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 >= 2 ? HZ / 100 : 2,
83 /* call for help after 10ms
85 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
86 CREATE_COOLDOWN = HZ, /* time to breath after fail */
87 TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */
90 * Rescue workers are used only on emergencies and shared by
93 RESCUER_NICE_LEVEL = -20,
97 * Structure fields follow one of the following exclusion rules.
99 * I: Modifiable by initialization/destruction paths and read-only for
102 * P: Preemption protected. Disabling preemption is enough and should
103 * only be modified and accessed from the local cpu.
105 * L: gcwq->lock protected. Access with gcwq->lock held.
107 * X: During normal operation, modification requires gcwq->lock and
108 * should be done only from local cpu. Either disabling preemption
109 * on local cpu or grabbing gcwq->lock is enough for read access.
110 * If GCWQ_DISASSOCIATED is set, it's identical to L.
112 * F: wq->flush_mutex protected.
114 * W: workqueue_lock protected.
120 * The poor guys doing the actual heavy lifting. All on-duty workers
121 * are either serving the manager role, on idle list or on busy hash.
124 /* on idle list while idle, on busy hash table while busy */
126 struct list_head entry; /* L: while idle */
127 struct hlist_node hentry; /* L: while busy */
130 struct work_struct *current_work; /* L: work being processed */
131 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
132 struct list_head scheduled; /* L: scheduled works */
133 struct task_struct *task; /* I: worker task */
134 struct global_cwq *gcwq; /* I: the associated gcwq */
135 /* 64 bytes boundary on 64bit, 32 on 32bit */
136 unsigned long last_active; /* L: last active timestamp */
137 unsigned int flags; /* X: flags */
138 int id; /* I: worker id */
139 struct work_struct rebind_work; /* L: rebind worker to cpu */
143 * Global per-cpu workqueue. There's one and only one for each cpu
144 * and all works are queued and processed here regardless of their
148 spinlock_t lock; /* the gcwq lock */
149 struct list_head worklist; /* L: list of pending works */
150 unsigned int cpu; /* I: the associated cpu */
151 unsigned int flags; /* L: GCWQ_* flags */
153 int nr_workers; /* L: total number of workers */
154 int nr_idle; /* L: currently idle ones */
156 /* workers are chained either in the idle_list or busy_hash */
157 struct list_head idle_list; /* X: list of idle workers */
158 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
159 /* L: hash of busy workers */
161 struct timer_list idle_timer; /* L: worker idle timeout */
162 struct timer_list mayday_timer; /* L: SOS timer for dworkers */
164 struct ida worker_ida; /* L: for worker IDs */
166 struct task_struct *trustee; /* L: for gcwq shutdown */
167 unsigned int trustee_state; /* L: trustee state */
168 wait_queue_head_t trustee_wait; /* trustee wait */
169 struct worker *first_idle; /* L: first idle worker */
170 } ____cacheline_aligned_in_smp;
173 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
174 * work_struct->data are used for flags and thus cwqs need to be
175 * aligned at two's power of the number of flag bits.
177 struct cpu_workqueue_struct {
178 struct global_cwq *gcwq; /* I: the associated gcwq */
179 struct workqueue_struct *wq; /* I: the owning workqueue */
180 int work_color; /* L: current color */
181 int flush_color; /* L: flushing color */
182 int nr_in_flight[WORK_NR_COLORS];
183 /* L: nr of in_flight works */
184 int nr_active; /* L: nr of active works */
185 int max_active; /* L: max active works */
186 struct list_head delayed_works; /* L: delayed works */
190 * Structure used to wait for workqueue flush.
193 struct list_head list; /* F: list of flushers */
194 int flush_color; /* F: flush color waiting for */
195 struct completion done; /* flush completion */
199 * All cpumasks are assumed to be always set on UP and thus can't be
200 * used to determine whether there's something to be done.
203 typedef cpumask_var_t mayday_mask_t;
204 #define mayday_test_and_set_cpu(cpu, mask) \
205 cpumask_test_and_set_cpu((cpu), (mask))
206 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
207 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
208 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
209 #define free_mayday_mask(mask) free_cpumask_var((mask))
211 typedef unsigned long mayday_mask_t;
212 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
213 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
214 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
215 #define alloc_mayday_mask(maskp, gfp) true
216 #define free_mayday_mask(mask) do { } while (0)
220 * The externally visible workqueue abstraction is an array of
221 * per-CPU workqueues:
223 struct workqueue_struct {
224 unsigned int flags; /* I: WQ_* flags */
226 struct cpu_workqueue_struct __percpu *pcpu;
227 struct cpu_workqueue_struct *single;
229 } cpu_wq; /* I: cwq's */
230 struct list_head list; /* W: list of all workqueues */
232 struct mutex flush_mutex; /* protects wq flushing */
233 int work_color; /* F: current work color */
234 int flush_color; /* F: current flush color */
235 atomic_t nr_cwqs_to_flush; /* flush in progress */
236 struct wq_flusher *first_flusher; /* F: first flusher */
237 struct list_head flusher_queue; /* F: flush waiters */
238 struct list_head flusher_overflow; /* F: flush overflow list */
240 mayday_mask_t mayday_mask; /* cpus requesting rescue */
241 struct worker *rescuer; /* I: rescue worker */
243 int saved_max_active; /* W: saved cwq max_active */
244 const char *name; /* I: workqueue name */
245 #ifdef CONFIG_LOCKDEP
246 struct lockdep_map lockdep_map;
250 struct workqueue_struct *system_wq __read_mostly;
251 struct workqueue_struct *system_long_wq __read_mostly;
252 struct workqueue_struct *system_nrt_wq __read_mostly;
253 struct workqueue_struct *system_unbound_wq __read_mostly;
254 struct workqueue_struct *system_freezable_wq __read_mostly;
255 struct workqueue_struct *system_nrt_freezable_wq __read_mostly;
256 EXPORT_SYMBOL_GPL(system_wq);
257 EXPORT_SYMBOL_GPL(system_long_wq);
258 EXPORT_SYMBOL_GPL(system_nrt_wq);
259 EXPORT_SYMBOL_GPL(system_unbound_wq);
260 EXPORT_SYMBOL_GPL(system_freezable_wq);
261 EXPORT_SYMBOL_GPL(system_nrt_freezable_wq);
263 #define CREATE_TRACE_POINTS
264 #include <trace/events/workqueue.h>
266 #define for_each_busy_worker(worker, i, pos, gcwq) \
267 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
268 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
270 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
273 if (cpu < nr_cpu_ids) {
275 cpu = cpumask_next(cpu, mask);
276 if (cpu < nr_cpu_ids)
280 return WORK_CPU_UNBOUND;
282 return WORK_CPU_NONE;
285 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
286 struct workqueue_struct *wq)
288 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
294 * An extra gcwq is defined for an invalid cpu number
295 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
296 * specific CPU. The following iterators are similar to
297 * for_each_*_cpu() iterators but also considers the unbound gcwq.
299 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
300 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
301 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
302 * WORK_CPU_UNBOUND for unbound workqueues
304 #define for_each_gcwq_cpu(cpu) \
305 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
306 (cpu) < WORK_CPU_NONE; \
307 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
309 #define for_each_online_gcwq_cpu(cpu) \
310 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
311 (cpu) < WORK_CPU_NONE; \
312 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
314 #define for_each_cwq_cpu(cpu, wq) \
315 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
316 (cpu) < WORK_CPU_NONE; \
317 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
319 #ifdef CONFIG_DEBUG_OBJECTS_WORK
321 static struct debug_obj_descr work_debug_descr;
323 static void *work_debug_hint(void *addr)
325 return ((struct work_struct *) addr)->func;
329 * fixup_init is called when:
330 * - an active object is initialized
332 static int work_fixup_init(void *addr, enum debug_obj_state state)
334 struct work_struct *work = addr;
337 case ODEBUG_STATE_ACTIVE:
338 cancel_work_sync(work);
339 debug_object_init(work, &work_debug_descr);
347 * fixup_activate is called when:
348 * - an active object is activated
349 * - an unknown object is activated (might be a statically initialized object)
351 static int work_fixup_activate(void *addr, enum debug_obj_state state)
353 struct work_struct *work = addr;
357 case ODEBUG_STATE_NOTAVAILABLE:
359 * This is not really a fixup. The work struct was
360 * statically initialized. We just make sure that it
361 * is tracked in the object tracker.
363 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
364 debug_object_init(work, &work_debug_descr);
365 debug_object_activate(work, &work_debug_descr);
371 case ODEBUG_STATE_ACTIVE:
380 * fixup_free is called when:
381 * - an active object is freed
383 static int work_fixup_free(void *addr, enum debug_obj_state state)
385 struct work_struct *work = addr;
388 case ODEBUG_STATE_ACTIVE:
389 cancel_work_sync(work);
390 debug_object_free(work, &work_debug_descr);
397 static struct debug_obj_descr work_debug_descr = {
398 .name = "work_struct",
399 .debug_hint = work_debug_hint,
400 .fixup_init = work_fixup_init,
401 .fixup_activate = work_fixup_activate,
402 .fixup_free = work_fixup_free,
405 static inline void debug_work_activate(struct work_struct *work)
407 debug_object_activate(work, &work_debug_descr);
410 static inline void debug_work_deactivate(struct work_struct *work)
412 debug_object_deactivate(work, &work_debug_descr);
415 void __init_work(struct work_struct *work, int onstack)
418 debug_object_init_on_stack(work, &work_debug_descr);
420 debug_object_init(work, &work_debug_descr);
422 EXPORT_SYMBOL_GPL(__init_work);
424 void destroy_work_on_stack(struct work_struct *work)
426 debug_object_free(work, &work_debug_descr);
428 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
431 static inline void debug_work_activate(struct work_struct *work) { }
432 static inline void debug_work_deactivate(struct work_struct *work) { }
435 /* Serializes the accesses to the list of workqueues. */
436 static DEFINE_SPINLOCK(workqueue_lock);
437 static LIST_HEAD(workqueues);
438 static bool workqueue_freezing; /* W: have wqs started freezing? */
441 * The almighty global cpu workqueues. nr_running is the only field
442 * which is expected to be used frequently by other cpus via
443 * try_to_wake_up(). Put it in a separate cacheline.
445 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
446 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
449 * Global cpu workqueue and nr_running counter for unbound gcwq. The
450 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
451 * workers have WORKER_UNBOUND set.
453 static struct global_cwq unbound_global_cwq;
454 static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */
456 static int worker_thread(void *__worker);
458 static struct global_cwq *get_gcwq(unsigned int cpu)
460 if (cpu != WORK_CPU_UNBOUND)
461 return &per_cpu(global_cwq, cpu);
463 return &unbound_global_cwq;
466 static atomic_t *get_gcwq_nr_running(unsigned int cpu)
468 if (cpu != WORK_CPU_UNBOUND)
469 return &per_cpu(gcwq_nr_running, cpu);
471 return &unbound_gcwq_nr_running;
474 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
475 struct workqueue_struct *wq)
477 if (!(wq->flags & WQ_UNBOUND)) {
478 if (likely(cpu < nr_cpu_ids)) {
480 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
482 return wq->cpu_wq.single;
485 } else if (likely(cpu == WORK_CPU_UNBOUND))
486 return wq->cpu_wq.single;
490 static unsigned int work_color_to_flags(int color)
492 return color << WORK_STRUCT_COLOR_SHIFT;
495 static int get_work_color(struct work_struct *work)
497 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
498 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
501 static int work_next_color(int color)
503 return (color + 1) % WORK_NR_COLORS;
507 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
508 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
509 * cleared and the work data contains the cpu number it was last on.
511 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
512 * cwq, cpu or clear work->data. These functions should only be
513 * called while the work is owned - ie. while the PENDING bit is set.
515 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
516 * corresponding to a work. gcwq is available once the work has been
517 * queued anywhere after initialization. cwq is available only from
518 * queueing until execution starts.
520 static inline void set_work_data(struct work_struct *work, unsigned long data,
523 BUG_ON(!work_pending(work));
524 atomic_long_set(&work->data, data | flags | work_static(work));
527 static void set_work_cwq(struct work_struct *work,
528 struct cpu_workqueue_struct *cwq,
529 unsigned long extra_flags)
531 set_work_data(work, (unsigned long)cwq,
532 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
535 static void set_work_cpu(struct work_struct *work, unsigned int cpu)
537 set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
540 static void clear_work_data(struct work_struct *work)
542 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
545 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
547 unsigned long data = atomic_long_read(&work->data);
549 if (data & WORK_STRUCT_CWQ)
550 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
555 static struct global_cwq *get_work_gcwq(struct work_struct *work)
557 unsigned long data = atomic_long_read(&work->data);
560 if (data & WORK_STRUCT_CWQ)
561 return ((struct cpu_workqueue_struct *)
562 (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
564 cpu = data >> WORK_STRUCT_FLAG_BITS;
565 if (cpu == WORK_CPU_NONE)
568 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
569 return get_gcwq(cpu);
573 * Policy functions. These define the policies on how the global
574 * worker pool is managed. Unless noted otherwise, these functions
575 * assume that they're being called with gcwq->lock held.
578 static bool __need_more_worker(struct global_cwq *gcwq)
580 return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) ||
581 gcwq->flags & GCWQ_HIGHPRI_PENDING;
585 * Need to wake up a worker? Called from anything but currently
588 static bool need_more_worker(struct global_cwq *gcwq)
590 return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
593 /* Can I start working? Called from busy but !running workers. */
594 static bool may_start_working(struct global_cwq *gcwq)
596 return gcwq->nr_idle;
599 /* Do I need to keep working? Called from currently running workers. */
600 static bool keep_working(struct global_cwq *gcwq)
602 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
604 return !list_empty(&gcwq->worklist) &&
605 (atomic_read(nr_running) <= 1 ||
606 gcwq->flags & GCWQ_HIGHPRI_PENDING);
609 /* Do we need a new worker? Called from manager. */
610 static bool need_to_create_worker(struct global_cwq *gcwq)
612 return need_more_worker(gcwq) && !may_start_working(gcwq);
615 /* Do I need to be the manager? */
616 static bool need_to_manage_workers(struct global_cwq *gcwq)
618 return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS;
621 /* Do we have too many workers and should some go away? */
622 static bool too_many_workers(struct global_cwq *gcwq)
624 bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
625 int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
626 int nr_busy = gcwq->nr_workers - nr_idle;
628 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
635 /* Return the first worker. Safe with preemption disabled */
636 static struct worker *first_worker(struct global_cwq *gcwq)
638 if (unlikely(list_empty(&gcwq->idle_list)))
641 return list_first_entry(&gcwq->idle_list, struct worker, entry);
645 * wake_up_worker - wake up an idle worker
646 * @gcwq: gcwq to wake worker for
648 * Wake up the first idle worker of @gcwq.
651 * spin_lock_irq(gcwq->lock).
653 static void wake_up_worker(struct global_cwq *gcwq)
655 struct worker *worker = first_worker(gcwq);
658 wake_up_process(worker->task);
662 * wq_worker_waking_up - a worker is waking up
663 * @task: task waking up
664 * @cpu: CPU @task is waking up to
666 * This function is called during try_to_wake_up() when a worker is
670 * spin_lock_irq(rq->lock)
672 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
674 struct worker *worker = kthread_data(task);
676 if (!(worker->flags & WORKER_NOT_RUNNING))
677 atomic_inc(get_gcwq_nr_running(cpu));
681 * wq_worker_sleeping - a worker is going to sleep
682 * @task: task going to sleep
683 * @cpu: CPU in question, must be the current CPU number
685 * This function is called during schedule() when a busy worker is
686 * going to sleep. Worker on the same cpu can be woken up by
687 * returning pointer to its task.
690 * spin_lock_irq(rq->lock)
693 * Worker task on @cpu to wake up, %NULL if none.
695 struct task_struct *wq_worker_sleeping(struct task_struct *task,
698 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
699 struct global_cwq *gcwq = get_gcwq(cpu);
700 atomic_t *nr_running = get_gcwq_nr_running(cpu);
702 if (worker->flags & WORKER_NOT_RUNNING)
705 /* this can only happen on the local cpu */
706 BUG_ON(cpu != raw_smp_processor_id());
709 * The counterpart of the following dec_and_test, implied mb,
710 * worklist not empty test sequence is in insert_work().
711 * Please read comment there.
713 * NOT_RUNNING is clear. This means that trustee is not in
714 * charge and we're running on the local cpu w/ rq lock held
715 * and preemption disabled, which in turn means that none else
716 * could be manipulating idle_list, so dereferencing idle_list
717 * without gcwq lock is safe.
719 if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
720 to_wakeup = first_worker(gcwq);
721 return to_wakeup ? to_wakeup->task : NULL;
725 * worker_set_flags - set worker flags and adjust nr_running accordingly
727 * @flags: flags to set
728 * @wakeup: wakeup an idle worker if necessary
730 * Set @flags in @worker->flags and adjust nr_running accordingly. If
731 * nr_running becomes zero and @wakeup is %true, an idle worker is
735 * spin_lock_irq(gcwq->lock)
737 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
740 struct global_cwq *gcwq = worker->gcwq;
742 WARN_ON_ONCE(worker->task != current);
745 * If transitioning into NOT_RUNNING, adjust nr_running and
746 * wake up an idle worker as necessary if requested by
749 if ((flags & WORKER_NOT_RUNNING) &&
750 !(worker->flags & WORKER_NOT_RUNNING)) {
751 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
754 if (atomic_dec_and_test(nr_running) &&
755 !list_empty(&gcwq->worklist))
756 wake_up_worker(gcwq);
758 atomic_dec(nr_running);
761 worker->flags |= flags;
765 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
767 * @flags: flags to clear
769 * Clear @flags in @worker->flags and adjust nr_running accordingly.
772 * spin_lock_irq(gcwq->lock)
774 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
776 struct global_cwq *gcwq = worker->gcwq;
777 unsigned int oflags = worker->flags;
779 WARN_ON_ONCE(worker->task != current);
781 worker->flags &= ~flags;
784 * If transitioning out of NOT_RUNNING, increment nr_running. Note
785 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
786 * of multiple flags, not a single flag.
788 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
789 if (!(worker->flags & WORKER_NOT_RUNNING))
790 atomic_inc(get_gcwq_nr_running(gcwq->cpu));
794 * busy_worker_head - return the busy hash head for a work
795 * @gcwq: gcwq of interest
796 * @work: work to be hashed
798 * Return hash head of @gcwq for @work.
801 * spin_lock_irq(gcwq->lock).
804 * Pointer to the hash head.
806 static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
807 struct work_struct *work)
809 const int base_shift = ilog2(sizeof(struct work_struct));
810 unsigned long v = (unsigned long)work;
812 /* simple shift and fold hash, do we need something better? */
814 v += v >> BUSY_WORKER_HASH_ORDER;
815 v &= BUSY_WORKER_HASH_MASK;
817 return &gcwq->busy_hash[v];
821 * __find_worker_executing_work - find worker which is executing a work
822 * @gcwq: gcwq of interest
823 * @bwh: hash head as returned by busy_worker_head()
824 * @work: work to find worker for
826 * Find a worker which is executing @work on @gcwq. @bwh should be
827 * the hash head obtained by calling busy_worker_head() with the same
831 * spin_lock_irq(gcwq->lock).
834 * Pointer to worker which is executing @work if found, NULL
837 static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
838 struct hlist_head *bwh,
839 struct work_struct *work)
841 struct worker *worker;
842 struct hlist_node *tmp;
844 hlist_for_each_entry(worker, tmp, bwh, hentry)
845 if (worker->current_work == work)
851 * find_worker_executing_work - find worker which is executing a work
852 * @gcwq: gcwq of interest
853 * @work: work to find worker for
855 * Find a worker which is executing @work on @gcwq. This function is
856 * identical to __find_worker_executing_work() except that this
857 * function calculates @bwh itself.
860 * spin_lock_irq(gcwq->lock).
863 * Pointer to worker which is executing @work if found, NULL
866 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
867 struct work_struct *work)
869 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
874 * gcwq_determine_ins_pos - find insertion position
875 * @gcwq: gcwq of interest
876 * @cwq: cwq a work is being queued for
878 * A work for @cwq is about to be queued on @gcwq, determine insertion
879 * position for the work. If @cwq is for HIGHPRI wq, the work is
880 * queued at the head of the queue but in FIFO order with respect to
881 * other HIGHPRI works; otherwise, at the end of the queue. This
882 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
883 * there are HIGHPRI works pending.
886 * spin_lock_irq(gcwq->lock).
889 * Pointer to inserstion position.
891 static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq,
892 struct cpu_workqueue_struct *cwq)
894 struct work_struct *twork;
896 if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
897 return &gcwq->worklist;
899 list_for_each_entry(twork, &gcwq->worklist, entry) {
900 struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
902 if (!(tcwq->wq->flags & WQ_HIGHPRI))
906 gcwq->flags |= GCWQ_HIGHPRI_PENDING;
907 return &twork->entry;
911 * insert_work - insert a work into gcwq
912 * @cwq: cwq @work belongs to
913 * @work: work to insert
914 * @head: insertion point
915 * @extra_flags: extra WORK_STRUCT_* flags to set
917 * Insert @work which belongs to @cwq into @gcwq after @head.
918 * @extra_flags is or'd to work_struct flags.
921 * spin_lock_irq(gcwq->lock).
923 static void insert_work(struct cpu_workqueue_struct *cwq,
924 struct work_struct *work, struct list_head *head,
925 unsigned int extra_flags)
927 struct global_cwq *gcwq = cwq->gcwq;
929 /* we own @work, set data and link */
930 set_work_cwq(work, cwq, extra_flags);
933 * Ensure that we get the right work->data if we see the
934 * result of list_add() below, see try_to_grab_pending().
938 list_add_tail(&work->entry, head);
941 * Ensure either worker_sched_deactivated() sees the above
942 * list_add_tail() or we see zero nr_running to avoid workers
943 * lying around lazily while there are works to be processed.
947 if (__need_more_worker(gcwq))
948 wake_up_worker(gcwq);
952 * Test whether @work is being queued from another work executing on the
953 * same workqueue. This is rather expensive and should only be used from
956 static bool is_chained_work(struct workqueue_struct *wq)
961 for_each_gcwq_cpu(cpu) {
962 struct global_cwq *gcwq = get_gcwq(cpu);
963 struct worker *worker;
964 struct hlist_node *pos;
967 spin_lock_irqsave(&gcwq->lock, flags);
968 for_each_busy_worker(worker, i, pos, gcwq) {
969 if (worker->task != current)
971 spin_unlock_irqrestore(&gcwq->lock, flags);
973 * I'm @worker, no locking necessary. See if @work
974 * is headed to the same workqueue.
976 return worker->current_cwq->wq == wq;
978 spin_unlock_irqrestore(&gcwq->lock, flags);
983 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
984 struct work_struct *work)
986 struct global_cwq *gcwq;
987 struct cpu_workqueue_struct *cwq;
988 struct list_head *worklist;
989 unsigned int work_flags;
992 debug_work_activate(work);
994 /* if dying, only works from the same workqueue are allowed */
995 if (unlikely(wq->flags & WQ_DYING) &&
996 WARN_ON_ONCE(!is_chained_work(wq)))
999 /* determine gcwq to use */
1000 if (!(wq->flags & WQ_UNBOUND)) {
1001 struct global_cwq *last_gcwq;
1003 if (unlikely(cpu == WORK_CPU_UNBOUND))
1004 cpu = raw_smp_processor_id();
1007 * It's multi cpu. If @wq is non-reentrant and @work
1008 * was previously on a different cpu, it might still
1009 * be running there, in which case the work needs to
1010 * be queued on that cpu to guarantee non-reentrance.
1012 gcwq = get_gcwq(cpu);
1013 if (wq->flags & WQ_NON_REENTRANT &&
1014 (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
1015 struct worker *worker;
1017 spin_lock_irqsave(&last_gcwq->lock, flags);
1019 worker = find_worker_executing_work(last_gcwq, work);
1021 if (worker && worker->current_cwq->wq == wq)
1024 /* meh... not running there, queue here */
1025 spin_unlock_irqrestore(&last_gcwq->lock, flags);
1026 spin_lock_irqsave(&gcwq->lock, flags);
1029 spin_lock_irqsave(&gcwq->lock, flags);
1031 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1032 spin_lock_irqsave(&gcwq->lock, flags);
1035 /* gcwq determined, get cwq and queue */
1036 cwq = get_cwq(gcwq->cpu, wq);
1037 trace_workqueue_queue_work(cpu, cwq, work);
1039 BUG_ON(!list_empty(&work->entry));
1041 cwq->nr_in_flight[cwq->work_color]++;
1042 work_flags = work_color_to_flags(cwq->work_color);
1044 if (likely(cwq->nr_active < cwq->max_active)) {
1045 trace_workqueue_activate_work(work);
1047 worklist = gcwq_determine_ins_pos(gcwq, cwq);
1049 work_flags |= WORK_STRUCT_DELAYED;
1050 worklist = &cwq->delayed_works;
1053 insert_work(cwq, work, worklist, work_flags);
1055 spin_unlock_irqrestore(&gcwq->lock, flags);
1059 * queue_work - queue work on a workqueue
1060 * @wq: workqueue to use
1061 * @work: work to queue
1063 * Returns 0 if @work was already on a queue, non-zero otherwise.
1065 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1066 * it can be processed by another CPU.
1068 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
1072 ret = queue_work_on(get_cpu(), wq, work);
1077 EXPORT_SYMBOL_GPL(queue_work);
1080 * queue_work_on - queue work on specific cpu
1081 * @cpu: CPU number to execute work on
1082 * @wq: workqueue to use
1083 * @work: work to queue
1085 * Returns 0 if @work was already on a queue, non-zero otherwise.
1087 * We queue the work to a specific CPU, the caller must ensure it
1091 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
1095 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1096 __queue_work(cpu, wq, work);
1101 EXPORT_SYMBOL_GPL(queue_work_on);
1103 static void delayed_work_timer_fn(unsigned long __data)
1105 struct delayed_work *dwork = (struct delayed_work *)__data;
1106 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1108 __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
1112 * queue_delayed_work - queue work on a workqueue after delay
1113 * @wq: workqueue to use
1114 * @dwork: delayable work to queue
1115 * @delay: number of jiffies to wait before queueing
1117 * Returns 0 if @work was already on a queue, non-zero otherwise.
1119 int queue_delayed_work(struct workqueue_struct *wq,
1120 struct delayed_work *dwork, unsigned long delay)
1123 return queue_work(wq, &dwork->work);
1125 return queue_delayed_work_on(-1, wq, dwork, delay);
1127 EXPORT_SYMBOL_GPL(queue_delayed_work);
1130 * queue_delayed_work_on - queue work on specific CPU after delay
1131 * @cpu: CPU number to execute work on
1132 * @wq: workqueue to use
1133 * @dwork: work to queue
1134 * @delay: number of jiffies to wait before queueing
1136 * Returns 0 if @work was already on a queue, non-zero otherwise.
1138 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1139 struct delayed_work *dwork, unsigned long delay)
1142 struct timer_list *timer = &dwork->timer;
1143 struct work_struct *work = &dwork->work;
1145 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1148 BUG_ON(timer_pending(timer));
1149 BUG_ON(!list_empty(&work->entry));
1151 timer_stats_timer_set_start_info(&dwork->timer);
1154 * This stores cwq for the moment, for the timer_fn.
1155 * Note that the work's gcwq is preserved to allow
1156 * reentrance detection for delayed works.
1158 if (!(wq->flags & WQ_UNBOUND)) {
1159 struct global_cwq *gcwq = get_work_gcwq(work);
1161 if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1164 lcpu = raw_smp_processor_id();
1166 lcpu = WORK_CPU_UNBOUND;
1168 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1170 timer->expires = jiffies + delay;
1171 timer->data = (unsigned long)dwork;
1172 timer->function = delayed_work_timer_fn;
1174 if (unlikely(cpu >= 0))
1175 add_timer_on(timer, cpu);
1182 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1185 * worker_enter_idle - enter idle state
1186 * @worker: worker which is entering idle state
1188 * @worker is entering idle state. Update stats and idle timer if
1192 * spin_lock_irq(gcwq->lock).
1194 static void worker_enter_idle(struct worker *worker)
1196 struct global_cwq *gcwq = worker->gcwq;
1198 BUG_ON(worker->flags & WORKER_IDLE);
1199 BUG_ON(!list_empty(&worker->entry) &&
1200 (worker->hentry.next || worker->hentry.pprev));
1202 /* can't use worker_set_flags(), also called from start_worker() */
1203 worker->flags |= WORKER_IDLE;
1205 worker->last_active = jiffies;
1207 /* idle_list is LIFO */
1208 list_add(&worker->entry, &gcwq->idle_list);
1210 if (likely(!(worker->flags & WORKER_ROGUE))) {
1211 if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1212 mod_timer(&gcwq->idle_timer,
1213 jiffies + IDLE_WORKER_TIMEOUT);
1215 wake_up_all(&gcwq->trustee_wait);
1217 /* sanity check nr_running */
1218 WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle &&
1219 atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1223 * worker_leave_idle - leave idle state
1224 * @worker: worker which is leaving idle state
1226 * @worker is leaving idle state. Update stats.
1229 * spin_lock_irq(gcwq->lock).
1231 static void worker_leave_idle(struct worker *worker)
1233 struct global_cwq *gcwq = worker->gcwq;
1235 BUG_ON(!(worker->flags & WORKER_IDLE));
1236 worker_clr_flags(worker, WORKER_IDLE);
1238 list_del_init(&worker->entry);
1242 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1245 * Works which are scheduled while the cpu is online must at least be
1246 * scheduled to a worker which is bound to the cpu so that if they are
1247 * flushed from cpu callbacks while cpu is going down, they are
1248 * guaranteed to execute on the cpu.
1250 * This function is to be used by rogue workers and rescuers to bind
1251 * themselves to the target cpu and may race with cpu going down or
1252 * coming online. kthread_bind() can't be used because it may put the
1253 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1254 * verbatim as it's best effort and blocking and gcwq may be
1255 * [dis]associated in the meantime.
1257 * This function tries set_cpus_allowed() and locks gcwq and verifies
1258 * the binding against GCWQ_DISASSOCIATED which is set during
1259 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1260 * idle state or fetches works without dropping lock, it can guarantee
1261 * the scheduling requirement described in the first paragraph.
1264 * Might sleep. Called without any lock but returns with gcwq->lock
1268 * %true if the associated gcwq is online (@worker is successfully
1269 * bound), %false if offline.
1271 static bool worker_maybe_bind_and_lock(struct worker *worker)
1272 __acquires(&gcwq->lock)
1274 struct global_cwq *gcwq = worker->gcwq;
1275 struct task_struct *task = worker->task;
1279 * The following call may fail, succeed or succeed
1280 * without actually migrating the task to the cpu if
1281 * it races with cpu hotunplug operation. Verify
1282 * against GCWQ_DISASSOCIATED.
1284 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1285 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1287 spin_lock_irq(&gcwq->lock);
1288 if (gcwq->flags & GCWQ_DISASSOCIATED)
1290 if (task_cpu(task) == gcwq->cpu &&
1291 cpumask_equal(¤t->cpus_allowed,
1292 get_cpu_mask(gcwq->cpu)))
1294 spin_unlock_irq(&gcwq->lock);
1297 * We've raced with CPU hot[un]plug. Give it a breather
1298 * and retry migration. cond_resched() is required here;
1299 * otherwise, we might deadlock against cpu_stop trying to
1300 * bring down the CPU on non-preemptive kernel.
1308 * Function for worker->rebind_work used to rebind rogue busy workers
1309 * to the associated cpu which is coming back online. This is
1310 * scheduled by cpu up but can race with other cpu hotplug operations
1311 * and may be executed twice without intervening cpu down.
1313 static void worker_rebind_fn(struct work_struct *work)
1315 struct worker *worker = container_of(work, struct worker, rebind_work);
1316 struct global_cwq *gcwq = worker->gcwq;
1318 if (worker_maybe_bind_and_lock(worker))
1319 worker_clr_flags(worker, WORKER_REBIND);
1321 spin_unlock_irq(&gcwq->lock);
1324 static struct worker *alloc_worker(void)
1326 struct worker *worker;
1328 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1330 INIT_LIST_HEAD(&worker->entry);
1331 INIT_LIST_HEAD(&worker->scheduled);
1332 INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1333 /* on creation a worker is in !idle && prep state */
1334 worker->flags = WORKER_PREP;
1340 * create_worker - create a new workqueue worker
1341 * @gcwq: gcwq the new worker will belong to
1342 * @bind: whether to set affinity to @cpu or not
1344 * Create a new worker which is bound to @gcwq. The returned worker
1345 * can be started by calling start_worker() or destroyed using
1349 * Might sleep. Does GFP_KERNEL allocations.
1352 * Pointer to the newly created worker.
1354 static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
1356 bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1357 struct worker *worker = NULL;
1360 spin_lock_irq(&gcwq->lock);
1361 while (ida_get_new(&gcwq->worker_ida, &id)) {
1362 spin_unlock_irq(&gcwq->lock);
1363 if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1365 spin_lock_irq(&gcwq->lock);
1367 spin_unlock_irq(&gcwq->lock);
1369 worker = alloc_worker();
1373 worker->gcwq = gcwq;
1376 if (!on_unbound_cpu)
1377 worker->task = kthread_create_on_node(worker_thread,
1379 cpu_to_node(gcwq->cpu),
1380 "kworker/%u:%d", gcwq->cpu, id);
1382 worker->task = kthread_create(worker_thread, worker,
1383 "kworker/u:%d", id);
1384 if (IS_ERR(worker->task))
1388 * A rogue worker will become a regular one if CPU comes
1389 * online later on. Make sure every worker has
1390 * PF_THREAD_BOUND set.
1392 if (bind && !on_unbound_cpu)
1393 kthread_bind(worker->task, gcwq->cpu);
1395 worker->task->flags |= PF_THREAD_BOUND;
1397 worker->flags |= WORKER_UNBOUND;
1403 spin_lock_irq(&gcwq->lock);
1404 ida_remove(&gcwq->worker_ida, id);
1405 spin_unlock_irq(&gcwq->lock);
1412 * start_worker - start a newly created worker
1413 * @worker: worker to start
1415 * Make the gcwq aware of @worker and start it.
1418 * spin_lock_irq(gcwq->lock).
1420 static void start_worker(struct worker *worker)
1422 worker->flags |= WORKER_STARTED;
1423 worker->gcwq->nr_workers++;
1424 worker_enter_idle(worker);
1425 wake_up_process(worker->task);
1429 * destroy_worker - destroy a workqueue worker
1430 * @worker: worker to be destroyed
1432 * Destroy @worker and adjust @gcwq stats accordingly.
1435 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1437 static void destroy_worker(struct worker *worker)
1439 struct global_cwq *gcwq = worker->gcwq;
1440 int id = worker->id;
1442 /* sanity check frenzy */
1443 BUG_ON(worker->current_work);
1444 BUG_ON(!list_empty(&worker->scheduled));
1446 if (worker->flags & WORKER_STARTED)
1448 if (worker->flags & WORKER_IDLE)
1451 list_del_init(&worker->entry);
1452 worker->flags |= WORKER_DIE;
1454 spin_unlock_irq(&gcwq->lock);
1456 kthread_stop(worker->task);
1459 spin_lock_irq(&gcwq->lock);
1460 ida_remove(&gcwq->worker_ida, id);
1463 static void idle_worker_timeout(unsigned long __gcwq)
1465 struct global_cwq *gcwq = (void *)__gcwq;
1467 spin_lock_irq(&gcwq->lock);
1469 if (too_many_workers(gcwq)) {
1470 struct worker *worker;
1471 unsigned long expires;
1473 /* idle_list is kept in LIFO order, check the last one */
1474 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1475 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1477 if (time_before(jiffies, expires))
1478 mod_timer(&gcwq->idle_timer, expires);
1480 /* it's been idle for too long, wake up manager */
1481 gcwq->flags |= GCWQ_MANAGE_WORKERS;
1482 wake_up_worker(gcwq);
1486 spin_unlock_irq(&gcwq->lock);
1489 static bool send_mayday(struct work_struct *work)
1491 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1492 struct workqueue_struct *wq = cwq->wq;
1495 if (!(wq->flags & WQ_RESCUER))
1498 /* mayday mayday mayday */
1499 cpu = cwq->gcwq->cpu;
1500 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1501 if (cpu == WORK_CPU_UNBOUND)
1503 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1504 wake_up_process(wq->rescuer->task);
1508 static void gcwq_mayday_timeout(unsigned long __gcwq)
1510 struct global_cwq *gcwq = (void *)__gcwq;
1511 struct work_struct *work;
1513 spin_lock_irq(&gcwq->lock);
1515 if (need_to_create_worker(gcwq)) {
1517 * We've been trying to create a new worker but
1518 * haven't been successful. We might be hitting an
1519 * allocation deadlock. Send distress signals to
1522 list_for_each_entry(work, &gcwq->worklist, entry)
1526 spin_unlock_irq(&gcwq->lock);
1528 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1532 * maybe_create_worker - create a new worker if necessary
1533 * @gcwq: gcwq to create a new worker for
1535 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1536 * have at least one idle worker on return from this function. If
1537 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1538 * sent to all rescuers with works scheduled on @gcwq to resolve
1539 * possible allocation deadlock.
1541 * On return, need_to_create_worker() is guaranteed to be false and
1542 * may_start_working() true.
1545 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1546 * multiple times. Does GFP_KERNEL allocations. Called only from
1550 * false if no action was taken and gcwq->lock stayed locked, true
1553 static bool maybe_create_worker(struct global_cwq *gcwq)
1554 __releases(&gcwq->lock)
1555 __acquires(&gcwq->lock)
1557 if (!need_to_create_worker(gcwq))
1560 spin_unlock_irq(&gcwq->lock);
1562 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1563 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1566 struct worker *worker;
1568 worker = create_worker(gcwq, true);
1570 del_timer_sync(&gcwq->mayday_timer);
1571 spin_lock_irq(&gcwq->lock);
1572 start_worker(worker);
1573 BUG_ON(need_to_create_worker(gcwq));
1577 if (!need_to_create_worker(gcwq))
1580 __set_current_state(TASK_INTERRUPTIBLE);
1581 schedule_timeout(CREATE_COOLDOWN);
1583 if (!need_to_create_worker(gcwq))
1587 del_timer_sync(&gcwq->mayday_timer);
1588 spin_lock_irq(&gcwq->lock);
1589 if (need_to_create_worker(gcwq))
1595 * maybe_destroy_worker - destroy workers which have been idle for a while
1596 * @gcwq: gcwq to destroy workers for
1598 * Destroy @gcwq workers which have been idle for longer than
1599 * IDLE_WORKER_TIMEOUT.
1602 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1603 * multiple times. Called only from manager.
1606 * false if no action was taken and gcwq->lock stayed locked, true
1609 static bool maybe_destroy_workers(struct global_cwq *gcwq)
1613 while (too_many_workers(gcwq)) {
1614 struct worker *worker;
1615 unsigned long expires;
1617 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1618 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1620 if (time_before(jiffies, expires)) {
1621 mod_timer(&gcwq->idle_timer, expires);
1625 destroy_worker(worker);
1633 * manage_workers - manage worker pool
1636 * Assume the manager role and manage gcwq worker pool @worker belongs
1637 * to. At any given time, there can be only zero or one manager per
1638 * gcwq. The exclusion is handled automatically by this function.
1640 * The caller can safely start processing works on false return. On
1641 * true return, it's guaranteed that need_to_create_worker() is false
1642 * and may_start_working() is true.
1645 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1646 * multiple times. Does GFP_KERNEL allocations.
1649 * false if no action was taken and gcwq->lock stayed locked, true if
1650 * some action was taken.
1652 static bool manage_workers(struct worker *worker)
1654 struct global_cwq *gcwq = worker->gcwq;
1657 if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1660 gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1661 gcwq->flags |= GCWQ_MANAGING_WORKERS;
1664 * Destroy and then create so that may_start_working() is true
1667 ret |= maybe_destroy_workers(gcwq);
1668 ret |= maybe_create_worker(gcwq);
1670 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1673 * The trustee might be waiting to take over the manager
1674 * position, tell it we're done.
1676 if (unlikely(gcwq->trustee))
1677 wake_up_all(&gcwq->trustee_wait);
1683 * move_linked_works - move linked works to a list
1684 * @work: start of series of works to be scheduled
1685 * @head: target list to append @work to
1686 * @nextp: out paramter for nested worklist walking
1688 * Schedule linked works starting from @work to @head. Work series to
1689 * be scheduled starts at @work and includes any consecutive work with
1690 * WORK_STRUCT_LINKED set in its predecessor.
1692 * If @nextp is not NULL, it's updated to point to the next work of
1693 * the last scheduled work. This allows move_linked_works() to be
1694 * nested inside outer list_for_each_entry_safe().
1697 * spin_lock_irq(gcwq->lock).
1699 static void move_linked_works(struct work_struct *work, struct list_head *head,
1700 struct work_struct **nextp)
1702 struct work_struct *n;
1705 * Linked worklist will always end before the end of the list,
1706 * use NULL for list head.
1708 list_for_each_entry_safe_from(work, n, NULL, entry) {
1709 list_move_tail(&work->entry, head);
1710 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1715 * If we're already inside safe list traversal and have moved
1716 * multiple works to the scheduled queue, the next position
1717 * needs to be updated.
1723 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1725 struct work_struct *work = list_first_entry(&cwq->delayed_works,
1726 struct work_struct, entry);
1727 struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1729 trace_workqueue_activate_work(work);
1730 move_linked_works(work, pos, NULL);
1731 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1736 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1737 * @cwq: cwq of interest
1738 * @color: color of work which left the queue
1739 * @delayed: for a delayed work
1741 * A work either has completed or is removed from pending queue,
1742 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1745 * spin_lock_irq(gcwq->lock).
1747 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1750 /* ignore uncolored works */
1751 if (color == WORK_NO_COLOR)
1754 cwq->nr_in_flight[color]--;
1758 if (!list_empty(&cwq->delayed_works)) {
1759 /* one down, submit a delayed one */
1760 if (cwq->nr_active < cwq->max_active)
1761 cwq_activate_first_delayed(cwq);
1765 /* is flush in progress and are we at the flushing tip? */
1766 if (likely(cwq->flush_color != color))
1769 /* are there still in-flight works? */
1770 if (cwq->nr_in_flight[color])
1773 /* this cwq is done, clear flush_color */
1774 cwq->flush_color = -1;
1777 * If this was the last cwq, wake up the first flusher. It
1778 * will handle the rest.
1780 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1781 complete(&cwq->wq->first_flusher->done);
1785 * process_one_work - process single work
1787 * @work: work to process
1789 * Process @work. This function contains all the logics necessary to
1790 * process a single work including synchronization against and
1791 * interaction with other workers on the same cpu, queueing and
1792 * flushing. As long as context requirement is met, any worker can
1793 * call this function to process a work.
1796 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1798 static void process_one_work(struct worker *worker, struct work_struct *work)
1799 __releases(&gcwq->lock)
1800 __acquires(&gcwq->lock)
1802 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1803 struct global_cwq *gcwq = cwq->gcwq;
1804 struct hlist_head *bwh = busy_worker_head(gcwq, work);
1805 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1806 work_func_t f = work->func;
1808 struct worker *collision;
1809 #ifdef CONFIG_LOCKDEP
1811 * It is permissible to free the struct work_struct from
1812 * inside the function that is called from it, this we need to
1813 * take into account for lockdep too. To avoid bogus "held
1814 * lock freed" warnings as well as problems when looking into
1815 * work->lockdep_map, make a copy and use that here.
1817 struct lockdep_map lockdep_map = work->lockdep_map;
1820 * A single work shouldn't be executed concurrently by
1821 * multiple workers on a single cpu. Check whether anyone is
1822 * already processing the work. If so, defer the work to the
1823 * currently executing one.
1825 collision = __find_worker_executing_work(gcwq, bwh, work);
1826 if (unlikely(collision)) {
1827 move_linked_works(work, &collision->scheduled, NULL);
1831 /* claim and process */
1832 debug_work_deactivate(work);
1833 hlist_add_head(&worker->hentry, bwh);
1834 worker->current_work = work;
1835 worker->current_cwq = cwq;
1836 work_color = get_work_color(work);
1838 /* record the current cpu number in the work data and dequeue */
1839 set_work_cpu(work, gcwq->cpu);
1840 list_del_init(&work->entry);
1843 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1844 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1846 if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1847 struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1848 struct work_struct, entry);
1850 if (!list_empty(&gcwq->worklist) &&
1851 get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1852 wake_up_worker(gcwq);
1854 gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1858 * CPU intensive works don't participate in concurrency
1859 * management. They're the scheduler's responsibility.
1861 if (unlikely(cpu_intensive))
1862 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1864 spin_unlock_irq(&gcwq->lock);
1866 work_clear_pending(work);
1867 lock_map_acquire_read(&cwq->wq->lockdep_map);
1868 lock_map_acquire(&lockdep_map);
1869 trace_workqueue_execute_start(work);
1872 * While we must be careful to not use "work" after this, the trace
1873 * point will only record its address.
1875 trace_workqueue_execute_end(work);
1876 lock_map_release(&lockdep_map);
1877 lock_map_release(&cwq->wq->lockdep_map);
1879 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1880 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1882 current->comm, preempt_count(), task_pid_nr(current));
1883 printk(KERN_ERR " last function: ");
1884 print_symbol("%s\n", (unsigned long)f);
1885 debug_show_held_locks(current);
1889 spin_lock_irq(&gcwq->lock);
1891 /* clear cpu intensive status */
1892 if (unlikely(cpu_intensive))
1893 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1895 /* we're done with it, release */
1896 hlist_del_init(&worker->hentry);
1897 worker->current_work = NULL;
1898 worker->current_cwq = NULL;
1899 cwq_dec_nr_in_flight(cwq, work_color, false);
1903 * process_scheduled_works - process scheduled works
1906 * Process all scheduled works. Please note that the scheduled list
1907 * may change while processing a work, so this function repeatedly
1908 * fetches a work from the top and executes it.
1911 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1914 static void process_scheduled_works(struct worker *worker)
1916 while (!list_empty(&worker->scheduled)) {
1917 struct work_struct *work = list_first_entry(&worker->scheduled,
1918 struct work_struct, entry);
1919 process_one_work(worker, work);
1924 * worker_thread - the worker thread function
1927 * The gcwq worker thread function. There's a single dynamic pool of
1928 * these per each cpu. These workers process all works regardless of
1929 * their specific target workqueue. The only exception is works which
1930 * belong to workqueues with a rescuer which will be explained in
1933 static int worker_thread(void *__worker)
1935 struct worker *worker = __worker;
1936 struct global_cwq *gcwq = worker->gcwq;
1938 /* tell the scheduler that this is a workqueue worker */
1939 worker->task->flags |= PF_WQ_WORKER;
1941 spin_lock_irq(&gcwq->lock);
1943 /* DIE can be set only while we're idle, checking here is enough */
1944 if (worker->flags & WORKER_DIE) {
1945 spin_unlock_irq(&gcwq->lock);
1946 worker->task->flags &= ~PF_WQ_WORKER;
1950 worker_leave_idle(worker);
1952 /* no more worker necessary? */
1953 if (!need_more_worker(gcwq))
1956 /* do we need to manage? */
1957 if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1961 * ->scheduled list can only be filled while a worker is
1962 * preparing to process a work or actually processing it.
1963 * Make sure nobody diddled with it while I was sleeping.
1965 BUG_ON(!list_empty(&worker->scheduled));
1968 * When control reaches this point, we're guaranteed to have
1969 * at least one idle worker or that someone else has already
1970 * assumed the manager role.
1972 worker_clr_flags(worker, WORKER_PREP);
1975 struct work_struct *work =
1976 list_first_entry(&gcwq->worklist,
1977 struct work_struct, entry);
1979 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1980 /* optimization path, not strictly necessary */
1981 process_one_work(worker, work);
1982 if (unlikely(!list_empty(&worker->scheduled)))
1983 process_scheduled_works(worker);
1985 move_linked_works(work, &worker->scheduled, NULL);
1986 process_scheduled_works(worker);
1988 } while (keep_working(gcwq));
1990 worker_set_flags(worker, WORKER_PREP, false);
1992 if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1996 * gcwq->lock is held and there's no work to process and no
1997 * need to manage, sleep. Workers are woken up only while
1998 * holding gcwq->lock or from local cpu, so setting the
1999 * current state before releasing gcwq->lock is enough to
2000 * prevent losing any event.
2002 worker_enter_idle(worker);
2003 __set_current_state(TASK_INTERRUPTIBLE);
2004 spin_unlock_irq(&gcwq->lock);
2010 * rescuer_thread - the rescuer thread function
2011 * @__wq: the associated workqueue
2013 * Workqueue rescuer thread function. There's one rescuer for each
2014 * workqueue which has WQ_RESCUER set.
2016 * Regular work processing on a gcwq may block trying to create a new
2017 * worker which uses GFP_KERNEL allocation which has slight chance of
2018 * developing into deadlock if some works currently on the same queue
2019 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2020 * the problem rescuer solves.
2022 * When such condition is possible, the gcwq summons rescuers of all
2023 * workqueues which have works queued on the gcwq and let them process
2024 * those works so that forward progress can be guaranteed.
2026 * This should happen rarely.
2028 static int rescuer_thread(void *__wq)
2030 struct workqueue_struct *wq = __wq;
2031 struct worker *rescuer = wq->rescuer;
2032 struct list_head *scheduled = &rescuer->scheduled;
2033 bool is_unbound = wq->flags & WQ_UNBOUND;
2036 set_user_nice(current, RESCUER_NICE_LEVEL);
2038 set_current_state(TASK_INTERRUPTIBLE);
2040 if (kthread_should_stop())
2044 * See whether any cpu is asking for help. Unbounded
2045 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2047 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2048 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2049 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2050 struct global_cwq *gcwq = cwq->gcwq;
2051 struct work_struct *work, *n;
2053 __set_current_state(TASK_RUNNING);
2054 mayday_clear_cpu(cpu, wq->mayday_mask);
2056 /* migrate to the target cpu if possible */
2057 rescuer->gcwq = gcwq;
2058 worker_maybe_bind_and_lock(rescuer);
2061 * Slurp in all works issued via this workqueue and
2064 BUG_ON(!list_empty(&rescuer->scheduled));
2065 list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2066 if (get_work_cwq(work) == cwq)
2067 move_linked_works(work, scheduled, &n);
2069 process_scheduled_works(rescuer);
2072 * Leave this gcwq. If keep_working() is %true, notify a
2073 * regular worker; otherwise, we end up with 0 concurrency
2074 * and stalling the execution.
2076 if (keep_working(gcwq))
2077 wake_up_worker(gcwq);
2079 spin_unlock_irq(&gcwq->lock);
2087 struct work_struct work;
2088 struct completion done;
2091 static void wq_barrier_func(struct work_struct *work)
2093 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2094 complete(&barr->done);
2098 * insert_wq_barrier - insert a barrier work
2099 * @cwq: cwq to insert barrier into
2100 * @barr: wq_barrier to insert
2101 * @target: target work to attach @barr to
2102 * @worker: worker currently executing @target, NULL if @target is not executing
2104 * @barr is linked to @target such that @barr is completed only after
2105 * @target finishes execution. Please note that the ordering
2106 * guarantee is observed only with respect to @target and on the local
2109 * Currently, a queued barrier can't be canceled. This is because
2110 * try_to_grab_pending() can't determine whether the work to be
2111 * grabbed is at the head of the queue and thus can't clear LINKED
2112 * flag of the previous work while there must be a valid next work
2113 * after a work with LINKED flag set.
2115 * Note that when @worker is non-NULL, @target may be modified
2116 * underneath us, so we can't reliably determine cwq from @target.
2119 * spin_lock_irq(gcwq->lock).
2121 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2122 struct wq_barrier *barr,
2123 struct work_struct *target, struct worker *worker)
2125 struct list_head *head;
2126 unsigned int linked = 0;
2129 * debugobject calls are safe here even with gcwq->lock locked
2130 * as we know for sure that this will not trigger any of the
2131 * checks and call back into the fixup functions where we
2134 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2135 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2136 init_completion(&barr->done);
2139 * If @target is currently being executed, schedule the
2140 * barrier to the worker; otherwise, put it after @target.
2143 head = worker->scheduled.next;
2145 unsigned long *bits = work_data_bits(target);
2147 head = target->entry.next;
2148 /* there can already be other linked works, inherit and set */
2149 linked = *bits & WORK_STRUCT_LINKED;
2150 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2153 debug_work_activate(&barr->work);
2154 insert_work(cwq, &barr->work, head,
2155 work_color_to_flags(WORK_NO_COLOR) | linked);
2159 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2160 * @wq: workqueue being flushed
2161 * @flush_color: new flush color, < 0 for no-op
2162 * @work_color: new work color, < 0 for no-op
2164 * Prepare cwqs for workqueue flushing.
2166 * If @flush_color is non-negative, flush_color on all cwqs should be
2167 * -1. If no cwq has in-flight commands at the specified color, all
2168 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2169 * has in flight commands, its cwq->flush_color is set to
2170 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2171 * wakeup logic is armed and %true is returned.
2173 * The caller should have initialized @wq->first_flusher prior to
2174 * calling this function with non-negative @flush_color. If
2175 * @flush_color is negative, no flush color update is done and %false
2178 * If @work_color is non-negative, all cwqs should have the same
2179 * work_color which is previous to @work_color and all will be
2180 * advanced to @work_color.
2183 * mutex_lock(wq->flush_mutex).
2186 * %true if @flush_color >= 0 and there's something to flush. %false
2189 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2190 int flush_color, int work_color)
2195 if (flush_color >= 0) {
2196 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2197 atomic_set(&wq->nr_cwqs_to_flush, 1);
2200 for_each_cwq_cpu(cpu, wq) {
2201 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2202 struct global_cwq *gcwq = cwq->gcwq;
2204 spin_lock_irq(&gcwq->lock);
2206 if (flush_color >= 0) {
2207 BUG_ON(cwq->flush_color != -1);
2209 if (cwq->nr_in_flight[flush_color]) {
2210 cwq->flush_color = flush_color;
2211 atomic_inc(&wq->nr_cwqs_to_flush);
2216 if (work_color >= 0) {
2217 BUG_ON(work_color != work_next_color(cwq->work_color));
2218 cwq->work_color = work_color;
2221 spin_unlock_irq(&gcwq->lock);
2224 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2225 complete(&wq->first_flusher->done);
2231 * flush_workqueue - ensure that any scheduled work has run to completion.
2232 * @wq: workqueue to flush
2234 * Forces execution of the workqueue and blocks until its completion.
2235 * This is typically used in driver shutdown handlers.
2237 * We sleep until all works which were queued on entry have been handled,
2238 * but we are not livelocked by new incoming ones.
2240 void flush_workqueue(struct workqueue_struct *wq)
2242 struct wq_flusher this_flusher = {
2243 .list = LIST_HEAD_INIT(this_flusher.list),
2245 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2249 lock_map_acquire(&wq->lockdep_map);
2250 lock_map_release(&wq->lockdep_map);
2252 mutex_lock(&wq->flush_mutex);
2255 * Start-to-wait phase
2257 next_color = work_next_color(wq->work_color);
2259 if (next_color != wq->flush_color) {
2261 * Color space is not full. The current work_color
2262 * becomes our flush_color and work_color is advanced
2265 BUG_ON(!list_empty(&wq->flusher_overflow));
2266 this_flusher.flush_color = wq->work_color;
2267 wq->work_color = next_color;
2269 if (!wq->first_flusher) {
2270 /* no flush in progress, become the first flusher */
2271 BUG_ON(wq->flush_color != this_flusher.flush_color);
2273 wq->first_flusher = &this_flusher;
2275 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2277 /* nothing to flush, done */
2278 wq->flush_color = next_color;
2279 wq->first_flusher = NULL;
2284 BUG_ON(wq->flush_color == this_flusher.flush_color);
2285 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2286 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2290 * Oops, color space is full, wait on overflow queue.
2291 * The next flush completion will assign us
2292 * flush_color and transfer to flusher_queue.
2294 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2297 mutex_unlock(&wq->flush_mutex);
2299 wait_for_completion(&this_flusher.done);
2302 * Wake-up-and-cascade phase
2304 * First flushers are responsible for cascading flushes and
2305 * handling overflow. Non-first flushers can simply return.
2307 if (wq->first_flusher != &this_flusher)
2310 mutex_lock(&wq->flush_mutex);
2312 /* we might have raced, check again with mutex held */
2313 if (wq->first_flusher != &this_flusher)
2316 wq->first_flusher = NULL;
2318 BUG_ON(!list_empty(&this_flusher.list));
2319 BUG_ON(wq->flush_color != this_flusher.flush_color);
2322 struct wq_flusher *next, *tmp;
2324 /* complete all the flushers sharing the current flush color */
2325 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2326 if (next->flush_color != wq->flush_color)
2328 list_del_init(&next->list);
2329 complete(&next->done);
2332 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2333 wq->flush_color != work_next_color(wq->work_color));
2335 /* this flush_color is finished, advance by one */
2336 wq->flush_color = work_next_color(wq->flush_color);
2338 /* one color has been freed, handle overflow queue */
2339 if (!list_empty(&wq->flusher_overflow)) {
2341 * Assign the same color to all overflowed
2342 * flushers, advance work_color and append to
2343 * flusher_queue. This is the start-to-wait
2344 * phase for these overflowed flushers.
2346 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2347 tmp->flush_color = wq->work_color;
2349 wq->work_color = work_next_color(wq->work_color);
2351 list_splice_tail_init(&wq->flusher_overflow,
2352 &wq->flusher_queue);
2353 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2356 if (list_empty(&wq->flusher_queue)) {
2357 BUG_ON(wq->flush_color != wq->work_color);
2362 * Need to flush more colors. Make the next flusher
2363 * the new first flusher and arm cwqs.
2365 BUG_ON(wq->flush_color == wq->work_color);
2366 BUG_ON(wq->flush_color != next->flush_color);
2368 list_del_init(&next->list);
2369 wq->first_flusher = next;
2371 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2375 * Meh... this color is already done, clear first
2376 * flusher and repeat cascading.
2378 wq->first_flusher = NULL;
2382 mutex_unlock(&wq->flush_mutex);
2384 EXPORT_SYMBOL_GPL(flush_workqueue);
2386 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2387 bool wait_executing)
2389 struct worker *worker = NULL;
2390 struct global_cwq *gcwq;
2391 struct cpu_workqueue_struct *cwq;
2394 gcwq = get_work_gcwq(work);
2398 spin_lock_irq(&gcwq->lock);
2399 if (!list_empty(&work->entry)) {
2401 * See the comment near try_to_grab_pending()->smp_rmb().
2402 * If it was re-queued to a different gcwq under us, we
2403 * are not going to wait.
2406 cwq = get_work_cwq(work);
2407 if (unlikely(!cwq || gcwq != cwq->gcwq))
2409 } else if (wait_executing) {
2410 worker = find_worker_executing_work(gcwq, work);
2413 cwq = worker->current_cwq;
2417 insert_wq_barrier(cwq, barr, work, worker);
2418 spin_unlock_irq(&gcwq->lock);
2421 * If @max_active is 1 or rescuer is in use, flushing another work
2422 * item on the same workqueue may lead to deadlock. Make sure the
2423 * flusher is not running on the same workqueue by verifying write
2426 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2427 lock_map_acquire(&cwq->wq->lockdep_map);
2429 lock_map_acquire_read(&cwq->wq->lockdep_map);
2430 lock_map_release(&cwq->wq->lockdep_map);
2434 spin_unlock_irq(&gcwq->lock);
2439 * flush_work - wait for a work to finish executing the last queueing instance
2440 * @work: the work to flush
2442 * Wait until @work has finished execution. This function considers
2443 * only the last queueing instance of @work. If @work has been
2444 * enqueued across different CPUs on a non-reentrant workqueue or on
2445 * multiple workqueues, @work might still be executing on return on
2446 * some of the CPUs from earlier queueing.
2448 * If @work was queued only on a non-reentrant, ordered or unbound
2449 * workqueue, @work is guaranteed to be idle on return if it hasn't
2450 * been requeued since flush started.
2453 * %true if flush_work() waited for the work to finish execution,
2454 * %false if it was already idle.
2456 bool flush_work(struct work_struct *work)
2458 struct wq_barrier barr;
2460 if (start_flush_work(work, &barr, true)) {
2461 wait_for_completion(&barr.done);
2462 destroy_work_on_stack(&barr.work);
2467 EXPORT_SYMBOL_GPL(flush_work);
2469 static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2471 struct wq_barrier barr;
2472 struct worker *worker;
2474 spin_lock_irq(&gcwq->lock);
2476 worker = find_worker_executing_work(gcwq, work);
2477 if (unlikely(worker))
2478 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2480 spin_unlock_irq(&gcwq->lock);
2482 if (unlikely(worker)) {
2483 wait_for_completion(&barr.done);
2484 destroy_work_on_stack(&barr.work);
2490 static bool wait_on_work(struct work_struct *work)
2497 lock_map_acquire(&work->lockdep_map);
2498 lock_map_release(&work->lockdep_map);
2500 for_each_gcwq_cpu(cpu)
2501 ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2506 * flush_work_sync - wait until a work has finished execution
2507 * @work: the work to flush
2509 * Wait until @work has finished execution. On return, it's
2510 * guaranteed that all queueing instances of @work which happened
2511 * before this function is called are finished. In other words, if
2512 * @work hasn't been requeued since this function was called, @work is
2513 * guaranteed to be idle on return.
2516 * %true if flush_work_sync() waited for the work to finish execution,
2517 * %false if it was already idle.
2519 bool flush_work_sync(struct work_struct *work)
2521 struct wq_barrier barr;
2522 bool pending, waited;
2524 /* we'll wait for executions separately, queue barr only if pending */
2525 pending = start_flush_work(work, &barr, false);
2527 /* wait for executions to finish */
2528 waited = wait_on_work(work);
2530 /* wait for the pending one */
2532 wait_for_completion(&barr.done);
2533 destroy_work_on_stack(&barr.work);
2536 return pending || waited;
2538 EXPORT_SYMBOL_GPL(flush_work_sync);
2541 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2542 * so this work can't be re-armed in any way.
2544 static int try_to_grab_pending(struct work_struct *work)
2546 struct global_cwq *gcwq;
2549 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2553 * The queueing is in progress, or it is already queued. Try to
2554 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2556 gcwq = get_work_gcwq(work);
2560 spin_lock_irq(&gcwq->lock);
2561 if (!list_empty(&work->entry)) {
2563 * This work is queued, but perhaps we locked the wrong gcwq.
2564 * In that case we must see the new value after rmb(), see
2565 * insert_work()->wmb().
2568 if (gcwq == get_work_gcwq(work)) {
2569 debug_work_deactivate(work);
2570 list_del_init(&work->entry);
2571 cwq_dec_nr_in_flight(get_work_cwq(work),
2572 get_work_color(work),
2573 *work_data_bits(work) & WORK_STRUCT_DELAYED);
2577 spin_unlock_irq(&gcwq->lock);
2582 static bool __cancel_work_timer(struct work_struct *work,
2583 struct timer_list* timer)
2588 ret = (timer && likely(del_timer(timer)));
2590 ret = try_to_grab_pending(work);
2592 } while (unlikely(ret < 0));
2594 clear_work_data(work);
2599 * cancel_work_sync - cancel a work and wait for it to finish
2600 * @work: the work to cancel
2602 * Cancel @work and wait for its execution to finish. This function
2603 * can be used even if the work re-queues itself or migrates to
2604 * another workqueue. On return from this function, @work is
2605 * guaranteed to be not pending or executing on any CPU.
2607 * cancel_work_sync(&delayed_work->work) must not be used for
2608 * delayed_work's. Use cancel_delayed_work_sync() instead.
2610 * The caller must ensure that the workqueue on which @work was last
2611 * queued can't be destroyed before this function returns.
2614 * %true if @work was pending, %false otherwise.
2616 bool cancel_work_sync(struct work_struct *work)
2618 return __cancel_work_timer(work, NULL);
2620 EXPORT_SYMBOL_GPL(cancel_work_sync);
2623 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2624 * @dwork: the delayed work to flush
2626 * Delayed timer is cancelled and the pending work is queued for
2627 * immediate execution. Like flush_work(), this function only
2628 * considers the last queueing instance of @dwork.
2631 * %true if flush_work() waited for the work to finish execution,
2632 * %false if it was already idle.
2634 bool flush_delayed_work(struct delayed_work *dwork)
2636 if (del_timer_sync(&dwork->timer))
2637 __queue_work(raw_smp_processor_id(),
2638 get_work_cwq(&dwork->work)->wq, &dwork->work);
2639 return flush_work(&dwork->work);
2641 EXPORT_SYMBOL(flush_delayed_work);
2644 * flush_delayed_work_sync - wait for a dwork to finish
2645 * @dwork: the delayed work to flush
2647 * Delayed timer is cancelled and the pending work is queued for
2648 * execution immediately. Other than timer handling, its behavior
2649 * is identical to flush_work_sync().
2652 * %true if flush_work_sync() waited for the work to finish execution,
2653 * %false if it was already idle.
2655 bool flush_delayed_work_sync(struct delayed_work *dwork)
2657 if (del_timer_sync(&dwork->timer))
2658 __queue_work(raw_smp_processor_id(),
2659 get_work_cwq(&dwork->work)->wq, &dwork->work);
2660 return flush_work_sync(&dwork->work);
2662 EXPORT_SYMBOL(flush_delayed_work_sync);
2665 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2666 * @dwork: the delayed work cancel
2668 * This is cancel_work_sync() for delayed works.
2671 * %true if @dwork was pending, %false otherwise.
2673 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2675 return __cancel_work_timer(&dwork->work, &dwork->timer);
2677 EXPORT_SYMBOL(cancel_delayed_work_sync);
2680 * schedule_work - put work task in global workqueue
2681 * @work: job to be done
2683 * Returns zero if @work was already on the kernel-global workqueue and
2684 * non-zero otherwise.
2686 * This puts a job in the kernel-global workqueue if it was not already
2687 * queued and leaves it in the same position on the kernel-global
2688 * workqueue otherwise.
2690 int schedule_work(struct work_struct *work)
2692 return queue_work(system_wq, work);
2694 EXPORT_SYMBOL(schedule_work);
2697 * schedule_work_on - put work task on a specific cpu
2698 * @cpu: cpu to put the work task on
2699 * @work: job to be done
2701 * This puts a job on a specific cpu
2703 int schedule_work_on(int cpu, struct work_struct *work)
2705 return queue_work_on(cpu, system_wq, work);
2707 EXPORT_SYMBOL(schedule_work_on);
2710 * schedule_delayed_work - put work task in global workqueue after delay
2711 * @dwork: job to be done
2712 * @delay: number of jiffies to wait or 0 for immediate execution
2714 * After waiting for a given time this puts a job in the kernel-global
2717 int schedule_delayed_work(struct delayed_work *dwork,
2718 unsigned long delay)
2720 return queue_delayed_work(system_wq, dwork, delay);
2722 EXPORT_SYMBOL(schedule_delayed_work);
2725 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2727 * @dwork: job to be done
2728 * @delay: number of jiffies to wait
2730 * After waiting for a given time this puts a job in the kernel-global
2731 * workqueue on the specified CPU.
2733 int schedule_delayed_work_on(int cpu,
2734 struct delayed_work *dwork, unsigned long delay)
2736 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2738 EXPORT_SYMBOL(schedule_delayed_work_on);
2741 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2742 * @func: the function to call
2744 * schedule_on_each_cpu() executes @func on each online CPU using the
2745 * system workqueue and blocks until all CPUs have completed.
2746 * schedule_on_each_cpu() is very slow.
2749 * 0 on success, -errno on failure.
2751 int schedule_on_each_cpu(work_func_t func)
2754 struct work_struct __percpu *works;
2756 works = alloc_percpu(struct work_struct);
2762 for_each_online_cpu(cpu) {
2763 struct work_struct *work = per_cpu_ptr(works, cpu);
2765 INIT_WORK(work, func);
2766 schedule_work_on(cpu, work);
2769 for_each_online_cpu(cpu)
2770 flush_work(per_cpu_ptr(works, cpu));
2778 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2780 * Forces execution of the kernel-global workqueue and blocks until its
2783 * Think twice before calling this function! It's very easy to get into
2784 * trouble if you don't take great care. Either of the following situations
2785 * will lead to deadlock:
2787 * One of the work items currently on the workqueue needs to acquire
2788 * a lock held by your code or its caller.
2790 * Your code is running in the context of a work routine.
2792 * They will be detected by lockdep when they occur, but the first might not
2793 * occur very often. It depends on what work items are on the workqueue and
2794 * what locks they need, which you have no control over.
2796 * In most situations flushing the entire workqueue is overkill; you merely
2797 * need to know that a particular work item isn't queued and isn't running.
2798 * In such cases you should use cancel_delayed_work_sync() or
2799 * cancel_work_sync() instead.
2801 void flush_scheduled_work(void)
2803 flush_workqueue(system_wq);
2805 EXPORT_SYMBOL(flush_scheduled_work);
2808 * execute_in_process_context - reliably execute the routine with user context
2809 * @fn: the function to execute
2810 * @ew: guaranteed storage for the execute work structure (must
2811 * be available when the work executes)
2813 * Executes the function immediately if process context is available,
2814 * otherwise schedules the function for delayed execution.
2816 * Returns: 0 - function was executed
2817 * 1 - function was scheduled for execution
2819 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2821 if (!in_interrupt()) {
2826 INIT_WORK(&ew->work, fn);
2827 schedule_work(&ew->work);
2831 EXPORT_SYMBOL_GPL(execute_in_process_context);
2833 int keventd_up(void)
2835 return system_wq != NULL;
2838 static int alloc_cwqs(struct workqueue_struct *wq)
2841 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2842 * Make sure that the alignment isn't lower than that of
2843 * unsigned long long.
2845 const size_t size = sizeof(struct cpu_workqueue_struct);
2846 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2847 __alignof__(unsigned long long));
2849 bool percpu = !(wq->flags & WQ_UNBOUND);
2851 bool percpu = false;
2855 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2860 * Allocate enough room to align cwq and put an extra
2861 * pointer at the end pointing back to the originally
2862 * allocated pointer which will be used for free.
2864 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2866 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2867 *(void **)(wq->cpu_wq.single + 1) = ptr;
2871 /* just in case, make sure it's actually aligned */
2872 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2873 return wq->cpu_wq.v ? 0 : -ENOMEM;
2876 static void free_cwqs(struct workqueue_struct *wq)
2879 bool percpu = !(wq->flags & WQ_UNBOUND);
2881 bool percpu = false;
2885 free_percpu(wq->cpu_wq.pcpu);
2886 else if (wq->cpu_wq.single) {
2887 /* the pointer to free is stored right after the cwq */
2888 kfree(*(void **)(wq->cpu_wq.single + 1));
2892 static int wq_clamp_max_active(int max_active, unsigned int flags,
2895 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2897 if (max_active < 1 || max_active > lim)
2898 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2899 "is out of range, clamping between %d and %d\n",
2900 max_active, name, 1, lim);
2902 return clamp_val(max_active, 1, lim);
2905 struct workqueue_struct *__alloc_workqueue_key(const char *name,
2908 struct lock_class_key *key,
2909 const char *lock_name)
2911 struct workqueue_struct *wq;
2915 * Workqueues which may be used during memory reclaim should
2916 * have a rescuer to guarantee forward progress.
2918 if (flags & WQ_MEM_RECLAIM)
2919 flags |= WQ_RESCUER;
2922 * Unbound workqueues aren't concurrency managed and should be
2923 * dispatched to workers immediately.
2925 if (flags & WQ_UNBOUND)
2926 flags |= WQ_HIGHPRI;
2928 max_active = max_active ?: WQ_DFL_ACTIVE;
2929 max_active = wq_clamp_max_active(max_active, flags, name);
2931 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
2936 wq->saved_max_active = max_active;
2937 mutex_init(&wq->flush_mutex);
2938 atomic_set(&wq->nr_cwqs_to_flush, 0);
2939 INIT_LIST_HEAD(&wq->flusher_queue);
2940 INIT_LIST_HEAD(&wq->flusher_overflow);
2943 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
2944 INIT_LIST_HEAD(&wq->list);
2946 if (alloc_cwqs(wq) < 0)
2949 for_each_cwq_cpu(cpu, wq) {
2950 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2951 struct global_cwq *gcwq = get_gcwq(cpu);
2953 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
2956 cwq->flush_color = -1;
2957 cwq->max_active = max_active;
2958 INIT_LIST_HEAD(&cwq->delayed_works);
2961 if (flags & WQ_RESCUER) {
2962 struct worker *rescuer;
2964 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
2967 wq->rescuer = rescuer = alloc_worker();
2971 rescuer->task = kthread_create(rescuer_thread, wq, "%s", name);
2972 if (IS_ERR(rescuer->task))
2975 rescuer->task->flags |= PF_THREAD_BOUND;
2976 wake_up_process(rescuer->task);
2980 * workqueue_lock protects global freeze state and workqueues
2981 * list. Grab it, set max_active accordingly and add the new
2982 * workqueue to workqueues list.
2984 spin_lock(&workqueue_lock);
2986 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
2987 for_each_cwq_cpu(cpu, wq)
2988 get_cwq(cpu, wq)->max_active = 0;
2990 list_add(&wq->list, &workqueues);
2992 spin_unlock(&workqueue_lock);
2998 free_mayday_mask(wq->mayday_mask);
3004 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3007 * destroy_workqueue - safely terminate a workqueue
3008 * @wq: target workqueue
3010 * Safely destroy a workqueue. All work currently pending will be done first.
3012 void destroy_workqueue(struct workqueue_struct *wq)
3014 unsigned int flush_cnt = 0;
3018 * Mark @wq dying and drain all pending works. Once WQ_DYING is
3019 * set, only chain queueing is allowed. IOW, only currently
3020 * pending or running work items on @wq can queue further work
3021 * items on it. @wq is flushed repeatedly until it becomes empty.
3022 * The number of flushing is detemined by the depth of chaining and
3023 * should be relatively short. Whine if it takes too long.
3025 wq->flags |= WQ_DYING;
3027 flush_workqueue(wq);
3029 for_each_cwq_cpu(cpu, wq) {
3030 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3033 spin_lock_irq(&cwq->gcwq->lock);
3034 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
3035 spin_unlock_irq(&cwq->gcwq->lock);
3040 if (++flush_cnt == 10 ||
3041 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
3042 printk(KERN_WARNING "workqueue %s: flush on "
3043 "destruction isn't complete after %u tries\n",
3044 wq->name, flush_cnt);
3049 * wq list is used to freeze wq, remove from list after
3050 * flushing is complete in case freeze races us.
3052 spin_lock(&workqueue_lock);
3053 list_del(&wq->list);
3054 spin_unlock(&workqueue_lock);
3057 for_each_cwq_cpu(cpu, wq) {
3058 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3061 for (i = 0; i < WORK_NR_COLORS; i++)
3062 BUG_ON(cwq->nr_in_flight[i]);
3063 BUG_ON(cwq->nr_active);
3064 BUG_ON(!list_empty(&cwq->delayed_works));
3067 if (wq->flags & WQ_RESCUER) {
3068 kthread_stop(wq->rescuer->task);
3069 free_mayday_mask(wq->mayday_mask);
3076 EXPORT_SYMBOL_GPL(destroy_workqueue);
3079 * workqueue_set_max_active - adjust max_active of a workqueue
3080 * @wq: target workqueue
3081 * @max_active: new max_active value.
3083 * Set max_active of @wq to @max_active.
3086 * Don't call from IRQ context.
3088 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3092 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3094 spin_lock(&workqueue_lock);
3096 wq->saved_max_active = max_active;
3098 for_each_cwq_cpu(cpu, wq) {
3099 struct global_cwq *gcwq = get_gcwq(cpu);
3101 spin_lock_irq(&gcwq->lock);
3103 if (!(wq->flags & WQ_FREEZABLE) ||
3104 !(gcwq->flags & GCWQ_FREEZING))
3105 get_cwq(gcwq->cpu, wq)->max_active = max_active;
3107 spin_unlock_irq(&gcwq->lock);
3110 spin_unlock(&workqueue_lock);
3112 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3115 * workqueue_congested - test whether a workqueue is congested
3116 * @cpu: CPU in question
3117 * @wq: target workqueue
3119 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3120 * no synchronization around this function and the test result is
3121 * unreliable and only useful as advisory hints or for debugging.
3124 * %true if congested, %false otherwise.
3126 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3128 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3130 return !list_empty(&cwq->delayed_works);
3132 EXPORT_SYMBOL_GPL(workqueue_congested);
3135 * work_cpu - return the last known associated cpu for @work
3136 * @work: the work of interest
3139 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3141 unsigned int work_cpu(struct work_struct *work)
3143 struct global_cwq *gcwq = get_work_gcwq(work);
3145 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3147 EXPORT_SYMBOL_GPL(work_cpu);
3150 * work_busy - test whether a work is currently pending or running
3151 * @work: the work to be tested
3153 * Test whether @work is currently pending or running. There is no
3154 * synchronization around this function and the test result is
3155 * unreliable and only useful as advisory hints or for debugging.
3156 * Especially for reentrant wqs, the pending state might hide the
3160 * OR'd bitmask of WORK_BUSY_* bits.
3162 unsigned int work_busy(struct work_struct *work)
3164 struct global_cwq *gcwq = get_work_gcwq(work);
3165 unsigned long flags;
3166 unsigned int ret = 0;
3171 spin_lock_irqsave(&gcwq->lock, flags);
3173 if (work_pending(work))
3174 ret |= WORK_BUSY_PENDING;
3175 if (find_worker_executing_work(gcwq, work))
3176 ret |= WORK_BUSY_RUNNING;
3178 spin_unlock_irqrestore(&gcwq->lock, flags);
3182 EXPORT_SYMBOL_GPL(work_busy);
3187 * There are two challenges in supporting CPU hotplug. Firstly, there
3188 * are a lot of assumptions on strong associations among work, cwq and
3189 * gcwq which make migrating pending and scheduled works very
3190 * difficult to implement without impacting hot paths. Secondly,
3191 * gcwqs serve mix of short, long and very long running works making
3192 * blocked draining impractical.
3194 * This is solved by allowing a gcwq to be detached from CPU, running
3195 * it with unbound (rogue) workers and allowing it to be reattached
3196 * later if the cpu comes back online. A separate thread is created
3197 * to govern a gcwq in such state and is called the trustee of the
3200 * Trustee states and their descriptions.
3202 * START Command state used on startup. On CPU_DOWN_PREPARE, a
3203 * new trustee is started with this state.
3205 * IN_CHARGE Once started, trustee will enter this state after
3206 * assuming the manager role and making all existing
3207 * workers rogue. DOWN_PREPARE waits for trustee to
3208 * enter this state. After reaching IN_CHARGE, trustee
3209 * tries to execute the pending worklist until it's empty
3210 * and the state is set to BUTCHER, or the state is set
3213 * BUTCHER Command state which is set by the cpu callback after
3214 * the cpu has went down. Once this state is set trustee
3215 * knows that there will be no new works on the worklist
3216 * and once the worklist is empty it can proceed to
3217 * killing idle workers.
3219 * RELEASE Command state which is set by the cpu callback if the
3220 * cpu down has been canceled or it has come online
3221 * again. After recognizing this state, trustee stops
3222 * trying to drain or butcher and clears ROGUE, rebinds
3223 * all remaining workers back to the cpu and releases
3226 * DONE Trustee will enter this state after BUTCHER or RELEASE
3229 * trustee CPU draining
3230 * took over down complete
3231 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3233 * | CPU is back online v return workers |
3234 * ----------------> RELEASE --------------
3238 * trustee_wait_event_timeout - timed event wait for trustee
3239 * @cond: condition to wait for
3240 * @timeout: timeout in jiffies
3242 * wait_event_timeout() for trustee to use. Handles locking and
3243 * checks for RELEASE request.
3246 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3247 * multiple times. To be used by trustee.
3250 * Positive indicating left time if @cond is satisfied, 0 if timed
3251 * out, -1 if canceled.
3253 #define trustee_wait_event_timeout(cond, timeout) ({ \
3254 long __ret = (timeout); \
3255 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3257 spin_unlock_irq(&gcwq->lock); \
3258 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3259 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3261 spin_lock_irq(&gcwq->lock); \
3263 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3267 * trustee_wait_event - event wait for trustee
3268 * @cond: condition to wait for
3270 * wait_event() for trustee to use. Automatically handles locking and
3271 * checks for CANCEL request.
3274 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3275 * multiple times. To be used by trustee.
3278 * 0 if @cond is satisfied, -1 if canceled.
3280 #define trustee_wait_event(cond) ({ \
3282 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3283 __ret1 < 0 ? -1 : 0; \
3286 static int __cpuinit trustee_thread(void *__gcwq)
3288 struct global_cwq *gcwq = __gcwq;
3289 struct worker *worker;
3290 struct work_struct *work;
3291 struct hlist_node *pos;
3295 BUG_ON(gcwq->cpu != smp_processor_id());
3297 spin_lock_irq(&gcwq->lock);
3299 * Claim the manager position and make all workers rogue.
3300 * Trustee must be bound to the target cpu and can't be
3303 BUG_ON(gcwq->cpu != smp_processor_id());
3304 rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3307 gcwq->flags |= GCWQ_MANAGING_WORKERS;
3309 list_for_each_entry(worker, &gcwq->idle_list, entry)
3310 worker->flags |= WORKER_ROGUE;
3312 for_each_busy_worker(worker, i, pos, gcwq)
3313 worker->flags |= WORKER_ROGUE;
3316 * Call schedule() so that we cross rq->lock and thus can
3317 * guarantee sched callbacks see the rogue flag. This is
3318 * necessary as scheduler callbacks may be invoked from other
3321 spin_unlock_irq(&gcwq->lock);
3323 spin_lock_irq(&gcwq->lock);
3326 * Sched callbacks are disabled now. Zap nr_running. After
3327 * this, nr_running stays zero and need_more_worker() and
3328 * keep_working() are always true as long as the worklist is
3331 atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3333 spin_unlock_irq(&gcwq->lock);
3334 del_timer_sync(&gcwq->idle_timer);
3335 spin_lock_irq(&gcwq->lock);
3338 * We're now in charge. Notify and proceed to drain. We need
3339 * to keep the gcwq running during the whole CPU down
3340 * procedure as other cpu hotunplug callbacks may need to
3341 * flush currently running tasks.
3343 gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3344 wake_up_all(&gcwq->trustee_wait);
3347 * The original cpu is in the process of dying and may go away
3348 * anytime now. When that happens, we and all workers would
3349 * be migrated to other cpus. Try draining any left work. We
3350 * want to get it over with ASAP - spam rescuers, wake up as
3351 * many idlers as necessary and create new ones till the
3352 * worklist is empty. Note that if the gcwq is frozen, there
3353 * may be frozen works in freezable cwqs. Don't declare
3354 * completion while frozen.
3356 while (gcwq->nr_workers != gcwq->nr_idle ||
3357 gcwq->flags & GCWQ_FREEZING ||
3358 gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3361 list_for_each_entry(work, &gcwq->worklist, entry) {
3366 list_for_each_entry(worker, &gcwq->idle_list, entry) {
3369 wake_up_process(worker->task);
3372 if (need_to_create_worker(gcwq)) {
3373 spin_unlock_irq(&gcwq->lock);
3374 worker = create_worker(gcwq, false);
3375 spin_lock_irq(&gcwq->lock);
3377 worker->flags |= WORKER_ROGUE;
3378 start_worker(worker);
3382 /* give a breather */
3383 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3388 * Either all works have been scheduled and cpu is down, or
3389 * cpu down has already been canceled. Wait for and butcher
3390 * all workers till we're canceled.
3393 rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3394 while (!list_empty(&gcwq->idle_list))
3395 destroy_worker(list_first_entry(&gcwq->idle_list,
3396 struct worker, entry));
3397 } while (gcwq->nr_workers && rc >= 0);
3400 * At this point, either draining has completed and no worker
3401 * is left, or cpu down has been canceled or the cpu is being
3402 * brought back up. There shouldn't be any idle one left.
3403 * Tell the remaining busy ones to rebind once it finishes the
3404 * currently scheduled works by scheduling the rebind_work.
3406 WARN_ON(!list_empty(&gcwq->idle_list));
3408 for_each_busy_worker(worker, i, pos, gcwq) {
3409 struct work_struct *rebind_work = &worker->rebind_work;
3412 * Rebind_work may race with future cpu hotplug
3413 * operations. Use a separate flag to mark that
3414 * rebinding is scheduled.
3416 worker->flags |= WORKER_REBIND;
3417 worker->flags &= ~WORKER_ROGUE;
3419 /* queue rebind_work, wq doesn't matter, use the default one */
3420 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3421 work_data_bits(rebind_work)))
3424 debug_work_activate(rebind_work);
3425 insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3426 worker->scheduled.next,
3427 work_color_to_flags(WORK_NO_COLOR));
3430 /* relinquish manager role */
3431 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3433 /* notify completion */
3434 gcwq->trustee = NULL;
3435 gcwq->trustee_state = TRUSTEE_DONE;
3436 wake_up_all(&gcwq->trustee_wait);
3437 spin_unlock_irq(&gcwq->lock);
3442 * wait_trustee_state - wait for trustee to enter the specified state
3443 * @gcwq: gcwq the trustee of interest belongs to
3444 * @state: target state to wait for
3446 * Wait for the trustee to reach @state. DONE is already matched.
3449 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3450 * multiple times. To be used by cpu_callback.
3452 static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3453 __releases(&gcwq->lock)
3454 __acquires(&gcwq->lock)
3456 if (!(gcwq->trustee_state == state ||
3457 gcwq->trustee_state == TRUSTEE_DONE)) {
3458 spin_unlock_irq(&gcwq->lock);
3459 __wait_event(gcwq->trustee_wait,
3460 gcwq->trustee_state == state ||
3461 gcwq->trustee_state == TRUSTEE_DONE);
3462 spin_lock_irq(&gcwq->lock);
3466 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3467 unsigned long action,
3470 unsigned int cpu = (unsigned long)hcpu;
3471 struct global_cwq *gcwq = get_gcwq(cpu);
3472 struct task_struct *new_trustee = NULL;
3473 struct worker *uninitialized_var(new_worker);
3474 unsigned long flags;
3476 action &= ~CPU_TASKS_FROZEN;
3479 case CPU_DOWN_PREPARE:
3480 new_trustee = kthread_create(trustee_thread, gcwq,
3481 "workqueue_trustee/%d\n", cpu);
3482 if (IS_ERR(new_trustee))
3483 return notifier_from_errno(PTR_ERR(new_trustee));
3484 kthread_bind(new_trustee, cpu);
3486 case CPU_UP_PREPARE:
3487 BUG_ON(gcwq->first_idle);
3488 new_worker = create_worker(gcwq, false);
3491 kthread_stop(new_trustee);
3496 /* some are called w/ irq disabled, don't disturb irq status */
3497 spin_lock_irqsave(&gcwq->lock, flags);
3500 case CPU_DOWN_PREPARE:
3501 /* initialize trustee and tell it to acquire the gcwq */
3502 BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
3503 gcwq->trustee = new_trustee;
3504 gcwq->trustee_state = TRUSTEE_START;
3505 wake_up_process(gcwq->trustee);
3506 wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3508 case CPU_UP_PREPARE:
3509 BUG_ON(gcwq->first_idle);
3510 gcwq->first_idle = new_worker;
3515 * Before this, the trustee and all workers except for
3516 * the ones which are still executing works from
3517 * before the last CPU down must be on the cpu. After
3518 * this, they'll all be diasporas.
3520 gcwq->flags |= GCWQ_DISASSOCIATED;
3524 gcwq->trustee_state = TRUSTEE_BUTCHER;
3526 case CPU_UP_CANCELED:
3527 destroy_worker(gcwq->first_idle);
3528 gcwq->first_idle = NULL;
3531 case CPU_DOWN_FAILED:
3533 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3534 if (gcwq->trustee_state != TRUSTEE_DONE) {
3535 gcwq->trustee_state = TRUSTEE_RELEASE;
3536 wake_up_process(gcwq->trustee);
3537 wait_trustee_state(gcwq, TRUSTEE_DONE);
3541 * Trustee is done and there might be no worker left.
3542 * Put the first_idle in and request a real manager to
3545 spin_unlock_irq(&gcwq->lock);
3546 kthread_bind(gcwq->first_idle->task, cpu);
3547 spin_lock_irq(&gcwq->lock);
3548 gcwq->flags |= GCWQ_MANAGE_WORKERS;
3549 start_worker(gcwq->first_idle);
3550 gcwq->first_idle = NULL;
3554 spin_unlock_irqrestore(&gcwq->lock, flags);
3556 return notifier_from_errno(0);
3561 struct work_for_cpu {
3562 struct completion completion;
3568 static int do_work_for_cpu(void *_wfc)
3570 struct work_for_cpu *wfc = _wfc;
3571 wfc->ret = wfc->fn(wfc->arg);
3572 complete(&wfc->completion);
3577 * work_on_cpu - run a function in user context on a particular cpu
3578 * @cpu: the cpu to run on
3579 * @fn: the function to run
3580 * @arg: the function arg
3582 * This will return the value @fn returns.
3583 * It is up to the caller to ensure that the cpu doesn't go offline.
3584 * The caller must not hold any locks which would prevent @fn from completing.
3586 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3588 struct task_struct *sub_thread;
3589 struct work_for_cpu wfc = {
3590 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3595 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3596 if (IS_ERR(sub_thread))
3597 return PTR_ERR(sub_thread);
3598 kthread_bind(sub_thread, cpu);
3599 wake_up_process(sub_thread);
3600 wait_for_completion(&wfc.completion);
3603 EXPORT_SYMBOL_GPL(work_on_cpu);
3604 #endif /* CONFIG_SMP */
3606 #ifdef CONFIG_FREEZER
3609 * freeze_workqueues_begin - begin freezing workqueues
3611 * Start freezing workqueues. After this function returns, all freezable
3612 * workqueues will queue new works to their frozen_works list instead of
3616 * Grabs and releases workqueue_lock and gcwq->lock's.
3618 void freeze_workqueues_begin(void)
3622 spin_lock(&workqueue_lock);
3624 BUG_ON(workqueue_freezing);
3625 workqueue_freezing = true;
3627 for_each_gcwq_cpu(cpu) {
3628 struct global_cwq *gcwq = get_gcwq(cpu);
3629 struct workqueue_struct *wq;
3631 spin_lock_irq(&gcwq->lock);
3633 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3634 gcwq->flags |= GCWQ_FREEZING;
3636 list_for_each_entry(wq, &workqueues, list) {
3637 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3639 if (cwq && wq->flags & WQ_FREEZABLE)
3640 cwq->max_active = 0;
3643 spin_unlock_irq(&gcwq->lock);
3646 spin_unlock(&workqueue_lock);
3650 * freeze_workqueues_busy - are freezable workqueues still busy?
3652 * Check whether freezing is complete. This function must be called
3653 * between freeze_workqueues_begin() and thaw_workqueues().
3656 * Grabs and releases workqueue_lock.
3659 * %true if some freezable workqueues are still busy. %false if freezing
3662 bool freeze_workqueues_busy(void)
3667 spin_lock(&workqueue_lock);
3669 BUG_ON(!workqueue_freezing);
3671 for_each_gcwq_cpu(cpu) {
3672 struct workqueue_struct *wq;
3674 * nr_active is monotonically decreasing. It's safe
3675 * to peek without lock.
3677 list_for_each_entry(wq, &workqueues, list) {
3678 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3680 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3683 BUG_ON(cwq->nr_active < 0);
3684 if (cwq->nr_active) {
3691 spin_unlock(&workqueue_lock);
3696 * thaw_workqueues - thaw workqueues
3698 * Thaw workqueues. Normal queueing is restored and all collected
3699 * frozen works are transferred to their respective gcwq worklists.
3702 * Grabs and releases workqueue_lock and gcwq->lock's.
3704 void thaw_workqueues(void)
3708 spin_lock(&workqueue_lock);
3710 if (!workqueue_freezing)
3713 for_each_gcwq_cpu(cpu) {
3714 struct global_cwq *gcwq = get_gcwq(cpu);
3715 struct workqueue_struct *wq;
3717 spin_lock_irq(&gcwq->lock);
3719 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3720 gcwq->flags &= ~GCWQ_FREEZING;
3722 list_for_each_entry(wq, &workqueues, list) {
3723 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3725 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3728 /* restore max_active and repopulate worklist */
3729 cwq->max_active = wq->saved_max_active;
3731 while (!list_empty(&cwq->delayed_works) &&
3732 cwq->nr_active < cwq->max_active)
3733 cwq_activate_first_delayed(cwq);
3736 wake_up_worker(gcwq);
3738 spin_unlock_irq(&gcwq->lock);
3741 workqueue_freezing = false;
3743 spin_unlock(&workqueue_lock);
3745 #endif /* CONFIG_FREEZER */
3747 static int __init init_workqueues(void)
3752 cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
3754 /* initialize gcwqs */
3755 for_each_gcwq_cpu(cpu) {
3756 struct global_cwq *gcwq = get_gcwq(cpu);
3758 spin_lock_init(&gcwq->lock);
3759 INIT_LIST_HEAD(&gcwq->worklist);
3761 gcwq->flags |= GCWQ_DISASSOCIATED;
3763 INIT_LIST_HEAD(&gcwq->idle_list);
3764 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3765 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3767 init_timer_deferrable(&gcwq->idle_timer);
3768 gcwq->idle_timer.function = idle_worker_timeout;
3769 gcwq->idle_timer.data = (unsigned long)gcwq;
3771 setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3772 (unsigned long)gcwq);
3774 ida_init(&gcwq->worker_ida);
3776 gcwq->trustee_state = TRUSTEE_DONE;
3777 init_waitqueue_head(&gcwq->trustee_wait);
3780 /* create the initial worker */
3781 for_each_online_gcwq_cpu(cpu) {
3782 struct global_cwq *gcwq = get_gcwq(cpu);
3783 struct worker *worker;
3785 if (cpu != WORK_CPU_UNBOUND)
3786 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3787 worker = create_worker(gcwq, true);
3789 spin_lock_irq(&gcwq->lock);
3790 start_worker(worker);
3791 spin_unlock_irq(&gcwq->lock);
3794 system_wq = alloc_workqueue("events", 0, 0);
3795 system_long_wq = alloc_workqueue("events_long", 0, 0);
3796 system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3797 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3798 WQ_UNBOUND_MAX_ACTIVE);
3799 system_freezable_wq = alloc_workqueue("events_freezable",
3801 system_nrt_freezable_wq = alloc_workqueue("events_nrt_freezable",
3802 WQ_NON_REENTRANT | WQ_FREEZABLE, 0);
3803 BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq ||
3804 !system_unbound_wq || !system_freezable_wq ||
3805 !system_nrt_freezable_wq);
3808 early_initcall(init_workqueues);