * worker_leave_idle - leave idle state
1167 * @worker: worker which is leaving idle state
1169 * @worker is leaving idle state. Update stats.
1172 * spin_lock_irq(gcwq->lock).
1174 static void worker_leave_idle(struct worker *worker)
1176 struct global_cwq *gcwq = worker->gcwq;
1178 BUG_ON(!(worker->flags & WORKER_IDLE));
1179 worker_clr_flags(worker, WORKER_IDLE);
1181 list_del_init(&worker->entry);
1185 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1188 * Works which are scheduled while the cpu is online must at least be
1189 * scheduled to a worker which is bound to the cpu so that if they are
1190 * flushed from cpu callbacks while cpu is going down, they are
1191 * guaranteed to execute on the cpu.
1193 * This function is to be used by rogue workers and rescuers to bind
1194 * themselves to the target cpu and may race with cpu going down or
1195 * coming online. kthread_bind() can't be used because it may put the
1196 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1197 * verbatim as it's best effort and blocking and gcwq may be
1198 * [dis]associated in the meantime.
1200 * This function tries set_cpus_allowed() and locks gcwq and verifies
1201 * the binding against GCWQ_DISASSOCIATED which is set during
1202 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1203 * idle state or fetches works without dropping lock, it can guarantee
1204 * the scheduling requirement described in the first paragraph.
1207 * Might sleep. Called without any lock but returns with gcwq->lock
1211 * %true if the associated gcwq is online (@worker is successfully
1212 * bound), %false if offline.
1214 static bool worker_maybe_bind_and_lock(struct worker *worker)
1215 __acquires(&gcwq->lock)
1217 struct global_cwq *gcwq = worker->gcwq;
1218 struct task_struct *task = worker->task;
1222 * The following call may fail, succeed or succeed
1223 * without actually migrating the task to the cpu if
1224 * it races with cpu hotunplug operation. Verify
1225 * against GCWQ_DISASSOCIATED.
1227 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1228 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1230 spin_lock_irq(&gcwq->lock);
1231 if (gcwq->flags & GCWQ_DISASSOCIATED)
1233 if (task_cpu(task) == gcwq->cpu &&
1234 cpumask_equal(¤t->cpus_allowed,
1235 get_cpu_mask(gcwq->cpu)))
1237 spin_unlock_irq(&gcwq->lock);
1239 /* CPU has come up inbetween, retry migration */
1245 * Function for worker->rebind_work used to rebind rogue busy workers
1246 * to the associated cpu which is coming back online. This is
1247 * scheduled by cpu up but can race with other cpu hotplug operations
1248 * and may be executed twice without intervening cpu down.
1250 static void worker_rebind_fn(struct work_struct *work)
1252 struct worker *worker = container_of(work, struct worker, rebind_work);
1253 struct global_cwq *gcwq = worker->gcwq;
1255 if (worker_maybe_bind_and_lock(worker))
1256 worker_clr_flags(worker, WORKER_REBIND);
1258 spin_unlock_irq(&gcwq->lock);
1261 static struct worker *alloc_worker(void)
1263 struct worker *worker;
1265 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1267 INIT_LIST_HEAD(&worker->entry);
1268 INIT_LIST_HEAD(&worker->scheduled);
1269 INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1270 /* on creation a worker is in !idle && prep state */
1271 worker->flags = WORKER_PREP;
1277 * create_worker - create a new workqueue worker
1278 * @gcwq: gcwq the new worker will belong to
1279 * @bind: whether to set affinity to @cpu or not
1281 * Create a new worker which is bound to @gcwq. The returned worker
1282 * can be started by calling start_worker() or destroyed using
1286 * Might sleep. Does GFP_KERNEL allocations.
1289 * Pointer to the newly created worker.
1291 static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
1293 bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1294 struct worker *worker = NULL;
1297 spin_lock_irq(&gcwq->lock);
1298 while (ida_get_new(&gcwq->worker_ida, &id)) {
1299 spin_unlock_irq(&gcwq->lock);
1300 if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1302 spin_lock_irq(&gcwq->lock);
1304 spin_unlock_irq(&gcwq->lock);
1306 worker = alloc_worker();
1310 worker->gcwq = gcwq;
1313 if (!on_unbound_cpu)
1314 worker->task = kthread_create(worker_thread, worker,
1315 "kworker/%u:%d", gcwq->cpu, id);
1317 worker->task = kthread_create(worker_thread, worker,
1318 "kworker/u:%d", id);
1319 if (IS_ERR(worker->task))
1323 * A rogue worker will become a regular one if CPU comes
1324 * online later on. Make sure every worker has
1325 * PF_THREAD_BOUND set.
1327 if (bind && !on_unbound_cpu)
1328 kthread_bind(worker->task, gcwq->cpu);
1330 worker->task->flags |= PF_THREAD_BOUND;
1332 worker->flags |= WORKER_UNBOUND;
1338 spin_lock_irq(&gcwq->lock);
1339 ida_remove(&gcwq->worker_ida, id);
1340 spin_unlock_irq(&gcwq->lock);
1347 * start_worker - start a newly created worker
1348 * @worker: worker to start
1350 * Make the gcwq aware of @worker and start it.
1353 * spin_lock_irq(gcwq->lock).
1355 static void start_worker(struct worker *worker)
1357 worker->flags |= WORKER_STARTED;
1358 worker->gcwq->nr_workers++;
1359 worker_enter_idle(worker);
1360 wake_up_process(worker->task);
1364 * destroy_worker - destroy a workqueue worker
1365 * @worker: worker to be destroyed
1367 * Destroy @worker and adjust @gcwq stats accordingly.
1370 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1372 static void destroy_worker(struct worker *worker)
1374 struct global_cwq *gcwq = worker->gcwq;
1375 int id = worker->id;
1377 /* sanity check frenzy */
1378 BUG_ON(worker->current_work);
1379 BUG_ON(!list_empty(&worker->scheduled));
1381 if (worker->flags & WORKER_STARTED)
1383 if (worker->flags & WORKER_IDLE)
1386 list_del_init(&worker->entry);
1387 worker->flags |= WORKER_DIE;
1389 spin_unlock_irq(&gcwq->lock);
1391 kthread_stop(worker->task);
1394 spin_lock_irq(&gcwq->lock);
1395 ida_remove(&gcwq->worker_ida, id);
1398 static void idle_worker_timeout(unsigned long __gcwq)
1400 struct global_cwq *gcwq = (void *)__gcwq;
1402 spin_lock_irq(&gcwq->lock);
1404 if (too_many_workers(gcwq)) {
1405 struct worker *worker;
1406 unsigned long expires;
1408 /* idle_list is kept in LIFO order, check the last one */
1409 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1410 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1412 if (time_before(jiffies, expires))
1413 mod_timer(&gcwq->idle_timer, expires);
1415 /* it's been idle for too long, wake up manager */
1416 gcwq->flags |= GCWQ_MANAGE_WORKERS;
1417 wake_up_worker(gcwq);
1421 spin_unlock_irq(&gcwq->lock);
1424 static bool send_mayday(struct work_struct *work)
1426 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1427 struct workqueue_struct *wq = cwq->wq;
1430 if (!(wq->flags & WQ_RESCUER))
1433 /* mayday mayday mayday */
1434 cpu = cwq->gcwq->cpu;
1435 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1436 if (cpu == WORK_CPU_UNBOUND)
1438 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1439 wake_up_process(wq->rescuer->task);
1443 static void gcwq_mayday_timeout(unsigned long __gcwq)
1445 struct global_cwq *gcwq = (void *)__gcwq;
1446 struct work_struct *work;
1448 spin_lock_irq(&gcwq->lock);
1450 if (need_to_create_worker(gcwq)) {
1452 * We've been trying to create a new worker but
1453 * haven't been successful. We might be hitting an
1454 * allocation deadlock. Send distress signals to
1457 list_for_each_entry(work, &gcwq->worklist, entry)
1461 spin_unlock_irq(&gcwq->lock);
1463 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1467 * maybe_create_worker - create a new worker if necessary
1468 * @gcwq: gcwq to create a new worker for
1470 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1471 * have at least one idle worker on return from this function. If
1472 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1473 * sent to all rescuers with works scheduled on @gcwq to resolve
1474 * possible allocation deadlock.
1476 * On return, need_to_create_worker() is guaranteed to be false and
1477 * may_start_working() true.
1480 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1481 * multiple times. Does GFP_KERNEL allocations. Called only from
1485 * false if no action was taken and gcwq->lock stayed locked, true
1488 static bool maybe_create_worker(struct global_cwq *gcwq)
1489 __releases(&gcwq->lock)
1490 __acquires(&gcwq->lock)
1492 if (!need_to_create_worker(gcwq))
1495 spin_unlock_irq(&gcwq->lock);
1497 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1498 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1501 struct worker *worker;
1503 worker = create_worker(gcwq, true);
1505 del_timer_sync(&gcwq->mayday_timer);
1506 spin_lock_irq(&gcwq->lock);
1507 start_worker(worker);
1508 BUG_ON(need_to_create_worker(gcwq));
1512 if (!need_to_create_worker(gcwq))
1515 __set_current_state(TASK_INTERRUPTIBLE);
1516 schedule_timeout(CREATE_COOLDOWN);
1518 if (!need_to_create_worker(gcwq))
1522 del_timer_sync(&gcwq->mayday_timer);
1523 spin_lock_irq(&gcwq->lock);
1524 if (need_to_create_worker(gcwq))
1530 * maybe_destroy_worker - destroy workers which have been idle for a while
1531 * @gcwq: gcwq to destroy workers for
1533 * Destroy @gcwq workers which have been idle for longer than
1534 * IDLE_WORKER_TIMEOUT.
1537 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1538 * multiple times. Called only from manager.
1541 * false if no action was taken and gcwq->lock stayed locked, true
1544 static bool maybe_destroy_workers(struct global_cwq *gcwq)
1548 while (too_many_workers(gcwq)) {
1549 struct worker *worker;
1550 unsigned long expires;
1552 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1553 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1555 if (time_before(jiffies, expires)) {
1556 mod_timer(&gcwq->idle_timer, expires);
1560 destroy_worker(worker);
1568 * manage_workers - manage worker pool
1571 * Assume the manager role and manage gcwq worker pool @worker belongs
1572 * to. At any given time, there can be only zero or one manager per
1573 * gcwq. The exclusion is handled automatically by this function.
1575 * The caller can safely start processing works on false return. On
1576 * true return, it's guaranteed that need_to_create_worker() is false
1577 * and may_start_working() is true.
1580 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1581 * multiple times. Does GFP_KERNEL allocations.
1584 * false if no action was taken and gcwq->lock stayed locked, true if
1585 * some action was taken.
1587 static bool manage_workers(struct worker *worker)
1589 struct global_cwq *gcwq = worker->gcwq;
1592 if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1595 gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1596 gcwq->flags |= GCWQ_MANAGING_WORKERS;
1599 * Destroy and then create so that may_start_working() is true
1602 ret |= maybe_destroy_workers(gcwq);
1603 ret |= maybe_create_worker(gcwq);
1605 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1608 * The trustee might be waiting to take over the manager
1609 * position, tell it we're done.
1611 if (unlikely(gcwq->trustee))
1612 wake_up_all(&gcwq->trustee_wait);
1618 * move_linked_works - move linked works to a list
1619 * @work: start of series of works to be scheduled
1620 * @head: target list to append @work to
1621 * @nextp: out paramter for nested worklist walking
1623 * Schedule linked works starting from @work to @head. Work series to
1624 * be scheduled starts at @work and includes any consecutive work with
1625 * WORK_STRUCT_LINKED set in its predecessor.
1627 * If @nextp is not NULL, it's updated to point to the next work of
1628 * the last scheduled work. This allows move_linked_works() to be
1629 * nested inside outer list_for_each_entry_safe().
1632 * spin_lock_irq(gcwq->lock).
1634 static void move_linked_works(struct work_struct *work, struct list_head *head,
1635 struct work_struct **nextp)
1637 struct work_struct *n;
1640 * Linked worklist will always end before the end of the list,
1641 * use NULL for list head.
1643 list_for_each_entry_safe_from(work, n, NULL, entry) {
1644 list_move_tail(&work->entry, head);
1645 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1650 * If we're already inside safe list traversal and have moved
1651 * multiple works to the scheduled queue, the next position
1652 * needs to be updated.
1658 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1660 struct work_struct *work = list_first_entry(&cwq->delayed_works,
1661 struct work_struct, entry);
1662 struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1664 move_linked_works(work, pos, NULL);
1669 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1670 * @cwq: cwq of interest
1671 * @color: color of work which left the queue
1673 * A work either has completed or is removed from pending queue,
1674 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1677 * spin_lock_irq(gcwq->lock).
1679 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
1681 /* ignore uncolored works */
1682 if (color == WORK_NO_COLOR)
1685 cwq->nr_in_flight[color]--;
1688 if (!list_empty(&cwq->delayed_works)) {
1689 /* one down, submit a delayed one */
1690 if (cwq->nr_active < cwq->max_active)
1691 cwq_activate_first_delayed(cwq);
1694 /* is flush in progress and are we at the flushing tip? */
1695 if (likely(cwq->flush_color != color))
1698 /* are there still in-flight works? */
1699 if (cwq->nr_in_flight[color])
1702 /* this cwq is done, clear flush_color */
1703 cwq->flush_color = -1;
1706 * If this was the last cwq, wake up the first flusher. It
1707 * will handle the rest.
1709 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1710 complete(&cwq->wq->first_flusher->done);
1714 * process_one_work - process single work
1716 * @work: work to process
1718 * Process @work. This function contains all the logics necessary to
1719 * process a single work including synchronization against and
1720 * interaction with other workers on the same cpu, queueing and
1721 * flushing. As long as context requirement is met, any worker can
1722 * call this function to process a work.
1725 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1727 static void process_one_work(struct worker *worker, struct work_struct *work)
1728 __releases(&gcwq->lock)
1729 __acquires(&gcwq->lock)
1731 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1732 struct global_cwq *gcwq = cwq->gcwq;
1733 struct hlist_head *bwh = busy_worker_head(gcwq, work);
1734 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1735 work_func_t f = work->func;
1737 struct worker *collision;
1738 #ifdef CONFIG_LOCKDEP
1740 * It is permissible to free the struct work_struct from
1741 * inside the function that is called from it, this we need to
1742 * take into account for lockdep too. To avoid bogus "held
1743 * lock freed" warnings as well as problems when looking into
1744 * work->lockdep_map, make a copy and use that here.
1746 struct lockdep_map lockdep_map = work->lockdep_map;
1749 * A single work shouldn't be executed concurrently by
1750 * multiple workers on a single cpu. Check whether anyone is
1751 * already processing the work. If so, defer the work to the
1752 * currently executing one.
1754 collision = __find_worker_executing_work(gcwq, bwh, work);
1755 if (unlikely(collision)) {
1756 move_linked_works(work, &collision->scheduled, NULL);
1760 /* claim and process */
1761 debug_work_deactivate(work);
1762 hlist_add_head(&worker->hentry, bwh);
1763 worker->current_work = work;
1764 worker->current_cwq = cwq;
1765 work_color = get_work_color(work);
1767 /* record the current cpu number in the work data and dequeue */
1768 set_work_cpu(work, gcwq->cpu);
1769 list_del_init(&work->entry);
1772 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1773 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1775 if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1776 struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1777 struct work_struct, entry);
1779 if (!list_empty(&gcwq->worklist) &&
1780 get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1781 wake_up_worker(gcwq);
1783 gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1787 * CPU intensive works don't participate in concurrency
1788 * management. They're the scheduler's responsibility.
1790 if (unlikely(cpu_intensive))
1791 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1793 spin_unlock_irq(&gcwq->lock);
1795 work_clear_pending(work);
1796 lock_map_acquire(&cwq->wq->lockdep_map);
1797 lock_map_acquire(&lockdep_map);
1799 lock_map_release(&lockdep_map);
1800 lock_map_release(&cwq->wq->lockdep_map);
1802 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1803 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1805 current->comm, preempt_count(), task_pid_nr(current));
1806 printk(KERN_ERR " last function: ");
1807 print_symbol("%s\n", (unsigned long)f);
1808 debug_show_held_locks(current);
1812 spin_lock_irq(&gcwq->lock);
1814 /* clear cpu intensive status */
1815 if (unlikely(cpu_intensive))
1816 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1818 /* we're done with it, release */
1819 hlist_del_init(&worker->hentry);
1820 worker->current_work = NULL;
1821 worker->current_cwq = NULL;
1822 cwq_dec_nr_in_flight(cwq, work_color);
1826 * process_scheduled_works - process scheduled works
1829 * Process all scheduled works. Please note that the scheduled list
1830 * may change while processing a work, so this function repeatedly
1831 * fetches a work from the top and executes it.
1834 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1837 static void process_scheduled_works(struct worker *worker)
1839 while (!list_empty(&worker->scheduled)) {
1840 struct work_struct *work = list_first_entry(&worker->scheduled,
1841 struct work_struct, entry);
1842 process_one_work(worker, work);
1847 * worker_thread - the worker thread function
1850 * The gcwq worker thread function. There's a single dynamic pool of
1851 * these per each cpu. These workers process all works regardless of
1852 * their specific target workqueue. The only exception is works which
1853 * belong to workqueues with a rescuer which will be explained in
1856 static int worker_thread(void *__worker)
1858 struct worker *worker = __worker;
1859 struct global_cwq *gcwq = worker->gcwq;
1861 /* tell the scheduler that this is a workqueue worker */
1862 worker->task->flags |= PF_WQ_WORKER;
1864 spin_lock_irq(&gcwq->lock);
1866 /* DIE can be set only while we're idle, checking here is enough */
1867 if (worker->flags & WORKER_DIE) {
1868 spin_unlock_irq(&gcwq->lock);
1869 worker->task->flags &= ~PF_WQ_WORKER;
1873 worker_leave_idle(worker);
1875 /* no more worker necessary? */
1876 if (!need_more_worker(gcwq))
1879 /* do we need to manage? */
1880 if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1884 * ->scheduled list can only be filled while a worker is
1885 * preparing to process a work or actually processing it.
1886 * Make sure nobody diddled with it while I was sleeping.
1888 BUG_ON(!list_empty(&worker->scheduled));
1891 * When control reaches this point, we're guaranteed to have
1892 * at least one idle worker or that someone else has already
1893 * assumed the manager role.
1895 worker_clr_flags(worker, WORKER_PREP);
1898 struct work_struct *work =
1899 list_first_entry(&gcwq->worklist,
1900 struct work_struct, entry);
1902 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1903 /* optimization path, not strictly necessary */
1904 process_one_work(worker, work);
1905 if (unlikely(!list_empty(&worker->scheduled)))
1906 process_scheduled_works(worker);
1908 move_linked_works(work, &worker->scheduled, NULL);
1909 process_scheduled_works(worker);
1911 } while (keep_working(gcwq));
1913 worker_set_flags(worker, WORKER_PREP, false);
1915 if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1919 * gcwq->lock is held and there's no work to process and no
1920 * need to manage, sleep. Workers are woken up only while
1921 * holding gcwq->lock or from local cpu, so setting the
1922 * current state before releasing gcwq->lock is enough to
1923 * prevent losing any event.
1925 worker_enter_idle(worker);
1926 __set_current_state(TASK_INTERRUPTIBLE);
1927 spin_unlock_irq(&gcwq->lock);
1933 * rescuer_thread - the rescuer thread function
1934 * @__wq: the associated workqueue
1936 * Workqueue rescuer thread function. There's one rescuer for each
1937 * workqueue which has WQ_RESCUER set.
1939 * Regular work processing on a gcwq may block trying to create a new
1940 * worker which uses GFP_KERNEL allocation which has slight chance of
1941 * developing into deadlock if some works currently on the same queue
1942 * need to be processed to satisfy the GFP_KERNEL allocation. This is
1943 * the problem rescuer solves.
1945 * When such condition is possible, the gcwq summons rescuers of all
1946 * workqueues which have works queued on the gcwq and let them process
1947 * those works so that forward progress can be guaranteed.
1949 * This should happen rarely.
1951 static int rescuer_thread(void *__wq)
1953 struct workqueue_struct *wq = __wq;
1954 struct worker *rescuer = wq->rescuer;
1955 struct list_head *scheduled = &rescuer->scheduled;
1956 bool is_unbound = wq->flags & WQ_UNBOUND;
1959 set_user_nice(current, RESCUER_NICE_LEVEL);
1961 set_current_state(TASK_INTERRUPTIBLE);
1963 if (kthread_should_stop())
1967 * See whether any cpu is asking for help. Unbounded
1968 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
1970 for_each_mayday_cpu(cpu, wq->mayday_mask) {
1971 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
1972 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
1973 struct global_cwq *gcwq = cwq->gcwq;
1974 struct work_struct *work, *n;
1976 __set_current_state(TASK_RUNNING);
1977 mayday_clear_cpu(cpu, wq->mayday_mask);
1979 /* migrate to the target cpu if possible */
1980 rescuer->gcwq = gcwq;
1981 worker_maybe_bind_and_lock(rescuer);
1984 * Slurp in all works issued via this workqueue and
1987 BUG_ON(!list_empty(&rescuer->scheduled));
1988 list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
1989 if (get_work_cwq(work) == cwq)
1990 move_linked_works(work, scheduled, &n);
1992 process_scheduled_works(rescuer);
1993 spin_unlock_irq(&gcwq->lock);
2001 struct work_struct work;
2002 struct completion done;
2005 static void wq_barrier_func(struct work_struct *work)
2007 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2008 complete(&barr->done);
2012 * insert_wq_barrier - insert a barrier work
2013 * @cwq: cwq to insert barrier into
2014 * @barr: wq_barrier to insert
2015 * @target: target work to attach @barr to
2016 * @worker: worker currently executing @target, NULL if @target is not executing
2018 * @barr is linked to @target such that @barr is completed only after
2019 * @target finishes execution. Please note that the ordering
2020 * guarantee is observed only with respect to @target and on the local
2023 * Currently, a queued barrier can't be canceled. This is because
2024 * try_to_grab_pending() can't determine whether the work to be
2025 * grabbed is at the head of the queue and thus can't clear LINKED
2026 * flag of the previous work while there must be a valid next work
2027 * after a work with LINKED flag set.
2029 * Note that when @worker is non-NULL, @target may be modified
2030 * underneath us, so we can't reliably determine cwq from @target.
2033 * spin_lock_irq(gcwq->lock).
2035 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2036 struct wq_barrier *barr,
2037 struct work_struct *target, struct worker *worker)
2039 struct list_head *head;
2040 unsigned int linked = 0;
2043 * debugobject calls are safe here even with gcwq->lock locked
2044 * as we know for sure that this will not trigger any of the
2045 * checks and call back into the fixup functions where we
2048 INIT_WORK_ON_STACK(&barr->work, wq_barrier_func);
2049 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2050 init_completion(&barr->done);
2053 * If @target is currently being executed, schedule the
2054 * barrier to the worker; otherwise, put it after @target.
2057 head = worker->scheduled.next;
2059 unsigned long *bits = work_data_bits(target);
2061 head = target->entry.next;
2062 /* there can already be other linked works, inherit and set */
2063 linked = *bits & WORK_STRUCT_LINKED;
2064 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2067 debug_work_activate(&barr->work);
2068 insert_work(cwq, &barr->work, head,
2069 work_color_to_flags(WORK_NO_COLOR) | linked);
2073 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2074 * @wq: workqueue being flushed
2075 * @flush_color: new flush color, < 0 for no-op
2076 * @work_color: new work color, < 0 for no-op
2078 * Prepare cwqs for workqueue flushing.
2080 * If @flush_color is non-negative, flush_color on all cwqs should be
2081 * -1. If no cwq has in-flight commands at the specified color, all
2082 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2083 * has in flight commands, its cwq->flush_color is set to
2084 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2085 * wakeup logic is armed and %true is returned.
2087 * The caller should have initialized @wq->first_flusher prior to
2088 * calling this function with non-negative @flush_color. If
2089 * @flush_color is negative, no flush color update is done and %false
2092 * If @work_color is non-negative, all cwqs should have the same
2093 * work_color which is previous to @work_color and all will be
2094 * advanced to @work_color.
2097 * mutex_lock(wq->flush_mutex).
2100 * %true if @flush_color >= 0 and there's something to flush. %false
2103 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2104 int flush_color, int work_color)
2109 if (flush_color >= 0) {
2110 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2111 atomic_set(&wq->nr_cwqs_to_flush, 1);
2114 for_each_cwq_cpu(cpu, wq) {
2115 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2116 struct global_cwq *gcwq = cwq->gcwq;
2118 spin_lock_irq(&gcwq->lock);
2120 if (flush_color >= 0) {
2121 BUG_ON(cwq->flush_color != -1);
2123 if (cwq->nr_in_flight[flush_color]) {
2124 cwq->flush_color = flush_color;
2125 atomic_inc(&wq->nr_cwqs_to_flush);
2130 if (work_color >= 0) {
2131 BUG_ON(work_color != work_next_color(cwq->work_color));
2132 cwq->work_color = work_color;
2135 spin_unlock_irq(&gcwq->lock);
2138 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2139 complete(&wq->first_flusher->done);
2145 * flush_workqueue - ensure that any scheduled work has run to completion.
2146 * @wq: workqueue to flush
2148 * Forces execution of the workqueue and blocks until its completion.
2149 * This is typically used in driver shutdown handlers.
2151 * We sleep until all works which were queued on entry have been handled,
2152 * but we are not livelocked by new incoming ones.
2154 void flush_workqueue(struct workqueue_struct *wq)
2156 struct wq_flusher this_flusher = {
2157 .list = LIST_HEAD_INIT(this_flusher.list),
2159 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2163 lock_map_acquire(&wq->lockdep_map);
2164 lock_map_release(&wq->lockdep_map);
2166 mutex_lock(&wq->flush_mutex);
2169 * Start-to-wait phase
2171 next_color = work_next_color(wq->work_color);
2173 if (next_color != wq->flush_color) {
2175 * Color space is not full. The current work_color
2176 * becomes our flush_color and work_color is advanced
2179 BUG_ON(!list_empty(&wq->flusher_overflow));
2180 this_flusher.flush_color = wq->work_color;
2181 wq->work_color = next_color;
2183 if (!wq->first_flusher) {
2184 /* no flush in progress, become the first flusher */
2185 BUG_ON(wq->flush_color != this_flusher.flush_color);
2187 wq->first_flusher = &this_flusher;
2189 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2191 /* nothing to flush, done */
2192 wq->flush_color = next_color;
2193 wq->first_flusher = NULL;
2198 BUG_ON(wq->flush_color == this_flusher.flush_color);
2199 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2200 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2204 * Oops, color space is full, wait on overflow queue.
2205 * The next flush completion will assign us
2206 * flush_color and transfer to flusher_queue.
2208 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2211 mutex_unlock(&wq->flush_mutex);
2213 wait_for_completion(&this_flusher.done);
2216 * Wake-up-and-cascade phase
2218 * First flushers are responsible for cascading flushes and
2219 * handling overflow. Non-first flushers can simply return.
2221 if (wq->first_flusher != &this_flusher)
2224 mutex_lock(&wq->flush_mutex);
2226 /* we might have raced, check again with mutex held */
2227 if (wq->first_flusher != &this_flusher)
2230 wq->first_flusher = NULL;
2232 BUG_ON(!list_empty(&this_flusher.list));
2233 BUG_ON(wq->flush_color != this_flusher.flush_color);
2236 struct wq_flusher *next, *tmp;
2238 /* complete all the flushers sharing the current flush color */
2239 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2240 if (next->flush_color != wq->flush_color)
2242 list_del_init(&next->list);
2243 complete(&next->done);
2246 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2247 wq->flush_color != work_next_color(wq->work_color));
2249 /* this flush_color is finished, advance by one */
2250 wq->flush_color = work_next_color(wq->flush_color);
2252 /* one color has been freed, handle overflow queue */
2253 if (!list_empty(&wq->flusher_overflow)) {
2255 * Assign the same color to all overflowed
2256 * flushers, advance work_color and append to
2257 * flusher_queue. This is the start-to-wait
2258 * phase for these overflowed flushers.
2260 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2261 tmp->flush_color = wq->work_color;
2263 wq->work_color = work_next_color(wq->work_color);
2265 list_splice_tail_init(&wq->flusher_overflow,
2266 &wq->flusher_queue);
2267 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2270 if (list_empty(&wq->flusher_queue)) {
2271 BUG_ON(wq->flush_color != wq->work_color);
2276 * Need to flush more colors. Make the next flusher
2277 * the new first flusher and arm cwqs.
2279 BUG_ON(wq->flush_color == wq->work_color);
2280 BUG_ON(wq->flush_color != next->flush_color);
2282 list_del_init(&next->list);
2283 wq->first_flusher = next;
2285 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2289 * Meh... this color is already done, clear first
2290 * flusher and repeat cascading.
2292 wq->first_flusher = NULL;
2296 mutex_unlock(&wq->flush_mutex);
2298 EXPORT_SYMBOL_GPL(flush_workqueue);
2301 * flush_work - block until a work_struct's callback has terminated
2302 * @work: the work which is to be flushed
2304 * Returns false if @work has already terminated.
2306 * It is expected that, prior to calling flush_work(), the caller has
2307 * arranged for the work to not be requeued, otherwise it doesn't make
2308 * sense to use this function.
2310 int flush_work(struct work_struct *work)
2312 struct worker *worker = NULL;
2313 struct global_cwq *gcwq;
2314 struct cpu_workqueue_struct *cwq;
2315 struct wq_barrier barr;
2318 gcwq = get_work_gcwq(work);
2322 spin_lock_irq(&gcwq->lock);
2323 if (!list_empty(&work->entry)) {
2325 * See the comment near try_to_grab_pending()->smp_rmb().
2326 * If it was re-queued to a different gcwq under us, we
2327 * are not going to wait.
2330 cwq = get_work_cwq(work);
2331 if (unlikely(!cwq || gcwq != cwq->gcwq))
2334 worker = find_worker_executing_work(gcwq, work);
2337 cwq = worker->current_cwq;
2340 insert_wq_barrier(cwq, &barr, work, worker);
2341 spin_unlock_irq(&gcwq->lock);
2343 lock_map_acquire(&cwq->wq->lockdep_map);
2344 lock_map_release(&cwq->wq->lockdep_map);
2346 wait_for_completion(&barr.done);
2347 destroy_work_on_stack(&barr.work);
2350 spin_unlock_irq(&gcwq->lock);
2353 EXPORT_SYMBOL_GPL(flush_work);
2356 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2357 * so this work can't be re-armed in any way.
2359 static int try_to_grab_pending(struct work_struct *work)
2361 struct global_cwq *gcwq;
2364 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2368 * The queueing is in progress, or it is already queued. Try to
2369 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2371 gcwq = get_work_gcwq(work);
2375 spin_lock_irq(&gcwq->lock);
2376 if (!list_empty(&work->entry)) {
2378 * This work is queued, but perhaps we locked the wrong gcwq.
2379 * In that case we must see the new value after rmb(), see
2380 * insert_work()->wmb().
2383 if (gcwq == get_work_gcwq(work)) {
2384 debug_work_deactivate(work);
2385 list_del_init(&work->entry);
2386 cwq_dec_nr_in_flight(get_work_cwq(work),
2387 get_work_color(work));
2391 spin_unlock_irq(&gcwq->lock);
2396 static void wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2398 struct wq_barrier barr;
2399 struct worker *worker;
2401 spin_lock_irq(&gcwq->lock);
2403 worker = find_worker_executing_work(gcwq, work);
2404 if (unlikely(worker))
2405 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2407 spin_unlock_irq(&gcwq->lock);
2409 if (unlikely(worker)) {
2410 wait_for_completion(&barr.done);
2411 destroy_work_on_stack(&barr.work);
2415 static void wait_on_work(struct work_struct *work)
2421 lock_map_acquire(&work->lockdep_map);
2422 lock_map_release(&work->lockdep_map);
2424 for_each_gcwq_cpu(cpu)
2425 wait_on_cpu_work(get_gcwq(cpu), work);
2428 static int __cancel_work_timer(struct work_struct *work,
2429 struct timer_list* timer)
2434 ret = (timer && likely(del_timer(timer)));
2436 ret = try_to_grab_pending(work);
2438 } while (unlikely(ret < 0));
2440 clear_work_data(work);
2445 * cancel_work_sync - block until a work_struct's callback has terminated
2446 * @work: the work which is to be flushed
2448 * Returns true if @work was pending.
2450 * cancel_work_sync() will cancel the work if it is queued. If the work's
2451 * callback appears to be running, cancel_work_sync() will block until it
2454 * It is possible to use this function if the work re-queues itself. It can
2455 * cancel the work even if it migrates to another workqueue, however in that
2456 * case it only guarantees that work->func() has completed on the last queued
2459 * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
2460 * pending, otherwise it goes into a busy-wait loop until the timer expires.
2462 * The caller must ensure that workqueue_struct on which this work was last
2463 * queued can't be destroyed before this function returns.
2465 int cancel_work_sync(struct work_struct *work)
2467 return __cancel_work_timer(work, NULL);
2469 EXPORT_SYMBOL_GPL(cancel_work_sync);
2472 * cancel_delayed_work_sync - reliably kill off a delayed work.
2473 * @dwork: the delayed work struct
2475 * Returns true if @dwork was pending.
2477 * It is possible to use this function if @dwork rearms itself via queue_work()
2478 * or queue_delayed_work(). See also the comment for cancel_work_sync().
2480 int cancel_delayed_work_sync(struct delayed_work *dwork)
2482 return __cancel_work_timer(&dwork->work, &dwork->timer);
2484 EXPORT_SYMBOL(cancel_delayed_work_sync);
2487 * schedule_work - put work task in global workqueue
2488 * @work: job to be done
2490 * Returns zero if @work was already on the kernel-global workqueue and
2491 * non-zero otherwise.
2493 * This puts a job in the kernel-global workqueue if it was not already
2494 * queued and leaves it in the same position on the kernel-global
2495 * workqueue otherwise.
2497 int schedule_work(struct work_struct *work)
2499 return queue_work(system_wq, work);
2501 EXPORT_SYMBOL(schedule_work);
2504 * schedule_work_on - put work task on a specific cpu
2505 * @cpu: cpu to put the work task on
2506 * @work: job to be done
2508 * This puts a job on a specific cpu
2510 int schedule_work_on(int cpu, struct work_struct *work)
2512 return queue_work_on(cpu, system_wq, work);
2514 EXPORT_SYMBOL(schedule_work_on);
2517 * schedule_delayed_work - put work task in global workqueue after delay
2518 * @dwork: job to be done
2519 * @delay: number of jiffies to wait or 0 for immediate execution
2521 * After waiting for a given time this puts a job in the kernel-global
2524 int schedule_delayed_work(struct delayed_work *dwork,
2525 unsigned long delay)
2527 return queue_delayed_work(system_wq, dwork, delay);
2529 EXPORT_SYMBOL(schedule_delayed_work);
2532 * flush_delayed_work - block until a dwork_struct's callback has terminated
2533 * @dwork: the delayed work which is to be flushed
2535 * Any timeout is cancelled, and any pending work is run immediately.
2537 void flush_delayed_work(struct delayed_work *dwork)
2539 if (del_timer_sync(&dwork->timer)) {
2540 __queue_work(get_cpu(), get_work_cwq(&dwork->work)->wq,
2544 flush_work(&dwork->work);
2546 EXPORT_SYMBOL(flush_delayed_work);
2549 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2551 * @dwork: job to be done
2552 * @delay: number of jiffies to wait
2554 * After waiting for a given time this puts a job in the kernel-global
2555 * workqueue on the specified CPU.
2557 int schedule_delayed_work_on(int cpu,
2558 struct delayed_work *dwork, unsigned long delay)
2560 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2562 EXPORT_SYMBOL(schedule_delayed_work_on);
2565 * schedule_on_each_cpu - call a function on each online CPU from keventd
2566 * @func: the function to call
2568 * Returns zero on success.
2569 * Returns -ve errno on failure.
2571 * schedule_on_each_cpu() is very slow.
2573 int schedule_on_each_cpu(work_func_t func)
2576 struct work_struct __percpu *works;
2578 works = alloc_percpu(struct work_struct);
2584 for_each_online_cpu(cpu) {
2585 struct work_struct *work = per_cpu_ptr(works, cpu);
2587 INIT_WORK(work, func);
2588 schedule_work_on(cpu, work);
2591 for_each_online_cpu(cpu)
2592 flush_work(per_cpu_ptr(works, cpu));
2600 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2602 * Forces execution of the kernel-global workqueue and blocks until its
2605 * Think twice before calling this function! It's very easy to get into
2606 * trouble if you don't take great care. Either of the following situations
2607 * will lead to deadlock:
2609 * One of the work items currently on the workqueue needs to acquire
2610 * a lock held by your code or its caller.
2612 * Your code is running in the context of a work routine.
2614 * They will be detected by lockdep when they occur, but the first might not
2615 * occur very often. It depends on what work items are on the workqueue and
2616 * what locks they need, which you have no control over.
2618 * In most situations flushing the entire workqueue is overkill; you merely
2619 * need to know that a particular work item isn't queued and isn't running.
2620 * In such cases you should use cancel_delayed_work_sync() or
2621 * cancel_work_sync() instead.
2623 void flush_scheduled_work(void)
2625 flush_workqueue(system_wq);
2627 EXPORT_SYMBOL(flush_scheduled_work);
2630 * execute_in_process_context - reliably execute the routine with user context
2631 * @fn: the function to execute
2632 * @ew: guaranteed storage for the execute work structure (must
2633 * be available when the work executes)
2635 * Executes the function immediately if process context is available,
2636 * otherwise schedules the function for delayed execution.
2638 * Returns: 0 - function was executed
2639 * 1 - function was scheduled for execution
2641 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2643 if (!in_interrupt()) {
2648 INIT_WORK(&ew->work, fn);
2649 schedule_work(&ew->work);
2653 EXPORT_SYMBOL_GPL(execute_in_process_context);
2655 int keventd_up(void)
2657 return system_wq != NULL;
2660 static int alloc_cwqs(struct workqueue_struct *wq)
2663 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2664 * Make sure that the alignment isn't lower than that of
2665 * unsigned long long.
2667 const size_t size = sizeof(struct cpu_workqueue_struct);
2668 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2669 __alignof__(unsigned long long));
2671 bool percpu = !(wq->flags & WQ_UNBOUND);
2673 bool percpu = false;
2677 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2682 * Allocate enough room to align cwq and put an extra
2683 * pointer at the end pointing back to the originally
2684 * allocated pointer which will be used for free.
2686 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2688 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2689 *(void **)(wq->cpu_wq.single + 1) = ptr;
2693 /* just in case, make sure it's actually aligned */
2694 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2695 return wq->cpu_wq.v ? 0 : -ENOMEM;
2698 static void free_cwqs(struct workqueue_struct *wq)
2701 bool percpu = !(wq->flags & WQ_UNBOUND);
2703 bool percpu = false;
2707 free_percpu(wq->cpu_wq.pcpu);
2708 else if (wq->cpu_wq.single) {
2709 /* the pointer to free is stored right after the cwq */
2710 kfree(*(void **)(wq->cpu_wq.single + 1));
2714 static int wq_clamp_max_active(int max_active, unsigned int flags,
2717 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2719 if (max_active < 1 || max_active > lim)
2720 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2721 "is out of range, clamping between %d and %d\n",
2722 max_active, name, 1, lim);
2724 return clamp_val(max_active, 1, lim);
2727 struct workqueue_struct *__alloc_workqueue_key(const char *name,
2730 struct lock_class_key *key,
2731 const char *lock_name)
2733 struct workqueue_struct *wq;
2737 * Unbound workqueues aren't concurrency managed and should be
2738 * dispatched to workers immediately.
2740 if (flags & WQ_UNBOUND)
2741 flags |= WQ_HIGHPRI;
2743 max_active = max_active ?: WQ_DFL_ACTIVE;
2744 max_active = wq_clamp_max_active(max_active, flags, name);
2746 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
2751 wq->saved_max_active = max_active;
2752 mutex_init(&wq->flush_mutex);
2753 atomic_set(&wq->nr_cwqs_to_flush, 0);
2754 INIT_LIST_HEAD(&wq->flusher_queue);
2755 INIT_LIST_HEAD(&wq->flusher_overflow);
2758 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
2759 INIT_LIST_HEAD(&wq->list);
2761 if (alloc_cwqs(wq) < 0)
2764 for_each_cwq_cpu(cpu, wq) {
2765 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2766 struct global_cwq *gcwq = get_gcwq(cpu);
2768 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
2771 cwq->flush_color = -1;
2772 cwq->max_active = max_active;
2773 INIT_LIST_HEAD(&cwq->delayed_works);
2776 if (flags & WQ_RESCUER) {
2777 struct worker *rescuer;
2779 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
2782 wq->rescuer = rescuer = alloc_worker();
2786 rescuer->task = kthread_create(rescuer_thread, wq, "%s", name);
2787 if (IS_ERR(rescuer->task))
2790 rescuer->task->flags |= PF_THREAD_BOUND;
2791 wake_up_process(rescuer->task);
2795 * workqueue_lock protects global freeze state and workqueues
2796 * list. Grab it, set max_active accordingly and add the new
2797 * workqueue to workqueues list.
2799 spin_lock(&workqueue_lock);
2801 if (workqueue_freezing && wq->flags & WQ_FREEZEABLE)
2802 for_each_cwq_cpu(cpu, wq)
2803 get_cwq(cpu, wq)->max_active = 0;
2805 list_add(&wq->list, &workqueues);
2807 spin_unlock(&workqueue_lock);
2813 free_mayday_mask(wq->mayday_mask);
2819 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
2822 * destroy_workqueue - safely terminate a workqueue
2823 * @wq: target workqueue
2825 * Safely destroy a workqueue. All work currently pending will be done first.
2827 void destroy_workqueue(struct workqueue_struct *wq)
2831 flush_workqueue(wq);
2834 * wq list is used to freeze wq, remove from list after
2835 * flushing is complete in case freeze races us.
2837 spin_lock(&workqueue_lock);
2838 list_del(&wq->list);
2839 spin_unlock(&workqueue_lock);
2842 for_each_cwq_cpu(cpu, wq) {
2843 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2846 for (i = 0; i < WORK_NR_COLORS; i++)
2847 BUG_ON(cwq->nr_in_flight[i]);
2848 BUG_ON(cwq->nr_active);
2849 BUG_ON(!list_empty(&cwq->delayed_works));
2852 if (wq->flags & WQ_RESCUER) {
2853 kthread_stop(wq->rescuer->task);
2854 free_mayday_mask(wq->mayday_mask);
2861 EXPORT_SYMBOL_GPL(destroy_workqueue);
2864 * workqueue_set_max_active - adjust max_active of a workqueue
2865 * @wq: target workqueue
2866 * @max_active: new max_active value.
2868 * Set max_active of @wq to @max_active.
2871 * Don't call from IRQ context.
2873 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
2877 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
2879 spin_lock(&workqueue_lock);
2881 wq->saved_max_active = max_active;
2883 for_each_cwq_cpu(cpu, wq) {
2884 struct global_cwq *gcwq = get_gcwq(cpu);
2886 spin_lock_irq(&gcwq->lock);
2888 if (!(wq->flags & WQ_FREEZEABLE) ||
2889 !(gcwq->flags & GCWQ_FREEZING))
2890 get_cwq(gcwq->cpu, wq)->max_active = max_active;
2892 spin_unlock_irq(&gcwq->lock);
2895 spin_unlock(&workqueue_lock);
2897 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
2900 * workqueue_congested - test whether a workqueue is congested
2901 * @cpu: CPU in question
2902 * @wq: target workqueue
2904 * Test whether @wq's cpu workqueue for @cpu is congested. There is
2905 * no synchronization around this function and the test result is
2906 * unreliable and only useful as advisory hints or for debugging.
2909 * %true if congested, %false otherwise.
2911 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
2913 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2915 return !list_empty(&cwq->delayed_works);
2917 EXPORT_SYMBOL_GPL(workqueue_congested);
2920 * work_cpu - return the last known associated cpu for @work
2921 * @work: the work of interest
2924 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
2926 unsigned int work_cpu(struct work_struct *work)
2928 struct global_cwq *gcwq = get_work_gcwq(work);
2930 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
2932 EXPORT_SYMBOL_GPL(work_cpu);
2935 * work_busy - test whether a work is currently pending or running
2936 * @work: the work to be tested
2938 * Test whether @work is currently pending or running. There is no
2939 * synchronization around this function and the test result is
2940 * unreliable and only useful as advisory hints or for debugging.
2941 * Especially for reentrant wqs, the pending state might hide the
2945 * OR'd bitmask of WORK_BUSY_* bits.
2947 unsigned int work_busy(struct work_struct *work)
2949 struct global_cwq *gcwq = get_work_gcwq(work);
2950 unsigned long flags;
2951 unsigned int ret = 0;
2956 spin_lock_irqsave(&gcwq->lock, flags);
2958 if (work_pending(work))
2959 ret |= WORK_BUSY_PENDING;
2960 if (find_worker_executing_work(gcwq, work))
2961 ret |= WORK_BUSY_RUNNING;
2963 spin_unlock_irqrestore(&gcwq->lock, flags);
2967 EXPORT_SYMBOL_GPL(work_busy);
2972 * There are two challenges in supporting CPU hotplug. Firstly, there
2973 * are a lot of assumptions on strong associations among work, cwq and
2974 * gcwq which make migrating pending and scheduled works very
2975 * difficult to implement without impacting hot paths. Secondly,
2976 * gcwqs serve mix of short, long and very long running works making
2977 * blocked draining impractical.
2979 * This is solved by allowing a gcwq to be detached from CPU, running
2980 * it with unbound (rogue) workers and allowing it to be reattached
2981 * later if the cpu comes back online. A separate thread is created
2982 * to govern a gcwq in such state and is called the trustee of the
2985 * Trustee states and their descriptions.
2987 * START Command state used on startup. On CPU_DOWN_PREPARE, a
2988 * new trustee is started with this state.
2990 * IN_CHARGE Once started, trustee will enter this state after
2991 * assuming the manager role and making all existing
2992 * workers rogue. DOWN_PREPARE waits for trustee to
2993 * enter this state. After reaching IN_CHARGE, trustee
2994 * tries to execute the pending worklist until it's empty
2995 * and the state is set to BUTCHER, or the state is set
2998 * BUTCHER Command state which is set by the cpu callback after
2999 * the cpu has went down. Once this state is set trustee
3000 * knows that there will be no new works on the worklist
3001 * and once the worklist is empty it can proceed to
3002 * killing idle workers.
3004 * RELEASE Command state which is set by the cpu callback if the
3005 * cpu down has been canceled or it has come online
3006 * again. After recognizing this state, trustee stops
3007 * trying to drain or butcher and clears ROGUE, rebinds
3008 * all remaining workers back to the cpu and releases
3011 * DONE Trustee will enter this state after BUTCHER or RELEASE
3014 * trustee CPU draining
3015 * took over down complete
3016 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3018 * | CPU is back online v return workers |
3019 * ----------------> RELEASE --------------
3023 * trustee_wait_event_timeout - timed event wait for trustee
3024 * @cond: condition to wait for
3025 * @timeout: timeout in jiffies
3027 * wait_event_timeout() for trustee to use. Handles locking and
3028 * checks for RELEASE request.
3031 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3032 * multiple times. To be used by trustee.
3035 * Positive indicating left time if @cond is satisfied, 0 if timed
3036 * out, -1 if canceled.
3038 #define trustee_wait_event_timeout(cond, timeout) ({ \
3039 long __ret = (timeout); \
3040 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3042 spin_unlock_irq(&gcwq->lock); \
3043 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3044 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3046 spin_lock_irq(&gcwq->lock); \
3048 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3052 * trustee_wait_event - event wait for trustee
3053 * @cond: condition to wait for
3055 * wait_event() for trustee to use. Automatically handles locking and
3056 * checks for CANCEL request.
3059 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3060 * multiple times. To be used by trustee.
3063 * 0 if @cond is satisfied, -1 if canceled.
3065 #define trustee_wait_event(cond) ({ \
3067 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3068 __ret1 < 0 ? -1 : 0; \
3071 static int __cpuinit trustee_thread(void *__gcwq)
3073 struct global_cwq *gcwq = __gcwq;
3074 struct worker *worker;
3075 struct work_struct *work;
3076 struct hlist_node *pos;
3080 BUG_ON(gcwq->cpu != smp_processor_id());
3082 spin_lock_irq(&gcwq->lock);
3084 * Claim the manager position and make all workers rogue.
3085 * Trustee must be bound to the target cpu and can't be
3088 BUG_ON(gcwq->cpu != smp_processor_id());
3089 rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3092 gcwq->flags |= GCWQ_MANAGING_WORKERS;
3094 list_for_each_entry(worker, &gcwq->idle_list, entry)
3095 worker->flags |= WORKER_ROGUE;
3097 for_each_busy_worker(worker, i, pos, gcwq)
3098 worker->flags |= WORKER_ROGUE;
3101 * Call schedule() so that we cross rq->lock and thus can
3102 * guarantee sched callbacks see the rogue flag. This is
3103 * necessary as scheduler callbacks may be invoked from other
3106 spin_unlock_irq(&gcwq->lock);
3108 spin_lock_irq(&gcwq->lock);
3111 * Sched callbacks are disabled now. Zap nr_running. After
3112 * this, nr_running stays zero and need_more_worker() and
3113 * keep_working() are always true as long as the worklist is
3116 atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3118 spin_unlock_irq(&gcwq->lock);
3119 del_timer_sync(&gcwq->idle_timer);
3120 spin_lock_irq(&gcwq->lock);
3123 * We're now in charge. Notify and proceed to drain. We need
3124 * to keep the gcwq running during the whole CPU down
3125 * procedure as other cpu hotunplug callbacks may need to
3126 * flush currently running tasks.
3128 gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3129 wake_up_all(&gcwq->trustee_wait);
3132 * The original cpu is in the process of dying and may go away
3133 * anytime now. When that happens, we and all workers would
3134 * be migrated to other cpus. Try draining any left work. We
3135 * want to get it over with ASAP - spam rescuers, wake up as
3136 * many idlers as necessary and create new ones till the
3137 * worklist is empty. Note that if the gcwq is frozen, there
3138 * may be frozen works in freezeable cwqs. Don't declare
3139 * completion while frozen.
3141 while (gcwq->nr_workers != gcwq->nr_idle ||
3142 gcwq->flags & GCWQ_FREEZING ||
3143 gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3146 list_for_each_entry(work, &gcwq->worklist, entry) {
3151 list_for_each_entry(worker, &gcwq->idle_list, entry) {
3154 wake_up_process(worker->task);
3157 if (need_to_create_worker(gcwq)) {
3158 spin_unlock_irq(&gcwq->lock);
3159 worker = create_worker(gcwq, false);
3160 spin_lock_irq(&gcwq->lock);
3162 worker->flags |= WORKER_ROGUE;
3163 start_worker(worker);
3167 /* give a breather */
3168 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3173 * Either all works have been scheduled and cpu is down, or
3174 * cpu down has already been canceled. Wait for and butcher
3175 * all workers till we're canceled.
3178 rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3179 while (!list_empty(&gcwq->idle_list))
3180 destroy_worker(list_first_entry(&gcwq->idle_list,
3181 struct worker, entry));
3182 } while (gcwq->nr_workers && rc >= 0);
3185 * At this point, either draining has completed and no worker
3186 * is left, or cpu down has been canceled or the cpu is being
3187 * brought back up. There shouldn't be any idle one left.
3188 * Tell the remaining busy ones to rebind once it finishes the
3189 * currently scheduled works by scheduling the rebind_work.
3191 WARN_ON(!list_empty(&gcwq->idle_list));
3193 for_each_busy_worker(worker, i, pos, gcwq) {
3194 struct work_struct *rebind_work = &worker->rebind_work;
3197 * Rebind_work may race with future cpu hotplug
3198 * operations. Use a separate flag to mark that
3199 * rebinding is scheduled.
3201 worker->flags |= WORKER_REBIND;
3202 worker->flags &= ~WORKER_ROGUE;
3204 /* queue rebind_work, wq doesn't matter, use the default one */
3205 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3206 work_data_bits(rebind_work)))
3209 debug_work_activate(rebind_work);
3210 insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3211 worker->scheduled.next,
3212 work_color_to_flags(WORK_NO_COLOR));
3215 /* relinquish manager role */
3216 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3218 /* notify completion */
3219 gcwq->trustee = NULL;
3220 gcwq->trustee_state = TRUSTEE_DONE;
3221 wake_up_all(&gcwq->trustee_wait);
3222 spin_unlock_irq(&gcwq->lock);
3227 * wait_trustee_state - wait for trustee to enter the specified state
3228 * @gcwq: gcwq the trustee of interest belongs to
3229 * @state: target state to wait for
3231 * Wait for the trustee to reach @state. DONE is already matched.
3234 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3235 * multiple times. To be used by cpu_callback.
3237 static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3238 __releases(&gcwq->lock)
3239 __acquires(&gcwq->lock)
3241 if (!(gcwq->trustee_state == state ||
3242 gcwq->trustee_state == TRUSTEE_DONE)) {
3243 spin_unlock_irq(&gcwq->lock);
3244 __wait_event(gcwq->trustee_wait,
3245 gcwq->trustee_state == state ||
3246 gcwq->trustee_state == TRUSTEE_DONE);
3247 spin_lock_irq(&gcwq->lock);
3251 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3252 unsigned long action,
3255 unsigned int cpu = (unsigned long)hcpu;
3256 struct global_cwq *gcwq = get_gcwq(cpu);
3257 struct task_struct *new_trustee = NULL;
3258 struct worker *uninitialized_var(new_worker);
3259 unsigned long flags;
3261 action &= ~CPU_TASKS_FROZEN;
3264 case CPU_DOWN_PREPARE:
3265 new_trustee = kthread_create(trustee_thread, gcwq,
3266 "workqueue_trustee/%d\n", cpu);
3267 if (IS_ERR(new_trustee))
3268 return notifier_from_errno(PTR_ERR(new_trustee));
3269 kthread_bind(new_trustee, cpu);
3271 case CPU_UP_PREPARE:
3272 BUG_ON(gcwq->first_idle);
3273 new_worker = create_worker(gcwq, false);
3276 kthread_stop(new_trustee);
3281 /* some are called w/ irq disabled, don't disturb irq status */
3282 spin_lock_irqsave(&gcwq->lock, flags);
3285 case CPU_DOWN_PREPARE:
3286 /* initialize trustee and tell it to acquire the gcwq */
3287 BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
3288 gcwq->trustee = new_trustee;
3289 gcwq->trustee_state = TRUSTEE_START;
3290 wake_up_process(gcwq->trustee);
3291 wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3293 case CPU_UP_PREPARE:
3294 BUG_ON(gcwq->first_idle);
3295 gcwq->first_idle = new_worker;
3300 * Before this, the trustee and all workers except for
3301 * the ones which are still executing works from
3302 * before the last CPU down must be on the cpu. After
3303 * this, they'll all be diasporas.
3305 gcwq->flags |= GCWQ_DISASSOCIATED;
3309 gcwq->trustee_state = TRUSTEE_BUTCHER;
3311 case CPU_UP_CANCELED:
3312 destroy_worker(gcwq->first_idle);
3313 gcwq->first_idle = NULL;
3316 case CPU_DOWN_FAILED:
3318 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3319 if (gcwq->trustee_state != TRUSTEE_DONE) {
3320 gcwq->trustee_state = TRUSTEE_RELEASE;
3321 wake_up_process(gcwq->trustee);
3322 wait_trustee_state(gcwq, TRUSTEE_DONE);
3326 * Trustee is done and there might be no worker left.
3327 * Put the first_idle in and request a real manager to
3330 spin_unlock_irq(&gcwq->lock);
3331 kthread_bind(gcwq->first_idle->task, cpu);
3332 spin_lock_irq(&gcwq->lock);
3333 gcwq->flags |= GCWQ_MANAGE_WORKERS;
3334 start_worker(gcwq->first_idle);
3335 gcwq->first_idle = NULL;
3339 spin_unlock_irqrestore(&gcwq->lock, flags);
3341 return notifier_from_errno(0);
3346 struct work_for_cpu {
3347 struct completion completion;
3353 static int do_work_for_cpu(void *_wfc)
3355 struct work_for_cpu *wfc = _wfc;
3356 wfc->ret = wfc->fn(wfc->arg);
3357 complete(&wfc->completion);
3362 * work_on_cpu - run a function in user context on a particular cpu
3363 * @cpu: the cpu to run on
3364 * @fn: the function to run
3365 * @arg: the function arg
3367 * This will return the value @fn returns.
3368 * It is up to the caller to ensure that the cpu doesn't go offline.
3369 * The caller must not hold any locks which would prevent @fn from completing.
3371 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3373 struct task_struct *sub_thread;
3374 struct work_for_cpu wfc = {
3375 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3380 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3381 if (IS_ERR(sub_thread))
3382 return PTR_ERR(sub_thread);
3383 kthread_bind(sub_thread, cpu);
3384 wake_up_process(sub_thread);
3385 wait_for_completion(&wfc.completion);
3388 EXPORT_SYMBOL_GPL(work_on_cpu);
3389 #endif /* CONFIG_SMP */
3391 #ifdef CONFIG_FREEZER
3394 * freeze_workqueues_begin - begin freezing workqueues
3396 * Start freezing workqueues. After this function returns, all
3397 * freezeable workqueues will queue new works to their frozen_works
3398 * list instead of gcwq->worklist.
3401 * Grabs and releases workqueue_lock and gcwq->lock's.
3403 void freeze_workqueues_begin(void)
3407 spin_lock(&workqueue_lock);
3409 BUG_ON(workqueue_freezing);
3410 workqueue_freezing = true;
3412 for_each_gcwq_cpu(cpu) {
3413 struct global_cwq *gcwq = get_gcwq(cpu);
3414 struct workqueue_struct *wq;
3416 spin_lock_irq(&gcwq->lock);
3418 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3419 gcwq->flags |= GCWQ_FREEZING;
3421 list_for_each_entry(wq, &workqueues, list) {
3422 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3424 if (cwq && wq->flags & WQ_FREEZEABLE)
3425 cwq->max_active = 0;
3428 spin_unlock_irq(&gcwq->lock);
3431 spin_unlock(&workqueue_lock);
3435 * freeze_workqueues_busy - are freezeable workqueues still busy?
3437 * Check whether freezing is complete. This function must be called
3438 * between freeze_workqueues_begin() and thaw_workqueues().
3441 * Grabs and releases workqueue_lock.
3444 * %true if some freezeable workqueues are still busy. %false if
3445 * freezing is complete.
3447 bool freeze_workqueues_busy(void)
3452 spin_lock(&workqueue_lock);
3454 BUG_ON(!workqueue_freezing);
3456 for_each_gcwq_cpu(cpu) {
3457 struct workqueue_struct *wq;
3459 * nr_active is monotonically decreasing. It's safe
3460 * to peek without lock.
3462 list_for_each_entry(wq, &workqueues, list) {
3463 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3465 if (!cwq || !(wq->flags & WQ_FREEZEABLE))
3468 BUG_ON(cwq->nr_active < 0);
3469 if (cwq->nr_active) {
3476 spin_unlock(&workqueue_lock);
3481 * thaw_workqueues - thaw workqueues
3483 * Thaw workqueues. Normal queueing is restored and all collected
3484 * frozen works are transferred to their respective gcwq worklists.
3487 * Grabs and releases workqueue_lock and gcwq->lock's.
3489 void thaw_workqueues(void)
3493 spin_lock(&workqueue_lock);
3495 if (!workqueue_freezing)
3498 for_each_gcwq_cpu(cpu) {
3499 struct global_cwq *gcwq = get_gcwq(cpu);
3500 struct workqueue_struct *wq;
3502 spin_lock_irq(&gcwq->lock);
3504 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3505 gcwq->flags &= ~GCWQ_FREEZING;
3507 list_for_each_entry(wq, &workqueues, list) {
3508 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3510 if (!cwq || !(wq->flags & WQ_FREEZEABLE))
3513 /* restore max_active and repopulate worklist */
3514 cwq->max_active = wq->saved_max_active;
3516 while (!list_empty(&cwq->delayed_works) &&
3517 cwq->nr_active < cwq->max_active)
3518 cwq_activate_first_delayed(cwq);
3521 wake_up_worker(gcwq);
3523 spin_unlock_irq(&gcwq->lock);
3526 workqueue_freezing = false;
3528 spin_unlock(&workqueue_lock);
3530 #endif /* CONFIG_FREEZER */
3532 static int __init init_workqueues(void)
3537 cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
3539 /* initialize gcwqs */
3540 for_each_gcwq_cpu(cpu) {
3541 struct global_cwq *gcwq = get_gcwq(cpu);
3543 spin_lock_init(&gcwq->lock);
3544 INIT_LIST_HEAD(&gcwq->worklist);
3546 if (cpu == WORK_CPU_UNBOUND)
3547 gcwq->flags |= GCWQ_DISASSOCIATED;
3549 INIT_LIST_HEAD(&gcwq->idle_list);
3550 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3551 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3553 init_timer_deferrable(&gcwq->idle_timer);
3554 gcwq->idle_timer.function = idle_worker_timeout;
3555 gcwq->idle_timer.data = (unsigned long)gcwq;
3557 setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3558 (unsigned long)gcwq);
3560 ida_init(&gcwq->worker_ida);
3562 gcwq->trustee_state = TRUSTEE_DONE;
3563 init_waitqueue_head(&gcwq->trustee_wait);
3566 /* create the initial worker */
3567 for_each_online_gcwq_cpu(cpu) {
3568 struct global_cwq *gcwq = get_gcwq(cpu);
3569 struct worker *worker;
3571 worker = create_worker(gcwq, true);
3573 spin_lock_irq(&gcwq->lock);
3574 start_worker(worker);
3575 spin_unlock_irq(&gcwq->lock);
3578 system_wq = alloc_workqueue("events", 0, 0);
3579 system_long_wq = alloc_workqueue("events_long", 0, 0);
3580 system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3581 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3582 WQ_UNBOUND_MAX_ACTIVE);
3583 BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq);
3586 early_initcall(init_workqueues);