2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/module.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
45 #include "workqueue_sched.h"
48 /* global_cwq flags */
49 GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
50 GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */
51 GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
52 GCWQ_FREEZING = 1 << 3, /* freeze in progress */
53 GCWQ_HIGHPRI_PENDING = 1 << 4, /* highpri works on queue */
56 WORKER_STARTED = 1 << 0, /* started */
57 WORKER_DIE = 1 << 1, /* die die die */
58 WORKER_IDLE = 1 << 2, /* is idle */
59 WORKER_PREP = 1 << 3, /* preparing to run works */
60 WORKER_ROGUE = 1 << 4, /* not bound to any cpu */
61 WORKER_REBIND = 1 << 5, /* mom is home, come back */
62 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
63 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
65 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND |
66 WORKER_CPU_INTENSIVE | WORKER_UNBOUND,
68 /* gcwq->trustee_state */
69 TRUSTEE_START = 0, /* start */
70 TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */
71 TRUSTEE_BUTCHER = 2, /* butcher workers */
72 TRUSTEE_RELEASE = 3, /* release workers */
73 TRUSTEE_DONE = 4, /* trustee is done */
75 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
76 BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
77 BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
79 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
80 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
82 MAYDAY_INITIAL_TIMEOUT = HZ / 100, /* call for help after 10ms */
83 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
84 CREATE_COOLDOWN = HZ, /* time to breath after fail */
85 TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */
88 * Rescue workers are used only on emergencies and shared by
91 RESCUER_NICE_LEVEL = -20,
95 * Structure fields follow one of the following exclusion rules.
97 * I: Modifiable by initialization/destruction paths and read-only for
100 * P: Preemption protected. Disabling preemption is enough and should
101 * only be modified and accessed from the local cpu.
103 * L: gcwq->lock protected. Access with gcwq->lock held.
105 * X: During normal operation, modification requires gcwq->lock and
106 * should be done only from local cpu. Either disabling preemption
107 * on local cpu or grabbing gcwq->lock is enough for read access.
108 * If GCWQ_DISASSOCIATED is set, it's identical to L.
110 * F: wq->flush_mutex protected.
112 * W: workqueue_lock protected.
118 * The poor guys doing the actual heavy lifting. All on-duty workers
119 * are either serving the manager role, on idle list or on busy hash.
122 /* on idle list while idle, on busy hash table while busy */
124 struct list_head entry; /* L: while idle */
125 struct hlist_node hentry; /* L: while busy */
128 struct work_struct *current_work; /* L: work being processed */
129 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
130 struct list_head scheduled; /* L: scheduled works */
131 struct task_struct *task; /* I: worker task */
132 struct global_cwq *gcwq; /* I: the associated gcwq */
133 /* 64 bytes boundary on 64bit, 32 on 32bit */
134 unsigned long last_active; /* L: last active timestamp */
135 unsigned int flags; /* X: flags */
136 int id; /* I: worker id */
137 struct work_struct rebind_work; /* L: rebind worker to cpu */
141 * Global per-cpu workqueue. There's one and only one for each cpu
142 * and all works are queued and processed here regardless of their
146 spinlock_t lock; /* the gcwq lock */
147 struct list_head worklist; /* L: list of pending works */
148 unsigned int cpu; /* I: the associated cpu */
149 unsigned int flags; /* L: GCWQ_* flags */
151 int nr_workers; /* L: total number of workers */
152 int nr_idle; /* L: currently idle ones */
154 /* workers are chained either in the idle_list or busy_hash */
155 struct list_head idle_list; /* X: list of idle workers */
156 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
157 /* L: hash of busy workers */
159 struct timer_list idle_timer; /* L: worker idle timeout */
160 struct timer_list mayday_timer; /* L: SOS timer for dworkers */
162 struct ida worker_ida; /* L: for worker IDs */
164 struct task_struct *trustee; /* L: for gcwq shutdown */
165 unsigned int trustee_state; /* L: trustee state */
166 wait_queue_head_t trustee_wait; /* trustee wait */
167 struct worker *first_idle; /* L: first idle worker */
168 } ____cacheline_aligned_in_smp;
171 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
172 * work_struct->data are used for flags and thus cwqs need to be
173 * aligned at two's power of the number of flag bits.
175 struct cpu_workqueue_struct {
176 struct global_cwq *gcwq; /* I: the associated gcwq */
177 struct workqueue_struct *wq; /* I: the owning workqueue */
178 int work_color; /* L: current color */
179 int flush_color; /* L: flushing color */
180 int nr_in_flight[WORK_NR_COLORS];
181 /* L: nr of in_flight works */
182 int nr_active; /* L: nr of active works */
183 int max_active; /* L: max active works */
184 struct list_head delayed_works; /* L: delayed works */
188 * Structure used to wait for workqueue flush.
191 struct list_head list; /* F: list of flushers */
192 int flush_color; /* F: flush color waiting for */
193 struct completion done; /* flush completion */
197 * All cpumasks are assumed to be always set on UP and thus can't be
198 * used to determine whether there's something to be done.
201 typedef cpumask_var_t mayday_mask_t;
202 #define mayday_test_and_set_cpu(cpu, mask) \
203 cpumask_test_and_set_cpu((cpu), (mask))
204 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
205 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
206 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
207 #define free_mayday_mask(mask) free_cpumask_var((mask))
209 typedef unsigned long mayday_mask_t;
210 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
211 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
212 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
213 #define alloc_mayday_mask(maskp, gfp) true
214 #define free_mayday_mask(mask) do { } while (0)
218 * The externally visible workqueue abstraction is an array of
219 * per-CPU workqueues:
221 struct workqueue_struct {
222 unsigned int flags; /* I: WQ_* flags */
224 struct cpu_workqueue_struct __percpu *pcpu;
225 struct cpu_workqueue_struct *single;
227 } cpu_wq; /* I: cwq's */
228 struct list_head list; /* W: list of all workqueues */
230 struct mutex flush_mutex; /* protects wq flushing */
231 int work_color; /* F: current work color */
232 int flush_color; /* F: current flush color */
233 atomic_t nr_cwqs_to_flush; /* flush in progress */
234 struct wq_flusher *first_flusher; /* F: first flusher */
235 struct list_head flusher_queue; /* F: flush waiters */
236 struct list_head flusher_overflow; /* F: flush overflow list */
238 mayday_mask_t mayday_mask; /* cpus requesting rescue */
239 struct worker *rescuer; /* I: rescue worker */
241 int saved_max_active; /* W: saved cwq max_active */
242 const char *name; /* I: workqueue name */
243 #ifdef CONFIG_LOCKDEP
244 struct lockdep_map lockdep_map;
248 struct workqueue_struct *system_wq __read_mostly;
249 struct workqueue_struct *system_long_wq __read_mostly;
250 struct workqueue_struct *system_nrt_wq __read_mostly;
251 struct workqueue_struct *system_unbound_wq __read_mostly;
252 struct workqueue_struct *system_freezeable_wq __read_mostly;
253 EXPORT_SYMBOL_GPL(system_wq);
254 EXPORT_SYMBOL_GPL(system_long_wq);
255 EXPORT_SYMBOL_GPL(system_nrt_wq);
256 EXPORT_SYMBOL_GPL(system_unbound_wq);
257 EXPORT_SYMBOL_GPL(system_freezeable_wq);
259 #define CREATE_TRACE_POINTS
260 #include <trace/events/workqueue.h>
262 #define for_each_busy_worker(worker, i, pos, gcwq) \
263 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
264 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
266 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
269 if (cpu < nr_cpu_ids) {
271 cpu = cpumask_next(cpu, mask);
272 if (cpu < nr_cpu_ids)
276 return WORK_CPU_UNBOUND;
278 return WORK_CPU_NONE;
281 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
282 struct workqueue_struct *wq)
284 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
290 * An extra gcwq is defined for an invalid cpu number
291 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
292 * specific CPU. The following iterators are similar to
293 * for_each_*_cpu() iterators but also considers the unbound gcwq.
295 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
296 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
297 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
298 * WORK_CPU_UNBOUND for unbound workqueues
300 #define for_each_gcwq_cpu(cpu) \
301 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
302 (cpu) < WORK_CPU_NONE; \
303 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
305 #define for_each_online_gcwq_cpu(cpu) \
306 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
307 (cpu) < WORK_CPU_NONE; \
308 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
310 #define for_each_cwq_cpu(cpu, wq) \
311 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
312 (cpu) < WORK_CPU_NONE; \
313 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
315 #ifdef CONFIG_DEBUG_OBJECTS_WORK
317 static struct debug_obj_descr work_debug_descr;
320 * fixup_init is called when:
321 * - an active object is initialized
323 static int work_fixup_init(void *addr, enum debug_obj_state state)
325 struct work_struct *work = addr;
328 case ODEBUG_STATE_ACTIVE:
329 cancel_work_sync(work);
330 debug_object_init(work, &work_debug_descr);
338 * fixup_activate is called when:
339 * - an active object is activated
340 * - an unknown object is activated (might be a statically initialized object)
342 static int work_fixup_activate(void *addr, enum debug_obj_state state)
344 struct work_struct *work = addr;
348 case ODEBUG_STATE_NOTAVAILABLE:
350 * This is not really a fixup. The work struct was
351 * statically initialized. We just make sure that it
352 * is tracked in the object tracker.
354 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
355 debug_object_init(work, &work_debug_descr);
356 debug_object_activate(work, &work_debug_descr);
362 case ODEBUG_STATE_ACTIVE:
371 * fixup_free is called when:
372 * - an active object is freed
374 static int work_fixup_free(void *addr, enum debug_obj_state state)
376 struct work_struct *work = addr;
379 case ODEBUG_STATE_ACTIVE:
380 cancel_work_sync(work);
381 debug_object_free(work, &work_debug_descr);
388 static struct debug_obj_descr work_debug_descr = {
389 .name = "work_struct",
390 .fixup_init = work_fixup_init,
391 .fixup_activate = work_fixup_activate,
392 .fixup_free = work_fixup_free,
395 static inline void debug_work_activate(struct work_struct *work)
397 debug_object_activate(work, &work_debug_descr);
400 static inline void debug_work_deactivate(struct work_struct *work)
402 debug_object_deactivate(work, &work_debug_descr);
405 void __init_work(struct work_struct *work, int onstack)
408 debug_object_init_on_stack(work, &work_debug_descr);
410 debug_object_init(work, &work_debug_descr);
412 EXPORT_SYMBOL_GPL(__init_work);
414 void destroy_work_on_stack(struct work_struct *work)
416 debug_object_free(work, &work_debug_descr);
418 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
421 static inline void debug_work_activate(struct work_struct *work) { }
422 static inline void debug_work_deactivate(struct work_struct *work) { }
425 /* Serializes the accesses to the list of workqueues. */
426 static DEFINE_SPINLOCK(workqueue_lock);
427 static LIST_HEAD(workqueues);
428 static bool workqueue_freezing; /* W: have wqs started freezing? */
431 * The almighty global cpu workqueues. nr_running is the only field
432 * which is expected to be used frequently by other cpus via
433 * try_to_wake_up(). Put it in a separate cacheline.
435 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
436 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
439 * Global cpu workqueue and nr_running counter for unbound gcwq. The
440 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
441 * workers have WORKER_UNBOUND set.
443 static struct global_cwq unbound_global_cwq;
444 static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */
446 static int worker_thread(void *__worker);
448 static struct global_cwq *get_gcwq(unsigned int cpu)
450 if (cpu != WORK_CPU_UNBOUND)
451 return &per_cpu(global_cwq, cpu);
453 return &unbound_global_cwq;
456 static atomic_t *get_gcwq_nr_running(unsigned int cpu)
458 if (cpu != WORK_CPU_UNBOUND)
459 return &per_cpu(gcwq_nr_running, cpu);
461 return &unbound_gcwq_nr_running;
464 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
465 struct workqueue_struct *wq)
467 if (!(wq->flags & WQ_UNBOUND)) {
468 if (likely(cpu < nr_cpu_ids)) {
470 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
472 return wq->cpu_wq.single;
475 } else if (likely(cpu == WORK_CPU_UNBOUND))
476 return wq->cpu_wq.single;
480 static unsigned int work_color_to_flags(int color)
482 return color << WORK_STRUCT_COLOR_SHIFT;
485 static int get_work_color(struct work_struct *work)
487 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
488 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
491 static int work_next_color(int color)
493 return (color + 1) % WORK_NR_COLORS;
497 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
498 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
499 * cleared and the work data contains the cpu number it was last on.
501 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
502 * cwq, cpu or clear work->data. These functions should only be
503 * called while the work is owned - ie. while the PENDING bit is set.
505 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
506 * corresponding to a work. gcwq is available once the work has been
507 * queued anywhere after initialization. cwq is available only from
508 * queueing until execution starts.
510 static inline void set_work_data(struct work_struct *work, unsigned long data,
513 BUG_ON(!work_pending(work));
514 atomic_long_set(&work->data, data | flags | work_static(work));
517 static void set_work_cwq(struct work_struct *work,
518 struct cpu_workqueue_struct *cwq,
519 unsigned long extra_flags)
521 set_work_data(work, (unsigned long)cwq,
522 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
525 static void set_work_cpu(struct work_struct *work, unsigned int cpu)
527 set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
530 static void clear_work_data(struct work_struct *work)
532 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
535 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
537 unsigned long data = atomic_long_read(&work->data);
539 if (data & WORK_STRUCT_CWQ)
540 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
545 static struct global_cwq *get_work_gcwq(struct work_struct *work)
547 unsigned long data = atomic_long_read(&work->data);
550 if (data & WORK_STRUCT_CWQ)
551 return ((struct cpu_workqueue_struct *)
552 (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
554 cpu = data >> WORK_STRUCT_FLAG_BITS;
555 if (cpu == WORK_CPU_NONE)
558 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
559 return get_gcwq(cpu);
563 * Policy functions. These define the policies on how the global
564 * worker pool is managed. Unless noted otherwise, these functions
565 * assume that they're being called with gcwq->lock held.
568 static bool __need_more_worker(struct global_cwq *gcwq)
570 return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) ||
571 gcwq->flags & GCWQ_HIGHPRI_PENDING;
575 * Need to wake up a worker? Called from anything but currently
578 static bool need_more_worker(struct global_cwq *gcwq)
580 return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
583 /* Can I start working? Called from busy but !running workers. */
584 static bool may_start_working(struct global_cwq *gcwq)
586 return gcwq->nr_idle;
589 /* Do I need to keep working? Called from currently running workers. */
590 static bool keep_working(struct global_cwq *gcwq)
592 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
594 return !list_empty(&gcwq->worklist) &&
595 (atomic_read(nr_running) <= 1 ||
596 gcwq->flags & GCWQ_HIGHPRI_PENDING);
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 (!(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 (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;
774 * If transitioning out of NOT_RUNNING, increment nr_running. Note
775 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
776 * of multiple flags, not a single flag.
778 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
779 if (!(worker->flags & WORKER_NOT_RUNNING))
780 atomic_inc(get_gcwq_nr_running(gcwq->cpu));
784 * busy_worker_head - return the busy hash head for a work
785 * @gcwq: gcwq of interest
786 * @work: work to be hashed
788 * Return hash head of @gcwq for @work.
791 * spin_lock_irq(gcwq->lock).
794 * Pointer to the hash head.
796 static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
797 struct work_struct *work)
799 const int base_shift = ilog2(sizeof(struct work_struct));
800 unsigned long v = (unsigned long)work;
802 /* simple shift and fold hash, do we need something better? */
804 v += v >> BUSY_WORKER_HASH_ORDER;
805 v &= BUSY_WORKER_HASH_MASK;
807 return &gcwq->busy_hash[v];
811 * __find_worker_executing_work - find worker which is executing a work
812 * @gcwq: gcwq of interest
813 * @bwh: hash head as returned by busy_worker_head()
814 * @work: work to find worker for
816 * Find a worker which is executing @work on @gcwq. @bwh should be
817 * the hash head obtained by calling busy_worker_head() with the same
821 * spin_lock_irq(gcwq->lock).
824 * Pointer to worker which is executing @work if found, NULL
827 static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
828 struct hlist_head *bwh,
829 struct work_struct *work)
831 struct worker *worker;
832 struct hlist_node *tmp;
834 hlist_for_each_entry(worker, tmp, bwh, hentry)
835 if (worker->current_work == work)
841 * find_worker_executing_work - find worker which is executing a work
842 * @gcwq: gcwq of interest
843 * @work: work to find worker for
845 * Find a worker which is executing @work on @gcwq. This function is
846 * identical to __find_worker_executing_work() except that this
847 * function calculates @bwh itself.
850 * spin_lock_irq(gcwq->lock).
853 * Pointer to worker which is executing @work if found, NULL
856 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
857 struct work_struct *work)
859 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
864 * gcwq_determine_ins_pos - find insertion position
865 * @gcwq: gcwq of interest
866 * @cwq: cwq a work is being queued for
868 * A work for @cwq is about to be queued on @gcwq, determine insertion
869 * position for the work. If @cwq is for HIGHPRI wq, the work is
870 * queued at the head of the queue but in FIFO order with respect to
871 * other HIGHPRI works; otherwise, at the end of the queue. This
872 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
873 * there are HIGHPRI works pending.
876 * spin_lock_irq(gcwq->lock).
879 * Pointer to inserstion position.
881 static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq,
882 struct cpu_workqueue_struct *cwq)
884 struct work_struct *twork;
886 if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
887 return &gcwq->worklist;
889 list_for_each_entry(twork, &gcwq->worklist, entry) {
890 struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
892 if (!(tcwq->wq->flags & WQ_HIGHPRI))
896 gcwq->flags |= GCWQ_HIGHPRI_PENDING;
897 return &twork->entry;
901 * insert_work - insert a work into gcwq
902 * @cwq: cwq @work belongs to
903 * @work: work to insert
904 * @head: insertion point
905 * @extra_flags: extra WORK_STRUCT_* flags to set
907 * Insert @work which belongs to @cwq into @gcwq after @head.
908 * @extra_flags is or'd to work_struct flags.
911 * spin_lock_irq(gcwq->lock).
913 static void insert_work(struct cpu_workqueue_struct *cwq,
914 struct work_struct *work, struct list_head *head,
915 unsigned int extra_flags)
917 struct global_cwq *gcwq = cwq->gcwq;
919 /* we own @work, set data and link */
920 set_work_cwq(work, cwq, extra_flags);
923 * Ensure that we get the right work->data if we see the
924 * result of list_add() below, see try_to_grab_pending().
928 list_add_tail(&work->entry, head);
931 * Ensure either worker_sched_deactivated() sees the above
932 * list_add_tail() or we see zero nr_running to avoid workers
933 * lying around lazily while there are works to be processed.
937 if (__need_more_worker(gcwq))
938 wake_up_worker(gcwq);
942 * Test whether @work is being queued from another work executing on the
943 * same workqueue. This is rather expensive and should only be used from
946 static bool is_chained_work(struct workqueue_struct *wq)
951 for_each_gcwq_cpu(cpu) {
952 struct global_cwq *gcwq = get_gcwq(cpu);
953 struct worker *worker;
954 struct hlist_node *pos;
957 spin_lock_irqsave(&gcwq->lock, flags);
958 for_each_busy_worker(worker, i, pos, gcwq) {
959 if (worker->task != current)
961 spin_unlock_irqrestore(&gcwq->lock, flags);
963 * I'm @worker, no locking necessary. See if @work
964 * is headed to the same workqueue.
966 return worker->current_cwq->wq == wq;
968 spin_unlock_irqrestore(&gcwq->lock, flags);
973 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
974 struct work_struct *work)
976 struct global_cwq *gcwq;
977 struct cpu_workqueue_struct *cwq;
978 struct list_head *worklist;
979 unsigned int work_flags;
982 debug_work_activate(work);
984 /* if dying, only works from the same workqueue are allowed */
985 if (unlikely(wq->flags & WQ_DYING) &&
986 WARN_ON_ONCE(!is_chained_work(wq)))
989 /* determine gcwq to use */
990 if (!(wq->flags & WQ_UNBOUND)) {
991 struct global_cwq *last_gcwq;
993 if (unlikely(cpu == WORK_CPU_UNBOUND))
994 cpu = raw_smp_processor_id();
997 * It's multi cpu. If @wq is non-reentrant and @work
998 * was previously on a different cpu, it might still
999 * be running there, in which case the work needs to
1000 * be queued on that cpu to guarantee non-reentrance.
1002 gcwq = get_gcwq(cpu);
1003 if (wq->flags & WQ_NON_REENTRANT &&
1004 (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
1005 struct worker *worker;
1007 spin_lock_irqsave(&last_gcwq->lock, flags);
1009 worker = find_worker_executing_work(last_gcwq, work);
1011 if (worker && worker->current_cwq->wq == wq)
1014 /* meh... not running there, queue here */
1015 spin_unlock_irqrestore(&last_gcwq->lock, flags);
1016 spin_lock_irqsave(&gcwq->lock, flags);
1019 spin_lock_irqsave(&gcwq->lock, flags);
1021 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1022 spin_lock_irqsave(&gcwq->lock, flags);
1025 /* gcwq determined, get cwq and queue */
1026 cwq = get_cwq(gcwq->cpu, wq);
1027 trace_workqueue_queue_work(cpu, cwq, work);
1029 BUG_ON(!list_empty(&work->entry));
1031 cwq->nr_in_flight[cwq->work_color]++;
1032 work_flags = work_color_to_flags(cwq->work_color);
1034 if (likely(cwq->nr_active < cwq->max_active)) {
1035 trace_workqueue_activate_work(work);
1037 worklist = gcwq_determine_ins_pos(gcwq, cwq);
1039 work_flags |= WORK_STRUCT_DELAYED;
1040 worklist = &cwq->delayed_works;
1043 insert_work(cwq, work, worklist, work_flags);
1045 spin_unlock_irqrestore(&gcwq->lock, flags);
1049 * queue_work - queue work on a workqueue
1050 * @wq: workqueue to use
1051 * @work: work to queue
1053 * Returns 0 if @work was already on a queue, non-zero otherwise.
1055 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1056 * it can be processed by another CPU.
1058 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
1062 ret = queue_work_on(get_cpu(), wq, work);
1067 EXPORT_SYMBOL_GPL(queue_work);
1070 * queue_work_on - queue work on specific cpu
1071 * @cpu: CPU number to execute work on
1072 * @wq: workqueue to use
1073 * @work: work to queue
1075 * Returns 0 if @work was already on a queue, non-zero otherwise.
1077 * We queue the work to a specific CPU, the caller must ensure it
1081 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
1085 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1086 __queue_work(cpu, wq, work);
1091 EXPORT_SYMBOL_GPL(queue_work_on);
1093 static void delayed_work_timer_fn(unsigned long __data)
1095 struct delayed_work *dwork = (struct delayed_work *)__data;
1096 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1098 __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
1102 * queue_delayed_work - queue work on a workqueue after delay
1103 * @wq: workqueue to use
1104 * @dwork: delayable work to queue
1105 * @delay: number of jiffies to wait before queueing
1107 * Returns 0 if @work was already on a queue, non-zero otherwise.
1109 int queue_delayed_work(struct workqueue_struct *wq,
1110 struct delayed_work *dwork, unsigned long delay)
1113 return queue_work(wq, &dwork->work);
1115 return queue_delayed_work_on(-1, wq, dwork, delay);
1117 EXPORT_SYMBOL_GPL(queue_delayed_work);
1120 * queue_delayed_work_on - queue work on specific CPU after delay
1121 * @cpu: CPU number to execute work on
1122 * @wq: workqueue to use
1123 * @dwork: work to queue
1124 * @delay: number of jiffies to wait before queueing
1126 * Returns 0 if @work was already on a queue, non-zero otherwise.
1128 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1129 struct delayed_work *dwork, unsigned long delay)
1132 struct timer_list *timer = &dwork->timer;
1133 struct work_struct *work = &dwork->work;
1135 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1138 BUG_ON(timer_pending(timer));
1139 BUG_ON(!list_empty(&work->entry));
1141 timer_stats_timer_set_start_info(&dwork->timer);
1144 * This stores cwq for the moment, for the timer_fn.
1145 * Note that the work's gcwq is preserved to allow
1146 * reentrance detection for delayed works.
1148 if (!(wq->flags & WQ_UNBOUND)) {
1149 struct global_cwq *gcwq = get_work_gcwq(work);
1151 if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1154 lcpu = raw_smp_processor_id();
1156 lcpu = WORK_CPU_UNBOUND;
1158 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1160 timer->expires = jiffies + delay;
1161 timer->data = (unsigned long)dwork;
1162 timer->function = delayed_work_timer_fn;
1164 if (unlikely(cpu >= 0))
1165 add_timer_on(timer, cpu);
1172 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1175 * worker_enter_idle - enter idle state
1176 * @worker: worker which is entering idle state
1178 * @worker is entering idle state. Update stats and idle timer if
1182 * spin_lock_irq(gcwq->lock).
1184 static void worker_enter_idle(struct worker *worker)
1186 struct global_cwq *gcwq = worker->gcwq;
1188 BUG_ON(worker->flags & WORKER_IDLE);
1189 BUG_ON(!list_empty(&worker->entry) &&
1190 (worker->hentry.next || worker->hentry.pprev));
1192 /* can't use worker_set_flags(), also called from start_worker() */
1193 worker->flags |= WORKER_IDLE;
1195 worker->last_active = jiffies;
1197 /* idle_list is LIFO */
1198 list_add(&worker->entry, &gcwq->idle_list);
1200 if (likely(!(worker->flags & WORKER_ROGUE))) {
1201 if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1202 mod_timer(&gcwq->idle_timer,
1203 jiffies + IDLE_WORKER_TIMEOUT);
1205 wake_up_all(&gcwq->trustee_wait);
1207 /* sanity check nr_running */
1208 WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle &&
1209 atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1213 * worker_leave_idle - leave idle state
1214 * @worker: worker which is leaving idle state
1216 * @worker is leaving idle state. Update stats.
1219 * spin_lock_irq(gcwq->lock).
1221 static void worker_leave_idle(struct worker *worker)
1223 struct global_cwq *gcwq = worker->gcwq;
1225 BUG_ON(!(worker->flags & WORKER_IDLE));
1226 worker_clr_flags(worker, WORKER_IDLE);
1228 list_del_init(&worker->entry);
1232 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1235 * Works which are scheduled while the cpu is online must at least be
1236 * scheduled to a worker which is bound to the cpu so that if they are
1237 * flushed from cpu callbacks while cpu is going down, they are
1238 * guaranteed to execute on the cpu.
1240 * This function is to be used by rogue workers and rescuers to bind
1241 * themselves to the target cpu and may race with cpu going down or
1242 * coming online. kthread_bind() can't be used because it may put the
1243 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1244 * verbatim as it's best effort and blocking and gcwq may be
1245 * [dis]associated in the meantime.
1247 * This function tries set_cpus_allowed() and locks gcwq and verifies
1248 * the binding against GCWQ_DISASSOCIATED which is set during
1249 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1250 * idle state or fetches works without dropping lock, it can guarantee
1251 * the scheduling requirement described in the first paragraph.
1254 * Might sleep. Called without any lock but returns with gcwq->lock
1258 * %true if the associated gcwq is online (@worker is successfully
1259 * bound), %false if offline.
1261 static bool worker_maybe_bind_and_lock(struct worker *worker)
1262 __acquires(&gcwq->lock)
1264 struct global_cwq *gcwq = worker->gcwq;
1265 struct task_struct *task = worker->task;
1269 * The following call may fail, succeed or succeed
1270 * without actually migrating the task to the cpu if
1271 * it races with cpu hotunplug operation. Verify
1272 * against GCWQ_DISASSOCIATED.
1274 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1275 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1277 spin_lock_irq(&gcwq->lock);
1278 if (gcwq->flags & GCWQ_DISASSOCIATED)
1280 if (task_cpu(task) == gcwq->cpu &&
1281 cpumask_equal(¤t->cpus_allowed,
1282 get_cpu_mask(gcwq->cpu)))
1284 spin_unlock_irq(&gcwq->lock);
1286 /* CPU has come up inbetween, retry migration */
1292 * Function for worker->rebind_work used to rebind rogue busy workers
1293 * to the associated cpu which is coming back online. This is
1294 * scheduled by cpu up but can race with other cpu hotplug operations
1295 * and may be executed twice without intervening cpu down.
1297 static void worker_rebind_fn(struct work_struct *work)
1299 struct worker *worker = container_of(work, struct worker, rebind_work);
1300 struct global_cwq *gcwq = worker->gcwq;
1302 if (worker_maybe_bind_and_lock(worker))
1303 worker_clr_flags(worker, WORKER_REBIND);
1305 spin_unlock_irq(&gcwq->lock);
1308 static struct worker *alloc_worker(void)
1310 struct worker *worker;
1312 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1314 INIT_LIST_HEAD(&worker->entry);
1315 INIT_LIST_HEAD(&worker->scheduled);
1316 INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1317 /* on creation a worker is in !idle && prep state */
1318 worker->flags = WORKER_PREP;
1324 * create_worker - create a new workqueue worker
1325 * @gcwq: gcwq the new worker will belong to
1326 * @bind: whether to set affinity to @cpu or not
1328 * Create a new worker which is bound to @gcwq. The returned worker
1329 * can be started by calling start_worker() or destroyed using
1333 * Might sleep. Does GFP_KERNEL allocations.
1336 * Pointer to the newly created worker.
1338 static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
1340 bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1341 struct worker *worker = NULL;
1344 spin_lock_irq(&gcwq->lock);
1345 while (ida_get_new(&gcwq->worker_ida, &id)) {
1346 spin_unlock_irq(&gcwq->lock);
1347 if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1349 spin_lock_irq(&gcwq->lock);
1351 spin_unlock_irq(&gcwq->lock);
1353 worker = alloc_worker();
1357 worker->gcwq = gcwq;
1360 if (!on_unbound_cpu)
1361 worker->task = kthread_create(worker_thread, worker,
1362 "kworker/%u:%d", gcwq->cpu, id);
1364 worker->task = kthread_create(worker_thread, worker,
1365 "kworker/u:%d", id);
1366 if (IS_ERR(worker->task))
1370 * A rogue worker will become a regular one if CPU comes
1371 * online later on. Make sure every worker has
1372 * PF_THREAD_BOUND set.
1374 if (bind && !on_unbound_cpu)
1375 kthread_bind(worker->task, gcwq->cpu);
1377 worker->task->flags |= PF_THREAD_BOUND;
1379 worker->flags |= WORKER_UNBOUND;
1385 spin_lock_irq(&gcwq->lock);
1386 ida_remove(&gcwq->worker_ida, id);
1387 spin_unlock_irq(&gcwq->lock);
1394 * start_worker - start a newly created worker
1395 * @worker: worker to start
1397 * Make the gcwq aware of @worker and start it.
1400 * spin_lock_irq(gcwq->lock).
1402 static void start_worker(struct worker *worker)
1404 worker->flags |= WORKER_STARTED;
1405 worker->gcwq->nr_workers++;
1406 worker_enter_idle(worker);
1407 wake_up_process(worker->task);
1411 * destroy_worker - destroy a workqueue worker
1412 * @worker: worker to be destroyed
1414 * Destroy @worker and adjust @gcwq stats accordingly.
1417 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1419 static void destroy_worker(struct worker *worker)
1421 struct global_cwq *gcwq = worker->gcwq;
1422 int id = worker->id;
1424 /* sanity check frenzy */
1425 BUG_ON(worker->current_work);
1426 BUG_ON(!list_empty(&worker->scheduled));
1428 if (worker->flags & WORKER_STARTED)
1430 if (worker->flags & WORKER_IDLE)
1433 list_del_init(&worker->entry);
1434 worker->flags |= WORKER_DIE;
1436 spin_unlock_irq(&gcwq->lock);
1438 kthread_stop(worker->task);
1441 spin_lock_irq(&gcwq->lock);
1442 ida_remove(&gcwq->worker_ida, id);
1445 static void idle_worker_timeout(unsigned long __gcwq)
1447 struct global_cwq *gcwq = (void *)__gcwq;
1449 spin_lock_irq(&gcwq->lock);
1451 if (too_many_workers(gcwq)) {
1452 struct worker *worker;
1453 unsigned long expires;
1455 /* idle_list is kept in LIFO order, check the last one */
1456 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1457 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1459 if (time_before(jiffies, expires))
1460 mod_timer(&gcwq->idle_timer, expires);
1462 /* it's been idle for too long, wake up manager */
1463 gcwq->flags |= GCWQ_MANAGE_WORKERS;
1464 wake_up_worker(gcwq);
1468 spin_unlock_irq(&gcwq->lock);
1471 static bool send_mayday(struct work_struct *work)
1473 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1474 struct workqueue_struct *wq = cwq->wq;
1477 if (!(wq->flags & WQ_RESCUER))
1480 /* mayday mayday mayday */
1481 cpu = cwq->gcwq->cpu;
1482 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1483 if (cpu == WORK_CPU_UNBOUND)
1485 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1486 wake_up_process(wq->rescuer->task);
1490 static void gcwq_mayday_timeout(unsigned long __gcwq)
1492 struct global_cwq *gcwq = (void *)__gcwq;
1493 struct work_struct *work;
1495 spin_lock_irq(&gcwq->lock);
1497 if (need_to_create_worker(gcwq)) {
1499 * We've been trying to create a new worker but
1500 * haven't been successful. We might be hitting an
1501 * allocation deadlock. Send distress signals to
1504 list_for_each_entry(work, &gcwq->worklist, entry)
1508 spin_unlock_irq(&gcwq->lock);
1510 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1514 * maybe_create_worker - create a new worker if necessary
1515 * @gcwq: gcwq to create a new worker for
1517 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1518 * have at least one idle worker on return from this function. If
1519 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1520 * sent to all rescuers with works scheduled on @gcwq to resolve
1521 * possible allocation deadlock.
1523 * On return, need_to_create_worker() is guaranteed to be false and
1524 * may_start_working() true.
1527 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1528 * multiple times. Does GFP_KERNEL allocations. Called only from
1532 * false if no action was taken and gcwq->lock stayed locked, true
1535 static bool maybe_create_worker(struct global_cwq *gcwq)
1536 __releases(&gcwq->lock)
1537 __acquires(&gcwq->lock)
1539 if (!need_to_create_worker(gcwq))
1542 spin_unlock_irq(&gcwq->lock);
1544 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1545 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1548 struct worker *worker;
1550 worker = create_worker(gcwq, true);
1552 del_timer_sync(&gcwq->mayday_timer);
1553 spin_lock_irq(&gcwq->lock);
1554 start_worker(worker);
1555 BUG_ON(need_to_create_worker(gcwq));
1559 if (!need_to_create_worker(gcwq))
1562 __set_current_state(TASK_INTERRUPTIBLE);
1563 schedule_timeout(CREATE_COOLDOWN);
1565 if (!need_to_create_worker(gcwq))
1569 del_timer_sync(&gcwq->mayday_timer);
1570 spin_lock_irq(&gcwq->lock);
1571 if (need_to_create_worker(gcwq))
1577 * maybe_destroy_worker - destroy workers which have been idle for a while
1578 * @gcwq: gcwq to destroy workers for
1580 * Destroy @gcwq workers which have been idle for longer than
1581 * IDLE_WORKER_TIMEOUT.
1584 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1585 * multiple times. Called only from manager.
1588 * false if no action was taken and gcwq->lock stayed locked, true
1591 static bool maybe_destroy_workers(struct global_cwq *gcwq)
1595 while (too_many_workers(gcwq)) {
1596 struct worker *worker;
1597 unsigned long expires;
1599 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1600 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1602 if (time_before(jiffies, expires)) {
1603 mod_timer(&gcwq->idle_timer, expires);
1607 destroy_worker(worker);
1615 * manage_workers - manage worker pool
1618 * Assume the manager role and manage gcwq worker pool @worker belongs
1619 * to. At any given time, there can be only zero or one manager per
1620 * gcwq. The exclusion is handled automatically by this function.
1622 * The caller can safely start processing works on false return. On
1623 * true return, it's guaranteed that need_to_create_worker() is false
1624 * and may_start_working() is true.
1627 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1628 * multiple times. Does GFP_KERNEL allocations.
1631 * false if no action was taken and gcwq->lock stayed locked, true if
1632 * some action was taken.
1634 static bool manage_workers(struct worker *worker)
1636 struct global_cwq *gcwq = worker->gcwq;
1639 if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1642 gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1643 gcwq->flags |= GCWQ_MANAGING_WORKERS;
1646 * Destroy and then create so that may_start_working() is true
1649 ret |= maybe_destroy_workers(gcwq);
1650 ret |= maybe_create_worker(gcwq);
1652 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1655 * The trustee might be waiting to take over the manager
1656 * position, tell it we're done.
1658 if (unlikely(gcwq->trustee))
1659 wake_up_all(&gcwq->trustee_wait);
1665 * move_linked_works - move linked works to a list
1666 * @work: start of series of works to be scheduled
1667 * @head: target list to append @work to
1668 * @nextp: out paramter for nested worklist walking
1670 * Schedule linked works starting from @work to @head. Work series to
1671 * be scheduled starts at @work and includes any consecutive work with
1672 * WORK_STRUCT_LINKED set in its predecessor.
1674 * If @nextp is not NULL, it's updated to point to the next work of
1675 * the last scheduled work. This allows move_linked_works() to be
1676 * nested inside outer list_for_each_entry_safe().
1679 * spin_lock_irq(gcwq->lock).
1681 static void move_linked_works(struct work_struct *work, struct list_head *head,
1682 struct work_struct **nextp)
1684 struct work_struct *n;
1687 * Linked worklist will always end before the end of the list,
1688 * use NULL for list head.
1690 list_for_each_entry_safe_from(work, n, NULL, entry) {
1691 list_move_tail(&work->entry, head);
1692 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1697 * If we're already inside safe list traversal and have moved
1698 * multiple works to the scheduled queue, the next position
1699 * needs to be updated.
1705 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1707 struct work_struct *work = list_first_entry(&cwq->delayed_works,
1708 struct work_struct, entry);
1709 struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1711 trace_workqueue_activate_work(work);
1712 move_linked_works(work, pos, NULL);
1713 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1718 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1719 * @cwq: cwq of interest
1720 * @color: color of work which left the queue
1721 * @delayed: for a delayed work
1723 * A work either has completed or is removed from pending queue,
1724 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1727 * spin_lock_irq(gcwq->lock).
1729 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1732 /* ignore uncolored works */
1733 if (color == WORK_NO_COLOR)
1736 cwq->nr_in_flight[color]--;
1740 if (!list_empty(&cwq->delayed_works)) {
1741 /* one down, submit a delayed one */
1742 if (cwq->nr_active < cwq->max_active)
1743 cwq_activate_first_delayed(cwq);
1747 /* is flush in progress and are we at the flushing tip? */
1748 if (likely(cwq->flush_color != color))
1751 /* are there still in-flight works? */
1752 if (cwq->nr_in_flight[color])
1755 /* this cwq is done, clear flush_color */
1756 cwq->flush_color = -1;
1759 * If this was the last cwq, wake up the first flusher. It
1760 * will handle the rest.
1762 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1763 complete(&cwq->wq->first_flusher->done);
1767 * process_one_work - process single work
1769 * @work: work to process
1771 * Process @work. This function contains all the logics necessary to
1772 * process a single work including synchronization against and
1773 * interaction with other workers on the same cpu, queueing and
1774 * flushing. As long as context requirement is met, any worker can
1775 * call this function to process a work.
1778 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1780 static void process_one_work(struct worker *worker, struct work_struct *work)
1781 __releases(&gcwq->lock)
1782 __acquires(&gcwq->lock)
1784 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1785 struct global_cwq *gcwq = cwq->gcwq;
1786 struct hlist_head *bwh = busy_worker_head(gcwq, work);
1787 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1788 work_func_t f = work->func;
1790 struct worker *collision;
1791 #ifdef CONFIG_LOCKDEP
1793 * It is permissible to free the struct work_struct from
1794 * inside the function that is called from it, this we need to
1795 * take into account for lockdep too. To avoid bogus "held
1796 * lock freed" warnings as well as problems when looking into
1797 * work->lockdep_map, make a copy and use that here.
1799 struct lockdep_map lockdep_map = work->lockdep_map;
1802 * A single work shouldn't be executed concurrently by
1803 * multiple workers on a single cpu. Check whether anyone is
1804 * already processing the work. If so, defer the work to the
1805 * currently executing one.
1807 collision = __find_worker_executing_work(gcwq, bwh, work);
1808 if (unlikely(collision)) {
1809 move_linked_works(work, &collision->scheduled, NULL);
1813 /* claim and process */
1814 debug_work_deactivate(work);
1815 hlist_add_head(&worker->hentry, bwh);
1816 worker->current_work = work;
1817 worker->current_cwq = cwq;
1818 work_color = get_work_color(work);
1820 /* record the current cpu number in the work data and dequeue */
1821 set_work_cpu(work, gcwq->cpu);
1822 list_del_init(&work->entry);
1825 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1826 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1828 if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1829 struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1830 struct work_struct, entry);
1832 if (!list_empty(&gcwq->worklist) &&
1833 get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1834 wake_up_worker(gcwq);
1836 gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1840 * CPU intensive works don't participate in concurrency
1841 * management. They're the scheduler's responsibility.
1843 if (unlikely(cpu_intensive))
1844 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1846 spin_unlock_irq(&gcwq->lock);
1848 work_clear_pending(work);
1849 lock_map_acquire_read(&cwq->wq->lockdep_map);
1850 lock_map_acquire(&lockdep_map);
1851 trace_workqueue_execute_start(work);
1854 * While we must be careful to not use "work" after this, the trace
1855 * point will only record its address.
1857 trace_workqueue_execute_end(work);
1858 lock_map_release(&lockdep_map);
1859 lock_map_release(&cwq->wq->lockdep_map);
1861 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1862 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1864 current->comm, preempt_count(), task_pid_nr(current));
1865 printk(KERN_ERR " last function: ");
1866 print_symbol("%s\n", (unsigned long)f);
1867 debug_show_held_locks(current);
1871 spin_lock_irq(&gcwq->lock);
1873 /* clear cpu intensive status */
1874 if (unlikely(cpu_intensive))
1875 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1877 /* we're done with it, release */
1878 hlist_del_init(&worker->hentry);
1879 worker->current_work = NULL;
1880 worker->current_cwq = NULL;
1881 cwq_dec_nr_in_flight(cwq, work_color, false);
1885 * process_scheduled_works - process scheduled works
1888 * Process all scheduled works. Please note that the scheduled list
1889 * may change while processing a work, so this function repeatedly
1890 * fetches a work from the top and executes it.
1893 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1896 static void process_scheduled_works(struct worker *worker)
1898 while (!list_empty(&worker->scheduled)) {
1899 struct work_struct *work = list_first_entry(&worker->scheduled,
1900 struct work_struct, entry);
1901 process_one_work(worker, work);
1906 * worker_thread - the worker thread function
1909 * The gcwq worker thread function. There's a single dynamic pool of
1910 * these per each cpu. These workers process all works regardless of
1911 * their specific target workqueue. The only exception is works which
1912 * belong to workqueues with a rescuer which will be explained in
1915 static int worker_thread(void *__worker)
1917 struct worker *worker = __worker;
1918 struct global_cwq *gcwq = worker->gcwq;
1920 /* tell the scheduler that this is a workqueue worker */
1921 worker->task->flags |= PF_WQ_WORKER;
1923 spin_lock_irq(&gcwq->lock);
1925 /* DIE can be set only while we're idle, checking here is enough */
1926 if (worker->flags & WORKER_DIE) {
1927 spin_unlock_irq(&gcwq->lock);
1928 worker->task->flags &= ~PF_WQ_WORKER;
1932 worker_leave_idle(worker);
1934 /* no more worker necessary? */
1935 if (!need_more_worker(gcwq))
1938 /* do we need to manage? */
1939 if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1943 * ->scheduled list can only be filled while a worker is
1944 * preparing to process a work or actually processing it.
1945 * Make sure nobody diddled with it while I was sleeping.
1947 BUG_ON(!list_empty(&worker->scheduled));
1950 * When control reaches this point, we're guaranteed to have
1951 * at least one idle worker or that someone else has already
1952 * assumed the manager role.
1954 worker_clr_flags(worker, WORKER_PREP);
1957 struct work_struct *work =
1958 list_first_entry(&gcwq->worklist,
1959 struct work_struct, entry);
1961 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1962 /* optimization path, not strictly necessary */
1963 process_one_work(worker, work);
1964 if (unlikely(!list_empty(&worker->scheduled)))
1965 process_scheduled_works(worker);
1967 move_linked_works(work, &worker->scheduled, NULL);
1968 process_scheduled_works(worker);
1970 } while (keep_working(gcwq));
1972 worker_set_flags(worker, WORKER_PREP, false);
1974 if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1978 * gcwq->lock is held and there's no work to process and no
1979 * need to manage, sleep. Workers are woken up only while
1980 * holding gcwq->lock or from local cpu, so setting the
1981 * current state before releasing gcwq->lock is enough to
1982 * prevent losing any event.
1984 worker_enter_idle(worker);
1985 __set_current_state(TASK_INTERRUPTIBLE);
1986 spin_unlock_irq(&gcwq->lock);
1992 * rescuer_thread - the rescuer thread function
1993 * @__wq: the associated workqueue
1995 * Workqueue rescuer thread function. There's one rescuer for each
1996 * workqueue which has WQ_RESCUER set.
1998 * Regular work processing on a gcwq may block trying to create a new
1999 * worker which uses GFP_KERNEL allocation which has slight chance of
2000 * developing into deadlock if some works currently on the same queue
2001 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2002 * the problem rescuer solves.
2004 * When such condition is possible, the gcwq summons rescuers of all
2005 * workqueues which have works queued on the gcwq and let them process
2006 * those works so that forward progress can be guaranteed.
2008 * This should happen rarely.
2010 static int rescuer_thread(void *__wq)
2012 struct workqueue_struct *wq = __wq;
2013 struct worker *rescuer = wq->rescuer;
2014 struct list_head *scheduled = &rescuer->scheduled;
2015 bool is_unbound = wq->flags & WQ_UNBOUND;
2018 set_user_nice(current, RESCUER_NICE_LEVEL);
2020 set_current_state(TASK_INTERRUPTIBLE);
2022 if (kthread_should_stop())
2026 * See whether any cpu is asking for help. Unbounded
2027 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2029 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2030 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2031 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2032 struct global_cwq *gcwq = cwq->gcwq;
2033 struct work_struct *work, *n;
2035 __set_current_state(TASK_RUNNING);
2036 mayday_clear_cpu(cpu, wq->mayday_mask);
2038 /* migrate to the target cpu if possible */
2039 rescuer->gcwq = gcwq;
2040 worker_maybe_bind_and_lock(rescuer);
2043 * Slurp in all works issued via this workqueue and
2046 BUG_ON(!list_empty(&rescuer->scheduled));
2047 list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2048 if (get_work_cwq(work) == cwq)
2049 move_linked_works(work, scheduled, &n);
2051 process_scheduled_works(rescuer);
2052 spin_unlock_irq(&gcwq->lock);
2060 struct work_struct work;
2061 struct completion done;
2064 static void wq_barrier_func(struct work_struct *work)
2066 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2067 complete(&barr->done);
2071 * insert_wq_barrier - insert a barrier work
2072 * @cwq: cwq to insert barrier into
2073 * @barr: wq_barrier to insert
2074 * @target: target work to attach @barr to
2075 * @worker: worker currently executing @target, NULL if @target is not executing
2077 * @barr is linked to @target such that @barr is completed only after
2078 * @target finishes execution. Please note that the ordering
2079 * guarantee is observed only with respect to @target and on the local
2082 * Currently, a queued barrier can't be canceled. This is because
2083 * try_to_grab_pending() can't determine whether the work to be
2084 * grabbed is at the head of the queue and thus can't clear LINKED
2085 * flag of the previous work while there must be a valid next work
2086 * after a work with LINKED flag set.
2088 * Note that when @worker is non-NULL, @target may be modified
2089 * underneath us, so we can't reliably determine cwq from @target.
2092 * spin_lock_irq(gcwq->lock).
2094 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2095 struct wq_barrier *barr,
2096 struct work_struct *target, struct worker *worker)
2098 struct list_head *head;
2099 unsigned int linked = 0;
2102 * debugobject calls are safe here even with gcwq->lock locked
2103 * as we know for sure that this will not trigger any of the
2104 * checks and call back into the fixup functions where we
2107 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2108 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2109 init_completion(&barr->done);
2112 * If @target is currently being executed, schedule the
2113 * barrier to the worker; otherwise, put it after @target.
2116 head = worker->scheduled.next;
2118 unsigned long *bits = work_data_bits(target);
2120 head = target->entry.next;
2121 /* there can already be other linked works, inherit and set */
2122 linked = *bits & WORK_STRUCT_LINKED;
2123 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2126 debug_work_activate(&barr->work);
2127 insert_work(cwq, &barr->work, head,
2128 work_color_to_flags(WORK_NO_COLOR) | linked);
2132 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2133 * @wq: workqueue being flushed
2134 * @flush_color: new flush color, < 0 for no-op
2135 * @work_color: new work color, < 0 for no-op
2137 * Prepare cwqs for workqueue flushing.
2139 * If @flush_color is non-negative, flush_color on all cwqs should be
2140 * -1. If no cwq has in-flight commands at the specified color, all
2141 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2142 * has in flight commands, its cwq->flush_color is set to
2143 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2144 * wakeup logic is armed and %true is returned.
2146 * The caller should have initialized @wq->first_flusher prior to
2147 * calling this function with non-negative @flush_color. If
2148 * @flush_color is negative, no flush color update is done and %false
2151 * If @work_color is non-negative, all cwqs should have the same
2152 * work_color which is previous to @work_color and all will be
2153 * advanced to @work_color.
2156 * mutex_lock(wq->flush_mutex).
2159 * %true if @flush_color >= 0 and there's something to flush. %false
2162 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2163 int flush_color, int work_color)
2168 if (flush_color >= 0) {
2169 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2170 atomic_set(&wq->nr_cwqs_to_flush, 1);
2173 for_each_cwq_cpu(cpu, wq) {
2174 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2175 struct global_cwq *gcwq = cwq->gcwq;
2177 spin_lock_irq(&gcwq->lock);
2179 if (flush_color >= 0) {
2180 BUG_ON(cwq->flush_color != -1);
2182 if (cwq->nr_in_flight[flush_color]) {
2183 cwq->flush_color = flush_color;
2184 atomic_inc(&wq->nr_cwqs_to_flush);
2189 if (work_color >= 0) {
2190 BUG_ON(work_color != work_next_color(cwq->work_color));
2191 cwq->work_color = work_color;
2194 spin_unlock_irq(&gcwq->lock);
2197 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2198 complete(&wq->first_flusher->done);
2204 * flush_workqueue - ensure that any scheduled work has run to completion.
2205 * @wq: workqueue to flush
2207 * Forces execution of the workqueue and blocks until its completion.
2208 * This is typically used in driver shutdown handlers.
2210 * We sleep until all works which were queued on entry have been handled,
2211 * but we are not livelocked by new incoming ones.
2213 void flush_workqueue(struct workqueue_struct *wq)
2215 struct wq_flusher this_flusher = {
2216 .list = LIST_HEAD_INIT(this_flusher.list),
2218 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2222 lock_map_acquire(&wq->lockdep_map);
2223 lock_map_release(&wq->lockdep_map);
2225 mutex_lock(&wq->flush_mutex);
2228 * Start-to-wait phase
2230 next_color = work_next_color(wq->work_color);
2232 if (next_color != wq->flush_color) {
2234 * Color space is not full. The current work_color
2235 * becomes our flush_color and work_color is advanced
2238 BUG_ON(!list_empty(&wq->flusher_overflow));
2239 this_flusher.flush_color = wq->work_color;
2240 wq->work_color = next_color;
2242 if (!wq->first_flusher) {
2243 /* no flush in progress, become the first flusher */
2244 BUG_ON(wq->flush_color != this_flusher.flush_color);
2246 wq->first_flusher = &this_flusher;
2248 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2250 /* nothing to flush, done */
2251 wq->flush_color = next_color;
2252 wq->first_flusher = NULL;
2257 BUG_ON(wq->flush_color == this_flusher.flush_color);
2258 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2259 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2263 * Oops, color space is full, wait on overflow queue.
2264 * The next flush completion will assign us
2265 * flush_color and transfer to flusher_queue.
2267 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2270 mutex_unlock(&wq->flush_mutex);
2272 wait_for_completion(&this_flusher.done);
2275 * Wake-up-and-cascade phase
2277 * First flushers are responsible for cascading flushes and
2278 * handling overflow. Non-first flushers can simply return.
2280 if (wq->first_flusher != &this_flusher)
2283 mutex_lock(&wq->flush_mutex);
2285 /* we might have raced, check again with mutex held */
2286 if (wq->first_flusher != &this_flusher)
2289 wq->first_flusher = NULL;
2291 BUG_ON(!list_empty(&this_flusher.list));
2292 BUG_ON(wq->flush_color != this_flusher.flush_color);
2295 struct wq_flusher *next, *tmp;
2297 /* complete all the flushers sharing the current flush color */
2298 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2299 if (next->flush_color != wq->flush_color)
2301 list_del_init(&next->list);
2302 complete(&next->done);
2305 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2306 wq->flush_color != work_next_color(wq->work_color));
2308 /* this flush_color is finished, advance by one */
2309 wq->flush_color = work_next_color(wq->flush_color);
2311 /* one color has been freed, handle overflow queue */
2312 if (!list_empty(&wq->flusher_overflow)) {
2314 * Assign the same color to all overflowed
2315 * flushers, advance work_color and append to
2316 * flusher_queue. This is the start-to-wait
2317 * phase for these overflowed flushers.
2319 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2320 tmp->flush_color = wq->work_color;
2322 wq->work_color = work_next_color(wq->work_color);
2324 list_splice_tail_init(&wq->flusher_overflow,
2325 &wq->flusher_queue);
2326 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2329 if (list_empty(&wq->flusher_queue)) {
2330 BUG_ON(wq->flush_color != wq->work_color);
2335 * Need to flush more colors. Make the next flusher
2336 * the new first flusher and arm cwqs.
2338 BUG_ON(wq->flush_color == wq->work_color);
2339 BUG_ON(wq->flush_color != next->flush_color);
2341 list_del_init(&next->list);
2342 wq->first_flusher = next;
2344 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2348 * Meh... this color is already done, clear first
2349 * flusher and repeat cascading.
2351 wq->first_flusher = NULL;
2355 mutex_unlock(&wq->flush_mutex);
2357 EXPORT_SYMBOL_GPL(flush_workqueue);
2359 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2360 bool wait_executing)
2362 struct worker *worker = NULL;
2363 struct global_cwq *gcwq;
2364 struct cpu_workqueue_struct *cwq;
2367 gcwq = get_work_gcwq(work);
2371 spin_lock_irq(&gcwq->lock);
2372 if (!list_empty(&work->entry)) {
2374 * See the comment near try_to_grab_pending()->smp_rmb().
2375 * If it was re-queued to a different gcwq under us, we
2376 * are not going to wait.
2379 cwq = get_work_cwq(work);
2380 if (unlikely(!cwq || gcwq != cwq->gcwq))
2382 } else if (wait_executing) {
2383 worker = find_worker_executing_work(gcwq, work);
2386 cwq = worker->current_cwq;
2390 insert_wq_barrier(cwq, barr, work, worker);
2391 spin_unlock_irq(&gcwq->lock);
2394 * If @max_active is 1 or rescuer is in use, flushing another work
2395 * item on the same workqueue may lead to deadlock. Make sure the
2396 * flusher is not running on the same workqueue by verifying write
2399 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2400 lock_map_acquire(&cwq->wq->lockdep_map);
2402 lock_map_acquire_read(&cwq->wq->lockdep_map);
2403 lock_map_release(&cwq->wq->lockdep_map);
2407 spin_unlock_irq(&gcwq->lock);
2412 * flush_work - wait for a work to finish executing the last queueing instance
2413 * @work: the work to flush
2415 * Wait until @work has finished execution. This function considers
2416 * only the last queueing instance of @work. If @work has been
2417 * enqueued across different CPUs on a non-reentrant workqueue or on
2418 * multiple workqueues, @work might still be executing on return on
2419 * some of the CPUs from earlier queueing.
2421 * If @work was queued only on a non-reentrant, ordered or unbound
2422 * workqueue, @work is guaranteed to be idle on return if it hasn't
2423 * been requeued since flush started.
2426 * %true if flush_work() waited for the work to finish execution,
2427 * %false if it was already idle.
2429 bool flush_work(struct work_struct *work)
2431 struct wq_barrier barr;
2433 if (start_flush_work(work, &barr, true)) {
2434 wait_for_completion(&barr.done);
2435 destroy_work_on_stack(&barr.work);
2440 EXPORT_SYMBOL_GPL(flush_work);
2442 static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2444 struct wq_barrier barr;
2445 struct worker *worker;
2447 spin_lock_irq(&gcwq->lock);
2449 worker = find_worker_executing_work(gcwq, work);
2450 if (unlikely(worker))
2451 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2453 spin_unlock_irq(&gcwq->lock);
2455 if (unlikely(worker)) {
2456 wait_for_completion(&barr.done);
2457 destroy_work_on_stack(&barr.work);
2463 static bool wait_on_work(struct work_struct *work)
2470 lock_map_acquire(&work->lockdep_map);
2471 lock_map_release(&work->lockdep_map);
2473 for_each_gcwq_cpu(cpu)
2474 ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2479 * flush_work_sync - wait until a work has finished execution
2480 * @work: the work to flush
2482 * Wait until @work has finished execution. On return, it's
2483 * guaranteed that all queueing instances of @work which happened
2484 * before this function is called are finished. In other words, if
2485 * @work hasn't been requeued since this function was called, @work is
2486 * guaranteed to be idle on return.
2489 * %true if flush_work_sync() waited for the work to finish execution,
2490 * %false if it was already idle.
2492 bool flush_work_sync(struct work_struct *work)
2494 struct wq_barrier barr;
2495 bool pending, waited;
2497 /* we'll wait for executions separately, queue barr only if pending */
2498 pending = start_flush_work(work, &barr, false);
2500 /* wait for executions to finish */
2501 waited = wait_on_work(work);
2503 /* wait for the pending one */
2505 wait_for_completion(&barr.done);
2506 destroy_work_on_stack(&barr.work);
2509 return pending || waited;
2511 EXPORT_SYMBOL_GPL(flush_work_sync);
2514 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2515 * so this work can't be re-armed in any way.
2517 static int try_to_grab_pending(struct work_struct *work)
2519 struct global_cwq *gcwq;
2522 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2526 * The queueing is in progress, or it is already queued. Try to
2527 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2529 gcwq = get_work_gcwq(work);
2533 spin_lock_irq(&gcwq->lock);
2534 if (!list_empty(&work->entry)) {
2536 * This work is queued, but perhaps we locked the wrong gcwq.
2537 * In that case we must see the new value after rmb(), see
2538 * insert_work()->wmb().
2541 if (gcwq == get_work_gcwq(work)) {
2542 debug_work_deactivate(work);
2543 list_del_init(&work->entry);
2544 cwq_dec_nr_in_flight(get_work_cwq(work),
2545 get_work_color(work),
2546 *work_data_bits(work) & WORK_STRUCT_DELAYED);
2550 spin_unlock_irq(&gcwq->lock);
2555 static bool __cancel_work_timer(struct work_struct *work,
2556 struct timer_list* timer)
2561 ret = (timer && likely(del_timer(timer)));
2563 ret = try_to_grab_pending(work);
2565 } while (unlikely(ret < 0));
2567 clear_work_data(work);
2572 * cancel_work_sync - cancel a work and wait for it to finish
2573 * @work: the work to cancel
2575 * Cancel @work and wait for its execution to finish. This function
2576 * can be used even if the work re-queues itself or migrates to
2577 * another workqueue. On return from this function, @work is
2578 * guaranteed to be not pending or executing on any CPU.
2580 * cancel_work_sync(&delayed_work->work) must not be used for
2581 * delayed_work's. Use cancel_delayed_work_sync() instead.
2583 * The caller must ensure that the workqueue on which @work was last
2584 * queued can't be destroyed before this function returns.
2587 * %true if @work was pending, %false otherwise.
2589 bool cancel_work_sync(struct work_struct *work)
2591 return __cancel_work_timer(work, NULL);
2593 EXPORT_SYMBOL_GPL(cancel_work_sync);
2596 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2597 * @dwork: the delayed work to flush
2599 * Delayed timer is cancelled and the pending work is queued for
2600 * immediate execution. Like flush_work(), this function only
2601 * considers the last queueing instance of @dwork.
2604 * %true if flush_work() waited for the work to finish execution,
2605 * %false if it was already idle.
2607 bool flush_delayed_work(struct delayed_work *dwork)
2609 if (del_timer_sync(&dwork->timer))
2610 __queue_work(raw_smp_processor_id(),
2611 get_work_cwq(&dwork->work)->wq, &dwork->work);
2612 return flush_work(&dwork->work);
2614 EXPORT_SYMBOL(flush_delayed_work);
2617 * flush_delayed_work_sync - wait for a dwork to finish
2618 * @dwork: the delayed work to flush
2620 * Delayed timer is cancelled and the pending work is queued for
2621 * execution immediately. Other than timer handling, its behavior
2622 * is identical to flush_work_sync().
2625 * %true if flush_work_sync() waited for the work to finish execution,
2626 * %false if it was already idle.
2628 bool flush_delayed_work_sync(struct delayed_work *dwork)
2630 if (del_timer_sync(&dwork->timer))
2631 __queue_work(raw_smp_processor_id(),
2632 get_work_cwq(&dwork->work)->wq, &dwork->work);
2633 return flush_work_sync(&dwork->work);
2635 EXPORT_SYMBOL(flush_delayed_work_sync);
2638 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2639 * @dwork: the delayed work cancel
2641 * This is cancel_work_sync() for delayed works.
2644 * %true if @dwork was pending, %false otherwise.
2646 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2648 return __cancel_work_timer(&dwork->work, &dwork->timer);
2650 EXPORT_SYMBOL(cancel_delayed_work_sync);
2653 * schedule_work - put work task in global workqueue
2654 * @work: job to be done
2656 * Returns zero if @work was already on the kernel-global workqueue and
2657 * non-zero otherwise.
2659 * This puts a job in the kernel-global workqueue if it was not already
2660 * queued and leaves it in the same position on the kernel-global
2661 * workqueue otherwise.
2663 int schedule_work(struct work_struct *work)
2665 return queue_work(system_wq, work);
2667 EXPORT_SYMBOL(schedule_work);
2670 * schedule_work_on - put work task on a specific cpu
2671 * @cpu: cpu to put the work task on
2672 * @work: job to be done
2674 * This puts a job on a specific cpu
2676 int schedule_work_on(int cpu, struct work_struct *work)
2678 return queue_work_on(cpu, system_wq, work);
2680 EXPORT_SYMBOL(schedule_work_on);
2683 * schedule_delayed_work - put work task in global workqueue after delay
2684 * @dwork: job to be done
2685 * @delay: number of jiffies to wait or 0 for immediate execution
2687 * After waiting for a given time this puts a job in the kernel-global
2690 int schedule_delayed_work(struct delayed_work *dwork,
2691 unsigned long delay)
2693 return queue_delayed_work(system_wq, dwork, delay);
2695 EXPORT_SYMBOL(schedule_delayed_work);
2698 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2700 * @dwork: job to be done
2701 * @delay: number of jiffies to wait
2703 * After waiting for a given time this puts a job in the kernel-global
2704 * workqueue on the specified CPU.
2706 int schedule_delayed_work_on(int cpu,
2707 struct delayed_work *dwork, unsigned long delay)
2709 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2711 EXPORT_SYMBOL(schedule_delayed_work_on);
2714 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2715 * @func: the function to call
2717 * schedule_on_each_cpu() executes @func on each online CPU using the
2718 * system workqueue and blocks until all CPUs have completed.
2719 * schedule_on_each_cpu() is very slow.
2722 * 0 on success, -errno on failure.
2724 int schedule_on_each_cpu(work_func_t func)
2727 struct work_struct __percpu *works;
2729 works = alloc_percpu(struct work_struct);
2735 for_each_online_cpu(cpu) {
2736 struct work_struct *work = per_cpu_ptr(works, cpu);
2738 INIT_WORK(work, func);
2739 schedule_work_on(cpu, work);
2742 for_each_online_cpu(cpu)
2743 flush_work(per_cpu_ptr(works, cpu));
2751 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2753 * Forces execution of the kernel-global workqueue and blocks until its
2756 * Think twice before calling this function! It's very easy to get into
2757 * trouble if you don't take great care. Either of the following situations
2758 * will lead to deadlock:
2760 * One of the work items currently on the workqueue needs to acquire
2761 * a lock held by your code or its caller.
2763 * Your code is running in the context of a work routine.
2765 * They will be detected by lockdep when they occur, but the first might not
2766 * occur very often. It depends on what work items are on the workqueue and
2767 * what locks they need, which you have no control over.
2769 * In most situations flushing the entire workqueue is overkill; you merely
2770 * need to know that a particular work item isn't queued and isn't running.
2771 * In such cases you should use cancel_delayed_work_sync() or
2772 * cancel_work_sync() instead.
2774 void flush_scheduled_work(void)
2776 flush_workqueue(system_wq);
2778 EXPORT_SYMBOL(flush_scheduled_work);
2781 * execute_in_process_context - reliably execute the routine with user context
2782 * @fn: the function to execute
2783 * @ew: guaranteed storage for the execute work structure (must
2784 * be available when the work executes)
2786 * Executes the function immediately if process context is available,
2787 * otherwise schedules the function for delayed execution.
2789 * Returns: 0 - function was executed
2790 * 1 - function was scheduled for execution
2792 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2794 if (!in_interrupt()) {
2799 INIT_WORK(&ew->work, fn);
2800 schedule_work(&ew->work);
2804 EXPORT_SYMBOL_GPL(execute_in_process_context);
2806 int keventd_up(void)
2808 return system_wq != NULL;
2811 static int alloc_cwqs(struct workqueue_struct *wq)
2814 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2815 * Make sure that the alignment isn't lower than that of
2816 * unsigned long long.
2818 const size_t size = sizeof(struct cpu_workqueue_struct);
2819 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2820 __alignof__(unsigned long long));
2822 bool percpu = !(wq->flags & WQ_UNBOUND);
2824 bool percpu = false;
2828 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2833 * Allocate enough room to align cwq and put an extra
2834 * pointer at the end pointing back to the originally
2835 * allocated pointer which will be used for free.
2837 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2839 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2840 *(void **)(wq->cpu_wq.single + 1) = ptr;
2844 /* just in case, make sure it's actually aligned
2845 * - this is affected by PERCPU() alignment in vmlinux.lds.S
2847 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2848 return wq->cpu_wq.v ? 0 : -ENOMEM;
2851 static void free_cwqs(struct workqueue_struct *wq)
2854 bool percpu = !(wq->flags & WQ_UNBOUND);
2856 bool percpu = false;
2860 free_percpu(wq->cpu_wq.pcpu);
2861 else if (wq->cpu_wq.single) {
2862 /* the pointer to free is stored right after the cwq */
2863 kfree(*(void **)(wq->cpu_wq.single + 1));
2867 static int wq_clamp_max_active(int max_active, unsigned int flags,
2870 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2872 if (max_active < 1 || max_active > lim)
2873 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2874 "is out of range, clamping between %d and %d\n",
2875 max_active, name, 1, lim);
2877 return clamp_val(max_active, 1, lim);
2880 struct workqueue_struct *__alloc_workqueue_key(const char *name,
2883 struct lock_class_key *key,
2884 const char *lock_name)
2886 struct workqueue_struct *wq;
2890 * Workqueues which may be used during memory reclaim should
2891 * have a rescuer to guarantee forward progress.
2893 if (flags & WQ_MEM_RECLAIM)
2894 flags |= WQ_RESCUER;
2897 * Unbound workqueues aren't concurrency managed and should be
2898 * dispatched to workers immediately.
2900 if (flags & WQ_UNBOUND)
2901 flags |= WQ_HIGHPRI;
2903 max_active = max_active ?: WQ_DFL_ACTIVE;
2904 max_active = wq_clamp_max_active(max_active, flags, name);
2906 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
2911 wq->saved_max_active = max_active;
2912 mutex_init(&wq->flush_mutex);
2913 atomic_set(&wq->nr_cwqs_to_flush, 0);
2914 INIT_LIST_HEAD(&wq->flusher_queue);
2915 INIT_LIST_HEAD(&wq->flusher_overflow);
2918 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
2919 INIT_LIST_HEAD(&wq->list);
2921 if (alloc_cwqs(wq) < 0)
2924 for_each_cwq_cpu(cpu, wq) {
2925 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2926 struct global_cwq *gcwq = get_gcwq(cpu);
2928 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
2931 cwq->flush_color = -1;
2932 cwq->max_active = max_active;
2933 INIT_LIST_HEAD(&cwq->delayed_works);
2936 if (flags & WQ_RESCUER) {
2937 struct worker *rescuer;
2939 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
2942 wq->rescuer = rescuer = alloc_worker();
2946 rescuer->task = kthread_create(rescuer_thread, wq, "%s", name);
2947 if (IS_ERR(rescuer->task))
2950 rescuer->task->flags |= PF_THREAD_BOUND;
2951 wake_up_process(rescuer->task);
2955 * workqueue_lock protects global freeze state and workqueues
2956 * list. Grab it, set max_active accordingly and add the new
2957 * workqueue to workqueues list.
2959 spin_lock(&workqueue_lock);
2961 if (workqueue_freezing && wq->flags & WQ_FREEZEABLE)
2962 for_each_cwq_cpu(cpu, wq)
2963 get_cwq(cpu, wq)->max_active = 0;
2965 list_add(&wq->list, &workqueues);
2967 spin_unlock(&workqueue_lock);
2973 free_mayday_mask(wq->mayday_mask);
2979 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
2982 * destroy_workqueue - safely terminate a workqueue
2983 * @wq: target workqueue
2985 * Safely destroy a workqueue. All work currently pending will be done first.
2987 void destroy_workqueue(struct workqueue_struct *wq)
2989 unsigned int flush_cnt = 0;
2993 * Mark @wq dying and drain all pending works. Once WQ_DYING is
2994 * set, only chain queueing is allowed. IOW, only currently
2995 * pending or running work items on @wq can queue further work
2996 * items on it. @wq is flushed repeatedly until it becomes empty.
2997 * The number of flushing is detemined by the depth of chaining and
2998 * should be relatively short. Whine if it takes too long.
3000 wq->flags |= WQ_DYING;
3002 flush_workqueue(wq);
3004 for_each_cwq_cpu(cpu, wq) {
3005 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3007 if (!cwq->nr_active && list_empty(&cwq->delayed_works))
3010 if (++flush_cnt == 10 ||
3011 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
3012 printk(KERN_WARNING "workqueue %s: flush on "
3013 "destruction isn't complete after %u tries\n",
3014 wq->name, flush_cnt);
3019 * wq list is used to freeze wq, remove from list after
3020 * flushing is complete in case freeze races us.
3022 spin_lock(&workqueue_lock);
3023 list_del(&wq->list);
3024 spin_unlock(&workqueue_lock);
3027 for_each_cwq_cpu(cpu, wq) {
3028 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3031 for (i = 0; i < WORK_NR_COLORS; i++)
3032 BUG_ON(cwq->nr_in_flight[i]);
3033 BUG_ON(cwq->nr_active);
3034 BUG_ON(!list_empty(&cwq->delayed_works));
3037 if (wq->flags & WQ_RESCUER) {
3038 kthread_stop(wq->rescuer->task);
3039 free_mayday_mask(wq->mayday_mask);
3046 EXPORT_SYMBOL_GPL(destroy_workqueue);
3049 * workqueue_set_max_active - adjust max_active of a workqueue
3050 * @wq: target workqueue
3051 * @max_active: new max_active value.
3053 * Set max_active of @wq to @max_active.
3056 * Don't call from IRQ context.
3058 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3062 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3064 spin_lock(&workqueue_lock);
3066 wq->saved_max_active = max_active;
3068 for_each_cwq_cpu(cpu, wq) {
3069 struct global_cwq *gcwq = get_gcwq(cpu);
3071 spin_lock_irq(&gcwq->lock);
3073 if (!(wq->flags & WQ_FREEZEABLE) ||
3074 !(gcwq->flags & GCWQ_FREEZING))
3075 get_cwq(gcwq->cpu, wq)->max_active = max_active;
3077 spin_unlock_irq(&gcwq->lock);
3080 spin_unlock(&workqueue_lock);
3082 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3085 * workqueue_congested - test whether a workqueue is congested
3086 * @cpu: CPU in question
3087 * @wq: target workqueue
3089 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3090 * no synchronization around this function and the test result is
3091 * unreliable and only useful as advisory hints or for debugging.
3094 * %true if congested, %false otherwise.
3096 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3098 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3100 return !list_empty(&cwq->delayed_works);
3102 EXPORT_SYMBOL_GPL(workqueue_congested);
3105 * work_cpu - return the last known associated cpu for @work
3106 * @work: the work of interest
3109 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3111 unsigned int work_cpu(struct work_struct *work)
3113 struct global_cwq *gcwq = get_work_gcwq(work);
3115 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3117 EXPORT_SYMBOL_GPL(work_cpu);
3120 * work_busy - test whether a work is currently pending or running
3121 * @work: the work to be tested
3123 * Test whether @work is currently pending or running. There is no
3124 * synchronization around this function and the test result is
3125 * unreliable and only useful as advisory hints or for debugging.
3126 * Especially for reentrant wqs, the pending state might hide the
3130 * OR'd bitmask of WORK_BUSY_* bits.
3132 unsigned int work_busy(struct work_struct *work)
3134 struct global_cwq *gcwq = get_work_gcwq(work);
3135 unsigned long flags;
3136 unsigned int ret = 0;
3141 spin_lock_irqsave(&gcwq->lock, flags);
3143 if (work_pending(work))
3144 ret |= WORK_BUSY_PENDING;
3145 if (find_worker_executing_work(gcwq, work))
3146 ret |= WORK_BUSY_RUNNING;
3148 spin_unlock_irqrestore(&gcwq->lock, flags);
3152 EXPORT_SYMBOL_GPL(work_busy);
3157 * There are two challenges in supporting CPU hotplug. Firstly, there
3158 * are a lot of assumptions on strong associations among work, cwq and
3159 * gcwq which make migrating pending and scheduled works very
3160 * difficult to implement without impacting hot paths. Secondly,
3161 * gcwqs serve mix of short, long and very long running works making
3162 * blocked draining impractical.
3164 * This is solved by allowing a gcwq to be detached from CPU, running
3165 * it with unbound (rogue) workers and allowing it to be reattached
3166 * later if the cpu comes back online. A separate thread is created
3167 * to govern a gcwq in such state and is called the trustee of the
3170 * Trustee states and their descriptions.
3172 * START Command state used on startup. On CPU_DOWN_PREPARE, a
3173 * new trustee is started with this state.
3175 * IN_CHARGE Once started, trustee will enter this state after
3176 * assuming the manager role and making all existing
3177 * workers rogue. DOWN_PREPARE waits for trustee to
3178 * enter this state. After reaching IN_CHARGE, trustee
3179 * tries to execute the pending worklist until it's empty
3180 * and the state is set to BUTCHER, or the state is set
3183 * BUTCHER Command state which is set by the cpu callback after
3184 * the cpu has went down. Once this state is set trustee
3185 * knows that there will be no new works on the worklist
3186 * and once the worklist is empty it can proceed to
3187 * killing idle workers.
3189 * RELEASE Command state which is set by the cpu callback if the
3190 * cpu down has been canceled or it has come online
3191 * again. After recognizing this state, trustee stops
3192 * trying to drain or butcher and clears ROGUE, rebinds
3193 * all remaining workers back to the cpu and releases
3196 * DONE Trustee will enter this state after BUTCHER or RELEASE
3199 * trustee CPU draining
3200 * took over down complete
3201 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3203 * | CPU is back online v return workers |
3204 * ----------------> RELEASE --------------
3208 * trustee_wait_event_timeout - timed event wait for trustee
3209 * @cond: condition to wait for
3210 * @timeout: timeout in jiffies
3212 * wait_event_timeout() for trustee to use. Handles locking and
3213 * checks for RELEASE request.
3216 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3217 * multiple times. To be used by trustee.
3220 * Positive indicating left time if @cond is satisfied, 0 if timed
3221 * out, -1 if canceled.
3223 #define trustee_wait_event_timeout(cond, timeout) ({ \
3224 long __ret = (timeout); \
3225 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3227 spin_unlock_irq(&gcwq->lock); \
3228 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3229 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3231 spin_lock_irq(&gcwq->lock); \
3233 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3237 * trustee_wait_event - event wait for trustee
3238 * @cond: condition to wait for
3240 * wait_event() for trustee to use. Automatically handles locking and
3241 * checks for CANCEL request.
3244 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3245 * multiple times. To be used by trustee.
3248 * 0 if @cond is satisfied, -1 if canceled.
3250 #define trustee_wait_event(cond) ({ \
3252 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3253 __ret1 < 0 ? -1 : 0; \
3256 static int __cpuinit trustee_thread(void *__gcwq)
3258 struct global_cwq *gcwq = __gcwq;
3259 struct worker *worker;
3260 struct work_struct *work;
3261 struct hlist_node *pos;
3265 BUG_ON(gcwq->cpu != smp_processor_id());
3267 spin_lock_irq(&gcwq->lock);
3269 * Claim the manager position and make all workers rogue.
3270 * Trustee must be bound to the target cpu and can't be
3273 BUG_ON(gcwq->cpu != smp_processor_id());
3274 rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3277 gcwq->flags |= GCWQ_MANAGING_WORKERS;
3279 list_for_each_entry(worker, &gcwq->idle_list, entry)
3280 worker->flags |= WORKER_ROGUE;
3282 for_each_busy_worker(worker, i, pos, gcwq)
3283 worker->flags |= WORKER_ROGUE;
3286 * Call schedule() so that we cross rq->lock and thus can
3287 * guarantee sched callbacks see the rogue flag. This is
3288 * necessary as scheduler callbacks may be invoked from other
3291 spin_unlock_irq(&gcwq->lock);
3293 spin_lock_irq(&gcwq->lock);
3296 * Sched callbacks are disabled now. Zap nr_running. After
3297 * this, nr_running stays zero and need_more_worker() and
3298 * keep_working() are always true as long as the worklist is
3301 atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3303 spin_unlock_irq(&gcwq->lock);
3304 del_timer_sync(&gcwq->idle_timer);
3305 spin_lock_irq(&gcwq->lock);
3308 * We're now in charge. Notify and proceed to drain. We need
3309 * to keep the gcwq running during the whole CPU down
3310 * procedure as other cpu hotunplug callbacks may need to
3311 * flush currently running tasks.
3313 gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3314 wake_up_all(&gcwq->trustee_wait);
3317 * The original cpu is in the process of dying and may go away
3318 * anytime now. When that happens, we and all workers would
3319 * be migrated to other cpus. Try draining any left work. We
3320 * want to get it over with ASAP - spam rescuers, wake up as
3321 * many idlers as necessary and create new ones till the
3322 * worklist is empty. Note that if the gcwq is frozen, there
3323 * may be frozen works in freezeable cwqs. Don't declare
3324 * completion while frozen.
3326 while (gcwq->nr_workers != gcwq->nr_idle ||
3327 gcwq->flags & GCWQ_FREEZING ||
3328 gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3331 list_for_each_entry(work, &gcwq->worklist, entry) {
3336 list_for_each_entry(worker, &gcwq->idle_list, entry) {
3339 wake_up_process(worker->task);
3342 if (need_to_create_worker(gcwq)) {
3343 spin_unlock_irq(&gcwq->lock);
3344 worker = create_worker(gcwq, false);
3345 spin_lock_irq(&gcwq->lock);
3347 worker->flags |= WORKER_ROGUE;
3348 start_worker(worker);
3352 /* give a breather */
3353 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3358 * Either all works have been scheduled and cpu is down, or
3359 * cpu down has already been canceled. Wait for and butcher
3360 * all workers till we're canceled.
3363 rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3364 while (!list_empty(&gcwq->idle_list))
3365 destroy_worker(list_first_entry(&gcwq->idle_list,
3366 struct worker, entry));
3367 } while (gcwq->nr_workers && rc >= 0);
3370 * At this point, either draining has completed and no worker
3371 * is left, or cpu down has been canceled or the cpu is being
3372 * brought back up. There shouldn't be any idle one left.
3373 * Tell the remaining busy ones to rebind once it finishes the
3374 * currently scheduled works by scheduling the rebind_work.
3376 WARN_ON(!list_empty(&gcwq->idle_list));
3378 for_each_busy_worker(worker, i, pos, gcwq) {
3379 struct work_struct *rebind_work = &worker->rebind_work;
3382 * Rebind_work may race with future cpu hotplug
3383 * operations. Use a separate flag to mark that
3384 * rebinding is scheduled.
3386 worker->flags |= WORKER_REBIND;
3387 worker->flags &= ~WORKER_ROGUE;
3389 /* queue rebind_work, wq doesn't matter, use the default one */
3390 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3391 work_data_bits(rebind_work)))
3394 debug_work_activate(rebind_work);
3395 insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3396 worker->scheduled.next,
3397 work_color_to_flags(WORK_NO_COLOR));
3400 /* relinquish manager role */
3401 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3403 /* notify completion */
3404 gcwq->trustee = NULL;
3405 gcwq->trustee_state = TRUSTEE_DONE;
3406 wake_up_all(&gcwq->trustee_wait);
3407 spin_unlock_irq(&gcwq->lock);
3412 * wait_trustee_state - wait for trustee to enter the specified state
3413 * @gcwq: gcwq the trustee of interest belongs to
3414 * @state: target state to wait for
3416 * Wait for the trustee to reach @state. DONE is already matched.
3419 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3420 * multiple times. To be used by cpu_callback.
3422 static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3423 __releases(&gcwq->lock)
3424 __acquires(&gcwq->lock)
3426 if (!(gcwq->trustee_state == state ||
3427 gcwq->trustee_state == TRUSTEE_DONE)) {
3428 spin_unlock_irq(&gcwq->lock);
3429 __wait_event(gcwq->trustee_wait,
3430 gcwq->trustee_state == state ||
3431 gcwq->trustee_state == TRUSTEE_DONE);
3432 spin_lock_irq(&gcwq->lock);
3436 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3437 unsigned long action,
3440 unsigned int cpu = (unsigned long)hcpu;
3441 struct global_cwq *gcwq = get_gcwq(cpu);
3442 struct task_struct *new_trustee = NULL;
3443 struct worker *uninitialized_var(new_worker);
3444 unsigned long flags;
3446 action &= ~CPU_TASKS_FROZEN;
3449 case CPU_DOWN_PREPARE:
3450 new_trustee = kthread_create(trustee_thread, gcwq,
3451 "workqueue_trustee/%d\n", cpu);
3452 if (IS_ERR(new_trustee))
3453 return notifier_from_errno(PTR_ERR(new_trustee));
3454 kthread_bind(new_trustee, cpu);
3456 case CPU_UP_PREPARE:
3457 BUG_ON(gcwq->first_idle);
3458 new_worker = create_worker(gcwq, false);
3461 kthread_stop(new_trustee);
3466 /* some are called w/ irq disabled, don't disturb irq status */
3467 spin_lock_irqsave(&gcwq->lock, flags);
3470 case CPU_DOWN_PREPARE:
3471 /* initialize trustee and tell it to acquire the gcwq */
3472 BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
3473 gcwq->trustee = new_trustee;
3474 gcwq->trustee_state = TRUSTEE_START;
3475 wake_up_process(gcwq->trustee);
3476 wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3478 case CPU_UP_PREPARE:
3479 BUG_ON(gcwq->first_idle);
3480 gcwq->first_idle = new_worker;
3485 * Before this, the trustee and all workers except for
3486 * the ones which are still executing works from
3487 * before the last CPU down must be on the cpu. After
3488 * this, they'll all be diasporas.
3490 gcwq->flags |= GCWQ_DISASSOCIATED;
3494 gcwq->trustee_state = TRUSTEE_BUTCHER;
3496 case CPU_UP_CANCELED:
3497 destroy_worker(gcwq->first_idle);
3498 gcwq->first_idle = NULL;
3501 case CPU_DOWN_FAILED:
3503 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3504 if (gcwq->trustee_state != TRUSTEE_DONE) {
3505 gcwq->trustee_state = TRUSTEE_RELEASE;
3506 wake_up_process(gcwq->trustee);
3507 wait_trustee_state(gcwq, TRUSTEE_DONE);
3511 * Trustee is done and there might be no worker left.
3512 * Put the first_idle in and request a real manager to
3515 spin_unlock_irq(&gcwq->lock);
3516 kthread_bind(gcwq->first_idle->task, cpu);
3517 spin_lock_irq(&gcwq->lock);
3518 gcwq->flags |= GCWQ_MANAGE_WORKERS;
3519 start_worker(gcwq->first_idle);
3520 gcwq->first_idle = NULL;
3524 spin_unlock_irqrestore(&gcwq->lock, flags);
3526 return notifier_from_errno(0);
3531 struct work_for_cpu {
3532 struct completion completion;
3538 static int do_work_for_cpu(void *_wfc)
3540 struct work_for_cpu *wfc = _wfc;
3541 wfc->ret = wfc->fn(wfc->arg);
3542 complete(&wfc->completion);
3547 * work_on_cpu - run a function in user context on a particular cpu
3548 * @cpu: the cpu to run on
3549 * @fn: the function to run
3550 * @arg: the function arg
3552 * This will return the value @fn returns.
3553 * It is up to the caller to ensure that the cpu doesn't go offline.
3554 * The caller must not hold any locks which would prevent @fn from completing.
3556 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3558 struct task_struct *sub_thread;
3559 struct work_for_cpu wfc = {
3560 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3565 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3566 if (IS_ERR(sub_thread))
3567 return PTR_ERR(sub_thread);
3568 kthread_bind(sub_thread, cpu);
3569 wake_up_process(sub_thread);
3570 wait_for_completion(&wfc.completion);
3573 EXPORT_SYMBOL_GPL(work_on_cpu);
3574 #endif /* CONFIG_SMP */
3576 #ifdef CONFIG_FREEZER
3579 * freeze_workqueues_begin - begin freezing workqueues
3581 * Start freezing workqueues. After this function returns, all
3582 * freezeable workqueues will queue new works to their frozen_works
3583 * list instead of gcwq->worklist.
3586 * Grabs and releases workqueue_lock and gcwq->lock's.
3588 void freeze_workqueues_begin(void)
3592 spin_lock(&workqueue_lock);
3594 BUG_ON(workqueue_freezing);
3595 workqueue_freezing = true;
3597 for_each_gcwq_cpu(cpu) {
3598 struct global_cwq *gcwq = get_gcwq(cpu);
3599 struct workqueue_struct *wq;
3601 spin_lock_irq(&gcwq->lock);
3603 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3604 gcwq->flags |= GCWQ_FREEZING;
3606 list_for_each_entry(wq, &workqueues, list) {
3607 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3609 if (cwq && wq->flags & WQ_FREEZEABLE)
3610 cwq->max_active = 0;
3613 spin_unlock_irq(&gcwq->lock);
3616 spin_unlock(&workqueue_lock);
3620 * freeze_workqueues_busy - are freezeable workqueues still busy?
3622 * Check whether freezing is complete. This function must be called
3623 * between freeze_workqueues_begin() and thaw_workqueues().
3626 * Grabs and releases workqueue_lock.
3629 * %true if some freezeable workqueues are still busy. %false if
3630 * freezing is complete.
3632 bool freeze_workqueues_busy(void)
3637 spin_lock(&workqueue_lock);
3639 BUG_ON(!workqueue_freezing);
3641 for_each_gcwq_cpu(cpu) {
3642 struct workqueue_struct *wq;
3644 * nr_active is monotonically decreasing. It's safe
3645 * to peek without lock.
3647 list_for_each_entry(wq, &workqueues, list) {
3648 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3650 if (!cwq || !(wq->flags & WQ_FREEZEABLE))
3653 BUG_ON(cwq->nr_active < 0);
3654 if (cwq->nr_active) {
3661 spin_unlock(&workqueue_lock);
3666 * thaw_workqueues - thaw workqueues
3668 * Thaw workqueues. Normal queueing is restored and all collected
3669 * frozen works are transferred to their respective gcwq worklists.
3672 * Grabs and releases workqueue_lock and gcwq->lock's.
3674 void thaw_workqueues(void)
3678 spin_lock(&workqueue_lock);
3680 if (!workqueue_freezing)
3683 for_each_gcwq_cpu(cpu) {
3684 struct global_cwq *gcwq = get_gcwq(cpu);
3685 struct workqueue_struct *wq;
3687 spin_lock_irq(&gcwq->lock);
3689 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3690 gcwq->flags &= ~GCWQ_FREEZING;
3692 list_for_each_entry(wq, &workqueues, list) {
3693 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3695 if (!cwq || !(wq->flags & WQ_FREEZEABLE))
3698 /* restore max_active and repopulate worklist */
3699 cwq->max_active = wq->saved_max_active;
3701 while (!list_empty(&cwq->delayed_works) &&
3702 cwq->nr_active < cwq->max_active)
3703 cwq_activate_first_delayed(cwq);
3706 wake_up_worker(gcwq);
3708 spin_unlock_irq(&gcwq->lock);
3711 workqueue_freezing = false;
3713 spin_unlock(&workqueue_lock);
3715 #endif /* CONFIG_FREEZER */
3717 static int __init init_workqueues(void)
3722 cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
3724 /* initialize gcwqs */
3725 for_each_gcwq_cpu(cpu) {
3726 struct global_cwq *gcwq = get_gcwq(cpu);
3728 spin_lock_init(&gcwq->lock);
3729 INIT_LIST_HEAD(&gcwq->worklist);
3731 gcwq->flags |= GCWQ_DISASSOCIATED;
3733 INIT_LIST_HEAD(&gcwq->idle_list);
3734 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3735 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3737 init_timer_deferrable(&gcwq->idle_timer);
3738 gcwq->idle_timer.function = idle_worker_timeout;
3739 gcwq->idle_timer.data = (unsigned long)gcwq;
3741 setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3742 (unsigned long)gcwq);
3744 ida_init(&gcwq->worker_ida);
3746 gcwq->trustee_state = TRUSTEE_DONE;
3747 init_waitqueue_head(&gcwq->trustee_wait);
3750 /* create the initial worker */
3751 for_each_online_gcwq_cpu(cpu) {
3752 struct global_cwq *gcwq = get_gcwq(cpu);
3753 struct worker *worker;
3755 if (cpu != WORK_CPU_UNBOUND)
3756 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3757 worker = create_worker(gcwq, true);
3759 spin_lock_irq(&gcwq->lock);
3760 start_worker(worker);
3761 spin_unlock_irq(&gcwq->lock);
3764 system_wq = alloc_workqueue("events", 0, 0);
3765 system_long_wq = alloc_workqueue("events_long", 0, 0);
3766 system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3767 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3768 WQ_UNBOUND_MAX_ACTIVE);
3769 system_freezeable_wq = alloc_workqueue("events_freezeable",
3771 BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq ||
3772 !system_unbound_wq || !system_freezeable_wq);
3775 early_initcall(init_workqueues);