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 #define CREATE_TRACE_POINTS
46 #include <trace/events/workqueue.h>
48 #include "workqueue_sched.h"
51 /* global_cwq flags */
52 GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
53 GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */
54 GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
55 GCWQ_FREEZING = 1 << 3, /* freeze in progress */
56 GCWQ_HIGHPRI_PENDING = 1 << 4, /* highpri works on queue */
59 WORKER_STARTED = 1 << 0, /* started */
60 WORKER_DIE = 1 << 1, /* die die die */
61 WORKER_IDLE = 1 << 2, /* is idle */
62 WORKER_PREP = 1 << 3, /* preparing to run works */
63 WORKER_ROGUE = 1 << 4, /* not bound to any cpu */
64 WORKER_REBIND = 1 << 5, /* mom is home, come back */
65 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
66 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
68 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND |
69 WORKER_CPU_INTENSIVE | WORKER_UNBOUND,
71 /* gcwq->trustee_state */
72 TRUSTEE_START = 0, /* start */
73 TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */
74 TRUSTEE_BUTCHER = 2, /* butcher workers */
75 TRUSTEE_RELEASE = 3, /* release workers */
76 TRUSTEE_DONE = 4, /* trustee is done */
78 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
79 BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
80 BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
82 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
83 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
85 MAYDAY_INITIAL_TIMEOUT = HZ / 100, /* call for help after 10ms */
86 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
87 CREATE_COOLDOWN = HZ, /* time to breath after fail */
88 TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */
91 * Rescue workers are used only on emergencies and shared by
94 RESCUER_NICE_LEVEL = -20,
98 * Structure fields follow one of the following exclusion rules.
100 * I: Modifiable by initialization/destruction paths and read-only for
103 * P: Preemption protected. Disabling preemption is enough and should
104 * only be modified and accessed from the local cpu.
106 * L: gcwq->lock protected. Access with gcwq->lock held.
108 * X: During normal operation, modification requires gcwq->lock and
109 * should be done only from local cpu. Either disabling preemption
110 * on local cpu or grabbing gcwq->lock is enough for read access.
111 * If GCWQ_DISASSOCIATED is set, it's identical to L.
113 * F: wq->flush_mutex protected.
115 * W: workqueue_lock protected.
121 * The poor guys doing the actual heavy lifting. All on-duty workers
122 * are either serving the manager role, on idle list or on busy hash.
125 /* on idle list while idle, on busy hash table while busy */
127 struct list_head entry; /* L: while idle */
128 struct hlist_node hentry; /* L: while busy */
131 struct work_struct *current_work; /* L: work being processed */
132 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
133 struct list_head scheduled; /* L: scheduled works */
134 struct task_struct *task; /* I: worker task */
135 struct global_cwq *gcwq; /* I: the associated gcwq */
136 /* 64 bytes boundary on 64bit, 32 on 32bit */
137 unsigned long last_active; /* L: last active timestamp */
138 unsigned int flags; /* X: flags */
139 int id; /* I: worker id */
140 struct work_struct rebind_work; /* L: rebind worker to cpu */
144 * Global per-cpu workqueue. There's one and only one for each cpu
145 * and all works are queued and processed here regardless of their
149 spinlock_t lock; /* the gcwq lock */
150 struct list_head worklist; /* L: list of pending works */
151 unsigned int cpu; /* I: the associated cpu */
152 unsigned int flags; /* L: GCWQ_* flags */
154 int nr_workers; /* L: total number of workers */
155 int nr_idle; /* L: currently idle ones */
157 /* workers are chained either in the idle_list or busy_hash */
158 struct list_head idle_list; /* X: list of idle workers */
159 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
160 /* L: hash of busy workers */
162 struct timer_list idle_timer; /* L: worker idle timeout */
163 struct timer_list mayday_timer; /* L: SOS timer for dworkers */
165 struct ida worker_ida; /* L: for worker IDs */
167 struct task_struct *trustee; /* L: for gcwq shutdown */
168 unsigned int trustee_state; /* L: trustee state */
169 wait_queue_head_t trustee_wait; /* trustee wait */
170 struct worker *first_idle; /* L: first idle worker */
171 } ____cacheline_aligned_in_smp;
174 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
175 * work_struct->data are used for flags and thus cwqs need to be
176 * aligned at two's power of the number of flag bits.
178 struct cpu_workqueue_struct {
179 struct global_cwq *gcwq; /* I: the associated gcwq */
180 struct workqueue_struct *wq; /* I: the owning workqueue */
181 int work_color; /* L: current color */
182 int flush_color; /* L: flushing color */
183 int nr_in_flight[WORK_NR_COLORS];
184 /* L: nr of in_flight works */
185 int nr_active; /* L: nr of active works */
186 int max_active; /* L: max active works */
187 struct list_head delayed_works; /* L: delayed works */
191 * Structure used to wait for workqueue flush.
194 struct list_head list; /* F: list of flushers */
195 int flush_color; /* F: flush color waiting for */
196 struct completion done; /* flush completion */
200 * All cpumasks are assumed to be always set on UP and thus can't be
201 * used to determine whether there's something to be done.
204 typedef cpumask_var_t mayday_mask_t;
205 #define mayday_test_and_set_cpu(cpu, mask) \
206 cpumask_test_and_set_cpu((cpu), (mask))
207 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
208 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
209 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
210 #define free_mayday_mask(mask) free_cpumask_var((mask))
212 typedef unsigned long mayday_mask_t;
213 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
214 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
215 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
216 #define alloc_mayday_mask(maskp, gfp) true
217 #define free_mayday_mask(mask) do { } while (0)
221 * The externally visible workqueue abstraction is an array of
222 * per-CPU workqueues:
224 struct workqueue_struct {
225 unsigned int flags; /* I: WQ_* flags */
227 struct cpu_workqueue_struct __percpu *pcpu;
228 struct cpu_workqueue_struct *single;
230 } cpu_wq; /* I: cwq's */
231 struct list_head list; /* W: list of all workqueues */
233 struct mutex flush_mutex; /* protects wq flushing */
234 int work_color; /* F: current work color */
235 int flush_color; /* F: current flush color */
236 atomic_t nr_cwqs_to_flush; /* flush in progress */
237 struct wq_flusher *first_flusher; /* F: first flusher */
238 struct list_head flusher_queue; /* F: flush waiters */
239 struct list_head flusher_overflow; /* F: flush overflow list */
241 mayday_mask_t mayday_mask; /* cpus requesting rescue */
242 struct worker *rescuer; /* I: rescue worker */
244 int saved_max_active; /* W: saved cwq max_active */
245 const char *name; /* I: workqueue name */
246 #ifdef CONFIG_LOCKDEP
247 struct lockdep_map lockdep_map;
251 struct workqueue_struct *system_wq __read_mostly;
252 struct workqueue_struct *system_long_wq __read_mostly;
253 struct workqueue_struct *system_nrt_wq __read_mostly;
254 struct workqueue_struct *system_unbound_wq __read_mostly;
255 EXPORT_SYMBOL_GPL(system_wq);
256 EXPORT_SYMBOL_GPL(system_long_wq);
257 EXPORT_SYMBOL_GPL(system_nrt_wq);
258 EXPORT_SYMBOL_GPL(system_unbound_wq);
260 #define for_each_busy_worker(worker, i, pos, gcwq) \
261 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
262 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
264 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
267 if (cpu < nr_cpu_ids) {
269 cpu = cpumask_next(cpu, mask);
270 if (cpu < nr_cpu_ids)
274 return WORK_CPU_UNBOUND;
276 return WORK_CPU_NONE;
279 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
280 struct workqueue_struct *wq)
282 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
288 * An extra gcwq is defined for an invalid cpu number
289 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
290 * specific CPU. The following iterators are similar to
291 * for_each_*_cpu() iterators but also considers the unbound gcwq.
293 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
294 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
295 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
296 * WORK_CPU_UNBOUND for unbound workqueues
298 #define for_each_gcwq_cpu(cpu) \
299 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
300 (cpu) < WORK_CPU_NONE; \
301 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
303 #define for_each_online_gcwq_cpu(cpu) \
304 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
305 (cpu) < WORK_CPU_NONE; \
306 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
308 #define for_each_cwq_cpu(cpu, wq) \
309 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
310 (cpu) < WORK_CPU_NONE; \
311 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
313 #ifdef CONFIG_LOCKDEP
315 * in_workqueue_context() - in context of specified workqueue?
316 * @wq: the workqueue of interest
318 * Checks lockdep state to see if the current task is executing from
319 * within a workqueue item. This function exists only if lockdep is
322 int in_workqueue_context(struct workqueue_struct *wq)
324 return lock_is_held(&wq->lockdep_map);
328 #ifdef CONFIG_DEBUG_OBJECTS_WORK
330 static struct debug_obj_descr work_debug_descr;
333 * fixup_init is called when:
334 * - an active object is initialized
336 static int work_fixup_init(void *addr, enum debug_obj_state state)
338 struct work_struct *work = addr;
341 case ODEBUG_STATE_ACTIVE:
342 cancel_work_sync(work);
343 debug_object_init(work, &work_debug_descr);
351 * fixup_activate is called when:
352 * - an active object is activated
353 * - an unknown object is activated (might be a statically initialized object)
355 static int work_fixup_activate(void *addr, enum debug_obj_state state)
357 struct work_struct *work = addr;
361 case ODEBUG_STATE_NOTAVAILABLE:
363 * This is not really a fixup. The work struct was
364 * statically initialized. We just make sure that it
365 * is tracked in the object tracker.
367 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
368 debug_object_init(work, &work_debug_descr);
369 debug_object_activate(work, &work_debug_descr);
375 case ODEBUG_STATE_ACTIVE:
384 * fixup_free is called when:
385 * - an active object is freed
387 static int work_fixup_free(void *addr, enum debug_obj_state state)
389 struct work_struct *work = addr;
392 case ODEBUG_STATE_ACTIVE:
393 cancel_work_sync(work);
394 debug_object_free(work, &work_debug_descr);
401 static struct debug_obj_descr work_debug_descr = {
402 .name = "work_struct",
403 .fixup_init = work_fixup_init,
404 .fixup_activate = work_fixup_activate,
405 .fixup_free = work_fixup_free,
408 static inline void debug_work_activate(struct work_struct *work)
410 debug_object_activate(work, &work_debug_descr);
413 static inline void debug_work_deactivate(struct work_struct *work)
415 debug_object_deactivate(work, &work_debug_descr);
418 void __init_work(struct work_struct *work, int onstack)
421 debug_object_init_on_stack(work, &work_debug_descr);
423 debug_object_init(work, &work_debug_descr);
425 EXPORT_SYMBOL_GPL(__init_work);
427 void destroy_work_on_stack(struct work_struct *work)
429 debug_object_free(work, &work_debug_descr);
431 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
434 static inline void debug_work_activate(struct work_struct *work) { }
435 static inline void debug_work_deactivate(struct work_struct *work) { }
438 /* Serializes the accesses to the list of workqueues. */
439 static DEFINE_SPINLOCK(workqueue_lock);
440 static LIST_HEAD(workqueues);
441 static bool workqueue_freezing; /* W: have wqs started freezing? */
444 * The almighty global cpu workqueues. nr_running is the only field
445 * which is expected to be used frequently by other cpus via
446 * try_to_wake_up(). Put it in a separate cacheline.
448 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
449 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
452 * Global cpu workqueue and nr_running counter for unbound gcwq. The
453 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
454 * workers have WORKER_UNBOUND set.
456 static struct global_cwq unbound_global_cwq;
457 static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */
459 static int worker_thread(void *__worker);
461 static struct global_cwq *get_gcwq(unsigned int cpu)
463 if (cpu != WORK_CPU_UNBOUND)
464 return &per_cpu(global_cwq, cpu);
466 return &unbound_global_cwq;
469 static atomic_t *get_gcwq_nr_running(unsigned int cpu)
471 if (cpu != WORK_CPU_UNBOUND)
472 return &per_cpu(gcwq_nr_running, cpu);
474 return &unbound_gcwq_nr_running;
477 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
478 struct workqueue_struct *wq)
480 if (!(wq->flags & WQ_UNBOUND)) {
481 if (likely(cpu < nr_cpu_ids)) {
483 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
485 return wq->cpu_wq.single;
488 } else if (likely(cpu == WORK_CPU_UNBOUND))
489 return wq->cpu_wq.single;
493 static unsigned int work_color_to_flags(int color)
495 return color << WORK_STRUCT_COLOR_SHIFT;
498 static int get_work_color(struct work_struct *work)
500 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
501 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
504 static int work_next_color(int color)
506 return (color + 1) % WORK_NR_COLORS;
510 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
511 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
512 * cleared and the work data contains the cpu number it was last on.
514 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
515 * cwq, cpu or clear work->data. These functions should only be
516 * called while the work is owned - ie. while the PENDING bit is set.
518 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
519 * corresponding to a work. gcwq is available once the work has been
520 * queued anywhere after initialization. cwq is available only from
521 * queueing until execution starts.
523 static inline void set_work_data(struct work_struct *work, unsigned long data,
526 BUG_ON(!work_pending(work));
527 atomic_long_set(&work->data, data | flags | work_static(work));
530 static void set_work_cwq(struct work_struct *work,
531 struct cpu_workqueue_struct *cwq,
532 unsigned long extra_flags)
534 set_work_data(work, (unsigned long)cwq,
535 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
538 static void set_work_cpu(struct work_struct *work, unsigned int cpu)
540 set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
543 static void clear_work_data(struct work_struct *work)
545 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
548 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
550 unsigned long data = atomic_long_read(&work->data);
552 if (data & WORK_STRUCT_CWQ)
553 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
558 static struct global_cwq *get_work_gcwq(struct work_struct *work)
560 unsigned long data = atomic_long_read(&work->data);
563 if (data & WORK_STRUCT_CWQ)
564 return ((struct cpu_workqueue_struct *)
565 (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
567 cpu = data >> WORK_STRUCT_FLAG_BITS;
568 if (cpu == WORK_CPU_NONE)
571 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
572 return get_gcwq(cpu);
576 * Policy functions. These define the policies on how the global
577 * worker pool is managed. Unless noted otherwise, these functions
578 * assume that they're being called with gcwq->lock held.
581 static bool __need_more_worker(struct global_cwq *gcwq)
583 return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) ||
584 gcwq->flags & GCWQ_HIGHPRI_PENDING;
588 * Need to wake up a worker? Called from anything but currently
591 static bool need_more_worker(struct global_cwq *gcwq)
593 return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
596 /* Can I start working? Called from busy but !running workers. */
597 static bool may_start_working(struct global_cwq *gcwq)
599 return gcwq->nr_idle;
602 /* Do I need to keep working? Called from currently running workers. */
603 static bool keep_working(struct global_cwq *gcwq)
605 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
607 return !list_empty(&gcwq->worklist) && atomic_read(nr_running) <= 1;
610 /* Do we need a new worker? Called from manager. */
611 static bool need_to_create_worker(struct global_cwq *gcwq)
613 return need_more_worker(gcwq) && !may_start_working(gcwq);
616 /* Do I need to be the manager? */
617 static bool need_to_manage_workers(struct global_cwq *gcwq)
619 return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS;
622 /* Do we have too many workers and should some go away? */
623 static bool too_many_workers(struct global_cwq *gcwq)
625 bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
626 int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
627 int nr_busy = gcwq->nr_workers - nr_idle;
629 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
636 /* Return the first worker. Safe with preemption disabled */
637 static struct worker *first_worker(struct global_cwq *gcwq)
639 if (unlikely(list_empty(&gcwq->idle_list)))
642 return list_first_entry(&gcwq->idle_list, struct worker, entry);
646 * wake_up_worker - wake up an idle worker
647 * @gcwq: gcwq to wake worker for
649 * Wake up the first idle worker of @gcwq.
652 * spin_lock_irq(gcwq->lock).
654 static void wake_up_worker(struct global_cwq *gcwq)
656 struct worker *worker = first_worker(gcwq);
659 wake_up_process(worker->task);
663 * wq_worker_waking_up - a worker is waking up
664 * @task: task waking up
665 * @cpu: CPU @task is waking up to
667 * This function is called during try_to_wake_up() when a worker is
671 * spin_lock_irq(rq->lock)
673 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
675 struct worker *worker = kthread_data(task);
677 if (likely(!(worker->flags & WORKER_NOT_RUNNING)))
678 atomic_inc(get_gcwq_nr_running(cpu));
682 * wq_worker_sleeping - a worker is going to sleep
683 * @task: task going to sleep
684 * @cpu: CPU in question, must be the current CPU number
686 * This function is called during schedule() when a busy worker is
687 * going to sleep. Worker on the same cpu can be woken up by
688 * returning pointer to its task.
691 * spin_lock_irq(rq->lock)
694 * Worker task on @cpu to wake up, %NULL if none.
696 struct task_struct *wq_worker_sleeping(struct task_struct *task,
699 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
700 struct global_cwq *gcwq = get_gcwq(cpu);
701 atomic_t *nr_running = get_gcwq_nr_running(cpu);
703 if (unlikely(worker->flags & WORKER_NOT_RUNNING))
706 /* this can only happen on the local cpu */
707 BUG_ON(cpu != raw_smp_processor_id());
710 * The counterpart of the following dec_and_test, implied mb,
711 * worklist not empty test sequence is in insert_work().
712 * Please read comment there.
714 * NOT_RUNNING is clear. This means that trustee is not in
715 * charge and we're running on the local cpu w/ rq lock held
716 * and preemption disabled, which in turn means that none else
717 * could be manipulating idle_list, so dereferencing idle_list
718 * without gcwq lock is safe.
720 if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
721 to_wakeup = first_worker(gcwq);
722 return to_wakeup ? to_wakeup->task : NULL;
726 * worker_set_flags - set worker flags and adjust nr_running accordingly
728 * @flags: flags to set
729 * @wakeup: wakeup an idle worker if necessary
731 * Set @flags in @worker->flags and adjust nr_running accordingly. If
732 * nr_running becomes zero and @wakeup is %true, an idle worker is
736 * spin_lock_irq(gcwq->lock)
738 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
741 struct global_cwq *gcwq = worker->gcwq;
743 WARN_ON_ONCE(worker->task != current);
746 * If transitioning into NOT_RUNNING, adjust nr_running and
747 * wake up an idle worker as necessary if requested by
750 if ((flags & WORKER_NOT_RUNNING) &&
751 !(worker->flags & WORKER_NOT_RUNNING)) {
752 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
755 if (atomic_dec_and_test(nr_running) &&
756 !list_empty(&gcwq->worklist))
757 wake_up_worker(gcwq);
759 atomic_dec(nr_running);
762 worker->flags |= flags;
766 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
768 * @flags: flags to clear
770 * Clear @flags in @worker->flags and adjust nr_running accordingly.
773 * spin_lock_irq(gcwq->lock)
775 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
777 struct global_cwq *gcwq = worker->gcwq;
778 unsigned int oflags = worker->flags;
780 WARN_ON_ONCE(worker->task != current);
782 worker->flags &= ~flags;
784 /* if transitioning out of NOT_RUNNING, increment nr_running */
785 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
786 if (!(worker->flags & WORKER_NOT_RUNNING))
787 atomic_inc(get_gcwq_nr_running(gcwq->cpu));
791 * busy_worker_head - return the busy hash head for a work
792 * @gcwq: gcwq of interest
793 * @work: work to be hashed
795 * Return hash head of @gcwq for @work.
798 * spin_lock_irq(gcwq->lock).
801 * Pointer to the hash head.
803 static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
804 struct work_struct *work)
806 const int base_shift = ilog2(sizeof(struct work_struct));
807 unsigned long v = (unsigned long)work;
809 /* simple shift and fold hash, do we need something better? */
811 v += v >> BUSY_WORKER_HASH_ORDER;
812 v &= BUSY_WORKER_HASH_MASK;
814 return &gcwq->busy_hash[v];
818 * __find_worker_executing_work - find worker which is executing a work
819 * @gcwq: gcwq of interest
820 * @bwh: hash head as returned by busy_worker_head()
821 * @work: work to find worker for
823 * Find a worker which is executing @work on @gcwq. @bwh should be
824 * the hash head obtained by calling busy_worker_head() with the same
828 * spin_lock_irq(gcwq->lock).
831 * Pointer to worker which is executing @work if found, NULL
834 static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
835 struct hlist_head *bwh,
836 struct work_struct *work)
838 struct worker *worker;
839 struct hlist_node *tmp;
841 hlist_for_each_entry(worker, tmp, bwh, hentry)
842 if (worker->current_work == work)
848 * find_worker_executing_work - find worker which is executing a work
849 * @gcwq: gcwq of interest
850 * @work: work to find worker for
852 * Find a worker which is executing @work on @gcwq. This function is
853 * identical to __find_worker_executing_work() except that this
854 * function calculates @bwh itself.
857 * spin_lock_irq(gcwq->lock).
860 * Pointer to worker which is executing @work if found, NULL
863 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
864 struct work_struct *work)
866 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
871 * gcwq_determine_ins_pos - find insertion position
872 * @gcwq: gcwq of interest
873 * @cwq: cwq a work is being queued for
875 * A work for @cwq is about to be queued on @gcwq, determine insertion
876 * position for the work. If @cwq is for HIGHPRI wq, the work is
877 * queued at the head of the queue but in FIFO order with respect to
878 * other HIGHPRI works; otherwise, at the end of the queue. This
879 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
880 * there are HIGHPRI works pending.
883 * spin_lock_irq(gcwq->lock).
886 * Pointer to inserstion position.
888 static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq,
889 struct cpu_workqueue_struct *cwq)
891 struct work_struct *twork;
893 if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
894 return &gcwq->worklist;
896 list_for_each_entry(twork, &gcwq->worklist, entry) {
897 struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
899 if (!(tcwq->wq->flags & WQ_HIGHPRI))
903 gcwq->flags |= GCWQ_HIGHPRI_PENDING;
904 return &twork->entry;
908 * insert_work - insert a work into gcwq
909 * @cwq: cwq @work belongs to
910 * @work: work to insert
911 * @head: insertion point
912 * @extra_flags: extra WORK_STRUCT_* flags to set
914 * Insert @work which belongs to @cwq into @gcwq after @head.
915 * @extra_flags is or'd to work_struct flags.
918 * spin_lock_irq(gcwq->lock).
920 static void insert_work(struct cpu_workqueue_struct *cwq,
921 struct work_struct *work, struct list_head *head,
922 unsigned int extra_flags)
924 struct global_cwq *gcwq = cwq->gcwq;
926 /* we own @work, set data and link */
927 set_work_cwq(work, cwq, extra_flags);
930 * Ensure that we get the right work->data if we see the
931 * result of list_add() below, see try_to_grab_pending().
935 list_add_tail(&work->entry, head);
938 * Ensure either worker_sched_deactivated() sees the above
939 * list_add_tail() or we see zero nr_running to avoid workers
940 * lying around lazily while there are works to be processed.
944 if (__need_more_worker(gcwq))
945 wake_up_worker(gcwq);
948 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
949 struct work_struct *work)
951 struct global_cwq *gcwq;
952 struct cpu_workqueue_struct *cwq;
953 struct list_head *worklist;
954 unsigned int work_flags;
957 debug_work_activate(work);
959 if (WARN_ON_ONCE(wq->flags & WQ_DYING))
962 /* determine gcwq to use */
963 if (!(wq->flags & WQ_UNBOUND)) {
964 struct global_cwq *last_gcwq;
966 if (unlikely(cpu == WORK_CPU_UNBOUND))
967 cpu = raw_smp_processor_id();
970 * It's multi cpu. If @wq is non-reentrant and @work
971 * was previously on a different cpu, it might still
972 * be running there, in which case the work needs to
973 * be queued on that cpu to guarantee non-reentrance.
975 gcwq = get_gcwq(cpu);
976 if (wq->flags & WQ_NON_REENTRANT &&
977 (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
978 struct worker *worker;
980 spin_lock_irqsave(&last_gcwq->lock, flags);
982 worker = find_worker_executing_work(last_gcwq, work);
984 if (worker && worker->current_cwq->wq == wq)
987 /* meh... not running there, queue here */
988 spin_unlock_irqrestore(&last_gcwq->lock, flags);
989 spin_lock_irqsave(&gcwq->lock, flags);
992 spin_lock_irqsave(&gcwq->lock, flags);
994 gcwq = get_gcwq(WORK_CPU_UNBOUND);
995 spin_lock_irqsave(&gcwq->lock, flags);
998 /* gcwq determined, get cwq and queue */
999 cwq = get_cwq(gcwq->cpu, wq);
1001 BUG_ON(!list_empty(&work->entry));
1003 cwq->nr_in_flight[cwq->work_color]++;
1004 work_flags = work_color_to_flags(cwq->work_color);
1006 if (likely(cwq->nr_active < cwq->max_active)) {
1008 worklist = gcwq_determine_ins_pos(gcwq, cwq);
1010 work_flags |= WORK_STRUCT_DELAYED;
1011 worklist = &cwq->delayed_works;
1014 insert_work(cwq, work, worklist, work_flags);
1016 spin_unlock_irqrestore(&gcwq->lock, flags);
1020 * queue_work - queue work on a workqueue
1021 * @wq: workqueue to use
1022 * @work: work to queue
1024 * Returns 0 if @work was already on a queue, non-zero otherwise.
1026 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1027 * it can be processed by another CPU.
1029 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
1033 ret = queue_work_on(get_cpu(), wq, work);
1038 EXPORT_SYMBOL_GPL(queue_work);
1041 * queue_work_on - queue work on specific cpu
1042 * @cpu: CPU number to execute work on
1043 * @wq: workqueue to use
1044 * @work: work to queue
1046 * Returns 0 if @work was already on a queue, non-zero otherwise.
1048 * We queue the work to a specific CPU, the caller must ensure it
1052 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
1056 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1057 __queue_work(cpu, wq, work);
1062 EXPORT_SYMBOL_GPL(queue_work_on);
1064 static void delayed_work_timer_fn(unsigned long __data)
1066 struct delayed_work *dwork = (struct delayed_work *)__data;
1067 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1069 __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
1073 * queue_delayed_work - queue work on a workqueue after delay
1074 * @wq: workqueue to use
1075 * @dwork: delayable work to queue
1076 * @delay: number of jiffies to wait before queueing
1078 * Returns 0 if @work was already on a queue, non-zero otherwise.
1080 int queue_delayed_work(struct workqueue_struct *wq,
1081 struct delayed_work *dwork, unsigned long delay)
1084 return queue_work(wq, &dwork->work);
1086 return queue_delayed_work_on(-1, wq, dwork, delay);
1088 EXPORT_SYMBOL_GPL(queue_delayed_work);
1091 * queue_delayed_work_on - queue work on specific CPU after delay
1092 * @cpu: CPU number to execute work on
1093 * @wq: workqueue to use
1094 * @dwork: work to queue
1095 * @delay: number of jiffies to wait before queueing
1097 * Returns 0 if @work was already on a queue, non-zero otherwise.
1099 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1100 struct delayed_work *dwork, unsigned long delay)
1103 struct timer_list *timer = &dwork->timer;
1104 struct work_struct *work = &dwork->work;
1106 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1109 BUG_ON(timer_pending(timer));
1110 BUG_ON(!list_empty(&work->entry));
1112 timer_stats_timer_set_start_info(&dwork->timer);
1115 * This stores cwq for the moment, for the timer_fn.
1116 * Note that the work's gcwq is preserved to allow
1117 * reentrance detection for delayed works.
1119 if (!(wq->flags & WQ_UNBOUND)) {
1120 struct global_cwq *gcwq = get_work_gcwq(work);
1122 if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1125 lcpu = raw_smp_processor_id();
1127 lcpu = WORK_CPU_UNBOUND;
1129 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1131 timer->expires = jiffies + delay;
1132 timer->data = (unsigned long)dwork;
1133 timer->function = delayed_work_timer_fn;
1135 if (unlikely(cpu >= 0))
1136 add_timer_on(timer, cpu);
1143 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1146 * worker_enter_idle - enter idle state
1147 * @worker: worker which is entering idle state
1149 * @worker is entering idle state. Update stats and idle timer if
1153 * spin_lock_irq(gcwq->lock).
1155 static void worker_enter_idle(struct worker *worker)
1157 struct global_cwq *gcwq = worker->gcwq;
1159 BUG_ON(worker->flags & WORKER_IDLE);
1160 BUG_ON(!list_empty(&worker->entry) &&
1161 (worker->hentry.next || worker->hentry.pprev));
1163 /* can't use worker_set_flags(), also called from start_worker() */
1164 worker->flags |= WORKER_IDLE;
1166 worker->last_active = jiffies;
1168 /* idle_list is LIFO */
1169 list_add(&worker->entry, &gcwq->idle_list);
1171 if (likely(!(worker->flags & WORKER_ROGUE))) {
1172 if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1173 mod_timer(&gcwq->idle_timer,
1174 jiffies + IDLE_WORKER_TIMEOUT);
1176 wake_up_all(&gcwq->trustee_wait);
1178 /* sanity check nr_running */
1179 WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle &&
1180 atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1184 * worker_leave_idle - leave idle state
1185 * @worker: worker which is leaving idle state
1187 * @worker is leaving idle state. Update stats.
1190 * spin_lock_irq(gcwq->lock).
1192 static void worker_leave_idle(struct worker *worker)
1194 struct global_cwq *gcwq = worker->gcwq;
1196 BUG_ON(!(worker->flags & WORKER_IDLE));
1197 worker_clr_flags(worker, WORKER_IDLE);
1199 list_del_init(&worker->entry);
1203 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1206 * Works which are scheduled while the cpu is online must at least be
1207 * scheduled to a worker which is bound to the cpu so that if they are
1208 * flushed from cpu callbacks while cpu is going down, they are
1209 * guaranteed to execute on the cpu.
1211 * This function is to be used by rogue workers and rescuers to bind
1212 * themselves to the target cpu and may race with cpu going down or
1213 * coming online. kthread_bind() can't be used because it may put the
1214 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1215 * verbatim as it's best effort and blocking and gcwq may be
1216 * [dis]associated in the meantime.
1218 * This function tries set_cpus_allowed() and locks gcwq and verifies
1219 * the binding against GCWQ_DISASSOCIATED which is set during
1220 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1221 * idle state or fetches works without dropping lock, it can guarantee
1222 * the scheduling requirement described in the first paragraph.
1225 * Might sleep. Called without any lock but returns with gcwq->lock
1229 * %true if the associated gcwq is online (@worker is successfully
1230 * bound), %false if offline.
1232 static bool worker_maybe_bind_and_lock(struct worker *worker)
1233 __acquires(&gcwq->lock)
1235 struct global_cwq *gcwq = worker->gcwq;
1236 struct task_struct *task = worker->task;
1240 * The following call may fail, succeed or succeed
1241 * without actually migrating the task to the cpu if
1242 * it races with cpu hotunplug operation. Verify
1243 * against GCWQ_DISASSOCIATED.
1245 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1246 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1248 spin_lock_irq(&gcwq->lock);
1249 if (gcwq->flags & GCWQ_DISASSOCIATED)
1251 if (task_cpu(task) == gcwq->cpu &&
1252 cpumask_equal(¤t->cpus_allowed,
1253 get_cpu_mask(gcwq->cpu)))
1255 spin_unlock_irq(&gcwq->lock);
1257 /* CPU has come up inbetween, retry migration */
1263 * Function for worker->rebind_work used to rebind rogue busy workers
1264 * to the associated cpu which is coming back online. This is
1265 * scheduled by cpu up but can race with other cpu hotplug operations
1266 * and may be executed twice without intervening cpu down.
1268 static void worker_rebind_fn(struct work_struct *work)
1270 struct worker *worker = container_of(work, struct worker, rebind_work);
1271 struct global_cwq *gcwq = worker->gcwq;
1273 if (worker_maybe_bind_and_lock(worker))
1274 worker_clr_flags(worker, WORKER_REBIND);
1276 spin_unlock_irq(&gcwq->lock);
1279 static struct worker *alloc_worker(void)
1281 struct worker *worker;
1283 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1285 INIT_LIST_HEAD(&worker->entry);
1286 INIT_LIST_HEAD(&worker->scheduled);
1287 INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1288 /* on creation a worker is in !idle && prep state */
1289 worker->flags = WORKER_PREP;
1295 * create_worker - create a new workqueue worker
1296 * @gcwq: gcwq the new worker will belong to
1297 * @bind: whether to set affinity to @cpu or not
1299 * Create a new worker which is bound to @gcwq. The returned worker
1300 * can be started by calling start_worker() or destroyed using
1304 * Might sleep. Does GFP_KERNEL allocations.
1307 * Pointer to the newly created worker.
1309 static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
1311 bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1312 struct worker *worker = NULL;
1315 spin_lock_irq(&gcwq->lock);
1316 while (ida_get_new(&gcwq->worker_ida, &id)) {
1317 spin_unlock_irq(&gcwq->lock);
1318 if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1320 spin_lock_irq(&gcwq->lock);
1322 spin_unlock_irq(&gcwq->lock);
1324 worker = alloc_worker();
1328 worker->gcwq = gcwq;
1331 if (!on_unbound_cpu)
1332 worker->task = kthread_create(worker_thread, worker,
1333 "kworker/%u:%d", gcwq->cpu, id);
1335 worker->task = kthread_create(worker_thread, worker,
1336 "kworker/u:%d", id);
1337 if (IS_ERR(worker->task))
1341 * A rogue worker will become a regular one if CPU comes
1342 * online later on. Make sure every worker has
1343 * PF_THREAD_BOUND set.
1345 if (bind && !on_unbound_cpu)
1346 kthread_bind(worker->task, gcwq->cpu);
1348 worker->task->flags |= PF_THREAD_BOUND;
1350 worker->flags |= WORKER_UNBOUND;
1356 spin_lock_irq(&gcwq->lock);
1357 ida_remove(&gcwq->worker_ida, id);
1358 spin_unlock_irq(&gcwq->lock);
1365 * start_worker - start a newly created worker
1366 * @worker: worker to start
1368 * Make the gcwq aware of @worker and start it.
1371 * spin_lock_irq(gcwq->lock).
1373 static void start_worker(struct worker *worker)
1375 worker->flags |= WORKER_STARTED;
1376 worker->gcwq->nr_workers++;
1377 worker_enter_idle(worker);
1378 wake_up_process(worker->task);
1382 * destroy_worker - destroy a workqueue worker
1383 * @worker: worker to be destroyed
1385 * Destroy @worker and adjust @gcwq stats accordingly.
1388 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1390 static void destroy_worker(struct worker *worker)
1392 struct global_cwq *gcwq = worker->gcwq;
1393 int id = worker->id;
1395 /* sanity check frenzy */
1396 BUG_ON(worker->current_work);
1397 BUG_ON(!list_empty(&worker->scheduled));
1399 if (worker->flags & WORKER_STARTED)
1401 if (worker->flags & WORKER_IDLE)
1404 list_del_init(&worker->entry);
1405 worker->flags |= WORKER_DIE;
1407 spin_unlock_irq(&gcwq->lock);
1409 kthread_stop(worker->task);
1412 spin_lock_irq(&gcwq->lock);
1413 ida_remove(&gcwq->worker_ida, id);
1416 static void idle_worker_timeout(unsigned long __gcwq)
1418 struct global_cwq *gcwq = (void *)__gcwq;
1420 spin_lock_irq(&gcwq->lock);
1422 if (too_many_workers(gcwq)) {
1423 struct worker *worker;
1424 unsigned long expires;
1426 /* idle_list is kept in LIFO order, check the last one */
1427 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1428 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1430 if (time_before(jiffies, expires))
1431 mod_timer(&gcwq->idle_timer, expires);
1433 /* it's been idle for too long, wake up manager */
1434 gcwq->flags |= GCWQ_MANAGE_WORKERS;
1435 wake_up_worker(gcwq);
1439 spin_unlock_irq(&gcwq->lock);
1442 static bool send_mayday(struct work_struct *work)
1444 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1445 struct workqueue_struct *wq = cwq->wq;
1448 if (!(wq->flags & WQ_RESCUER))
1451 /* mayday mayday mayday */
1452 cpu = cwq->gcwq->cpu;
1453 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1454 if (cpu == WORK_CPU_UNBOUND)
1456 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1457 wake_up_process(wq->rescuer->task);
1461 static void gcwq_mayday_timeout(unsigned long __gcwq)
1463 struct global_cwq *gcwq = (void *)__gcwq;
1464 struct work_struct *work;
1466 spin_lock_irq(&gcwq->lock);
1468 if (need_to_create_worker(gcwq)) {
1470 * We've been trying to create a new worker but
1471 * haven't been successful. We might be hitting an
1472 * allocation deadlock. Send distress signals to
1475 list_for_each_entry(work, &gcwq->worklist, entry)
1479 spin_unlock_irq(&gcwq->lock);
1481 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1485 * maybe_create_worker - create a new worker if necessary
1486 * @gcwq: gcwq to create a new worker for
1488 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1489 * have at least one idle worker on return from this function. If
1490 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1491 * sent to all rescuers with works scheduled on @gcwq to resolve
1492 * possible allocation deadlock.
1494 * On return, need_to_create_worker() is guaranteed to be false and
1495 * may_start_working() true.
1498 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1499 * multiple times. Does GFP_KERNEL allocations. Called only from
1503 * false if no action was taken and gcwq->lock stayed locked, true
1506 static bool maybe_create_worker(struct global_cwq *gcwq)
1507 __releases(&gcwq->lock)
1508 __acquires(&gcwq->lock)
1510 if (!need_to_create_worker(gcwq))
1513 spin_unlock_irq(&gcwq->lock);
1515 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1516 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1519 struct worker *worker;
1521 worker = create_worker(gcwq, true);
1523 del_timer_sync(&gcwq->mayday_timer);
1524 spin_lock_irq(&gcwq->lock);
1525 start_worker(worker);
1526 BUG_ON(need_to_create_worker(gcwq));
1530 if (!need_to_create_worker(gcwq))
1533 __set_current_state(TASK_INTERRUPTIBLE);
1534 schedule_timeout(CREATE_COOLDOWN);
1536 if (!need_to_create_worker(gcwq))
1540 del_timer_sync(&gcwq->mayday_timer);
1541 spin_lock_irq(&gcwq->lock);
1542 if (need_to_create_worker(gcwq))
1548 * maybe_destroy_worker - destroy workers which have been idle for a while
1549 * @gcwq: gcwq to destroy workers for
1551 * Destroy @gcwq workers which have been idle for longer than
1552 * IDLE_WORKER_TIMEOUT.
1555 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1556 * multiple times. Called only from manager.
1559 * false if no action was taken and gcwq->lock stayed locked, true
1562 static bool maybe_destroy_workers(struct global_cwq *gcwq)
1566 while (too_many_workers(gcwq)) {
1567 struct worker *worker;
1568 unsigned long expires;
1570 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1571 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1573 if (time_before(jiffies, expires)) {
1574 mod_timer(&gcwq->idle_timer, expires);
1578 destroy_worker(worker);
1586 * manage_workers - manage worker pool
1589 * Assume the manager role and manage gcwq worker pool @worker belongs
1590 * to. At any given time, there can be only zero or one manager per
1591 * gcwq. The exclusion is handled automatically by this function.
1593 * The caller can safely start processing works on false return. On
1594 * true return, it's guaranteed that need_to_create_worker() is false
1595 * and may_start_working() is true.
1598 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1599 * multiple times. Does GFP_KERNEL allocations.
1602 * false if no action was taken and gcwq->lock stayed locked, true if
1603 * some action was taken.
1605 static bool manage_workers(struct worker *worker)
1607 struct global_cwq *gcwq = worker->gcwq;
1610 if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1613 gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1614 gcwq->flags |= GCWQ_MANAGING_WORKERS;
1617 * Destroy and then create so that may_start_working() is true
1620 ret |= maybe_destroy_workers(gcwq);
1621 ret |= maybe_create_worker(gcwq);
1623 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1626 * The trustee might be waiting to take over the manager
1627 * position, tell it we're done.
1629 if (unlikely(gcwq->trustee))
1630 wake_up_all(&gcwq->trustee_wait);
1636 * move_linked_works - move linked works to a list
1637 * @work: start of series of works to be scheduled
1638 * @head: target list to append @work to
1639 * @nextp: out paramter for nested worklist walking
1641 * Schedule linked works starting from @work to @head. Work series to
1642 * be scheduled starts at @work and includes any consecutive work with
1643 * WORK_STRUCT_LINKED set in its predecessor.
1645 * If @nextp is not NULL, it's updated to point to the next work of
1646 * the last scheduled work. This allows move_linked_works() to be
1647 * nested inside outer list_for_each_entry_safe().
1650 * spin_lock_irq(gcwq->lock).
1652 static void move_linked_works(struct work_struct *work, struct list_head *head,
1653 struct work_struct **nextp)
1655 struct work_struct *n;
1658 * Linked worklist will always end before the end of the list,
1659 * use NULL for list head.
1661 list_for_each_entry_safe_from(work, n, NULL, entry) {
1662 list_move_tail(&work->entry, head);
1663 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1668 * If we're already inside safe list traversal and have moved
1669 * multiple works to the scheduled queue, the next position
1670 * needs to be updated.
1676 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1678 struct work_struct *work = list_first_entry(&cwq->delayed_works,
1679 struct work_struct, entry);
1680 struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1682 move_linked_works(work, pos, NULL);
1683 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1688 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1689 * @cwq: cwq of interest
1690 * @color: color of work which left the queue
1691 * @delayed: for a delayed work
1693 * A work either has completed or is removed from pending queue,
1694 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1697 * spin_lock_irq(gcwq->lock).
1699 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1702 /* ignore uncolored works */
1703 if (color == WORK_NO_COLOR)
1706 cwq->nr_in_flight[color]--;
1710 if (!list_empty(&cwq->delayed_works)) {
1711 /* one down, submit a delayed one */
1712 if (cwq->nr_active < cwq->max_active)
1713 cwq_activate_first_delayed(cwq);
1717 /* is flush in progress and are we at the flushing tip? */
1718 if (likely(cwq->flush_color != color))
1721 /* are there still in-flight works? */
1722 if (cwq->nr_in_flight[color])
1725 /* this cwq is done, clear flush_color */
1726 cwq->flush_color = -1;
1729 * If this was the last cwq, wake up the first flusher. It
1730 * will handle the rest.
1732 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1733 complete(&cwq->wq->first_flusher->done);
1737 * process_one_work - process single work
1739 * @work: work to process
1741 * Process @work. This function contains all the logics necessary to
1742 * process a single work including synchronization against and
1743 * interaction with other workers on the same cpu, queueing and
1744 * flushing. As long as context requirement is met, any worker can
1745 * call this function to process a work.
1748 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1750 static void process_one_work(struct worker *worker, struct work_struct *work)
1751 __releases(&gcwq->lock)
1752 __acquires(&gcwq->lock)
1754 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1755 struct global_cwq *gcwq = cwq->gcwq;
1756 struct hlist_head *bwh = busy_worker_head(gcwq, work);
1757 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1758 work_func_t f = work->func;
1760 struct worker *collision;
1761 #ifdef CONFIG_LOCKDEP
1763 * It is permissible to free the struct work_struct from
1764 * inside the function that is called from it, this we need to
1765 * take into account for lockdep too. To avoid bogus "held
1766 * lock freed" warnings as well as problems when looking into
1767 * work->lockdep_map, make a copy and use that here.
1769 struct lockdep_map lockdep_map = work->lockdep_map;
1772 * A single work shouldn't be executed concurrently by
1773 * multiple workers on a single cpu. Check whether anyone is
1774 * already processing the work. If so, defer the work to the
1775 * currently executing one.
1777 collision = __find_worker_executing_work(gcwq, bwh, work);
1778 if (unlikely(collision)) {
1779 move_linked_works(work, &collision->scheduled, NULL);
1783 /* claim and process */
1784 debug_work_deactivate(work);
1785 hlist_add_head(&worker->hentry, bwh);
1786 worker->current_work = work;
1787 worker->current_cwq = cwq;
1788 work_color = get_work_color(work);
1790 /* record the current cpu number in the work data and dequeue */
1791 set_work_cpu(work, gcwq->cpu);
1792 list_del_init(&work->entry);
1795 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1796 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1798 if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1799 struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1800 struct work_struct, entry);
1802 if (!list_empty(&gcwq->worklist) &&
1803 get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1804 wake_up_worker(gcwq);
1806 gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1810 * CPU intensive works don't participate in concurrency
1811 * management. They're the scheduler's responsibility.
1813 if (unlikely(cpu_intensive))
1814 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1816 spin_unlock_irq(&gcwq->lock);
1818 work_clear_pending(work);
1819 lock_map_acquire(&cwq->wq->lockdep_map);
1820 lock_map_acquire(&lockdep_map);
1821 trace_workqueue_execute_start(work);
1824 * While we must be careful to not use "work" after this, the trace
1825 * point will only record its address.
1827 trace_workqueue_execute_end(work);
1828 lock_map_release(&lockdep_map);
1829 lock_map_release(&cwq->wq->lockdep_map);
1831 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1832 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1834 current->comm, preempt_count(), task_pid_nr(current));
1835 printk(KERN_ERR " last function: ");
1836 print_symbol("%s\n", (unsigned long)f);
1837 debug_show_held_locks(current);
1841 spin_lock_irq(&gcwq->lock);
1843 /* clear cpu intensive status */
1844 if (unlikely(cpu_intensive))
1845 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1847 /* we're done with it, release */
1848 hlist_del_init(&worker->hentry);
1849 worker->current_work = NULL;
1850 worker->current_cwq = NULL;
1851 cwq_dec_nr_in_flight(cwq, work_color, false);
1855 * process_scheduled_works - process scheduled works
1858 * Process all scheduled works. Please note that the scheduled list
1859 * may change while processing a work, so this function repeatedly
1860 * fetches a work from the top and executes it.
1863 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1866 static void process_scheduled_works(struct worker *worker)
1868 while (!list_empty(&worker->scheduled)) {
1869 struct work_struct *work = list_first_entry(&worker->scheduled,
1870 struct work_struct, entry);
1871 process_one_work(worker, work);
1876 * worker_thread - the worker thread function
1879 * The gcwq worker thread function. There's a single dynamic pool of
1880 * these per each cpu. These workers process all works regardless of
1881 * their specific target workqueue. The only exception is works which
1882 * belong to workqueues with a rescuer which will be explained in
1885 static int worker_thread(void *__worker)
1887 struct worker *worker = __worker;
1888 struct global_cwq *gcwq = worker->gcwq;
1890 /* tell the scheduler that this is a workqueue worker */
1891 worker->task->flags |= PF_WQ_WORKER;
1893 spin_lock_irq(&gcwq->lock);
1895 /* DIE can be set only while we're idle, checking here is enough */
1896 if (worker->flags & WORKER_DIE) {
1897 spin_unlock_irq(&gcwq->lock);
1898 worker->task->flags &= ~PF_WQ_WORKER;
1902 worker_leave_idle(worker);
1904 /* no more worker necessary? */
1905 if (!need_more_worker(gcwq))
1908 /* do we need to manage? */
1909 if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1913 * ->scheduled list can only be filled while a worker is
1914 * preparing to process a work or actually processing it.
1915 * Make sure nobody diddled with it while I was sleeping.
1917 BUG_ON(!list_empty(&worker->scheduled));
1920 * When control reaches this point, we're guaranteed to have
1921 * at least one idle worker or that someone else has already
1922 * assumed the manager role.
1924 worker_clr_flags(worker, WORKER_PREP);
1927 struct work_struct *work =
1928 list_first_entry(&gcwq->worklist,
1929 struct work_struct, entry);
1931 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1932 /* optimization path, not strictly necessary */
1933 process_one_work(worker, work);
1934 if (unlikely(!list_empty(&worker->scheduled)))
1935 process_scheduled_works(worker);
1937 move_linked_works(work, &worker->scheduled, NULL);
1938 process_scheduled_works(worker);
1940 } while (keep_working(gcwq));
1942 worker_set_flags(worker, WORKER_PREP, false);
1944 if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1948 * gcwq->lock is held and there's no work to process and no
1949 * need to manage, sleep. Workers are woken up only while
1950 * holding gcwq->lock or from local cpu, so setting the
1951 * current state before releasing gcwq->lock is enough to
1952 * prevent losing any event.
1954 worker_enter_idle(worker);
1955 __set_current_state(TASK_INTERRUPTIBLE);
1956 spin_unlock_irq(&gcwq->lock);
1962 * rescuer_thread - the rescuer thread function
1963 * @__wq: the associated workqueue
1965 * Workqueue rescuer thread function. There's one rescuer for each
1966 * workqueue which has WQ_RESCUER set.
1968 * Regular work processing on a gcwq may block trying to create a new
1969 * worker which uses GFP_KERNEL allocation which has slight chance of
1970 * developing into deadlock if some works currently on the same queue
1971 * need to be processed to satisfy the GFP_KERNEL allocation. This is
1972 * the problem rescuer solves.
1974 * When such condition is possible, the gcwq summons rescuers of all
1975 * workqueues which have works queued on the gcwq and let them process
1976 * those works so that forward progress can be guaranteed.
1978 * This should happen rarely.
1980 static int rescuer_thread(void *__wq)
1982 struct workqueue_struct *wq = __wq;
1983 struct worker *rescuer = wq->rescuer;
1984 struct list_head *scheduled = &rescuer->scheduled;
1985 bool is_unbound = wq->flags & WQ_UNBOUND;
1988 set_user_nice(current, RESCUER_NICE_LEVEL);
1990 set_current_state(TASK_INTERRUPTIBLE);
1992 if (kthread_should_stop())
1996 * See whether any cpu is asking for help. Unbounded
1997 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
1999 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2000 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2001 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2002 struct global_cwq *gcwq = cwq->gcwq;
2003 struct work_struct *work, *n;
2005 __set_current_state(TASK_RUNNING);
2006 mayday_clear_cpu(cpu, wq->mayday_mask);
2008 /* migrate to the target cpu if possible */
2009 rescuer->gcwq = gcwq;
2010 worker_maybe_bind_and_lock(rescuer);
2013 * Slurp in all works issued via this workqueue and
2016 BUG_ON(!list_empty(&rescuer->scheduled));
2017 list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2018 if (get_work_cwq(work) == cwq)
2019 move_linked_works(work, scheduled, &n);
2021 process_scheduled_works(rescuer);
2022 spin_unlock_irq(&gcwq->lock);
2030 struct work_struct work;
2031 struct completion done;
2034 static void wq_barrier_func(struct work_struct *work)
2036 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2037 complete(&barr->done);
2041 * insert_wq_barrier - insert a barrier work
2042 * @cwq: cwq to insert barrier into
2043 * @barr: wq_barrier to insert
2044 * @target: target work to attach @barr to
2045 * @worker: worker currently executing @target, NULL if @target is not executing
2047 * @barr is linked to @target such that @barr is completed only after
2048 * @target finishes execution. Please note that the ordering
2049 * guarantee is observed only with respect to @target and on the local
2052 * Currently, a queued barrier can't be canceled. This is because
2053 * try_to_grab_pending() can't determine whether the work to be
2054 * grabbed is at the head of the queue and thus can't clear LINKED
2055 * flag of the previous work while there must be a valid next work
2056 * after a work with LINKED flag set.
2058 * Note that when @worker is non-NULL, @target may be modified
2059 * underneath us, so we can't reliably determine cwq from @target.
2062 * spin_lock_irq(gcwq->lock).
2064 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2065 struct wq_barrier *barr,
2066 struct work_struct *target, struct worker *worker)
2068 struct list_head *head;
2069 unsigned int linked = 0;
2072 * debugobject calls are safe here even with gcwq->lock locked
2073 * as we know for sure that this will not trigger any of the
2074 * checks and call back into the fixup functions where we
2077 INIT_WORK_ON_STACK(&barr->work, wq_barrier_func);
2078 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2079 init_completion(&barr->done);
2082 * If @target is currently being executed, schedule the
2083 * barrier to the worker; otherwise, put it after @target.
2086 head = worker->scheduled.next;
2088 unsigned long *bits = work_data_bits(target);
2090 head = target->entry.next;
2091 /* there can already be other linked works, inherit and set */
2092 linked = *bits & WORK_STRUCT_LINKED;
2093 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2096 debug_work_activate(&barr->work);
2097 insert_work(cwq, &barr->work, head,
2098 work_color_to_flags(WORK_NO_COLOR) | linked);
2102 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2103 * @wq: workqueue being flushed
2104 * @flush_color: new flush color, < 0 for no-op
2105 * @work_color: new work color, < 0 for no-op
2107 * Prepare cwqs for workqueue flushing.
2109 * If @flush_color is non-negative, flush_color on all cwqs should be
2110 * -1. If no cwq has in-flight commands at the specified color, all
2111 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2112 * has in flight commands, its cwq->flush_color is set to
2113 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2114 * wakeup logic is armed and %true is returned.
2116 * The caller should have initialized @wq->first_flusher prior to
2117 * calling this function with non-negative @flush_color. If
2118 * @flush_color is negative, no flush color update is done and %false
2121 * If @work_color is non-negative, all cwqs should have the same
2122 * work_color which is previous to @work_color and all will be
2123 * advanced to @work_color.
2126 * mutex_lock(wq->flush_mutex).
2129 * %true if @flush_color >= 0 and there's something to flush. %false
2132 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2133 int flush_color, int work_color)
2138 if (flush_color >= 0) {
2139 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2140 atomic_set(&wq->nr_cwqs_to_flush, 1);
2143 for_each_cwq_cpu(cpu, wq) {
2144 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2145 struct global_cwq *gcwq = cwq->gcwq;
2147 spin_lock_irq(&gcwq->lock);
2149 if (flush_color >= 0) {
2150 BUG_ON(cwq->flush_color != -1);
2152 if (cwq->nr_in_flight[flush_color]) {
2153 cwq->flush_color = flush_color;
2154 atomic_inc(&wq->nr_cwqs_to_flush);
2159 if (work_color >= 0) {
2160 BUG_ON(work_color != work_next_color(cwq->work_color));
2161 cwq->work_color = work_color;
2164 spin_unlock_irq(&gcwq->lock);
2167 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2168 complete(&wq->first_flusher->done);
2174 * flush_workqueue - ensure that any scheduled work has run to completion.
2175 * @wq: workqueue to flush
2177 * Forces execution of the workqueue and blocks until its completion.
2178 * This is typically used in driver shutdown handlers.
2180 * We sleep until all works which were queued on entry have been handled,
2181 * but we are not livelocked by new incoming ones.
2183 void flush_workqueue(struct workqueue_struct *wq)
2185 struct wq_flusher this_flusher = {
2186 .list = LIST_HEAD_INIT(this_flusher.list),
2188 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2192 lock_map_acquire(&wq->lockdep_map);
2193 lock_map_release(&wq->lockdep_map);
2195 mutex_lock(&wq->flush_mutex);
2198 * Start-to-wait phase
2200 next_color = work_next_color(wq->work_color);
2202 if (next_color != wq->flush_color) {
2204 * Color space is not full. The current work_color
2205 * becomes our flush_color and work_color is advanced
2208 BUG_ON(!list_empty(&wq->flusher_overflow));
2209 this_flusher.flush_color = wq->work_color;
2210 wq->work_color = next_color;
2212 if (!wq->first_flusher) {
2213 /* no flush in progress, become the first flusher */
2214 BUG_ON(wq->flush_color != this_flusher.flush_color);
2216 wq->first_flusher = &this_flusher;
2218 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2220 /* nothing to flush, done */
2221 wq->flush_color = next_color;
2222 wq->first_flusher = NULL;
2227 BUG_ON(wq->flush_color == this_flusher.flush_color);
2228 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2229 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2233 * Oops, color space is full, wait on overflow queue.
2234 * The next flush completion will assign us
2235 * flush_color and transfer to flusher_queue.
2237 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2240 mutex_unlock(&wq->flush_mutex);
2242 wait_for_completion(&this_flusher.done);
2245 * Wake-up-and-cascade phase
2247 * First flushers are responsible for cascading flushes and
2248 * handling overflow. Non-first flushers can simply return.
2250 if (wq->first_flusher != &this_flusher)
2253 mutex_lock(&wq->flush_mutex);
2255 /* we might have raced, check again with mutex held */
2256 if (wq->first_flusher != &this_flusher)
2259 wq->first_flusher = NULL;
2261 BUG_ON(!list_empty(&this_flusher.list));
2262 BUG_ON(wq->flush_color != this_flusher.flush_color);
2265 struct wq_flusher *next, *tmp;
2267 /* complete all the flushers sharing the current flush color */
2268 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2269 if (next->flush_color != wq->flush_color)
2271 list_del_init(&next->list);
2272 complete(&next->done);
2275 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2276 wq->flush_color != work_next_color(wq->work_color));
2278 /* this flush_color is finished, advance by one */
2279 wq->flush_color = work_next_color(wq->flush_color);
2281 /* one color has been freed, handle overflow queue */
2282 if (!list_empty(&wq->flusher_overflow)) {
2284 * Assign the same color to all overflowed
2285 * flushers, advance work_color and append to
2286 * flusher_queue. This is the start-to-wait
2287 * phase for these overflowed flushers.
2289 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2290 tmp->flush_color = wq->work_color;
2292 wq->work_color = work_next_color(wq->work_color);
2294 list_splice_tail_init(&wq->flusher_overflow,
2295 &wq->flusher_queue);
2296 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2299 if (list_empty(&wq->flusher_queue)) {
2300 BUG_ON(wq->flush_color != wq->work_color);
2305 * Need to flush more colors. Make the next flusher
2306 * the new first flusher and arm cwqs.
2308 BUG_ON(wq->flush_color == wq->work_color);
2309 BUG_ON(wq->flush_color != next->flush_color);
2311 list_del_init(&next->list);
2312 wq->first_flusher = next;
2314 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2318 * Meh... this color is already done, clear first
2319 * flusher and repeat cascading.
2321 wq->first_flusher = NULL;
2325 mutex_unlock(&wq->flush_mutex);
2327 EXPORT_SYMBOL_GPL(flush_workqueue);
2329 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2330 bool wait_executing)
2332 struct worker *worker = NULL;
2333 struct global_cwq *gcwq;
2334 struct cpu_workqueue_struct *cwq;
2337 gcwq = get_work_gcwq(work);
2341 spin_lock_irq(&gcwq->lock);
2342 if (!list_empty(&work->entry)) {
2344 * See the comment near try_to_grab_pending()->smp_rmb().
2345 * If it was re-queued to a different gcwq under us, we
2346 * are not going to wait.
2349 cwq = get_work_cwq(work);
2350 if (unlikely(!cwq || gcwq != cwq->gcwq))
2352 } else if (wait_executing) {
2353 worker = find_worker_executing_work(gcwq, work);
2356 cwq = worker->current_cwq;
2360 insert_wq_barrier(cwq, barr, work, worker);
2361 spin_unlock_irq(&gcwq->lock);
2363 lock_map_acquire(&cwq->wq->lockdep_map);
2364 lock_map_release(&cwq->wq->lockdep_map);
2367 spin_unlock_irq(&gcwq->lock);
2372 * flush_work - wait for a work to finish executing the last queueing instance
2373 * @work: the work to flush
2375 * Wait until @work has finished execution. This function considers
2376 * only the last queueing instance of @work. If @work has been
2377 * enqueued across different CPUs on a non-reentrant workqueue or on
2378 * multiple workqueues, @work might still be executing on return on
2379 * some of the CPUs from earlier queueing.
2381 * If @work was queued only on a non-reentrant, ordered or unbound
2382 * workqueue, @work is guaranteed to be idle on return if it hasn't
2383 * been requeued since flush started.
2386 * %true if flush_work() waited for the work to finish execution,
2387 * %false if it was already idle.
2389 bool flush_work(struct work_struct *work)
2391 struct wq_barrier barr;
2393 if (start_flush_work(work, &barr, true)) {
2394 wait_for_completion(&barr.done);
2395 destroy_work_on_stack(&barr.work);
2400 EXPORT_SYMBOL_GPL(flush_work);
2402 static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2404 struct wq_barrier barr;
2405 struct worker *worker;
2407 spin_lock_irq(&gcwq->lock);
2409 worker = find_worker_executing_work(gcwq, work);
2410 if (unlikely(worker))
2411 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2413 spin_unlock_irq(&gcwq->lock);
2415 if (unlikely(worker)) {
2416 wait_for_completion(&barr.done);
2417 destroy_work_on_stack(&barr.work);
2423 static bool wait_on_work(struct work_struct *work)
2430 lock_map_acquire(&work->lockdep_map);
2431 lock_map_release(&work->lockdep_map);
2433 for_each_gcwq_cpu(cpu)
2434 ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2439 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2440 * so this work can't be re-armed in any way.
2442 static int try_to_grab_pending(struct work_struct *work)
2444 struct global_cwq *gcwq;
2447 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2451 * The queueing is in progress, or it is already queued. Try to
2452 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2454 gcwq = get_work_gcwq(work);
2458 spin_lock_irq(&gcwq->lock);
2459 if (!list_empty(&work->entry)) {
2461 * This work is queued, but perhaps we locked the wrong gcwq.
2462 * In that case we must see the new value after rmb(), see
2463 * insert_work()->wmb().
2466 if (gcwq == get_work_gcwq(work)) {
2467 debug_work_deactivate(work);
2468 list_del_init(&work->entry);
2469 cwq_dec_nr_in_flight(get_work_cwq(work),
2470 get_work_color(work),
2471 *work_data_bits(work) & WORK_STRUCT_DELAYED);
2475 spin_unlock_irq(&gcwq->lock);
2480 static bool __cancel_work_timer(struct work_struct *work,
2481 struct timer_list* timer)
2486 ret = (timer && likely(del_timer(timer)));
2488 ret = try_to_grab_pending(work);
2490 } while (unlikely(ret < 0));
2492 clear_work_data(work);
2497 * cancel_work_sync - cancel a work and wait for it to finish
2498 * @work: the work to cancel
2500 * Cancel @work and wait for its execution to finish. This function
2501 * can be used even if the work re-queues itself or migrates to
2502 * another workqueue. On return from this function, @work is
2503 * guaranteed to be not pending or executing on any CPU.
2505 * cancel_work_sync(&delayed_work->work) must not be used for
2506 * delayed_work's. Use cancel_delayed_work_sync() instead.
2508 * The caller must ensure that the workqueue on which @work was last
2509 * queued can't be destroyed before this function returns.
2512 * %true if @work was pending, %false otherwise.
2514 bool cancel_work_sync(struct work_struct *work)
2516 return __cancel_work_timer(work, NULL);
2518 EXPORT_SYMBOL_GPL(cancel_work_sync);
2521 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2522 * @dwork: the delayed work to flush
2524 * Delayed timer is cancelled and the pending work is queued for
2525 * immediate execution. Like flush_work(), this function only
2526 * considers the last queueing instance of @dwork.
2529 * %true if flush_work() waited for the work to finish execution,
2530 * %false if it was already idle.
2532 bool flush_delayed_work(struct delayed_work *dwork)
2534 if (del_timer_sync(&dwork->timer))
2535 __queue_work(raw_smp_processor_id(),
2536 get_work_cwq(&dwork->work)->wq, &dwork->work);
2537 return flush_work(&dwork->work);
2539 EXPORT_SYMBOL(flush_delayed_work);
2542 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2543 * @dwork: the delayed work cancel
2545 * This is cancel_work_sync() for delayed works.
2548 * %true if @dwork was pending, %false otherwise.
2550 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2552 return __cancel_work_timer(&dwork->work, &dwork->timer);
2554 EXPORT_SYMBOL(cancel_delayed_work_sync);
2557 * schedule_work - put work task in global workqueue
2558 * @work: job to be done
2560 * Returns zero if @work was already on the kernel-global workqueue and
2561 * non-zero otherwise.
2563 * This puts a job in the kernel-global workqueue if it was not already
2564 * queued and leaves it in the same position on the kernel-global
2565 * workqueue otherwise.
2567 int schedule_work(struct work_struct *work)
2569 return queue_work(system_wq, work);
2571 EXPORT_SYMBOL(schedule_work);
2574 * schedule_work_on - put work task on a specific cpu
2575 * @cpu: cpu to put the work task on
2576 * @work: job to be done
2578 * This puts a job on a specific cpu
2580 int schedule_work_on(int cpu, struct work_struct *work)
2582 return queue_work_on(cpu, system_wq, work);
2584 EXPORT_SYMBOL(schedule_work_on);
2587 * schedule_delayed_work - put work task in global workqueue after delay
2588 * @dwork: job to be done
2589 * @delay: number of jiffies to wait or 0 for immediate execution
2591 * After waiting for a given time this puts a job in the kernel-global
2594 int schedule_delayed_work(struct delayed_work *dwork,
2595 unsigned long delay)
2597 return queue_delayed_work(system_wq, dwork, delay);
2599 EXPORT_SYMBOL(schedule_delayed_work);
2602 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2604 * @dwork: job to be done
2605 * @delay: number of jiffies to wait
2607 * After waiting for a given time this puts a job in the kernel-global
2608 * workqueue on the specified CPU.
2610 int schedule_delayed_work_on(int cpu,
2611 struct delayed_work *dwork, unsigned long delay)
2613 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2615 EXPORT_SYMBOL(schedule_delayed_work_on);
2618 * schedule_on_each_cpu - call a function on each online CPU from keventd
2619 * @func: the function to call
2621 * Returns zero on success.
2622 * Returns -ve errno on failure.
2624 * schedule_on_each_cpu() is very slow.
2626 int schedule_on_each_cpu(work_func_t func)
2629 struct work_struct __percpu *works;
2631 works = alloc_percpu(struct work_struct);
2637 for_each_online_cpu(cpu) {
2638 struct work_struct *work = per_cpu_ptr(works, cpu);
2640 INIT_WORK(work, func);
2641 schedule_work_on(cpu, work);
2644 for_each_online_cpu(cpu)
2645 flush_work(per_cpu_ptr(works, cpu));
2653 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2655 * Forces execution of the kernel-global workqueue and blocks until its
2658 * Think twice before calling this function! It's very easy to get into
2659 * trouble if you don't take great care. Either of the following situations
2660 * will lead to deadlock:
2662 * One of the work items currently on the workqueue needs to acquire
2663 * a lock held by your code or its caller.
2665 * Your code is running in the context of a work routine.
2667 * They will be detected by lockdep when they occur, but the first might not
2668 * occur very often. It depends on what work items are on the workqueue and
2669 * what locks they need, which you have no control over.
2671 * In most situations flushing the entire workqueue is overkill; you merely
2672 * need to know that a particular work item isn't queued and isn't running.
2673 * In such cases you should use cancel_delayed_work_sync() or
2674 * cancel_work_sync() instead.
2676 void flush_scheduled_work(void)
2678 flush_workqueue(system_wq);
2680 EXPORT_SYMBOL(flush_scheduled_work);
2683 * execute_in_process_context - reliably execute the routine with user context
2684 * @fn: the function to execute
2685 * @ew: guaranteed storage for the execute work structure (must
2686 * be available when the work executes)
2688 * Executes the function immediately if process context is available,
2689 * otherwise schedules the function for delayed execution.
2691 * Returns: 0 - function was executed
2692 * 1 - function was scheduled for execution
2694 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2696 if (!in_interrupt()) {
2701 INIT_WORK(&ew->work, fn);
2702 schedule_work(&ew->work);
2706 EXPORT_SYMBOL_GPL(execute_in_process_context);
2708 int keventd_up(void)
2710 return system_wq != NULL;
2713 static int alloc_cwqs(struct workqueue_struct *wq)
2716 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2717 * Make sure that the alignment isn't lower than that of
2718 * unsigned long long.
2720 const size_t size = sizeof(struct cpu_workqueue_struct);
2721 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2722 __alignof__(unsigned long long));
2724 bool percpu = !(wq->flags & WQ_UNBOUND);
2726 bool percpu = false;
2730 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2735 * Allocate enough room to align cwq and put an extra
2736 * pointer at the end pointing back to the originally
2737 * allocated pointer which will be used for free.
2739 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2741 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2742 *(void **)(wq->cpu_wq.single + 1) = ptr;
2746 /* just in case, make sure it's actually aligned */
2747 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2748 return wq->cpu_wq.v ? 0 : -ENOMEM;
2751 static void free_cwqs(struct workqueue_struct *wq)
2754 bool percpu = !(wq->flags & WQ_UNBOUND);
2756 bool percpu = false;
2760 free_percpu(wq->cpu_wq.pcpu);
2761 else if (wq->cpu_wq.single) {
2762 /* the pointer to free is stored right after the cwq */
2763 kfree(*(void **)(wq->cpu_wq.single + 1));
2767 static int wq_clamp_max_active(int max_active, unsigned int flags,
2770 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2772 if (max_active < 1 || max_active > lim)
2773 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2774 "is out of range, clamping between %d and %d\n",
2775 max_active, name, 1, lim);
2777 return clamp_val(max_active, 1, lim);
2780 struct workqueue_struct *__alloc_workqueue_key(const char *name,
2783 struct lock_class_key *key,
2784 const char *lock_name)
2786 struct workqueue_struct *wq;
2790 * Unbound workqueues aren't concurrency managed and should be
2791 * dispatched to workers immediately.
2793 if (flags & WQ_UNBOUND)
2794 flags |= WQ_HIGHPRI;
2796 max_active = max_active ?: WQ_DFL_ACTIVE;
2797 max_active = wq_clamp_max_active(max_active, flags, name);
2799 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
2804 wq->saved_max_active = max_active;
2805 mutex_init(&wq->flush_mutex);
2806 atomic_set(&wq->nr_cwqs_to_flush, 0);
2807 INIT_LIST_HEAD(&wq->flusher_queue);
2808 INIT_LIST_HEAD(&wq->flusher_overflow);
2811 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
2812 INIT_LIST_HEAD(&wq->list);
2814 if (alloc_cwqs(wq) < 0)
2817 for_each_cwq_cpu(cpu, wq) {
2818 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2819 struct global_cwq *gcwq = get_gcwq(cpu);
2821 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
2824 cwq->flush_color = -1;
2825 cwq->max_active = max_active;
2826 INIT_LIST_HEAD(&cwq->delayed_works);
2829 if (flags & WQ_RESCUER) {
2830 struct worker *rescuer;
2832 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
2835 wq->rescuer = rescuer = alloc_worker();
2839 rescuer->task = kthread_create(rescuer_thread, wq, "%s", name);
2840 if (IS_ERR(rescuer->task))
2843 rescuer->task->flags |= PF_THREAD_BOUND;
2844 wake_up_process(rescuer->task);
2848 * workqueue_lock protects global freeze state and workqueues
2849 * list. Grab it, set max_active accordingly and add the new
2850 * workqueue to workqueues list.
2852 spin_lock(&workqueue_lock);
2854 if (workqueue_freezing && wq->flags & WQ_FREEZEABLE)
2855 for_each_cwq_cpu(cpu, wq)
2856 get_cwq(cpu, wq)->max_active = 0;
2858 list_add(&wq->list, &workqueues);
2860 spin_unlock(&workqueue_lock);
2866 free_mayday_mask(wq->mayday_mask);
2872 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
2875 * destroy_workqueue - safely terminate a workqueue
2876 * @wq: target workqueue
2878 * Safely destroy a workqueue. All work currently pending will be done first.
2880 void destroy_workqueue(struct workqueue_struct *wq)
2884 wq->flags |= WQ_DYING;
2885 flush_workqueue(wq);
2888 * wq list is used to freeze wq, remove from list after
2889 * flushing is complete in case freeze races us.
2891 spin_lock(&workqueue_lock);
2892 list_del(&wq->list);
2893 spin_unlock(&workqueue_lock);
2896 for_each_cwq_cpu(cpu, wq) {
2897 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2900 for (i = 0; i < WORK_NR_COLORS; i++)
2901 BUG_ON(cwq->nr_in_flight[i]);
2902 BUG_ON(cwq->nr_active);
2903 BUG_ON(!list_empty(&cwq->delayed_works));
2906 if (wq->flags & WQ_RESCUER) {
2907 kthread_stop(wq->rescuer->task);
2908 free_mayday_mask(wq->mayday_mask);
2915 EXPORT_SYMBOL_GPL(destroy_workqueue);
2918 * workqueue_set_max_active - adjust max_active of a workqueue
2919 * @wq: target workqueue
2920 * @max_active: new max_active value.
2922 * Set max_active of @wq to @max_active.
2925 * Don't call from IRQ context.
2927 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
2931 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
2933 spin_lock(&workqueue_lock);
2935 wq->saved_max_active = max_active;
2937 for_each_cwq_cpu(cpu, wq) {
2938 struct global_cwq *gcwq = get_gcwq(cpu);
2940 spin_lock_irq(&gcwq->lock);
2942 if (!(wq->flags & WQ_FREEZEABLE) ||
2943 !(gcwq->flags & GCWQ_FREEZING))
2944 get_cwq(gcwq->cpu, wq)->max_active = max_active;
2946 spin_unlock_irq(&gcwq->lock);
2949 spin_unlock(&workqueue_lock);
2951 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
2954 * workqueue_congested - test whether a workqueue is congested
2955 * @cpu: CPU in question
2956 * @wq: target workqueue
2958 * Test whether @wq's cpu workqueue for @cpu is congested. There is
2959 * no synchronization around this function and the test result is
2960 * unreliable and only useful as advisory hints or for debugging.
2963 * %true if congested, %false otherwise.
2965 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
2967 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2969 return !list_empty(&cwq->delayed_works);
2971 EXPORT_SYMBOL_GPL(workqueue_congested);
2974 * work_cpu - return the last known associated cpu for @work
2975 * @work: the work of interest
2978 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
2980 unsigned int work_cpu(struct work_struct *work)
2982 struct global_cwq *gcwq = get_work_gcwq(work);
2984 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
2986 EXPORT_SYMBOL_GPL(work_cpu);
2989 * work_busy - test whether a work is currently pending or running
2990 * @work: the work to be tested
2992 * Test whether @work is currently pending or running. There is no
2993 * synchronization around this function and the test result is
2994 * unreliable and only useful as advisory hints or for debugging.
2995 * Especially for reentrant wqs, the pending state might hide the
2999 * OR'd bitmask of WORK_BUSY_* bits.
3001 unsigned int work_busy(struct work_struct *work)
3003 struct global_cwq *gcwq = get_work_gcwq(work);
3004 unsigned long flags;
3005 unsigned int ret = 0;
3010 spin_lock_irqsave(&gcwq->lock, flags);
3012 if (work_pending(work))
3013 ret |= WORK_BUSY_PENDING;
3014 if (find_worker_executing_work(gcwq, work))
3015 ret |= WORK_BUSY_RUNNING;
3017 spin_unlock_irqrestore(&gcwq->lock, flags);
3021 EXPORT_SYMBOL_GPL(work_busy);
3026 * There are two challenges in supporting CPU hotplug. Firstly, there
3027 * are a lot of assumptions on strong associations among work, cwq and
3028 * gcwq which make migrating pending and scheduled works very
3029 * difficult to implement without impacting hot paths. Secondly,
3030 * gcwqs serve mix of short, long and very long running works making
3031 * blocked draining impractical.
3033 * This is solved by allowing a gcwq to be detached from CPU, running
3034 * it with unbound (rogue) workers and allowing it to be reattached
3035 * later if the cpu comes back online. A separate thread is created
3036 * to govern a gcwq in such state and is called the trustee of the
3039 * Trustee states and their descriptions.
3041 * START Command state used on startup. On CPU_DOWN_PREPARE, a
3042 * new trustee is started with this state.
3044 * IN_CHARGE Once started, trustee will enter this state after
3045 * assuming the manager role and making all existing
3046 * workers rogue. DOWN_PREPARE waits for trustee to
3047 * enter this state. After reaching IN_CHARGE, trustee
3048 * tries to execute the pending worklist until it's empty
3049 * and the state is set to BUTCHER, or the state is set
3052 * BUTCHER Command state which is set by the cpu callback after
3053 * the cpu has went down. Once this state is set trustee
3054 * knows that there will be no new works on the worklist
3055 * and once the worklist is empty it can proceed to
3056 * killing idle workers.
3058 * RELEASE Command state which is set by the cpu callback if the
3059 * cpu down has been canceled or it has come online
3060 * again. After recognizing this state, trustee stops
3061 * trying to drain or butcher and clears ROGUE, rebinds
3062 * all remaining workers back to the cpu and releases
3065 * DONE Trustee will enter this state after BUTCHER or RELEASE
3068 * trustee CPU draining
3069 * took over down complete
3070 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3072 * | CPU is back online v return workers |
3073 * ----------------> RELEASE --------------
3077 * trustee_wait_event_timeout - timed event wait for trustee
3078 * @cond: condition to wait for
3079 * @timeout: timeout in jiffies
3081 * wait_event_timeout() for trustee to use. Handles locking and
3082 * checks for RELEASE request.
3085 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3086 * multiple times. To be used by trustee.
3089 * Positive indicating left time if @cond is satisfied, 0 if timed
3090 * out, -1 if canceled.
3092 #define trustee_wait_event_timeout(cond, timeout) ({ \
3093 long __ret = (timeout); \
3094 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3096 spin_unlock_irq(&gcwq->lock); \
3097 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3098 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3100 spin_lock_irq(&gcwq->lock); \
3102 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3106 * trustee_wait_event - event wait for trustee
3107 * @cond: condition to wait for
3109 * wait_event() for trustee to use. Automatically handles locking and
3110 * checks for CANCEL request.
3113 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3114 * multiple times. To be used by trustee.
3117 * 0 if @cond is satisfied, -1 if canceled.
3119 #define trustee_wait_event(cond) ({ \
3121 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3122 __ret1 < 0 ? -1 : 0; \
3125 static int __cpuinit trustee_thread(void *__gcwq)
3127 struct global_cwq *gcwq = __gcwq;
3128 struct worker *worker;
3129 struct work_struct *work;
3130 struct hlist_node *pos;
3134 BUG_ON(gcwq->cpu != smp_processor_id());
3136 spin_lock_irq(&gcwq->lock);
3138 * Claim the manager position and make all workers rogue.
3139 * Trustee must be bound to the target cpu and can't be
3142 BUG_ON(gcwq->cpu != smp_processor_id());
3143 rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3146 gcwq->flags |= GCWQ_MANAGING_WORKERS;
3148 list_for_each_entry(worker, &gcwq->idle_list, entry)
3149 worker->flags |= WORKER_ROGUE;
3151 for_each_busy_worker(worker, i, pos, gcwq)
3152 worker->flags |= WORKER_ROGUE;
3155 * Call schedule() so that we cross rq->lock and thus can
3156 * guarantee sched callbacks see the rogue flag. This is
3157 * necessary as scheduler callbacks may be invoked from other
3160 spin_unlock_irq(&gcwq->lock);
3162 spin_lock_irq(&gcwq->lock);
3165 * Sched callbacks are disabled now. Zap nr_running. After
3166 * this, nr_running stays zero and need_more_worker() and
3167 * keep_working() are always true as long as the worklist is
3170 atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3172 spin_unlock_irq(&gcwq->lock);
3173 del_timer_sync(&gcwq->idle_timer);
3174 spin_lock_irq(&gcwq->lock);
3177 * We're now in charge. Notify and proceed to drain. We need
3178 * to keep the gcwq running during the whole CPU down
3179 * procedure as other cpu hotunplug callbacks may need to
3180 * flush currently running tasks.
3182 gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3183 wake_up_all(&gcwq->trustee_wait);
3186 * The original cpu is in the process of dying and may go away
3187 * anytime now. When that happens, we and all workers would
3188 * be migrated to other cpus. Try draining any left work. We
3189 * want to get it over with ASAP - spam rescuers, wake up as
3190 * many idlers as necessary and create new ones till the
3191 * worklist is empty. Note that if the gcwq is frozen, there
3192 * may be frozen works in freezeable cwqs. Don't declare
3193 * completion while frozen.
3195 while (gcwq->nr_workers != gcwq->nr_idle ||
3196 gcwq->flags & GCWQ_FREEZING ||
3197 gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3200 list_for_each_entry(work, &gcwq->worklist, entry) {
3205 list_for_each_entry(worker, &gcwq->idle_list, entry) {
3208 wake_up_process(worker->task);
3211 if (need_to_create_worker(gcwq)) {
3212 spin_unlock_irq(&gcwq->lock);
3213 worker = create_worker(gcwq, false);
3214 spin_lock_irq(&gcwq->lock);
3216 worker->flags |= WORKER_ROGUE;
3217 start_worker(worker);
3221 /* give a breather */
3222 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3227 * Either all works have been scheduled and cpu is down, or
3228 * cpu down has already been canceled. Wait for and butcher
3229 * all workers till we're canceled.
3232 rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3233 while (!list_empty(&gcwq->idle_list))
3234 destroy_worker(list_first_entry(&gcwq->idle_list,
3235 struct worker, entry));
3236 } while (gcwq->nr_workers && rc >= 0);
3239 * At this point, either draining has completed and no worker
3240 * is left, or cpu down has been canceled or the cpu is being
3241 * brought back up. There shouldn't be any idle one left.
3242 * Tell the remaining busy ones to rebind once it finishes the
3243 * currently scheduled works by scheduling the rebind_work.
3245 WARN_ON(!list_empty(&gcwq->idle_list));
3247 for_each_busy_worker(worker, i, pos, gcwq) {
3248 struct work_struct *rebind_work = &worker->rebind_work;
3251 * Rebind_work may race with future cpu hotplug
3252 * operations. Use a separate flag to mark that
3253 * rebinding is scheduled.
3255 worker->flags |= WORKER_REBIND;
3256 worker->flags &= ~WORKER_ROGUE;
3258 /* queue rebind_work, wq doesn't matter, use the default one */
3259 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3260 work_data_bits(rebind_work)))
3263 debug_work_activate(rebind_work);
3264 insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3265 worker->scheduled.next,
3266 work_color_to_flags(WORK_NO_COLOR));
3269 /* relinquish manager role */
3270 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3272 /* notify completion */
3273 gcwq->trustee = NULL;
3274 gcwq->trustee_state = TRUSTEE_DONE;
3275 wake_up_all(&gcwq->trustee_wait);
3276 spin_unlock_irq(&gcwq->lock);
3281 * wait_trustee_state - wait for trustee to enter the specified state
3282 * @gcwq: gcwq the trustee of interest belongs to
3283 * @state: target state to wait for
3285 * Wait for the trustee to reach @state. DONE is already matched.
3288 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3289 * multiple times. To be used by cpu_callback.
3291 static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3292 __releases(&gcwq->lock)
3293 __acquires(&gcwq->lock)
3295 if (!(gcwq->trustee_state == state ||
3296 gcwq->trustee_state == TRUSTEE_DONE)) {
3297 spin_unlock_irq(&gcwq->lock);
3298 __wait_event(gcwq->trustee_wait,
3299 gcwq->trustee_state == state ||
3300 gcwq->trustee_state == TRUSTEE_DONE);
3301 spin_lock_irq(&gcwq->lock);
3305 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3306 unsigned long action,
3309 unsigned int cpu = (unsigned long)hcpu;
3310 struct global_cwq *gcwq = get_gcwq(cpu);
3311 struct task_struct *new_trustee = NULL;
3312 struct worker *uninitialized_var(new_worker);
3313 unsigned long flags;
3315 action &= ~CPU_TASKS_FROZEN;
3318 case CPU_DOWN_PREPARE:
3319 new_trustee = kthread_create(trustee_thread, gcwq,
3320 "workqueue_trustee/%d\n", cpu);
3321 if (IS_ERR(new_trustee))
3322 return notifier_from_errno(PTR_ERR(new_trustee));
3323 kthread_bind(new_trustee, cpu);
3325 case CPU_UP_PREPARE:
3326 BUG_ON(gcwq->first_idle);
3327 new_worker = create_worker(gcwq, false);
3330 kthread_stop(new_trustee);
3335 /* some are called w/ irq disabled, don't disturb irq status */
3336 spin_lock_irqsave(&gcwq->lock, flags);
3339 case CPU_DOWN_PREPARE:
3340 /* initialize trustee and tell it to acquire the gcwq */
3341 BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
3342 gcwq->trustee = new_trustee;
3343 gcwq->trustee_state = TRUSTEE_START;
3344 wake_up_process(gcwq->trustee);
3345 wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3347 case CPU_UP_PREPARE:
3348 BUG_ON(gcwq->first_idle);
3349 gcwq->first_idle = new_worker;
3354 * Before this, the trustee and all workers except for
3355 * the ones which are still executing works from
3356 * before the last CPU down must be on the cpu. After
3357 * this, they'll all be diasporas.
3359 gcwq->flags |= GCWQ_DISASSOCIATED;
3363 gcwq->trustee_state = TRUSTEE_BUTCHER;
3365 case CPU_UP_CANCELED:
3366 destroy_worker(gcwq->first_idle);
3367 gcwq->first_idle = NULL;
3370 case CPU_DOWN_FAILED:
3372 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3373 if (gcwq->trustee_state != TRUSTEE_DONE) {
3374 gcwq->trustee_state = TRUSTEE_RELEASE;
3375 wake_up_process(gcwq->trustee);
3376 wait_trustee_state(gcwq, TRUSTEE_DONE);
3380 * Trustee is done and there might be no worker left.
3381 * Put the first_idle in and request a real manager to
3384 spin_unlock_irq(&gcwq->lock);
3385 kthread_bind(gcwq->first_idle->task, cpu);
3386 spin_lock_irq(&gcwq->lock);
3387 gcwq->flags |= GCWQ_MANAGE_WORKERS;
3388 start_worker(gcwq->first_idle);
3389 gcwq->first_idle = NULL;
3393 spin_unlock_irqrestore(&gcwq->lock, flags);
3395 return notifier_from_errno(0);
3400 struct work_for_cpu {
3401 struct completion completion;
3407 static int do_work_for_cpu(void *_wfc)
3409 struct work_for_cpu *wfc = _wfc;
3410 wfc->ret = wfc->fn(wfc->arg);
3411 complete(&wfc->completion);
3416 * work_on_cpu - run a function in user context on a particular cpu
3417 * @cpu: the cpu to run on
3418 * @fn: the function to run
3419 * @arg: the function arg
3421 * This will return the value @fn returns.
3422 * It is up to the caller to ensure that the cpu doesn't go offline.
3423 * The caller must not hold any locks which would prevent @fn from completing.
3425 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3427 struct task_struct *sub_thread;
3428 struct work_for_cpu wfc = {
3429 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3434 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3435 if (IS_ERR(sub_thread))
3436 return PTR_ERR(sub_thread);
3437 kthread_bind(sub_thread, cpu);
3438 wake_up_process(sub_thread);
3439 wait_for_completion(&wfc.completion);
3442 EXPORT_SYMBOL_GPL(work_on_cpu);
3443 #endif /* CONFIG_SMP */
3445 #ifdef CONFIG_FREEZER
3448 * freeze_workqueues_begin - begin freezing workqueues
3450 * Start freezing workqueues. After this function returns, all
3451 * freezeable workqueues will queue new works to their frozen_works
3452 * list instead of gcwq->worklist.
3455 * Grabs and releases workqueue_lock and gcwq->lock's.
3457 void freeze_workqueues_begin(void)
3461 spin_lock(&workqueue_lock);
3463 BUG_ON(workqueue_freezing);
3464 workqueue_freezing = true;
3466 for_each_gcwq_cpu(cpu) {
3467 struct global_cwq *gcwq = get_gcwq(cpu);
3468 struct workqueue_struct *wq;
3470 spin_lock_irq(&gcwq->lock);
3472 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3473 gcwq->flags |= GCWQ_FREEZING;
3475 list_for_each_entry(wq, &workqueues, list) {
3476 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3478 if (cwq && wq->flags & WQ_FREEZEABLE)
3479 cwq->max_active = 0;
3482 spin_unlock_irq(&gcwq->lock);
3485 spin_unlock(&workqueue_lock);
3489 * freeze_workqueues_busy - are freezeable workqueues still busy?
3491 * Check whether freezing is complete. This function must be called
3492 * between freeze_workqueues_begin() and thaw_workqueues().
3495 * Grabs and releases workqueue_lock.
3498 * %true if some freezeable workqueues are still busy. %false if
3499 * freezing is complete.
3501 bool freeze_workqueues_busy(void)
3506 spin_lock(&workqueue_lock);
3508 BUG_ON(!workqueue_freezing);
3510 for_each_gcwq_cpu(cpu) {
3511 struct workqueue_struct *wq;
3513 * nr_active is monotonically decreasing. It's safe
3514 * to peek without lock.
3516 list_for_each_entry(wq, &workqueues, list) {
3517 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3519 if (!cwq || !(wq->flags & WQ_FREEZEABLE))
3522 BUG_ON(cwq->nr_active < 0);
3523 if (cwq->nr_active) {
3530 spin_unlock(&workqueue_lock);
3535 * thaw_workqueues - thaw workqueues
3537 * Thaw workqueues. Normal queueing is restored and all collected
3538 * frozen works are transferred to their respective gcwq worklists.
3541 * Grabs and releases workqueue_lock and gcwq->lock's.
3543 void thaw_workqueues(void)
3547 spin_lock(&workqueue_lock);
3549 if (!workqueue_freezing)
3552 for_each_gcwq_cpu(cpu) {
3553 struct global_cwq *gcwq = get_gcwq(cpu);
3554 struct workqueue_struct *wq;
3556 spin_lock_irq(&gcwq->lock);
3558 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3559 gcwq->flags &= ~GCWQ_FREEZING;
3561 list_for_each_entry(wq, &workqueues, list) {
3562 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3564 if (!cwq || !(wq->flags & WQ_FREEZEABLE))
3567 /* restore max_active and repopulate worklist */
3568 cwq->max_active = wq->saved_max_active;
3570 while (!list_empty(&cwq->delayed_works) &&
3571 cwq->nr_active < cwq->max_active)
3572 cwq_activate_first_delayed(cwq);
3575 wake_up_worker(gcwq);
3577 spin_unlock_irq(&gcwq->lock);
3580 workqueue_freezing = false;
3582 spin_unlock(&workqueue_lock);
3584 #endif /* CONFIG_FREEZER */
3586 static int __init init_workqueues(void)
3591 cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
3593 /* initialize gcwqs */
3594 for_each_gcwq_cpu(cpu) {
3595 struct global_cwq *gcwq = get_gcwq(cpu);
3597 spin_lock_init(&gcwq->lock);
3598 INIT_LIST_HEAD(&gcwq->worklist);
3600 gcwq->flags |= GCWQ_DISASSOCIATED;
3602 INIT_LIST_HEAD(&gcwq->idle_list);
3603 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3604 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3606 init_timer_deferrable(&gcwq->idle_timer);
3607 gcwq->idle_timer.function = idle_worker_timeout;
3608 gcwq->idle_timer.data = (unsigned long)gcwq;
3610 setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3611 (unsigned long)gcwq);
3613 ida_init(&gcwq->worker_ida);
3615 gcwq->trustee_state = TRUSTEE_DONE;
3616 init_waitqueue_head(&gcwq->trustee_wait);
3619 /* create the initial worker */
3620 for_each_online_gcwq_cpu(cpu) {
3621 struct global_cwq *gcwq = get_gcwq(cpu);
3622 struct worker *worker;
3624 if (cpu != WORK_CPU_UNBOUND)
3625 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3626 worker = create_worker(gcwq, true);
3628 spin_lock_irq(&gcwq->lock);
3629 start_worker(worker);
3630 spin_unlock_irq(&gcwq->lock);
3633 system_wq = alloc_workqueue("events", 0, 0);
3634 system_long_wq = alloc_workqueue("events_long", 0, 0);
3635 system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3636 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3637 WQ_UNBOUND_MAX_ACTIVE);
3638 BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq);
3641 early_initcall(init_workqueues);