SECURITY: Move exec_permission RCU checks into security modules
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / kernel / workqueue.c
blob8859a41806dddae875e1ceae1e6ed94a4e3aebff
1 /*
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>
8 * Andrew Morton
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/module.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
45 #include "workqueue_sched.h"
47 enum {
48 /* global_cwq flags */
49 GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
50 GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */
51 GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
52 GCWQ_FREEZING = 1 << 3, /* freeze in progress */
53 GCWQ_HIGHPRI_PENDING = 1 << 4, /* highpri works on queue */
55 /* worker flags */
56 WORKER_STARTED = 1 << 0, /* started */
57 WORKER_DIE = 1 << 1, /* die die die */
58 WORKER_IDLE = 1 << 2, /* is idle */
59 WORKER_PREP = 1 << 3, /* preparing to run works */
60 WORKER_ROGUE = 1 << 4, /* not bound to any cpu */
61 WORKER_REBIND = 1 << 5, /* mom is home, come back */
62 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
63 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
65 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND |
66 WORKER_CPU_INTENSIVE | WORKER_UNBOUND,
68 /* gcwq->trustee_state */
69 TRUSTEE_START = 0, /* start */
70 TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */
71 TRUSTEE_BUTCHER = 2, /* butcher workers */
72 TRUSTEE_RELEASE = 3, /* release workers */
73 TRUSTEE_DONE = 4, /* trustee is done */
75 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
76 BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
77 BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
79 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
80 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
82 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
83 /* call for help after 10ms
84 (min two ticks) */
85 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
86 CREATE_COOLDOWN = HZ, /* time to breath after fail */
87 TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */
90 * Rescue workers are used only on emergencies and shared by
91 * all cpus. Give -20.
93 RESCUER_NICE_LEVEL = -20,
97 * Structure fields follow one of the following exclusion rules.
99 * I: Modifiable by initialization/destruction paths and read-only for
100 * everyone else.
102 * P: Preemption protected. Disabling preemption is enough and should
103 * only be modified and accessed from the local cpu.
105 * L: gcwq->lock protected. Access with gcwq->lock held.
107 * X: During normal operation, modification requires gcwq->lock and
108 * should be done only from local cpu. Either disabling preemption
109 * on local cpu or grabbing gcwq->lock is enough for read access.
110 * If GCWQ_DISASSOCIATED is set, it's identical to L.
112 * F: wq->flush_mutex protected.
114 * W: workqueue_lock protected.
117 struct global_cwq;
120 * The poor guys doing the actual heavy lifting. All on-duty workers
121 * are either serving the manager role, on idle list or on busy hash.
123 struct worker {
124 /* on idle list while idle, on busy hash table while busy */
125 union {
126 struct list_head entry; /* L: while idle */
127 struct hlist_node hentry; /* L: while busy */
130 struct work_struct *current_work; /* L: work being processed */
131 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
132 struct list_head scheduled; /* L: scheduled works */
133 struct task_struct *task; /* I: worker task */
134 struct global_cwq *gcwq; /* I: the associated gcwq */
135 /* 64 bytes boundary on 64bit, 32 on 32bit */
136 unsigned long last_active; /* L: last active timestamp */
137 unsigned int flags; /* X: flags */
138 int id; /* I: worker id */
139 struct work_struct rebind_work; /* L: rebind worker to cpu */
143 * Global per-cpu workqueue. There's one and only one for each cpu
144 * and all works are queued and processed here regardless of their
145 * target workqueues.
147 struct global_cwq {
148 spinlock_t lock; /* the gcwq lock */
149 struct list_head worklist; /* L: list of pending works */
150 unsigned int cpu; /* I: the associated cpu */
151 unsigned int flags; /* L: GCWQ_* flags */
153 int nr_workers; /* L: total number of workers */
154 int nr_idle; /* L: currently idle ones */
156 /* workers are chained either in the idle_list or busy_hash */
157 struct list_head idle_list; /* X: list of idle workers */
158 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
159 /* L: hash of busy workers */
161 struct timer_list idle_timer; /* L: worker idle timeout */
162 struct timer_list mayday_timer; /* L: SOS timer for dworkers */
164 struct ida worker_ida; /* L: for worker IDs */
166 struct task_struct *trustee; /* L: for gcwq shutdown */
167 unsigned int trustee_state; /* L: trustee state */
168 wait_queue_head_t trustee_wait; /* trustee wait */
169 struct worker *first_idle; /* L: first idle worker */
170 } ____cacheline_aligned_in_smp;
173 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
174 * work_struct->data are used for flags and thus cwqs need to be
175 * aligned at two's power of the number of flag bits.
177 struct cpu_workqueue_struct {
178 struct global_cwq *gcwq; /* I: the associated gcwq */
179 struct workqueue_struct *wq; /* I: the owning workqueue */
180 int work_color; /* L: current color */
181 int flush_color; /* L: flushing color */
182 int nr_in_flight[WORK_NR_COLORS];
183 /* L: nr of in_flight works */
184 int nr_active; /* L: nr of active works */
185 int max_active; /* L: max active works */
186 struct list_head delayed_works; /* L: delayed works */
190 * Structure used to wait for workqueue flush.
192 struct wq_flusher {
193 struct list_head list; /* F: list of flushers */
194 int flush_color; /* F: flush color waiting for */
195 struct completion done; /* flush completion */
199 * All cpumasks are assumed to be always set on UP and thus can't be
200 * used to determine whether there's something to be done.
202 #ifdef CONFIG_SMP
203 typedef cpumask_var_t mayday_mask_t;
204 #define mayday_test_and_set_cpu(cpu, mask) \
205 cpumask_test_and_set_cpu((cpu), (mask))
206 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
207 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
208 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
209 #define free_mayday_mask(mask) free_cpumask_var((mask))
210 #else
211 typedef unsigned long mayday_mask_t;
212 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
213 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
214 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
215 #define alloc_mayday_mask(maskp, gfp) true
216 #define free_mayday_mask(mask) do { } while (0)
217 #endif
220 * The externally visible workqueue abstraction is an array of
221 * per-CPU workqueues:
223 struct workqueue_struct {
224 unsigned int flags; /* I: WQ_* flags */
225 union {
226 struct cpu_workqueue_struct __percpu *pcpu;
227 struct cpu_workqueue_struct *single;
228 unsigned long v;
229 } cpu_wq; /* I: cwq's */
230 struct list_head list; /* W: list of all workqueues */
232 struct mutex flush_mutex; /* protects wq flushing */
233 int work_color; /* F: current work color */
234 int flush_color; /* F: current flush color */
235 atomic_t nr_cwqs_to_flush; /* flush in progress */
236 struct wq_flusher *first_flusher; /* F: first flusher */
237 struct list_head flusher_queue; /* F: flush waiters */
238 struct list_head flusher_overflow; /* F: flush overflow list */
240 mayday_mask_t mayday_mask; /* cpus requesting rescue */
241 struct worker *rescuer; /* I: rescue worker */
243 int saved_max_active; /* W: saved cwq max_active */
244 const char *name; /* I: workqueue name */
245 #ifdef CONFIG_LOCKDEP
246 struct lockdep_map lockdep_map;
247 #endif
250 struct workqueue_struct *system_wq __read_mostly;
251 struct workqueue_struct *system_long_wq __read_mostly;
252 struct workqueue_struct *system_nrt_wq __read_mostly;
253 struct workqueue_struct *system_unbound_wq __read_mostly;
254 struct workqueue_struct *system_freezable_wq __read_mostly;
255 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);
259 EXPORT_SYMBOL_GPL(system_freezable_wq);
261 #define CREATE_TRACE_POINTS
262 #include <trace/events/workqueue.h>
264 #define for_each_busy_worker(worker, i, pos, gcwq) \
265 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
266 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
268 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
269 unsigned int sw)
271 if (cpu < nr_cpu_ids) {
272 if (sw & 1) {
273 cpu = cpumask_next(cpu, mask);
274 if (cpu < nr_cpu_ids)
275 return cpu;
277 if (sw & 2)
278 return WORK_CPU_UNBOUND;
280 return WORK_CPU_NONE;
283 static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
284 struct workqueue_struct *wq)
286 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
290 * CPU iterators
292 * An extra gcwq is defined for an invalid cpu number
293 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
294 * specific CPU. The following iterators are similar to
295 * for_each_*_cpu() iterators but also considers the unbound gcwq.
297 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
298 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
299 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
300 * WORK_CPU_UNBOUND for unbound workqueues
302 #define for_each_gcwq_cpu(cpu) \
303 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
304 (cpu) < WORK_CPU_NONE; \
305 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
307 #define for_each_online_gcwq_cpu(cpu) \
308 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
309 (cpu) < WORK_CPU_NONE; \
310 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
312 #define for_each_cwq_cpu(cpu, wq) \
313 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
314 (cpu) < WORK_CPU_NONE; \
315 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
317 #ifdef CONFIG_DEBUG_OBJECTS_WORK
319 static struct debug_obj_descr work_debug_descr;
321 static void *work_debug_hint(void *addr)
323 return ((struct work_struct *) addr)->func;
327 * fixup_init is called when:
328 * - an active object is initialized
330 static int work_fixup_init(void *addr, enum debug_obj_state state)
332 struct work_struct *work = addr;
334 switch (state) {
335 case ODEBUG_STATE_ACTIVE:
336 cancel_work_sync(work);
337 debug_object_init(work, &work_debug_descr);
338 return 1;
339 default:
340 return 0;
345 * fixup_activate is called when:
346 * - an active object is activated
347 * - an unknown object is activated (might be a statically initialized object)
349 static int work_fixup_activate(void *addr, enum debug_obj_state state)
351 struct work_struct *work = addr;
353 switch (state) {
355 case ODEBUG_STATE_NOTAVAILABLE:
357 * This is not really a fixup. The work struct was
358 * statically initialized. We just make sure that it
359 * is tracked in the object tracker.
361 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
362 debug_object_init(work, &work_debug_descr);
363 debug_object_activate(work, &work_debug_descr);
364 return 0;
366 WARN_ON_ONCE(1);
367 return 0;
369 case ODEBUG_STATE_ACTIVE:
370 WARN_ON(1);
372 default:
373 return 0;
378 * fixup_free is called when:
379 * - an active object is freed
381 static int work_fixup_free(void *addr, enum debug_obj_state state)
383 struct work_struct *work = addr;
385 switch (state) {
386 case ODEBUG_STATE_ACTIVE:
387 cancel_work_sync(work);
388 debug_object_free(work, &work_debug_descr);
389 return 1;
390 default:
391 return 0;
395 static struct debug_obj_descr work_debug_descr = {
396 .name = "work_struct",
397 .debug_hint = work_debug_hint,
398 .fixup_init = work_fixup_init,
399 .fixup_activate = work_fixup_activate,
400 .fixup_free = work_fixup_free,
403 static inline void debug_work_activate(struct work_struct *work)
405 debug_object_activate(work, &work_debug_descr);
408 static inline void debug_work_deactivate(struct work_struct *work)
410 debug_object_deactivate(work, &work_debug_descr);
413 void __init_work(struct work_struct *work, int onstack)
415 if (onstack)
416 debug_object_init_on_stack(work, &work_debug_descr);
417 else
418 debug_object_init(work, &work_debug_descr);
420 EXPORT_SYMBOL_GPL(__init_work);
422 void destroy_work_on_stack(struct work_struct *work)
424 debug_object_free(work, &work_debug_descr);
426 EXPORT_SYMBOL_GPL(destroy_work_on_stack);
428 #else
429 static inline void debug_work_activate(struct work_struct *work) { }
430 static inline void debug_work_deactivate(struct work_struct *work) { }
431 #endif
433 /* Serializes the accesses to the list of workqueues. */
434 static DEFINE_SPINLOCK(workqueue_lock);
435 static LIST_HEAD(workqueues);
436 static bool workqueue_freezing; /* W: have wqs started freezing? */
439 * The almighty global cpu workqueues. nr_running is the only field
440 * which is expected to be used frequently by other cpus via
441 * try_to_wake_up(). Put it in a separate cacheline.
443 static DEFINE_PER_CPU(struct global_cwq, global_cwq);
444 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
447 * Global cpu workqueue and nr_running counter for unbound gcwq. The
448 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
449 * workers have WORKER_UNBOUND set.
451 static struct global_cwq unbound_global_cwq;
452 static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */
454 static int worker_thread(void *__worker);
456 static struct global_cwq *get_gcwq(unsigned int cpu)
458 if (cpu != WORK_CPU_UNBOUND)
459 return &per_cpu(global_cwq, cpu);
460 else
461 return &unbound_global_cwq;
464 static atomic_t *get_gcwq_nr_running(unsigned int cpu)
466 if (cpu != WORK_CPU_UNBOUND)
467 return &per_cpu(gcwq_nr_running, cpu);
468 else
469 return &unbound_gcwq_nr_running;
472 static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
473 struct workqueue_struct *wq)
475 if (!(wq->flags & WQ_UNBOUND)) {
476 if (likely(cpu < nr_cpu_ids)) {
477 #ifdef CONFIG_SMP
478 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
479 #else
480 return wq->cpu_wq.single;
481 #endif
483 } else if (likely(cpu == WORK_CPU_UNBOUND))
484 return wq->cpu_wq.single;
485 return NULL;
488 static unsigned int work_color_to_flags(int color)
490 return color << WORK_STRUCT_COLOR_SHIFT;
493 static int get_work_color(struct work_struct *work)
495 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
496 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
499 static int work_next_color(int color)
501 return (color + 1) % WORK_NR_COLORS;
505 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
506 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
507 * cleared and the work data contains the cpu number it was last on.
509 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
510 * cwq, cpu or clear work->data. These functions should only be
511 * called while the work is owned - ie. while the PENDING bit is set.
513 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
514 * corresponding to a work. gcwq is available once the work has been
515 * queued anywhere after initialization. cwq is available only from
516 * queueing until execution starts.
518 static inline void set_work_data(struct work_struct *work, unsigned long data,
519 unsigned long flags)
521 BUG_ON(!work_pending(work));
522 atomic_long_set(&work->data, data | flags | work_static(work));
525 static void set_work_cwq(struct work_struct *work,
526 struct cpu_workqueue_struct *cwq,
527 unsigned long extra_flags)
529 set_work_data(work, (unsigned long)cwq,
530 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
533 static void set_work_cpu(struct work_struct *work, unsigned int cpu)
535 set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
538 static void clear_work_data(struct work_struct *work)
540 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
543 static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
545 unsigned long data = atomic_long_read(&work->data);
547 if (data & WORK_STRUCT_CWQ)
548 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
549 else
550 return NULL;
553 static struct global_cwq *get_work_gcwq(struct work_struct *work)
555 unsigned long data = atomic_long_read(&work->data);
556 unsigned int cpu;
558 if (data & WORK_STRUCT_CWQ)
559 return ((struct cpu_workqueue_struct *)
560 (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
562 cpu = data >> WORK_STRUCT_FLAG_BITS;
563 if (cpu == WORK_CPU_NONE)
564 return NULL;
566 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
567 return get_gcwq(cpu);
571 * Policy functions. These define the policies on how the global
572 * worker pool is managed. Unless noted otherwise, these functions
573 * assume that they're being called with gcwq->lock held.
576 static bool __need_more_worker(struct global_cwq *gcwq)
578 return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) ||
579 gcwq->flags & GCWQ_HIGHPRI_PENDING;
583 * Need to wake up a worker? Called from anything but currently
584 * running workers.
586 static bool need_more_worker(struct global_cwq *gcwq)
588 return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
591 /* Can I start working? Called from busy but !running workers. */
592 static bool may_start_working(struct global_cwq *gcwq)
594 return gcwq->nr_idle;
597 /* Do I need to keep working? Called from currently running workers. */
598 static bool keep_working(struct global_cwq *gcwq)
600 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
602 return !list_empty(&gcwq->worklist) &&
603 (atomic_read(nr_running) <= 1 ||
604 gcwq->flags & GCWQ_HIGHPRI_PENDING);
607 /* Do we need a new worker? Called from manager. */
608 static bool need_to_create_worker(struct global_cwq *gcwq)
610 return need_more_worker(gcwq) && !may_start_working(gcwq);
613 /* Do I need to be the manager? */
614 static bool need_to_manage_workers(struct global_cwq *gcwq)
616 return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS;
619 /* Do we have too many workers and should some go away? */
620 static bool too_many_workers(struct global_cwq *gcwq)
622 bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
623 int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
624 int nr_busy = gcwq->nr_workers - nr_idle;
626 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
630 * Wake up functions.
633 /* Return the first worker. Safe with preemption disabled */
634 static struct worker *first_worker(struct global_cwq *gcwq)
636 if (unlikely(list_empty(&gcwq->idle_list)))
637 return NULL;
639 return list_first_entry(&gcwq->idle_list, struct worker, entry);
643 * wake_up_worker - wake up an idle worker
644 * @gcwq: gcwq to wake worker for
646 * Wake up the first idle worker of @gcwq.
648 * CONTEXT:
649 * spin_lock_irq(gcwq->lock).
651 static void wake_up_worker(struct global_cwq *gcwq)
653 struct worker *worker = first_worker(gcwq);
655 if (likely(worker))
656 wake_up_process(worker->task);
660 * wq_worker_waking_up - a worker is waking up
661 * @task: task waking up
662 * @cpu: CPU @task is waking up to
664 * This function is called during try_to_wake_up() when a worker is
665 * being awoken.
667 * CONTEXT:
668 * spin_lock_irq(rq->lock)
670 void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
672 struct worker *worker = kthread_data(task);
674 if (!(worker->flags & WORKER_NOT_RUNNING))
675 atomic_inc(get_gcwq_nr_running(cpu));
679 * wq_worker_sleeping - a worker is going to sleep
680 * @task: task going to sleep
681 * @cpu: CPU in question, must be the current CPU number
683 * This function is called during schedule() when a busy worker is
684 * going to sleep. Worker on the same cpu can be woken up by
685 * returning pointer to its task.
687 * CONTEXT:
688 * spin_lock_irq(rq->lock)
690 * RETURNS:
691 * Worker task on @cpu to wake up, %NULL if none.
693 struct task_struct *wq_worker_sleeping(struct task_struct *task,
694 unsigned int cpu)
696 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
697 struct global_cwq *gcwq = get_gcwq(cpu);
698 atomic_t *nr_running = get_gcwq_nr_running(cpu);
700 if (worker->flags & WORKER_NOT_RUNNING)
701 return NULL;
703 /* this can only happen on the local cpu */
704 BUG_ON(cpu != raw_smp_processor_id());
707 * The counterpart of the following dec_and_test, implied mb,
708 * worklist not empty test sequence is in insert_work().
709 * Please read comment there.
711 * NOT_RUNNING is clear. This means that trustee is not in
712 * charge and we're running on the local cpu w/ rq lock held
713 * and preemption disabled, which in turn means that none else
714 * could be manipulating idle_list, so dereferencing idle_list
715 * without gcwq lock is safe.
717 if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
718 to_wakeup = first_worker(gcwq);
719 return to_wakeup ? to_wakeup->task : NULL;
723 * worker_set_flags - set worker flags and adjust nr_running accordingly
724 * @worker: self
725 * @flags: flags to set
726 * @wakeup: wakeup an idle worker if necessary
728 * Set @flags in @worker->flags and adjust nr_running accordingly. If
729 * nr_running becomes zero and @wakeup is %true, an idle worker is
730 * woken up.
732 * CONTEXT:
733 * spin_lock_irq(gcwq->lock)
735 static inline void worker_set_flags(struct worker *worker, unsigned int flags,
736 bool wakeup)
738 struct global_cwq *gcwq = worker->gcwq;
740 WARN_ON_ONCE(worker->task != current);
743 * If transitioning into NOT_RUNNING, adjust nr_running and
744 * wake up an idle worker as necessary if requested by
745 * @wakeup.
747 if ((flags & WORKER_NOT_RUNNING) &&
748 !(worker->flags & WORKER_NOT_RUNNING)) {
749 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
751 if (wakeup) {
752 if (atomic_dec_and_test(nr_running) &&
753 !list_empty(&gcwq->worklist))
754 wake_up_worker(gcwq);
755 } else
756 atomic_dec(nr_running);
759 worker->flags |= flags;
763 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
764 * @worker: self
765 * @flags: flags to clear
767 * Clear @flags in @worker->flags and adjust nr_running accordingly.
769 * CONTEXT:
770 * spin_lock_irq(gcwq->lock)
772 static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
774 struct global_cwq *gcwq = worker->gcwq;
775 unsigned int oflags = worker->flags;
777 WARN_ON_ONCE(worker->task != current);
779 worker->flags &= ~flags;
782 * If transitioning out of NOT_RUNNING, increment nr_running. Note
783 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
784 * of multiple flags, not a single flag.
786 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
787 if (!(worker->flags & WORKER_NOT_RUNNING))
788 atomic_inc(get_gcwq_nr_running(gcwq->cpu));
792 * busy_worker_head - return the busy hash head for a work
793 * @gcwq: gcwq of interest
794 * @work: work to be hashed
796 * Return hash head of @gcwq for @work.
798 * CONTEXT:
799 * spin_lock_irq(gcwq->lock).
801 * RETURNS:
802 * Pointer to the hash head.
804 static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
805 struct work_struct *work)
807 const int base_shift = ilog2(sizeof(struct work_struct));
808 unsigned long v = (unsigned long)work;
810 /* simple shift and fold hash, do we need something better? */
811 v >>= base_shift;
812 v += v >> BUSY_WORKER_HASH_ORDER;
813 v &= BUSY_WORKER_HASH_MASK;
815 return &gcwq->busy_hash[v];
819 * __find_worker_executing_work - find worker which is executing a work
820 * @gcwq: gcwq of interest
821 * @bwh: hash head as returned by busy_worker_head()
822 * @work: work to find worker for
824 * Find a worker which is executing @work on @gcwq. @bwh should be
825 * the hash head obtained by calling busy_worker_head() with the same
826 * work.
828 * CONTEXT:
829 * spin_lock_irq(gcwq->lock).
831 * RETURNS:
832 * Pointer to worker which is executing @work if found, NULL
833 * otherwise.
835 static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
836 struct hlist_head *bwh,
837 struct work_struct *work)
839 struct worker *worker;
840 struct hlist_node *tmp;
842 hlist_for_each_entry(worker, tmp, bwh, hentry)
843 if (worker->current_work == work)
844 return worker;
845 return NULL;
849 * find_worker_executing_work - find worker which is executing a work
850 * @gcwq: gcwq of interest
851 * @work: work to find worker for
853 * Find a worker which is executing @work on @gcwq. This function is
854 * identical to __find_worker_executing_work() except that this
855 * function calculates @bwh itself.
857 * CONTEXT:
858 * spin_lock_irq(gcwq->lock).
860 * RETURNS:
861 * Pointer to worker which is executing @work if found, NULL
862 * otherwise.
864 static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
865 struct work_struct *work)
867 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
868 work);
872 * gcwq_determine_ins_pos - find insertion position
873 * @gcwq: gcwq of interest
874 * @cwq: cwq a work is being queued for
876 * A work for @cwq is about to be queued on @gcwq, determine insertion
877 * position for the work. If @cwq is for HIGHPRI wq, the work is
878 * queued at the head of the queue but in FIFO order with respect to
879 * other HIGHPRI works; otherwise, at the end of the queue. This
880 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
881 * there are HIGHPRI works pending.
883 * CONTEXT:
884 * spin_lock_irq(gcwq->lock).
886 * RETURNS:
887 * Pointer to inserstion position.
889 static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq,
890 struct cpu_workqueue_struct *cwq)
892 struct work_struct *twork;
894 if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
895 return &gcwq->worklist;
897 list_for_each_entry(twork, &gcwq->worklist, entry) {
898 struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
900 if (!(tcwq->wq->flags & WQ_HIGHPRI))
901 break;
904 gcwq->flags |= GCWQ_HIGHPRI_PENDING;
905 return &twork->entry;
909 * insert_work - insert a work into gcwq
910 * @cwq: cwq @work belongs to
911 * @work: work to insert
912 * @head: insertion point
913 * @extra_flags: extra WORK_STRUCT_* flags to set
915 * Insert @work which belongs to @cwq into @gcwq after @head.
916 * @extra_flags is or'd to work_struct flags.
918 * CONTEXT:
919 * spin_lock_irq(gcwq->lock).
921 static void insert_work(struct cpu_workqueue_struct *cwq,
922 struct work_struct *work, struct list_head *head,
923 unsigned int extra_flags)
925 struct global_cwq *gcwq = cwq->gcwq;
927 /* we own @work, set data and link */
928 set_work_cwq(work, cwq, extra_flags);
931 * Ensure that we get the right work->data if we see the
932 * result of list_add() below, see try_to_grab_pending().
934 smp_wmb();
936 list_add_tail(&work->entry, head);
939 * Ensure either worker_sched_deactivated() sees the above
940 * list_add_tail() or we see zero nr_running to avoid workers
941 * lying around lazily while there are works to be processed.
943 smp_mb();
945 if (__need_more_worker(gcwq))
946 wake_up_worker(gcwq);
950 * Test whether @work is being queued from another work executing on the
951 * same workqueue. This is rather expensive and should only be used from
952 * cold paths.
954 static bool is_chained_work(struct workqueue_struct *wq)
956 unsigned long flags;
957 unsigned int cpu;
959 for_each_gcwq_cpu(cpu) {
960 struct global_cwq *gcwq = get_gcwq(cpu);
961 struct worker *worker;
962 struct hlist_node *pos;
963 int i;
965 spin_lock_irqsave(&gcwq->lock, flags);
966 for_each_busy_worker(worker, i, pos, gcwq) {
967 if (worker->task != current)
968 continue;
969 spin_unlock_irqrestore(&gcwq->lock, flags);
971 * I'm @worker, no locking necessary. See if @work
972 * is headed to the same workqueue.
974 return worker->current_cwq->wq == wq;
976 spin_unlock_irqrestore(&gcwq->lock, flags);
978 return false;
981 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
982 struct work_struct *work)
984 struct global_cwq *gcwq;
985 struct cpu_workqueue_struct *cwq;
986 struct list_head *worklist;
987 unsigned int work_flags;
988 unsigned long flags;
990 debug_work_activate(work);
992 /* if dying, only works from the same workqueue are allowed */
993 if (unlikely(wq->flags & WQ_DYING) &&
994 WARN_ON_ONCE(!is_chained_work(wq)))
995 return;
997 /* determine gcwq to use */
998 if (!(wq->flags & WQ_UNBOUND)) {
999 struct global_cwq *last_gcwq;
1001 if (unlikely(cpu == WORK_CPU_UNBOUND))
1002 cpu = raw_smp_processor_id();
1005 * It's multi cpu. If @wq is non-reentrant and @work
1006 * was previously on a different cpu, it might still
1007 * be running there, in which case the work needs to
1008 * be queued on that cpu to guarantee non-reentrance.
1010 gcwq = get_gcwq(cpu);
1011 if (wq->flags & WQ_NON_REENTRANT &&
1012 (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
1013 struct worker *worker;
1015 spin_lock_irqsave(&last_gcwq->lock, flags);
1017 worker = find_worker_executing_work(last_gcwq, work);
1019 if (worker && worker->current_cwq->wq == wq)
1020 gcwq = last_gcwq;
1021 else {
1022 /* meh... not running there, queue here */
1023 spin_unlock_irqrestore(&last_gcwq->lock, flags);
1024 spin_lock_irqsave(&gcwq->lock, flags);
1026 } else
1027 spin_lock_irqsave(&gcwq->lock, flags);
1028 } else {
1029 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1030 spin_lock_irqsave(&gcwq->lock, flags);
1033 /* gcwq determined, get cwq and queue */
1034 cwq = get_cwq(gcwq->cpu, wq);
1035 trace_workqueue_queue_work(cpu, cwq, work);
1037 BUG_ON(!list_empty(&work->entry));
1039 cwq->nr_in_flight[cwq->work_color]++;
1040 work_flags = work_color_to_flags(cwq->work_color);
1042 if (likely(cwq->nr_active < cwq->max_active)) {
1043 trace_workqueue_activate_work(work);
1044 cwq->nr_active++;
1045 worklist = gcwq_determine_ins_pos(gcwq, cwq);
1046 } else {
1047 work_flags |= WORK_STRUCT_DELAYED;
1048 worklist = &cwq->delayed_works;
1051 insert_work(cwq, work, worklist, work_flags);
1053 spin_unlock_irqrestore(&gcwq->lock, flags);
1057 * queue_work - queue work on a workqueue
1058 * @wq: workqueue to use
1059 * @work: work to queue
1061 * Returns 0 if @work was already on a queue, non-zero otherwise.
1063 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1064 * it can be processed by another CPU.
1066 int queue_work(struct workqueue_struct *wq, struct work_struct *work)
1068 int ret;
1070 ret = queue_work_on(get_cpu(), wq, work);
1071 put_cpu();
1073 return ret;
1075 EXPORT_SYMBOL_GPL(queue_work);
1078 * queue_work_on - queue work on specific cpu
1079 * @cpu: CPU number to execute work on
1080 * @wq: workqueue to use
1081 * @work: work to queue
1083 * Returns 0 if @work was already on a queue, non-zero otherwise.
1085 * We queue the work to a specific CPU, the caller must ensure it
1086 * can't go away.
1089 queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
1091 int ret = 0;
1093 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1094 __queue_work(cpu, wq, work);
1095 ret = 1;
1097 return ret;
1099 EXPORT_SYMBOL_GPL(queue_work_on);
1101 static void delayed_work_timer_fn(unsigned long __data)
1103 struct delayed_work *dwork = (struct delayed_work *)__data;
1104 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1106 __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
1110 * queue_delayed_work - queue work on a workqueue after delay
1111 * @wq: workqueue to use
1112 * @dwork: delayable work to queue
1113 * @delay: number of jiffies to wait before queueing
1115 * Returns 0 if @work was already on a queue, non-zero otherwise.
1117 int queue_delayed_work(struct workqueue_struct *wq,
1118 struct delayed_work *dwork, unsigned long delay)
1120 if (delay == 0)
1121 return queue_work(wq, &dwork->work);
1123 return queue_delayed_work_on(-1, wq, dwork, delay);
1125 EXPORT_SYMBOL_GPL(queue_delayed_work);
1128 * queue_delayed_work_on - queue work on specific CPU after delay
1129 * @cpu: CPU number to execute work on
1130 * @wq: workqueue to use
1131 * @dwork: work to queue
1132 * @delay: number of jiffies to wait before queueing
1134 * Returns 0 if @work was already on a queue, non-zero otherwise.
1136 int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1137 struct delayed_work *dwork, unsigned long delay)
1139 int ret = 0;
1140 struct timer_list *timer = &dwork->timer;
1141 struct work_struct *work = &dwork->work;
1143 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1144 unsigned int lcpu;
1146 BUG_ON(timer_pending(timer));
1147 BUG_ON(!list_empty(&work->entry));
1149 timer_stats_timer_set_start_info(&dwork->timer);
1152 * This stores cwq for the moment, for the timer_fn.
1153 * Note that the work's gcwq is preserved to allow
1154 * reentrance detection for delayed works.
1156 if (!(wq->flags & WQ_UNBOUND)) {
1157 struct global_cwq *gcwq = get_work_gcwq(work);
1159 if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1160 lcpu = gcwq->cpu;
1161 else
1162 lcpu = raw_smp_processor_id();
1163 } else
1164 lcpu = WORK_CPU_UNBOUND;
1166 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1168 timer->expires = jiffies + delay;
1169 timer->data = (unsigned long)dwork;
1170 timer->function = delayed_work_timer_fn;
1172 if (unlikely(cpu >= 0))
1173 add_timer_on(timer, cpu);
1174 else
1175 add_timer(timer);
1176 ret = 1;
1178 return ret;
1180 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1183 * worker_enter_idle - enter idle state
1184 * @worker: worker which is entering idle state
1186 * @worker is entering idle state. Update stats and idle timer if
1187 * necessary.
1189 * LOCKING:
1190 * spin_lock_irq(gcwq->lock).
1192 static void worker_enter_idle(struct worker *worker)
1194 struct global_cwq *gcwq = worker->gcwq;
1196 BUG_ON(worker->flags & WORKER_IDLE);
1197 BUG_ON(!list_empty(&worker->entry) &&
1198 (worker->hentry.next || worker->hentry.pprev));
1200 /* can't use worker_set_flags(), also called from start_worker() */
1201 worker->flags |= WORKER_IDLE;
1202 gcwq->nr_idle++;
1203 worker->last_active = jiffies;
1205 /* idle_list is LIFO */
1206 list_add(&worker->entry, &gcwq->idle_list);
1208 if (likely(!(worker->flags & WORKER_ROGUE))) {
1209 if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1210 mod_timer(&gcwq->idle_timer,
1211 jiffies + IDLE_WORKER_TIMEOUT);
1212 } else
1213 wake_up_all(&gcwq->trustee_wait);
1215 /* sanity check nr_running */
1216 WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle &&
1217 atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1221 * worker_leave_idle - leave idle state
1222 * @worker: worker which is leaving idle state
1224 * @worker is leaving idle state. Update stats.
1226 * LOCKING:
1227 * spin_lock_irq(gcwq->lock).
1229 static void worker_leave_idle(struct worker *worker)
1231 struct global_cwq *gcwq = worker->gcwq;
1233 BUG_ON(!(worker->flags & WORKER_IDLE));
1234 worker_clr_flags(worker, WORKER_IDLE);
1235 gcwq->nr_idle--;
1236 list_del_init(&worker->entry);
1240 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1241 * @worker: self
1243 * Works which are scheduled while the cpu is online must at least be
1244 * scheduled to a worker which is bound to the cpu so that if they are
1245 * flushed from cpu callbacks while cpu is going down, they are
1246 * guaranteed to execute on the cpu.
1248 * This function is to be used by rogue workers and rescuers to bind
1249 * themselves to the target cpu and may race with cpu going down or
1250 * coming online. kthread_bind() can't be used because it may put the
1251 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1252 * verbatim as it's best effort and blocking and gcwq may be
1253 * [dis]associated in the meantime.
1255 * This function tries set_cpus_allowed() and locks gcwq and verifies
1256 * the binding against GCWQ_DISASSOCIATED which is set during
1257 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1258 * idle state or fetches works without dropping lock, it can guarantee
1259 * the scheduling requirement described in the first paragraph.
1261 * CONTEXT:
1262 * Might sleep. Called without any lock but returns with gcwq->lock
1263 * held.
1265 * RETURNS:
1266 * %true if the associated gcwq is online (@worker is successfully
1267 * bound), %false if offline.
1269 static bool worker_maybe_bind_and_lock(struct worker *worker)
1270 __acquires(&gcwq->lock)
1272 struct global_cwq *gcwq = worker->gcwq;
1273 struct task_struct *task = worker->task;
1275 while (true) {
1277 * The following call may fail, succeed or succeed
1278 * without actually migrating the task to the cpu if
1279 * it races with cpu hotunplug operation. Verify
1280 * against GCWQ_DISASSOCIATED.
1282 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1283 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1285 spin_lock_irq(&gcwq->lock);
1286 if (gcwq->flags & GCWQ_DISASSOCIATED)
1287 return false;
1288 if (task_cpu(task) == gcwq->cpu &&
1289 cpumask_equal(&current->cpus_allowed,
1290 get_cpu_mask(gcwq->cpu)))
1291 return true;
1292 spin_unlock_irq(&gcwq->lock);
1294 /* CPU has come up in between, retry migration */
1295 cpu_relax();
1300 * Function for worker->rebind_work used to rebind rogue busy workers
1301 * to the associated cpu which is coming back online. This is
1302 * scheduled by cpu up but can race with other cpu hotplug operations
1303 * and may be executed twice without intervening cpu down.
1305 static void worker_rebind_fn(struct work_struct *work)
1307 struct worker *worker = container_of(work, struct worker, rebind_work);
1308 struct global_cwq *gcwq = worker->gcwq;
1310 if (worker_maybe_bind_and_lock(worker))
1311 worker_clr_flags(worker, WORKER_REBIND);
1313 spin_unlock_irq(&gcwq->lock);
1316 static struct worker *alloc_worker(void)
1318 struct worker *worker;
1320 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1321 if (worker) {
1322 INIT_LIST_HEAD(&worker->entry);
1323 INIT_LIST_HEAD(&worker->scheduled);
1324 INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1325 /* on creation a worker is in !idle && prep state */
1326 worker->flags = WORKER_PREP;
1328 return worker;
1332 * create_worker - create a new workqueue worker
1333 * @gcwq: gcwq the new worker will belong to
1334 * @bind: whether to set affinity to @cpu or not
1336 * Create a new worker which is bound to @gcwq. The returned worker
1337 * can be started by calling start_worker() or destroyed using
1338 * destroy_worker().
1340 * CONTEXT:
1341 * Might sleep. Does GFP_KERNEL allocations.
1343 * RETURNS:
1344 * Pointer to the newly created worker.
1346 static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
1348 bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1349 struct worker *worker = NULL;
1350 int id = -1;
1352 spin_lock_irq(&gcwq->lock);
1353 while (ida_get_new(&gcwq->worker_ida, &id)) {
1354 spin_unlock_irq(&gcwq->lock);
1355 if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1356 goto fail;
1357 spin_lock_irq(&gcwq->lock);
1359 spin_unlock_irq(&gcwq->lock);
1361 worker = alloc_worker();
1362 if (!worker)
1363 goto fail;
1365 worker->gcwq = gcwq;
1366 worker->id = id;
1368 if (!on_unbound_cpu)
1369 worker->task = kthread_create_on_node(worker_thread,
1370 worker,
1371 cpu_to_node(gcwq->cpu),
1372 "kworker/%u:%d", gcwq->cpu, id);
1373 else
1374 worker->task = kthread_create(worker_thread, worker,
1375 "kworker/u:%d", id);
1376 if (IS_ERR(worker->task))
1377 goto fail;
1380 * A rogue worker will become a regular one if CPU comes
1381 * online later on. Make sure every worker has
1382 * PF_THREAD_BOUND set.
1384 if (bind && !on_unbound_cpu)
1385 kthread_bind(worker->task, gcwq->cpu);
1386 else {
1387 worker->task->flags |= PF_THREAD_BOUND;
1388 if (on_unbound_cpu)
1389 worker->flags |= WORKER_UNBOUND;
1392 return worker;
1393 fail:
1394 if (id >= 0) {
1395 spin_lock_irq(&gcwq->lock);
1396 ida_remove(&gcwq->worker_ida, id);
1397 spin_unlock_irq(&gcwq->lock);
1399 kfree(worker);
1400 return NULL;
1404 * start_worker - start a newly created worker
1405 * @worker: worker to start
1407 * Make the gcwq aware of @worker and start it.
1409 * CONTEXT:
1410 * spin_lock_irq(gcwq->lock).
1412 static void start_worker(struct worker *worker)
1414 worker->flags |= WORKER_STARTED;
1415 worker->gcwq->nr_workers++;
1416 worker_enter_idle(worker);
1417 wake_up_process(worker->task);
1421 * destroy_worker - destroy a workqueue worker
1422 * @worker: worker to be destroyed
1424 * Destroy @worker and adjust @gcwq stats accordingly.
1426 * CONTEXT:
1427 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1429 static void destroy_worker(struct worker *worker)
1431 struct global_cwq *gcwq = worker->gcwq;
1432 int id = worker->id;
1434 /* sanity check frenzy */
1435 BUG_ON(worker->current_work);
1436 BUG_ON(!list_empty(&worker->scheduled));
1438 if (worker->flags & WORKER_STARTED)
1439 gcwq->nr_workers--;
1440 if (worker->flags & WORKER_IDLE)
1441 gcwq->nr_idle--;
1443 list_del_init(&worker->entry);
1444 worker->flags |= WORKER_DIE;
1446 spin_unlock_irq(&gcwq->lock);
1448 kthread_stop(worker->task);
1449 kfree(worker);
1451 spin_lock_irq(&gcwq->lock);
1452 ida_remove(&gcwq->worker_ida, id);
1455 static void idle_worker_timeout(unsigned long __gcwq)
1457 struct global_cwq *gcwq = (void *)__gcwq;
1459 spin_lock_irq(&gcwq->lock);
1461 if (too_many_workers(gcwq)) {
1462 struct worker *worker;
1463 unsigned long expires;
1465 /* idle_list is kept in LIFO order, check the last one */
1466 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1467 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1469 if (time_before(jiffies, expires))
1470 mod_timer(&gcwq->idle_timer, expires);
1471 else {
1472 /* it's been idle for too long, wake up manager */
1473 gcwq->flags |= GCWQ_MANAGE_WORKERS;
1474 wake_up_worker(gcwq);
1478 spin_unlock_irq(&gcwq->lock);
1481 static bool send_mayday(struct work_struct *work)
1483 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1484 struct workqueue_struct *wq = cwq->wq;
1485 unsigned int cpu;
1487 if (!(wq->flags & WQ_RESCUER))
1488 return false;
1490 /* mayday mayday mayday */
1491 cpu = cwq->gcwq->cpu;
1492 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1493 if (cpu == WORK_CPU_UNBOUND)
1494 cpu = 0;
1495 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1496 wake_up_process(wq->rescuer->task);
1497 return true;
1500 static void gcwq_mayday_timeout(unsigned long __gcwq)
1502 struct global_cwq *gcwq = (void *)__gcwq;
1503 struct work_struct *work;
1505 spin_lock_irq(&gcwq->lock);
1507 if (need_to_create_worker(gcwq)) {
1509 * We've been trying to create a new worker but
1510 * haven't been successful. We might be hitting an
1511 * allocation deadlock. Send distress signals to
1512 * rescuers.
1514 list_for_each_entry(work, &gcwq->worklist, entry)
1515 send_mayday(work);
1518 spin_unlock_irq(&gcwq->lock);
1520 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1524 * maybe_create_worker - create a new worker if necessary
1525 * @gcwq: gcwq to create a new worker for
1527 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1528 * have at least one idle worker on return from this function. If
1529 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1530 * sent to all rescuers with works scheduled on @gcwq to resolve
1531 * possible allocation deadlock.
1533 * On return, need_to_create_worker() is guaranteed to be false and
1534 * may_start_working() true.
1536 * LOCKING:
1537 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1538 * multiple times. Does GFP_KERNEL allocations. Called only from
1539 * manager.
1541 * RETURNS:
1542 * false if no action was taken and gcwq->lock stayed locked, true
1543 * otherwise.
1545 static bool maybe_create_worker(struct global_cwq *gcwq)
1546 __releases(&gcwq->lock)
1547 __acquires(&gcwq->lock)
1549 if (!need_to_create_worker(gcwq))
1550 return false;
1551 restart:
1552 spin_unlock_irq(&gcwq->lock);
1554 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1555 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1557 while (true) {
1558 struct worker *worker;
1560 worker = create_worker(gcwq, true);
1561 if (worker) {
1562 del_timer_sync(&gcwq->mayday_timer);
1563 spin_lock_irq(&gcwq->lock);
1564 start_worker(worker);
1565 BUG_ON(need_to_create_worker(gcwq));
1566 return true;
1569 if (!need_to_create_worker(gcwq))
1570 break;
1572 __set_current_state(TASK_INTERRUPTIBLE);
1573 schedule_timeout(CREATE_COOLDOWN);
1575 if (!need_to_create_worker(gcwq))
1576 break;
1579 del_timer_sync(&gcwq->mayday_timer);
1580 spin_lock_irq(&gcwq->lock);
1581 if (need_to_create_worker(gcwq))
1582 goto restart;
1583 return true;
1587 * maybe_destroy_worker - destroy workers which have been idle for a while
1588 * @gcwq: gcwq to destroy workers for
1590 * Destroy @gcwq workers which have been idle for longer than
1591 * IDLE_WORKER_TIMEOUT.
1593 * LOCKING:
1594 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1595 * multiple times. Called only from manager.
1597 * RETURNS:
1598 * false if no action was taken and gcwq->lock stayed locked, true
1599 * otherwise.
1601 static bool maybe_destroy_workers(struct global_cwq *gcwq)
1603 bool ret = false;
1605 while (too_many_workers(gcwq)) {
1606 struct worker *worker;
1607 unsigned long expires;
1609 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1610 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1612 if (time_before(jiffies, expires)) {
1613 mod_timer(&gcwq->idle_timer, expires);
1614 break;
1617 destroy_worker(worker);
1618 ret = true;
1621 return ret;
1625 * manage_workers - manage worker pool
1626 * @worker: self
1628 * Assume the manager role and manage gcwq worker pool @worker belongs
1629 * to. At any given time, there can be only zero or one manager per
1630 * gcwq. The exclusion is handled automatically by this function.
1632 * The caller can safely start processing works on false return. On
1633 * true return, it's guaranteed that need_to_create_worker() is false
1634 * and may_start_working() is true.
1636 * CONTEXT:
1637 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1638 * multiple times. Does GFP_KERNEL allocations.
1640 * RETURNS:
1641 * false if no action was taken and gcwq->lock stayed locked, true if
1642 * some action was taken.
1644 static bool manage_workers(struct worker *worker)
1646 struct global_cwq *gcwq = worker->gcwq;
1647 bool ret = false;
1649 if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1650 return ret;
1652 gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1653 gcwq->flags |= GCWQ_MANAGING_WORKERS;
1656 * Destroy and then create so that may_start_working() is true
1657 * on return.
1659 ret |= maybe_destroy_workers(gcwq);
1660 ret |= maybe_create_worker(gcwq);
1662 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1665 * The trustee might be waiting to take over the manager
1666 * position, tell it we're done.
1668 if (unlikely(gcwq->trustee))
1669 wake_up_all(&gcwq->trustee_wait);
1671 return ret;
1675 * move_linked_works - move linked works to a list
1676 * @work: start of series of works to be scheduled
1677 * @head: target list to append @work to
1678 * @nextp: out paramter for nested worklist walking
1680 * Schedule linked works starting from @work to @head. Work series to
1681 * be scheduled starts at @work and includes any consecutive work with
1682 * WORK_STRUCT_LINKED set in its predecessor.
1684 * If @nextp is not NULL, it's updated to point to the next work of
1685 * the last scheduled work. This allows move_linked_works() to be
1686 * nested inside outer list_for_each_entry_safe().
1688 * CONTEXT:
1689 * spin_lock_irq(gcwq->lock).
1691 static void move_linked_works(struct work_struct *work, struct list_head *head,
1692 struct work_struct **nextp)
1694 struct work_struct *n;
1697 * Linked worklist will always end before the end of the list,
1698 * use NULL for list head.
1700 list_for_each_entry_safe_from(work, n, NULL, entry) {
1701 list_move_tail(&work->entry, head);
1702 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1703 break;
1707 * If we're already inside safe list traversal and have moved
1708 * multiple works to the scheduled queue, the next position
1709 * needs to be updated.
1711 if (nextp)
1712 *nextp = n;
1715 static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1717 struct work_struct *work = list_first_entry(&cwq->delayed_works,
1718 struct work_struct, entry);
1719 struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1721 trace_workqueue_activate_work(work);
1722 move_linked_works(work, pos, NULL);
1723 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1724 cwq->nr_active++;
1728 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1729 * @cwq: cwq of interest
1730 * @color: color of work which left the queue
1731 * @delayed: for a delayed work
1733 * A work either has completed or is removed from pending queue,
1734 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1736 * CONTEXT:
1737 * spin_lock_irq(gcwq->lock).
1739 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1740 bool delayed)
1742 /* ignore uncolored works */
1743 if (color == WORK_NO_COLOR)
1744 return;
1746 cwq->nr_in_flight[color]--;
1748 if (!delayed) {
1749 cwq->nr_active--;
1750 if (!list_empty(&cwq->delayed_works)) {
1751 /* one down, submit a delayed one */
1752 if (cwq->nr_active < cwq->max_active)
1753 cwq_activate_first_delayed(cwq);
1757 /* is flush in progress and are we at the flushing tip? */
1758 if (likely(cwq->flush_color != color))
1759 return;
1761 /* are there still in-flight works? */
1762 if (cwq->nr_in_flight[color])
1763 return;
1765 /* this cwq is done, clear flush_color */
1766 cwq->flush_color = -1;
1769 * If this was the last cwq, wake up the first flusher. It
1770 * will handle the rest.
1772 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1773 complete(&cwq->wq->first_flusher->done);
1777 * process_one_work - process single work
1778 * @worker: self
1779 * @work: work to process
1781 * Process @work. This function contains all the logics necessary to
1782 * process a single work including synchronization against and
1783 * interaction with other workers on the same cpu, queueing and
1784 * flushing. As long as context requirement is met, any worker can
1785 * call this function to process a work.
1787 * CONTEXT:
1788 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1790 static void process_one_work(struct worker *worker, struct work_struct *work)
1791 __releases(&gcwq->lock)
1792 __acquires(&gcwq->lock)
1794 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1795 struct global_cwq *gcwq = cwq->gcwq;
1796 struct hlist_head *bwh = busy_worker_head(gcwq, work);
1797 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1798 work_func_t f = work->func;
1799 int work_color;
1800 struct worker *collision;
1801 #ifdef CONFIG_LOCKDEP
1803 * It is permissible to free the struct work_struct from
1804 * inside the function that is called from it, this we need to
1805 * take into account for lockdep too. To avoid bogus "held
1806 * lock freed" warnings as well as problems when looking into
1807 * work->lockdep_map, make a copy and use that here.
1809 struct lockdep_map lockdep_map = work->lockdep_map;
1810 #endif
1812 * A single work shouldn't be executed concurrently by
1813 * multiple workers on a single cpu. Check whether anyone is
1814 * already processing the work. If so, defer the work to the
1815 * currently executing one.
1817 collision = __find_worker_executing_work(gcwq, bwh, work);
1818 if (unlikely(collision)) {
1819 move_linked_works(work, &collision->scheduled, NULL);
1820 return;
1823 /* claim and process */
1824 debug_work_deactivate(work);
1825 hlist_add_head(&worker->hentry, bwh);
1826 worker->current_work = work;
1827 worker->current_cwq = cwq;
1828 work_color = get_work_color(work);
1830 /* record the current cpu number in the work data and dequeue */
1831 set_work_cpu(work, gcwq->cpu);
1832 list_del_init(&work->entry);
1835 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1836 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1838 if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1839 struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1840 struct work_struct, entry);
1842 if (!list_empty(&gcwq->worklist) &&
1843 get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1844 wake_up_worker(gcwq);
1845 else
1846 gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1850 * CPU intensive works don't participate in concurrency
1851 * management. They're the scheduler's responsibility.
1853 if (unlikely(cpu_intensive))
1854 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1856 spin_unlock_irq(&gcwq->lock);
1858 work_clear_pending(work);
1859 lock_map_acquire_read(&cwq->wq->lockdep_map);
1860 lock_map_acquire(&lockdep_map);
1861 trace_workqueue_execute_start(work);
1862 f(work);
1864 * While we must be careful to not use "work" after this, the trace
1865 * point will only record its address.
1867 trace_workqueue_execute_end(work);
1868 lock_map_release(&lockdep_map);
1869 lock_map_release(&cwq->wq->lockdep_map);
1871 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1872 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1873 "%s/0x%08x/%d\n",
1874 current->comm, preempt_count(), task_pid_nr(current));
1875 printk(KERN_ERR " last function: ");
1876 print_symbol("%s\n", (unsigned long)f);
1877 debug_show_held_locks(current);
1878 dump_stack();
1881 spin_lock_irq(&gcwq->lock);
1883 /* clear cpu intensive status */
1884 if (unlikely(cpu_intensive))
1885 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1887 /* we're done with it, release */
1888 hlist_del_init(&worker->hentry);
1889 worker->current_work = NULL;
1890 worker->current_cwq = NULL;
1891 cwq_dec_nr_in_flight(cwq, work_color, false);
1895 * process_scheduled_works - process scheduled works
1896 * @worker: self
1898 * Process all scheduled works. Please note that the scheduled list
1899 * may change while processing a work, so this function repeatedly
1900 * fetches a work from the top and executes it.
1902 * CONTEXT:
1903 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1904 * multiple times.
1906 static void process_scheduled_works(struct worker *worker)
1908 while (!list_empty(&worker->scheduled)) {
1909 struct work_struct *work = list_first_entry(&worker->scheduled,
1910 struct work_struct, entry);
1911 process_one_work(worker, work);
1916 * worker_thread - the worker thread function
1917 * @__worker: self
1919 * The gcwq worker thread function. There's a single dynamic pool of
1920 * these per each cpu. These workers process all works regardless of
1921 * their specific target workqueue. The only exception is works which
1922 * belong to workqueues with a rescuer which will be explained in
1923 * rescuer_thread().
1925 static int worker_thread(void *__worker)
1927 struct worker *worker = __worker;
1928 struct global_cwq *gcwq = worker->gcwq;
1930 /* tell the scheduler that this is a workqueue worker */
1931 worker->task->flags |= PF_WQ_WORKER;
1932 woke_up:
1933 spin_lock_irq(&gcwq->lock);
1935 /* DIE can be set only while we're idle, checking here is enough */
1936 if (worker->flags & WORKER_DIE) {
1937 spin_unlock_irq(&gcwq->lock);
1938 worker->task->flags &= ~PF_WQ_WORKER;
1939 return 0;
1942 worker_leave_idle(worker);
1943 recheck:
1944 /* no more worker necessary? */
1945 if (!need_more_worker(gcwq))
1946 goto sleep;
1948 /* do we need to manage? */
1949 if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1950 goto recheck;
1953 * ->scheduled list can only be filled while a worker is
1954 * preparing to process a work or actually processing it.
1955 * Make sure nobody diddled with it while I was sleeping.
1957 BUG_ON(!list_empty(&worker->scheduled));
1960 * When control reaches this point, we're guaranteed to have
1961 * at least one idle worker or that someone else has already
1962 * assumed the manager role.
1964 worker_clr_flags(worker, WORKER_PREP);
1966 do {
1967 struct work_struct *work =
1968 list_first_entry(&gcwq->worklist,
1969 struct work_struct, entry);
1971 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1972 /* optimization path, not strictly necessary */
1973 process_one_work(worker, work);
1974 if (unlikely(!list_empty(&worker->scheduled)))
1975 process_scheduled_works(worker);
1976 } else {
1977 move_linked_works(work, &worker->scheduled, NULL);
1978 process_scheduled_works(worker);
1980 } while (keep_working(gcwq));
1982 worker_set_flags(worker, WORKER_PREP, false);
1983 sleep:
1984 if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1985 goto recheck;
1988 * gcwq->lock is held and there's no work to process and no
1989 * need to manage, sleep. Workers are woken up only while
1990 * holding gcwq->lock or from local cpu, so setting the
1991 * current state before releasing gcwq->lock is enough to
1992 * prevent losing any event.
1994 worker_enter_idle(worker);
1995 __set_current_state(TASK_INTERRUPTIBLE);
1996 spin_unlock_irq(&gcwq->lock);
1997 schedule();
1998 goto woke_up;
2002 * rescuer_thread - the rescuer thread function
2003 * @__wq: the associated workqueue
2005 * Workqueue rescuer thread function. There's one rescuer for each
2006 * workqueue which has WQ_RESCUER set.
2008 * Regular work processing on a gcwq may block trying to create a new
2009 * worker which uses GFP_KERNEL allocation which has slight chance of
2010 * developing into deadlock if some works currently on the same queue
2011 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2012 * the problem rescuer solves.
2014 * When such condition is possible, the gcwq summons rescuers of all
2015 * workqueues which have works queued on the gcwq and let them process
2016 * those works so that forward progress can be guaranteed.
2018 * This should happen rarely.
2020 static int rescuer_thread(void *__wq)
2022 struct workqueue_struct *wq = __wq;
2023 struct worker *rescuer = wq->rescuer;
2024 struct list_head *scheduled = &rescuer->scheduled;
2025 bool is_unbound = wq->flags & WQ_UNBOUND;
2026 unsigned int cpu;
2028 set_user_nice(current, RESCUER_NICE_LEVEL);
2029 repeat:
2030 set_current_state(TASK_INTERRUPTIBLE);
2032 if (kthread_should_stop())
2033 return 0;
2036 * See whether any cpu is asking for help. Unbounded
2037 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2039 for_each_mayday_cpu(cpu, wq->mayday_mask) {
2040 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2041 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2042 struct global_cwq *gcwq = cwq->gcwq;
2043 struct work_struct *work, *n;
2045 __set_current_state(TASK_RUNNING);
2046 mayday_clear_cpu(cpu, wq->mayday_mask);
2048 /* migrate to the target cpu if possible */
2049 rescuer->gcwq = gcwq;
2050 worker_maybe_bind_and_lock(rescuer);
2053 * Slurp in all works issued via this workqueue and
2054 * process'em.
2056 BUG_ON(!list_empty(&rescuer->scheduled));
2057 list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2058 if (get_work_cwq(work) == cwq)
2059 move_linked_works(work, scheduled, &n);
2061 process_scheduled_works(rescuer);
2064 * Leave this gcwq. If keep_working() is %true, notify a
2065 * regular worker; otherwise, we end up with 0 concurrency
2066 * and stalling the execution.
2068 if (keep_working(gcwq))
2069 wake_up_worker(gcwq);
2071 spin_unlock_irq(&gcwq->lock);
2074 schedule();
2075 goto repeat;
2078 struct wq_barrier {
2079 struct work_struct work;
2080 struct completion done;
2083 static void wq_barrier_func(struct work_struct *work)
2085 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2086 complete(&barr->done);
2090 * insert_wq_barrier - insert a barrier work
2091 * @cwq: cwq to insert barrier into
2092 * @barr: wq_barrier to insert
2093 * @target: target work to attach @barr to
2094 * @worker: worker currently executing @target, NULL if @target is not executing
2096 * @barr is linked to @target such that @barr is completed only after
2097 * @target finishes execution. Please note that the ordering
2098 * guarantee is observed only with respect to @target and on the local
2099 * cpu.
2101 * Currently, a queued barrier can't be canceled. This is because
2102 * try_to_grab_pending() can't determine whether the work to be
2103 * grabbed is at the head of the queue and thus can't clear LINKED
2104 * flag of the previous work while there must be a valid next work
2105 * after a work with LINKED flag set.
2107 * Note that when @worker is non-NULL, @target may be modified
2108 * underneath us, so we can't reliably determine cwq from @target.
2110 * CONTEXT:
2111 * spin_lock_irq(gcwq->lock).
2113 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2114 struct wq_barrier *barr,
2115 struct work_struct *target, struct worker *worker)
2117 struct list_head *head;
2118 unsigned int linked = 0;
2121 * debugobject calls are safe here even with gcwq->lock locked
2122 * as we know for sure that this will not trigger any of the
2123 * checks and call back into the fixup functions where we
2124 * might deadlock.
2126 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2127 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2128 init_completion(&barr->done);
2131 * If @target is currently being executed, schedule the
2132 * barrier to the worker; otherwise, put it after @target.
2134 if (worker)
2135 head = worker->scheduled.next;
2136 else {
2137 unsigned long *bits = work_data_bits(target);
2139 head = target->entry.next;
2140 /* there can already be other linked works, inherit and set */
2141 linked = *bits & WORK_STRUCT_LINKED;
2142 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2145 debug_work_activate(&barr->work);
2146 insert_work(cwq, &barr->work, head,
2147 work_color_to_flags(WORK_NO_COLOR) | linked);
2151 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2152 * @wq: workqueue being flushed
2153 * @flush_color: new flush color, < 0 for no-op
2154 * @work_color: new work color, < 0 for no-op
2156 * Prepare cwqs for workqueue flushing.
2158 * If @flush_color is non-negative, flush_color on all cwqs should be
2159 * -1. If no cwq has in-flight commands at the specified color, all
2160 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2161 * has in flight commands, its cwq->flush_color is set to
2162 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2163 * wakeup logic is armed and %true is returned.
2165 * The caller should have initialized @wq->first_flusher prior to
2166 * calling this function with non-negative @flush_color. If
2167 * @flush_color is negative, no flush color update is done and %false
2168 * is returned.
2170 * If @work_color is non-negative, all cwqs should have the same
2171 * work_color which is previous to @work_color and all will be
2172 * advanced to @work_color.
2174 * CONTEXT:
2175 * mutex_lock(wq->flush_mutex).
2177 * RETURNS:
2178 * %true if @flush_color >= 0 and there's something to flush. %false
2179 * otherwise.
2181 static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2182 int flush_color, int work_color)
2184 bool wait = false;
2185 unsigned int cpu;
2187 if (flush_color >= 0) {
2188 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2189 atomic_set(&wq->nr_cwqs_to_flush, 1);
2192 for_each_cwq_cpu(cpu, wq) {
2193 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2194 struct global_cwq *gcwq = cwq->gcwq;
2196 spin_lock_irq(&gcwq->lock);
2198 if (flush_color >= 0) {
2199 BUG_ON(cwq->flush_color != -1);
2201 if (cwq->nr_in_flight[flush_color]) {
2202 cwq->flush_color = flush_color;
2203 atomic_inc(&wq->nr_cwqs_to_flush);
2204 wait = true;
2208 if (work_color >= 0) {
2209 BUG_ON(work_color != work_next_color(cwq->work_color));
2210 cwq->work_color = work_color;
2213 spin_unlock_irq(&gcwq->lock);
2216 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2217 complete(&wq->first_flusher->done);
2219 return wait;
2223 * flush_workqueue - ensure that any scheduled work has run to completion.
2224 * @wq: workqueue to flush
2226 * Forces execution of the workqueue and blocks until its completion.
2227 * This is typically used in driver shutdown handlers.
2229 * We sleep until all works which were queued on entry have been handled,
2230 * but we are not livelocked by new incoming ones.
2232 void flush_workqueue(struct workqueue_struct *wq)
2234 struct wq_flusher this_flusher = {
2235 .list = LIST_HEAD_INIT(this_flusher.list),
2236 .flush_color = -1,
2237 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2239 int next_color;
2241 lock_map_acquire(&wq->lockdep_map);
2242 lock_map_release(&wq->lockdep_map);
2244 mutex_lock(&wq->flush_mutex);
2247 * Start-to-wait phase
2249 next_color = work_next_color(wq->work_color);
2251 if (next_color != wq->flush_color) {
2253 * Color space is not full. The current work_color
2254 * becomes our flush_color and work_color is advanced
2255 * by one.
2257 BUG_ON(!list_empty(&wq->flusher_overflow));
2258 this_flusher.flush_color = wq->work_color;
2259 wq->work_color = next_color;
2261 if (!wq->first_flusher) {
2262 /* no flush in progress, become the first flusher */
2263 BUG_ON(wq->flush_color != this_flusher.flush_color);
2265 wq->first_flusher = &this_flusher;
2267 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2268 wq->work_color)) {
2269 /* nothing to flush, done */
2270 wq->flush_color = next_color;
2271 wq->first_flusher = NULL;
2272 goto out_unlock;
2274 } else {
2275 /* wait in queue */
2276 BUG_ON(wq->flush_color == this_flusher.flush_color);
2277 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2278 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2280 } else {
2282 * Oops, color space is full, wait on overflow queue.
2283 * The next flush completion will assign us
2284 * flush_color and transfer to flusher_queue.
2286 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2289 mutex_unlock(&wq->flush_mutex);
2291 wait_for_completion(&this_flusher.done);
2294 * Wake-up-and-cascade phase
2296 * First flushers are responsible for cascading flushes and
2297 * handling overflow. Non-first flushers can simply return.
2299 if (wq->first_flusher != &this_flusher)
2300 return;
2302 mutex_lock(&wq->flush_mutex);
2304 /* we might have raced, check again with mutex held */
2305 if (wq->first_flusher != &this_flusher)
2306 goto out_unlock;
2308 wq->first_flusher = NULL;
2310 BUG_ON(!list_empty(&this_flusher.list));
2311 BUG_ON(wq->flush_color != this_flusher.flush_color);
2313 while (true) {
2314 struct wq_flusher *next, *tmp;
2316 /* complete all the flushers sharing the current flush color */
2317 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2318 if (next->flush_color != wq->flush_color)
2319 break;
2320 list_del_init(&next->list);
2321 complete(&next->done);
2324 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2325 wq->flush_color != work_next_color(wq->work_color));
2327 /* this flush_color is finished, advance by one */
2328 wq->flush_color = work_next_color(wq->flush_color);
2330 /* one color has been freed, handle overflow queue */
2331 if (!list_empty(&wq->flusher_overflow)) {
2333 * Assign the same color to all overflowed
2334 * flushers, advance work_color and append to
2335 * flusher_queue. This is the start-to-wait
2336 * phase for these overflowed flushers.
2338 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2339 tmp->flush_color = wq->work_color;
2341 wq->work_color = work_next_color(wq->work_color);
2343 list_splice_tail_init(&wq->flusher_overflow,
2344 &wq->flusher_queue);
2345 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2348 if (list_empty(&wq->flusher_queue)) {
2349 BUG_ON(wq->flush_color != wq->work_color);
2350 break;
2354 * Need to flush more colors. Make the next flusher
2355 * the new first flusher and arm cwqs.
2357 BUG_ON(wq->flush_color == wq->work_color);
2358 BUG_ON(wq->flush_color != next->flush_color);
2360 list_del_init(&next->list);
2361 wq->first_flusher = next;
2363 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2364 break;
2367 * Meh... this color is already done, clear first
2368 * flusher and repeat cascading.
2370 wq->first_flusher = NULL;
2373 out_unlock:
2374 mutex_unlock(&wq->flush_mutex);
2376 EXPORT_SYMBOL_GPL(flush_workqueue);
2378 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2379 bool wait_executing)
2381 struct worker *worker = NULL;
2382 struct global_cwq *gcwq;
2383 struct cpu_workqueue_struct *cwq;
2385 might_sleep();
2386 gcwq = get_work_gcwq(work);
2387 if (!gcwq)
2388 return false;
2390 spin_lock_irq(&gcwq->lock);
2391 if (!list_empty(&work->entry)) {
2393 * See the comment near try_to_grab_pending()->smp_rmb().
2394 * If it was re-queued to a different gcwq under us, we
2395 * are not going to wait.
2397 smp_rmb();
2398 cwq = get_work_cwq(work);
2399 if (unlikely(!cwq || gcwq != cwq->gcwq))
2400 goto already_gone;
2401 } else if (wait_executing) {
2402 worker = find_worker_executing_work(gcwq, work);
2403 if (!worker)
2404 goto already_gone;
2405 cwq = worker->current_cwq;
2406 } else
2407 goto already_gone;
2409 insert_wq_barrier(cwq, barr, work, worker);
2410 spin_unlock_irq(&gcwq->lock);
2413 * If @max_active is 1 or rescuer is in use, flushing another work
2414 * item on the same workqueue may lead to deadlock. Make sure the
2415 * flusher is not running on the same workqueue by verifying write
2416 * access.
2418 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2419 lock_map_acquire(&cwq->wq->lockdep_map);
2420 else
2421 lock_map_acquire_read(&cwq->wq->lockdep_map);
2422 lock_map_release(&cwq->wq->lockdep_map);
2424 return true;
2425 already_gone:
2426 spin_unlock_irq(&gcwq->lock);
2427 return false;
2431 * flush_work - wait for a work to finish executing the last queueing instance
2432 * @work: the work to flush
2434 * Wait until @work has finished execution. This function considers
2435 * only the last queueing instance of @work. If @work has been
2436 * enqueued across different CPUs on a non-reentrant workqueue or on
2437 * multiple workqueues, @work might still be executing on return on
2438 * some of the CPUs from earlier queueing.
2440 * If @work was queued only on a non-reentrant, ordered or unbound
2441 * workqueue, @work is guaranteed to be idle on return if it hasn't
2442 * been requeued since flush started.
2444 * RETURNS:
2445 * %true if flush_work() waited for the work to finish execution,
2446 * %false if it was already idle.
2448 bool flush_work(struct work_struct *work)
2450 struct wq_barrier barr;
2452 if (start_flush_work(work, &barr, true)) {
2453 wait_for_completion(&barr.done);
2454 destroy_work_on_stack(&barr.work);
2455 return true;
2456 } else
2457 return false;
2459 EXPORT_SYMBOL_GPL(flush_work);
2461 static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2463 struct wq_barrier barr;
2464 struct worker *worker;
2466 spin_lock_irq(&gcwq->lock);
2468 worker = find_worker_executing_work(gcwq, work);
2469 if (unlikely(worker))
2470 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2472 spin_unlock_irq(&gcwq->lock);
2474 if (unlikely(worker)) {
2475 wait_for_completion(&barr.done);
2476 destroy_work_on_stack(&barr.work);
2477 return true;
2478 } else
2479 return false;
2482 static bool wait_on_work(struct work_struct *work)
2484 bool ret = false;
2485 int cpu;
2487 might_sleep();
2489 lock_map_acquire(&work->lockdep_map);
2490 lock_map_release(&work->lockdep_map);
2492 for_each_gcwq_cpu(cpu)
2493 ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2494 return ret;
2498 * flush_work_sync - wait until a work has finished execution
2499 * @work: the work to flush
2501 * Wait until @work has finished execution. On return, it's
2502 * guaranteed that all queueing instances of @work which happened
2503 * before this function is called are finished. In other words, if
2504 * @work hasn't been requeued since this function was called, @work is
2505 * guaranteed to be idle on return.
2507 * RETURNS:
2508 * %true if flush_work_sync() waited for the work to finish execution,
2509 * %false if it was already idle.
2511 bool flush_work_sync(struct work_struct *work)
2513 struct wq_barrier barr;
2514 bool pending, waited;
2516 /* we'll wait for executions separately, queue barr only if pending */
2517 pending = start_flush_work(work, &barr, false);
2519 /* wait for executions to finish */
2520 waited = wait_on_work(work);
2522 /* wait for the pending one */
2523 if (pending) {
2524 wait_for_completion(&barr.done);
2525 destroy_work_on_stack(&barr.work);
2528 return pending || waited;
2530 EXPORT_SYMBOL_GPL(flush_work_sync);
2533 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2534 * so this work can't be re-armed in any way.
2536 static int try_to_grab_pending(struct work_struct *work)
2538 struct global_cwq *gcwq;
2539 int ret = -1;
2541 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2542 return 0;
2545 * The queueing is in progress, or it is already queued. Try to
2546 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2548 gcwq = get_work_gcwq(work);
2549 if (!gcwq)
2550 return ret;
2552 spin_lock_irq(&gcwq->lock);
2553 if (!list_empty(&work->entry)) {
2555 * This work is queued, but perhaps we locked the wrong gcwq.
2556 * In that case we must see the new value after rmb(), see
2557 * insert_work()->wmb().
2559 smp_rmb();
2560 if (gcwq == get_work_gcwq(work)) {
2561 debug_work_deactivate(work);
2562 list_del_init(&work->entry);
2563 cwq_dec_nr_in_flight(get_work_cwq(work),
2564 get_work_color(work),
2565 *work_data_bits(work) & WORK_STRUCT_DELAYED);
2566 ret = 1;
2569 spin_unlock_irq(&gcwq->lock);
2571 return ret;
2574 static bool __cancel_work_timer(struct work_struct *work,
2575 struct timer_list* timer)
2577 int ret;
2579 do {
2580 ret = (timer && likely(del_timer(timer)));
2581 if (!ret)
2582 ret = try_to_grab_pending(work);
2583 wait_on_work(work);
2584 } while (unlikely(ret < 0));
2586 clear_work_data(work);
2587 return ret;
2591 * cancel_work_sync - cancel a work and wait for it to finish
2592 * @work: the work to cancel
2594 * Cancel @work and wait for its execution to finish. This function
2595 * can be used even if the work re-queues itself or migrates to
2596 * another workqueue. On return from this function, @work is
2597 * guaranteed to be not pending or executing on any CPU.
2599 * cancel_work_sync(&delayed_work->work) must not be used for
2600 * delayed_work's. Use cancel_delayed_work_sync() instead.
2602 * The caller must ensure that the workqueue on which @work was last
2603 * queued can't be destroyed before this function returns.
2605 * RETURNS:
2606 * %true if @work was pending, %false otherwise.
2608 bool cancel_work_sync(struct work_struct *work)
2610 return __cancel_work_timer(work, NULL);
2612 EXPORT_SYMBOL_GPL(cancel_work_sync);
2615 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2616 * @dwork: the delayed work to flush
2618 * Delayed timer is cancelled and the pending work is queued for
2619 * immediate execution. Like flush_work(), this function only
2620 * considers the last queueing instance of @dwork.
2622 * RETURNS:
2623 * %true if flush_work() waited for the work to finish execution,
2624 * %false if it was already idle.
2626 bool flush_delayed_work(struct delayed_work *dwork)
2628 if (del_timer_sync(&dwork->timer))
2629 __queue_work(raw_smp_processor_id(),
2630 get_work_cwq(&dwork->work)->wq, &dwork->work);
2631 return flush_work(&dwork->work);
2633 EXPORT_SYMBOL(flush_delayed_work);
2636 * flush_delayed_work_sync - wait for a dwork to finish
2637 * @dwork: the delayed work to flush
2639 * Delayed timer is cancelled and the pending work is queued for
2640 * execution immediately. Other than timer handling, its behavior
2641 * is identical to flush_work_sync().
2643 * RETURNS:
2644 * %true if flush_work_sync() waited for the work to finish execution,
2645 * %false if it was already idle.
2647 bool flush_delayed_work_sync(struct delayed_work *dwork)
2649 if (del_timer_sync(&dwork->timer))
2650 __queue_work(raw_smp_processor_id(),
2651 get_work_cwq(&dwork->work)->wq, &dwork->work);
2652 return flush_work_sync(&dwork->work);
2654 EXPORT_SYMBOL(flush_delayed_work_sync);
2657 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2658 * @dwork: the delayed work cancel
2660 * This is cancel_work_sync() for delayed works.
2662 * RETURNS:
2663 * %true if @dwork was pending, %false otherwise.
2665 bool cancel_delayed_work_sync(struct delayed_work *dwork)
2667 return __cancel_work_timer(&dwork->work, &dwork->timer);
2669 EXPORT_SYMBOL(cancel_delayed_work_sync);
2672 * schedule_work - put work task in global workqueue
2673 * @work: job to be done
2675 * Returns zero if @work was already on the kernel-global workqueue and
2676 * non-zero otherwise.
2678 * This puts a job in the kernel-global workqueue if it was not already
2679 * queued and leaves it in the same position on the kernel-global
2680 * workqueue otherwise.
2682 int schedule_work(struct work_struct *work)
2684 return queue_work(system_wq, work);
2686 EXPORT_SYMBOL(schedule_work);
2689 * schedule_work_on - put work task on a specific cpu
2690 * @cpu: cpu to put the work task on
2691 * @work: job to be done
2693 * This puts a job on a specific cpu
2695 int schedule_work_on(int cpu, struct work_struct *work)
2697 return queue_work_on(cpu, system_wq, work);
2699 EXPORT_SYMBOL(schedule_work_on);
2702 * schedule_delayed_work - put work task in global workqueue after delay
2703 * @dwork: job to be done
2704 * @delay: number of jiffies to wait or 0 for immediate execution
2706 * After waiting for a given time this puts a job in the kernel-global
2707 * workqueue.
2709 int schedule_delayed_work(struct delayed_work *dwork,
2710 unsigned long delay)
2712 return queue_delayed_work(system_wq, dwork, delay);
2714 EXPORT_SYMBOL(schedule_delayed_work);
2717 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2718 * @cpu: cpu to use
2719 * @dwork: job to be done
2720 * @delay: number of jiffies to wait
2722 * After waiting for a given time this puts a job in the kernel-global
2723 * workqueue on the specified CPU.
2725 int schedule_delayed_work_on(int cpu,
2726 struct delayed_work *dwork, unsigned long delay)
2728 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2730 EXPORT_SYMBOL(schedule_delayed_work_on);
2733 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2734 * @func: the function to call
2736 * schedule_on_each_cpu() executes @func on each online CPU using the
2737 * system workqueue and blocks until all CPUs have completed.
2738 * schedule_on_each_cpu() is very slow.
2740 * RETURNS:
2741 * 0 on success, -errno on failure.
2743 int schedule_on_each_cpu(work_func_t func)
2745 int cpu;
2746 struct work_struct __percpu *works;
2748 works = alloc_percpu(struct work_struct);
2749 if (!works)
2750 return -ENOMEM;
2752 get_online_cpus();
2754 for_each_online_cpu(cpu) {
2755 struct work_struct *work = per_cpu_ptr(works, cpu);
2757 INIT_WORK(work, func);
2758 schedule_work_on(cpu, work);
2761 for_each_online_cpu(cpu)
2762 flush_work(per_cpu_ptr(works, cpu));
2764 put_online_cpus();
2765 free_percpu(works);
2766 return 0;
2770 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2772 * Forces execution of the kernel-global workqueue and blocks until its
2773 * completion.
2775 * Think twice before calling this function! It's very easy to get into
2776 * trouble if you don't take great care. Either of the following situations
2777 * will lead to deadlock:
2779 * One of the work items currently on the workqueue needs to acquire
2780 * a lock held by your code or its caller.
2782 * Your code is running in the context of a work routine.
2784 * They will be detected by lockdep when they occur, but the first might not
2785 * occur very often. It depends on what work items are on the workqueue and
2786 * what locks they need, which you have no control over.
2788 * In most situations flushing the entire workqueue is overkill; you merely
2789 * need to know that a particular work item isn't queued and isn't running.
2790 * In such cases you should use cancel_delayed_work_sync() or
2791 * cancel_work_sync() instead.
2793 void flush_scheduled_work(void)
2795 flush_workqueue(system_wq);
2797 EXPORT_SYMBOL(flush_scheduled_work);
2800 * execute_in_process_context - reliably execute the routine with user context
2801 * @fn: the function to execute
2802 * @ew: guaranteed storage for the execute work structure (must
2803 * be available when the work executes)
2805 * Executes the function immediately if process context is available,
2806 * otherwise schedules the function for delayed execution.
2808 * Returns: 0 - function was executed
2809 * 1 - function was scheduled for execution
2811 int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2813 if (!in_interrupt()) {
2814 fn(&ew->work);
2815 return 0;
2818 INIT_WORK(&ew->work, fn);
2819 schedule_work(&ew->work);
2821 return 1;
2823 EXPORT_SYMBOL_GPL(execute_in_process_context);
2825 int keventd_up(void)
2827 return system_wq != NULL;
2830 static int alloc_cwqs(struct workqueue_struct *wq)
2833 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2834 * Make sure that the alignment isn't lower than that of
2835 * unsigned long long.
2837 const size_t size = sizeof(struct cpu_workqueue_struct);
2838 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2839 __alignof__(unsigned long long));
2840 #ifdef CONFIG_SMP
2841 bool percpu = !(wq->flags & WQ_UNBOUND);
2842 #else
2843 bool percpu = false;
2844 #endif
2846 if (percpu)
2847 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2848 else {
2849 void *ptr;
2852 * Allocate enough room to align cwq and put an extra
2853 * pointer at the end pointing back to the originally
2854 * allocated pointer which will be used for free.
2856 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2857 if (ptr) {
2858 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2859 *(void **)(wq->cpu_wq.single + 1) = ptr;
2863 /* just in case, make sure it's actually aligned
2864 * - this is affected by PERCPU() alignment in vmlinux.lds.S
2866 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2867 return wq->cpu_wq.v ? 0 : -ENOMEM;
2870 static void free_cwqs(struct workqueue_struct *wq)
2872 #ifdef CONFIG_SMP
2873 bool percpu = !(wq->flags & WQ_UNBOUND);
2874 #else
2875 bool percpu = false;
2876 #endif
2878 if (percpu)
2879 free_percpu(wq->cpu_wq.pcpu);
2880 else if (wq->cpu_wq.single) {
2881 /* the pointer to free is stored right after the cwq */
2882 kfree(*(void **)(wq->cpu_wq.single + 1));
2886 static int wq_clamp_max_active(int max_active, unsigned int flags,
2887 const char *name)
2889 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2891 if (max_active < 1 || max_active > lim)
2892 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2893 "is out of range, clamping between %d and %d\n",
2894 max_active, name, 1, lim);
2896 return clamp_val(max_active, 1, lim);
2899 struct workqueue_struct *__alloc_workqueue_key(const char *name,
2900 unsigned int flags,
2901 int max_active,
2902 struct lock_class_key *key,
2903 const char *lock_name)
2905 struct workqueue_struct *wq;
2906 unsigned int cpu;
2909 * Workqueues which may be used during memory reclaim should
2910 * have a rescuer to guarantee forward progress.
2912 if (flags & WQ_MEM_RECLAIM)
2913 flags |= WQ_RESCUER;
2916 * Unbound workqueues aren't concurrency managed and should be
2917 * dispatched to workers immediately.
2919 if (flags & WQ_UNBOUND)
2920 flags |= WQ_HIGHPRI;
2922 max_active = max_active ?: WQ_DFL_ACTIVE;
2923 max_active = wq_clamp_max_active(max_active, flags, name);
2925 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
2926 if (!wq)
2927 goto err;
2929 wq->flags = flags;
2930 wq->saved_max_active = max_active;
2931 mutex_init(&wq->flush_mutex);
2932 atomic_set(&wq->nr_cwqs_to_flush, 0);
2933 INIT_LIST_HEAD(&wq->flusher_queue);
2934 INIT_LIST_HEAD(&wq->flusher_overflow);
2936 wq->name = name;
2937 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
2938 INIT_LIST_HEAD(&wq->list);
2940 if (alloc_cwqs(wq) < 0)
2941 goto err;
2943 for_each_cwq_cpu(cpu, wq) {
2944 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2945 struct global_cwq *gcwq = get_gcwq(cpu);
2947 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
2948 cwq->gcwq = gcwq;
2949 cwq->wq = wq;
2950 cwq->flush_color = -1;
2951 cwq->max_active = max_active;
2952 INIT_LIST_HEAD(&cwq->delayed_works);
2955 if (flags & WQ_RESCUER) {
2956 struct worker *rescuer;
2958 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
2959 goto err;
2961 wq->rescuer = rescuer = alloc_worker();
2962 if (!rescuer)
2963 goto err;
2965 rescuer->task = kthread_create(rescuer_thread, wq, "%s", name);
2966 if (IS_ERR(rescuer->task))
2967 goto err;
2969 rescuer->task->flags |= PF_THREAD_BOUND;
2970 wake_up_process(rescuer->task);
2974 * workqueue_lock protects global freeze state and workqueues
2975 * list. Grab it, set max_active accordingly and add the new
2976 * workqueue to workqueues list.
2978 spin_lock(&workqueue_lock);
2980 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
2981 for_each_cwq_cpu(cpu, wq)
2982 get_cwq(cpu, wq)->max_active = 0;
2984 list_add(&wq->list, &workqueues);
2986 spin_unlock(&workqueue_lock);
2988 return wq;
2989 err:
2990 if (wq) {
2991 free_cwqs(wq);
2992 free_mayday_mask(wq->mayday_mask);
2993 kfree(wq->rescuer);
2994 kfree(wq);
2996 return NULL;
2998 EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3001 * destroy_workqueue - safely terminate a workqueue
3002 * @wq: target workqueue
3004 * Safely destroy a workqueue. All work currently pending will be done first.
3006 void destroy_workqueue(struct workqueue_struct *wq)
3008 unsigned int flush_cnt = 0;
3009 unsigned int cpu;
3012 * Mark @wq dying and drain all pending works. Once WQ_DYING is
3013 * set, only chain queueing is allowed. IOW, only currently
3014 * pending or running work items on @wq can queue further work
3015 * items on it. @wq is flushed repeatedly until it becomes empty.
3016 * The number of flushing is detemined by the depth of chaining and
3017 * should be relatively short. Whine if it takes too long.
3019 wq->flags |= WQ_DYING;
3020 reflush:
3021 flush_workqueue(wq);
3023 for_each_cwq_cpu(cpu, wq) {
3024 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3026 if (!cwq->nr_active && list_empty(&cwq->delayed_works))
3027 continue;
3029 if (++flush_cnt == 10 ||
3030 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
3031 printk(KERN_WARNING "workqueue %s: flush on "
3032 "destruction isn't complete after %u tries\n",
3033 wq->name, flush_cnt);
3034 goto reflush;
3038 * wq list is used to freeze wq, remove from list after
3039 * flushing is complete in case freeze races us.
3041 spin_lock(&workqueue_lock);
3042 list_del(&wq->list);
3043 spin_unlock(&workqueue_lock);
3045 /* sanity check */
3046 for_each_cwq_cpu(cpu, wq) {
3047 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3048 int i;
3050 for (i = 0; i < WORK_NR_COLORS; i++)
3051 BUG_ON(cwq->nr_in_flight[i]);
3052 BUG_ON(cwq->nr_active);
3053 BUG_ON(!list_empty(&cwq->delayed_works));
3056 if (wq->flags & WQ_RESCUER) {
3057 kthread_stop(wq->rescuer->task);
3058 free_mayday_mask(wq->mayday_mask);
3059 kfree(wq->rescuer);
3062 free_cwqs(wq);
3063 kfree(wq);
3065 EXPORT_SYMBOL_GPL(destroy_workqueue);
3068 * workqueue_set_max_active - adjust max_active of a workqueue
3069 * @wq: target workqueue
3070 * @max_active: new max_active value.
3072 * Set max_active of @wq to @max_active.
3074 * CONTEXT:
3075 * Don't call from IRQ context.
3077 void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3079 unsigned int cpu;
3081 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3083 spin_lock(&workqueue_lock);
3085 wq->saved_max_active = max_active;
3087 for_each_cwq_cpu(cpu, wq) {
3088 struct global_cwq *gcwq = get_gcwq(cpu);
3090 spin_lock_irq(&gcwq->lock);
3092 if (!(wq->flags & WQ_FREEZABLE) ||
3093 !(gcwq->flags & GCWQ_FREEZING))
3094 get_cwq(gcwq->cpu, wq)->max_active = max_active;
3096 spin_unlock_irq(&gcwq->lock);
3099 spin_unlock(&workqueue_lock);
3101 EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3104 * workqueue_congested - test whether a workqueue is congested
3105 * @cpu: CPU in question
3106 * @wq: target workqueue
3108 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3109 * no synchronization around this function and the test result is
3110 * unreliable and only useful as advisory hints or for debugging.
3112 * RETURNS:
3113 * %true if congested, %false otherwise.
3115 bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3117 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3119 return !list_empty(&cwq->delayed_works);
3121 EXPORT_SYMBOL_GPL(workqueue_congested);
3124 * work_cpu - return the last known associated cpu for @work
3125 * @work: the work of interest
3127 * RETURNS:
3128 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3130 unsigned int work_cpu(struct work_struct *work)
3132 struct global_cwq *gcwq = get_work_gcwq(work);
3134 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3136 EXPORT_SYMBOL_GPL(work_cpu);
3139 * work_busy - test whether a work is currently pending or running
3140 * @work: the work to be tested
3142 * Test whether @work is currently pending or running. There is no
3143 * synchronization around this function and the test result is
3144 * unreliable and only useful as advisory hints or for debugging.
3145 * Especially for reentrant wqs, the pending state might hide the
3146 * running state.
3148 * RETURNS:
3149 * OR'd bitmask of WORK_BUSY_* bits.
3151 unsigned int work_busy(struct work_struct *work)
3153 struct global_cwq *gcwq = get_work_gcwq(work);
3154 unsigned long flags;
3155 unsigned int ret = 0;
3157 if (!gcwq)
3158 return false;
3160 spin_lock_irqsave(&gcwq->lock, flags);
3162 if (work_pending(work))
3163 ret |= WORK_BUSY_PENDING;
3164 if (find_worker_executing_work(gcwq, work))
3165 ret |= WORK_BUSY_RUNNING;
3167 spin_unlock_irqrestore(&gcwq->lock, flags);
3169 return ret;
3171 EXPORT_SYMBOL_GPL(work_busy);
3174 * CPU hotplug.
3176 * There are two challenges in supporting CPU hotplug. Firstly, there
3177 * are a lot of assumptions on strong associations among work, cwq and
3178 * gcwq which make migrating pending and scheduled works very
3179 * difficult to implement without impacting hot paths. Secondly,
3180 * gcwqs serve mix of short, long and very long running works making
3181 * blocked draining impractical.
3183 * This is solved by allowing a gcwq to be detached from CPU, running
3184 * it with unbound (rogue) workers and allowing it to be reattached
3185 * later if the cpu comes back online. A separate thread is created
3186 * to govern a gcwq in such state and is called the trustee of the
3187 * gcwq.
3189 * Trustee states and their descriptions.
3191 * START Command state used on startup. On CPU_DOWN_PREPARE, a
3192 * new trustee is started with this state.
3194 * IN_CHARGE Once started, trustee will enter this state after
3195 * assuming the manager role and making all existing
3196 * workers rogue. DOWN_PREPARE waits for trustee to
3197 * enter this state. After reaching IN_CHARGE, trustee
3198 * tries to execute the pending worklist until it's empty
3199 * and the state is set to BUTCHER, or the state is set
3200 * to RELEASE.
3202 * BUTCHER Command state which is set by the cpu callback after
3203 * the cpu has went down. Once this state is set trustee
3204 * knows that there will be no new works on the worklist
3205 * and once the worklist is empty it can proceed to
3206 * killing idle workers.
3208 * RELEASE Command state which is set by the cpu callback if the
3209 * cpu down has been canceled or it has come online
3210 * again. After recognizing this state, trustee stops
3211 * trying to drain or butcher and clears ROGUE, rebinds
3212 * all remaining workers back to the cpu and releases
3213 * manager role.
3215 * DONE Trustee will enter this state after BUTCHER or RELEASE
3216 * is complete.
3218 * trustee CPU draining
3219 * took over down complete
3220 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3221 * | | ^
3222 * | CPU is back online v return workers |
3223 * ----------------> RELEASE --------------
3227 * trustee_wait_event_timeout - timed event wait for trustee
3228 * @cond: condition to wait for
3229 * @timeout: timeout in jiffies
3231 * wait_event_timeout() for trustee to use. Handles locking and
3232 * checks for RELEASE request.
3234 * CONTEXT:
3235 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3236 * multiple times. To be used by trustee.
3238 * RETURNS:
3239 * Positive indicating left time if @cond is satisfied, 0 if timed
3240 * out, -1 if canceled.
3242 #define trustee_wait_event_timeout(cond, timeout) ({ \
3243 long __ret = (timeout); \
3244 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3245 __ret) { \
3246 spin_unlock_irq(&gcwq->lock); \
3247 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3248 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3249 __ret); \
3250 spin_lock_irq(&gcwq->lock); \
3252 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3256 * trustee_wait_event - event wait for trustee
3257 * @cond: condition to wait for
3259 * wait_event() for trustee to use. Automatically handles locking and
3260 * checks for CANCEL request.
3262 * CONTEXT:
3263 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3264 * multiple times. To be used by trustee.
3266 * RETURNS:
3267 * 0 if @cond is satisfied, -1 if canceled.
3269 #define trustee_wait_event(cond) ({ \
3270 long __ret1; \
3271 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3272 __ret1 < 0 ? -1 : 0; \
3275 static int __cpuinit trustee_thread(void *__gcwq)
3277 struct global_cwq *gcwq = __gcwq;
3278 struct worker *worker;
3279 struct work_struct *work;
3280 struct hlist_node *pos;
3281 long rc;
3282 int i;
3284 BUG_ON(gcwq->cpu != smp_processor_id());
3286 spin_lock_irq(&gcwq->lock);
3288 * Claim the manager position and make all workers rogue.
3289 * Trustee must be bound to the target cpu and can't be
3290 * cancelled.
3292 BUG_ON(gcwq->cpu != smp_processor_id());
3293 rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3294 BUG_ON(rc < 0);
3296 gcwq->flags |= GCWQ_MANAGING_WORKERS;
3298 list_for_each_entry(worker, &gcwq->idle_list, entry)
3299 worker->flags |= WORKER_ROGUE;
3301 for_each_busy_worker(worker, i, pos, gcwq)
3302 worker->flags |= WORKER_ROGUE;
3305 * Call schedule() so that we cross rq->lock and thus can
3306 * guarantee sched callbacks see the rogue flag. This is
3307 * necessary as scheduler callbacks may be invoked from other
3308 * cpus.
3310 spin_unlock_irq(&gcwq->lock);
3311 schedule();
3312 spin_lock_irq(&gcwq->lock);
3315 * Sched callbacks are disabled now. Zap nr_running. After
3316 * this, nr_running stays zero and need_more_worker() and
3317 * keep_working() are always true as long as the worklist is
3318 * not empty.
3320 atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3322 spin_unlock_irq(&gcwq->lock);
3323 del_timer_sync(&gcwq->idle_timer);
3324 spin_lock_irq(&gcwq->lock);
3327 * We're now in charge. Notify and proceed to drain. We need
3328 * to keep the gcwq running during the whole CPU down
3329 * procedure as other cpu hotunplug callbacks may need to
3330 * flush currently running tasks.
3332 gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3333 wake_up_all(&gcwq->trustee_wait);
3336 * The original cpu is in the process of dying and may go away
3337 * anytime now. When that happens, we and all workers would
3338 * be migrated to other cpus. Try draining any left work. We
3339 * want to get it over with ASAP - spam rescuers, wake up as
3340 * many idlers as necessary and create new ones till the
3341 * worklist is empty. Note that if the gcwq is frozen, there
3342 * may be frozen works in freezable cwqs. Don't declare
3343 * completion while frozen.
3345 while (gcwq->nr_workers != gcwq->nr_idle ||
3346 gcwq->flags & GCWQ_FREEZING ||
3347 gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3348 int nr_works = 0;
3350 list_for_each_entry(work, &gcwq->worklist, entry) {
3351 send_mayday(work);
3352 nr_works++;
3355 list_for_each_entry(worker, &gcwq->idle_list, entry) {
3356 if (!nr_works--)
3357 break;
3358 wake_up_process(worker->task);
3361 if (need_to_create_worker(gcwq)) {
3362 spin_unlock_irq(&gcwq->lock);
3363 worker = create_worker(gcwq, false);
3364 spin_lock_irq(&gcwq->lock);
3365 if (worker) {
3366 worker->flags |= WORKER_ROGUE;
3367 start_worker(worker);
3371 /* give a breather */
3372 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3373 break;
3377 * Either all works have been scheduled and cpu is down, or
3378 * cpu down has already been canceled. Wait for and butcher
3379 * all workers till we're canceled.
3381 do {
3382 rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3383 while (!list_empty(&gcwq->idle_list))
3384 destroy_worker(list_first_entry(&gcwq->idle_list,
3385 struct worker, entry));
3386 } while (gcwq->nr_workers && rc >= 0);
3389 * At this point, either draining has completed and no worker
3390 * is left, or cpu down has been canceled or the cpu is being
3391 * brought back up. There shouldn't be any idle one left.
3392 * Tell the remaining busy ones to rebind once it finishes the
3393 * currently scheduled works by scheduling the rebind_work.
3395 WARN_ON(!list_empty(&gcwq->idle_list));
3397 for_each_busy_worker(worker, i, pos, gcwq) {
3398 struct work_struct *rebind_work = &worker->rebind_work;
3401 * Rebind_work may race with future cpu hotplug
3402 * operations. Use a separate flag to mark that
3403 * rebinding is scheduled.
3405 worker->flags |= WORKER_REBIND;
3406 worker->flags &= ~WORKER_ROGUE;
3408 /* queue rebind_work, wq doesn't matter, use the default one */
3409 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3410 work_data_bits(rebind_work)))
3411 continue;
3413 debug_work_activate(rebind_work);
3414 insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3415 worker->scheduled.next,
3416 work_color_to_flags(WORK_NO_COLOR));
3419 /* relinquish manager role */
3420 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3422 /* notify completion */
3423 gcwq->trustee = NULL;
3424 gcwq->trustee_state = TRUSTEE_DONE;
3425 wake_up_all(&gcwq->trustee_wait);
3426 spin_unlock_irq(&gcwq->lock);
3427 return 0;
3431 * wait_trustee_state - wait for trustee to enter the specified state
3432 * @gcwq: gcwq the trustee of interest belongs to
3433 * @state: target state to wait for
3435 * Wait for the trustee to reach @state. DONE is already matched.
3437 * CONTEXT:
3438 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3439 * multiple times. To be used by cpu_callback.
3441 static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3442 __releases(&gcwq->lock)
3443 __acquires(&gcwq->lock)
3445 if (!(gcwq->trustee_state == state ||
3446 gcwq->trustee_state == TRUSTEE_DONE)) {
3447 spin_unlock_irq(&gcwq->lock);
3448 __wait_event(gcwq->trustee_wait,
3449 gcwq->trustee_state == state ||
3450 gcwq->trustee_state == TRUSTEE_DONE);
3451 spin_lock_irq(&gcwq->lock);
3455 static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3456 unsigned long action,
3457 void *hcpu)
3459 unsigned int cpu = (unsigned long)hcpu;
3460 struct global_cwq *gcwq = get_gcwq(cpu);
3461 struct task_struct *new_trustee = NULL;
3462 struct worker *uninitialized_var(new_worker);
3463 unsigned long flags;
3465 action &= ~CPU_TASKS_FROZEN;
3467 switch (action) {
3468 case CPU_DOWN_PREPARE:
3469 new_trustee = kthread_create(trustee_thread, gcwq,
3470 "workqueue_trustee/%d\n", cpu);
3471 if (IS_ERR(new_trustee))
3472 return notifier_from_errno(PTR_ERR(new_trustee));
3473 kthread_bind(new_trustee, cpu);
3474 /* fall through */
3475 case CPU_UP_PREPARE:
3476 BUG_ON(gcwq->first_idle);
3477 new_worker = create_worker(gcwq, false);
3478 if (!new_worker) {
3479 if (new_trustee)
3480 kthread_stop(new_trustee);
3481 return NOTIFY_BAD;
3485 /* some are called w/ irq disabled, don't disturb irq status */
3486 spin_lock_irqsave(&gcwq->lock, flags);
3488 switch (action) {
3489 case CPU_DOWN_PREPARE:
3490 /* initialize trustee and tell it to acquire the gcwq */
3491 BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
3492 gcwq->trustee = new_trustee;
3493 gcwq->trustee_state = TRUSTEE_START;
3494 wake_up_process(gcwq->trustee);
3495 wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3496 /* fall through */
3497 case CPU_UP_PREPARE:
3498 BUG_ON(gcwq->first_idle);
3499 gcwq->first_idle = new_worker;
3500 break;
3502 case CPU_DYING:
3504 * Before this, the trustee and all workers except for
3505 * the ones which are still executing works from
3506 * before the last CPU down must be on the cpu. After
3507 * this, they'll all be diasporas.
3509 gcwq->flags |= GCWQ_DISASSOCIATED;
3510 break;
3512 case CPU_POST_DEAD:
3513 gcwq->trustee_state = TRUSTEE_BUTCHER;
3514 /* fall through */
3515 case CPU_UP_CANCELED:
3516 destroy_worker(gcwq->first_idle);
3517 gcwq->first_idle = NULL;
3518 break;
3520 case CPU_DOWN_FAILED:
3521 case CPU_ONLINE:
3522 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3523 if (gcwq->trustee_state != TRUSTEE_DONE) {
3524 gcwq->trustee_state = TRUSTEE_RELEASE;
3525 wake_up_process(gcwq->trustee);
3526 wait_trustee_state(gcwq, TRUSTEE_DONE);
3530 * Trustee is done and there might be no worker left.
3531 * Put the first_idle in and request a real manager to
3532 * take a look.
3534 spin_unlock_irq(&gcwq->lock);
3535 kthread_bind(gcwq->first_idle->task, cpu);
3536 spin_lock_irq(&gcwq->lock);
3537 gcwq->flags |= GCWQ_MANAGE_WORKERS;
3538 start_worker(gcwq->first_idle);
3539 gcwq->first_idle = NULL;
3540 break;
3543 spin_unlock_irqrestore(&gcwq->lock, flags);
3545 return notifier_from_errno(0);
3548 #ifdef CONFIG_SMP
3550 struct work_for_cpu {
3551 struct completion completion;
3552 long (*fn)(void *);
3553 void *arg;
3554 long ret;
3557 static int do_work_for_cpu(void *_wfc)
3559 struct work_for_cpu *wfc = _wfc;
3560 wfc->ret = wfc->fn(wfc->arg);
3561 complete(&wfc->completion);
3562 return 0;
3566 * work_on_cpu - run a function in user context on a particular cpu
3567 * @cpu: the cpu to run on
3568 * @fn: the function to run
3569 * @arg: the function arg
3571 * This will return the value @fn returns.
3572 * It is up to the caller to ensure that the cpu doesn't go offline.
3573 * The caller must not hold any locks which would prevent @fn from completing.
3575 long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3577 struct task_struct *sub_thread;
3578 struct work_for_cpu wfc = {
3579 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3580 .fn = fn,
3581 .arg = arg,
3584 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3585 if (IS_ERR(sub_thread))
3586 return PTR_ERR(sub_thread);
3587 kthread_bind(sub_thread, cpu);
3588 wake_up_process(sub_thread);
3589 wait_for_completion(&wfc.completion);
3590 return wfc.ret;
3592 EXPORT_SYMBOL_GPL(work_on_cpu);
3593 #endif /* CONFIG_SMP */
3595 #ifdef CONFIG_FREEZER
3598 * freeze_workqueues_begin - begin freezing workqueues
3600 * Start freezing workqueues. After this function returns, all freezable
3601 * workqueues will queue new works to their frozen_works list instead of
3602 * gcwq->worklist.
3604 * CONTEXT:
3605 * Grabs and releases workqueue_lock and gcwq->lock's.
3607 void freeze_workqueues_begin(void)
3609 unsigned int cpu;
3611 spin_lock(&workqueue_lock);
3613 BUG_ON(workqueue_freezing);
3614 workqueue_freezing = true;
3616 for_each_gcwq_cpu(cpu) {
3617 struct global_cwq *gcwq = get_gcwq(cpu);
3618 struct workqueue_struct *wq;
3620 spin_lock_irq(&gcwq->lock);
3622 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3623 gcwq->flags |= GCWQ_FREEZING;
3625 list_for_each_entry(wq, &workqueues, list) {
3626 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3628 if (cwq && wq->flags & WQ_FREEZABLE)
3629 cwq->max_active = 0;
3632 spin_unlock_irq(&gcwq->lock);
3635 spin_unlock(&workqueue_lock);
3639 * freeze_workqueues_busy - are freezable workqueues still busy?
3641 * Check whether freezing is complete. This function must be called
3642 * between freeze_workqueues_begin() and thaw_workqueues().
3644 * CONTEXT:
3645 * Grabs and releases workqueue_lock.
3647 * RETURNS:
3648 * %true if some freezable workqueues are still busy. %false if freezing
3649 * is complete.
3651 bool freeze_workqueues_busy(void)
3653 unsigned int cpu;
3654 bool busy = false;
3656 spin_lock(&workqueue_lock);
3658 BUG_ON(!workqueue_freezing);
3660 for_each_gcwq_cpu(cpu) {
3661 struct workqueue_struct *wq;
3663 * nr_active is monotonically decreasing. It's safe
3664 * to peek without lock.
3666 list_for_each_entry(wq, &workqueues, list) {
3667 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3669 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3670 continue;
3672 BUG_ON(cwq->nr_active < 0);
3673 if (cwq->nr_active) {
3674 busy = true;
3675 goto out_unlock;
3679 out_unlock:
3680 spin_unlock(&workqueue_lock);
3681 return busy;
3685 * thaw_workqueues - thaw workqueues
3687 * Thaw workqueues. Normal queueing is restored and all collected
3688 * frozen works are transferred to their respective gcwq worklists.
3690 * CONTEXT:
3691 * Grabs and releases workqueue_lock and gcwq->lock's.
3693 void thaw_workqueues(void)
3695 unsigned int cpu;
3697 spin_lock(&workqueue_lock);
3699 if (!workqueue_freezing)
3700 goto out_unlock;
3702 for_each_gcwq_cpu(cpu) {
3703 struct global_cwq *gcwq = get_gcwq(cpu);
3704 struct workqueue_struct *wq;
3706 spin_lock_irq(&gcwq->lock);
3708 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3709 gcwq->flags &= ~GCWQ_FREEZING;
3711 list_for_each_entry(wq, &workqueues, list) {
3712 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3714 if (!cwq || !(wq->flags & WQ_FREEZABLE))
3715 continue;
3717 /* restore max_active and repopulate worklist */
3718 cwq->max_active = wq->saved_max_active;
3720 while (!list_empty(&cwq->delayed_works) &&
3721 cwq->nr_active < cwq->max_active)
3722 cwq_activate_first_delayed(cwq);
3725 wake_up_worker(gcwq);
3727 spin_unlock_irq(&gcwq->lock);
3730 workqueue_freezing = false;
3731 out_unlock:
3732 spin_unlock(&workqueue_lock);
3734 #endif /* CONFIG_FREEZER */
3736 static int __init init_workqueues(void)
3738 unsigned int cpu;
3739 int i;
3741 cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
3743 /* initialize gcwqs */
3744 for_each_gcwq_cpu(cpu) {
3745 struct global_cwq *gcwq = get_gcwq(cpu);
3747 spin_lock_init(&gcwq->lock);
3748 INIT_LIST_HEAD(&gcwq->worklist);
3749 gcwq->cpu = cpu;
3750 gcwq->flags |= GCWQ_DISASSOCIATED;
3752 INIT_LIST_HEAD(&gcwq->idle_list);
3753 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3754 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3756 init_timer_deferrable(&gcwq->idle_timer);
3757 gcwq->idle_timer.function = idle_worker_timeout;
3758 gcwq->idle_timer.data = (unsigned long)gcwq;
3760 setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3761 (unsigned long)gcwq);
3763 ida_init(&gcwq->worker_ida);
3765 gcwq->trustee_state = TRUSTEE_DONE;
3766 init_waitqueue_head(&gcwq->trustee_wait);
3769 /* create the initial worker */
3770 for_each_online_gcwq_cpu(cpu) {
3771 struct global_cwq *gcwq = get_gcwq(cpu);
3772 struct worker *worker;
3774 if (cpu != WORK_CPU_UNBOUND)
3775 gcwq->flags &= ~GCWQ_DISASSOCIATED;
3776 worker = create_worker(gcwq, true);
3777 BUG_ON(!worker);
3778 spin_lock_irq(&gcwq->lock);
3779 start_worker(worker);
3780 spin_unlock_irq(&gcwq->lock);
3783 system_wq = alloc_workqueue("events", 0, 0);
3784 system_long_wq = alloc_workqueue("events_long", 0, 0);
3785 system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3786 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3787 WQ_UNBOUND_MAX_ACTIVE);
3788 system_freezable_wq = alloc_workqueue("events_freezable",
3789 WQ_FREEZABLE, 0);
3790 BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq ||
3791 !system_unbound_wq || !system_freezable_wq);
3792 return 0;
3794 early_initcall(init_workqueues);