2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/module.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
45 #include "workqueue_sched.h"
48 /* global_cwq flags */
49 GCWQ_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
50 GCWQ_MANAGING_WORKERS
= 1 << 1, /* managing workers */
51 GCWQ_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
52 GCWQ_FREEZING
= 1 << 3, /* freeze in progress */
53 GCWQ_HIGHPRI_PENDING
= 1 << 4, /* highpri works on queue */
56 WORKER_STARTED
= 1 << 0, /* started */
57 WORKER_DIE
= 1 << 1, /* die die die */
58 WORKER_IDLE
= 1 << 2, /* is idle */
59 WORKER_PREP
= 1 << 3, /* preparing to run works */
60 WORKER_ROGUE
= 1 << 4, /* not bound to any cpu */
61 WORKER_REBIND
= 1 << 5, /* mom is home, come back */
62 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
63 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
65 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_ROGUE
| WORKER_REBIND
|
66 WORKER_CPU_INTENSIVE
| WORKER_UNBOUND
,
68 /* gcwq->trustee_state */
69 TRUSTEE_START
= 0, /* start */
70 TRUSTEE_IN_CHARGE
= 1, /* trustee in charge of gcwq */
71 TRUSTEE_BUTCHER
= 2, /* butcher workers */
72 TRUSTEE_RELEASE
= 3, /* release workers */
73 TRUSTEE_DONE
= 4, /* trustee is done */
75 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
76 BUSY_WORKER_HASH_SIZE
= 1 << BUSY_WORKER_HASH_ORDER
,
77 BUSY_WORKER_HASH_MASK
= BUSY_WORKER_HASH_SIZE
- 1,
79 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
80 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
82 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
83 /* call for help after 10ms
85 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
86 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
87 TRUSTEE_COOLDOWN
= HZ
/ 10, /* for trustee draining */
90 * Rescue workers are used only on emergencies and shared by
93 RESCUER_NICE_LEVEL
= -20,
97 * Structure fields follow one of the following exclusion rules.
99 * I: Modifiable by initialization/destruction paths and read-only for
102 * P: Preemption protected. Disabling preemption is enough and should
103 * only be modified and accessed from the local cpu.
105 * L: gcwq->lock protected. Access with gcwq->lock held.
107 * X: During normal operation, modification requires gcwq->lock and
108 * should be done only from local cpu. Either disabling preemption
109 * on local cpu or grabbing gcwq->lock is enough for read access.
110 * If GCWQ_DISASSOCIATED is set, it's identical to L.
112 * F: wq->flush_mutex protected.
114 * W: workqueue_lock protected.
120 * The poor guys doing the actual heavy lifting. All on-duty workers
121 * are either serving the manager role, on idle list or on busy hash.
124 /* on idle list while idle, on busy hash table while busy */
126 struct list_head entry
; /* L: while idle */
127 struct hlist_node hentry
; /* L: while busy */
130 struct work_struct
*current_work
; /* L: work being processed */
131 struct cpu_workqueue_struct
*current_cwq
; /* L: current_work's cwq */
132 struct list_head scheduled
; /* L: scheduled works */
133 struct task_struct
*task
; /* I: worker task */
134 struct global_cwq
*gcwq
; /* I: the associated gcwq */
135 /* 64 bytes boundary on 64bit, 32 on 32bit */
136 unsigned long last_active
; /* L: last active timestamp */
137 unsigned int flags
; /* X: flags */
138 int id
; /* I: worker id */
139 struct work_struct rebind_work
; /* L: rebind worker to cpu */
143 * Global per-cpu workqueue. There's one and only one for each cpu
144 * and all works are queued and processed here regardless of their
148 spinlock_t lock
; /* the gcwq lock */
149 struct list_head worklist
; /* L: list of pending works */
150 unsigned int cpu
; /* I: the associated cpu */
151 unsigned int flags
; /* L: GCWQ_* flags */
153 int nr_workers
; /* L: total number of workers */
154 int nr_idle
; /* L: currently idle ones */
156 /* workers are chained either in the idle_list or busy_hash */
157 struct list_head idle_list
; /* X: list of idle workers */
158 struct hlist_head busy_hash
[BUSY_WORKER_HASH_SIZE
];
159 /* L: hash of busy workers */
161 struct timer_list idle_timer
; /* L: worker idle timeout */
162 struct timer_list mayday_timer
; /* L: SOS timer for dworkers */
164 struct ida worker_ida
; /* L: for worker IDs */
166 struct task_struct
*trustee
; /* L: for gcwq shutdown */
167 unsigned int trustee_state
; /* L: trustee state */
168 wait_queue_head_t trustee_wait
; /* trustee wait */
169 struct worker
*first_idle
; /* L: first idle worker */
170 } ____cacheline_aligned_in_smp
;
173 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
174 * work_struct->data are used for flags and thus cwqs need to be
175 * aligned at two's power of the number of flag bits.
177 struct cpu_workqueue_struct
{
178 struct global_cwq
*gcwq
; /* I: the associated gcwq */
179 struct workqueue_struct
*wq
; /* I: the owning workqueue */
180 int work_color
; /* L: current color */
181 int flush_color
; /* L: flushing color */
182 int nr_in_flight
[WORK_NR_COLORS
];
183 /* L: nr of in_flight works */
184 int nr_active
; /* L: nr of active works */
185 int max_active
; /* L: max active works */
186 struct list_head delayed_works
; /* L: delayed works */
190 * Structure used to wait for workqueue flush.
193 struct list_head list
; /* F: list of flushers */
194 int flush_color
; /* F: flush color waiting for */
195 struct completion done
; /* flush completion */
199 * All cpumasks are assumed to be always set on UP and thus can't be
200 * used to determine whether there's something to be done.
203 typedef cpumask_var_t mayday_mask_t
;
204 #define mayday_test_and_set_cpu(cpu, mask) \
205 cpumask_test_and_set_cpu((cpu), (mask))
206 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
207 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
208 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
209 #define free_mayday_mask(mask) free_cpumask_var((mask))
211 typedef unsigned long mayday_mask_t
;
212 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
213 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
214 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
215 #define alloc_mayday_mask(maskp, gfp) true
216 #define free_mayday_mask(mask) do { } while (0)
220 * The externally visible workqueue abstraction is an array of
221 * per-CPU workqueues:
223 struct workqueue_struct
{
224 unsigned int flags
; /* I: WQ_* flags */
226 struct cpu_workqueue_struct __percpu
*pcpu
;
227 struct cpu_workqueue_struct
*single
;
229 } cpu_wq
; /* I: cwq's */
230 struct list_head list
; /* W: list of all workqueues */
232 struct mutex flush_mutex
; /* protects wq flushing */
233 int work_color
; /* F: current work color */
234 int flush_color
; /* F: current flush color */
235 atomic_t nr_cwqs_to_flush
; /* flush in progress */
236 struct wq_flusher
*first_flusher
; /* F: first flusher */
237 struct list_head flusher_queue
; /* F: flush waiters */
238 struct list_head flusher_overflow
; /* F: flush overflow list */
240 mayday_mask_t mayday_mask
; /* cpus requesting rescue */
241 struct worker
*rescuer
; /* I: rescue worker */
243 int saved_max_active
; /* W: saved cwq max_active */
244 const char *name
; /* I: workqueue name */
245 #ifdef CONFIG_LOCKDEP
246 struct lockdep_map lockdep_map
;
250 struct workqueue_struct
*system_wq __read_mostly
;
251 struct workqueue_struct
*system_long_wq __read_mostly
;
252 struct workqueue_struct
*system_nrt_wq __read_mostly
;
253 struct workqueue_struct
*system_unbound_wq __read_mostly
;
254 struct workqueue_struct
*system_freezable_wq __read_mostly
;
255 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
,
271 if (cpu
< nr_cpu_ids
) {
273 cpu
= cpumask_next(cpu
, mask
);
274 if (cpu
< nr_cpu_ids
)
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);
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
;
322 * fixup_init is called when:
323 * - an active object is initialized
325 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
327 struct work_struct
*work
= addr
;
330 case ODEBUG_STATE_ACTIVE
:
331 cancel_work_sync(work
);
332 debug_object_init(work
, &work_debug_descr
);
340 * fixup_activate is called when:
341 * - an active object is activated
342 * - an unknown object is activated (might be a statically initialized object)
344 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
346 struct work_struct
*work
= addr
;
350 case ODEBUG_STATE_NOTAVAILABLE
:
352 * This is not really a fixup. The work struct was
353 * statically initialized. We just make sure that it
354 * is tracked in the object tracker.
356 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
357 debug_object_init(work
, &work_debug_descr
);
358 debug_object_activate(work
, &work_debug_descr
);
364 case ODEBUG_STATE_ACTIVE
:
373 * fixup_free is called when:
374 * - an active object is freed
376 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
378 struct work_struct
*work
= addr
;
381 case ODEBUG_STATE_ACTIVE
:
382 cancel_work_sync(work
);
383 debug_object_free(work
, &work_debug_descr
);
390 static struct debug_obj_descr work_debug_descr
= {
391 .name
= "work_struct",
392 .fixup_init
= work_fixup_init
,
393 .fixup_activate
= work_fixup_activate
,
394 .fixup_free
= work_fixup_free
,
397 static inline void debug_work_activate(struct work_struct
*work
)
399 debug_object_activate(work
, &work_debug_descr
);
402 static inline void debug_work_deactivate(struct work_struct
*work
)
404 debug_object_deactivate(work
, &work_debug_descr
);
407 void __init_work(struct work_struct
*work
, int onstack
)
410 debug_object_init_on_stack(work
, &work_debug_descr
);
412 debug_object_init(work
, &work_debug_descr
);
414 EXPORT_SYMBOL_GPL(__init_work
);
416 void destroy_work_on_stack(struct work_struct
*work
)
418 debug_object_free(work
, &work_debug_descr
);
420 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
423 static inline void debug_work_activate(struct work_struct
*work
) { }
424 static inline void debug_work_deactivate(struct work_struct
*work
) { }
427 /* Serializes the accesses to the list of workqueues. */
428 static DEFINE_SPINLOCK(workqueue_lock
);
429 static LIST_HEAD(workqueues
);
430 static bool workqueue_freezing
; /* W: have wqs started freezing? */
433 * The almighty global cpu workqueues. nr_running is the only field
434 * which is expected to be used frequently by other cpus via
435 * try_to_wake_up(). Put it in a separate cacheline.
437 static DEFINE_PER_CPU(struct global_cwq
, global_cwq
);
438 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t
, gcwq_nr_running
);
441 * Global cpu workqueue and nr_running counter for unbound gcwq. The
442 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
443 * workers have WORKER_UNBOUND set.
445 static struct global_cwq unbound_global_cwq
;
446 static atomic_t unbound_gcwq_nr_running
= ATOMIC_INIT(0); /* always 0 */
448 static int worker_thread(void *__worker
);
450 static struct global_cwq
*get_gcwq(unsigned int cpu
)
452 if (cpu
!= WORK_CPU_UNBOUND
)
453 return &per_cpu(global_cwq
, cpu
);
455 return &unbound_global_cwq
;
458 static atomic_t
*get_gcwq_nr_running(unsigned int cpu
)
460 if (cpu
!= WORK_CPU_UNBOUND
)
461 return &per_cpu(gcwq_nr_running
, cpu
);
463 return &unbound_gcwq_nr_running
;
466 static struct cpu_workqueue_struct
*get_cwq(unsigned int cpu
,
467 struct workqueue_struct
*wq
)
469 if (!(wq
->flags
& WQ_UNBOUND
)) {
470 if (likely(cpu
< nr_cpu_ids
)) {
472 return per_cpu_ptr(wq
->cpu_wq
.pcpu
, cpu
);
474 return wq
->cpu_wq
.single
;
477 } else if (likely(cpu
== WORK_CPU_UNBOUND
))
478 return wq
->cpu_wq
.single
;
482 static unsigned int work_color_to_flags(int color
)
484 return color
<< WORK_STRUCT_COLOR_SHIFT
;
487 static int get_work_color(struct work_struct
*work
)
489 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
490 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
493 static int work_next_color(int color
)
495 return (color
+ 1) % WORK_NR_COLORS
;
499 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
500 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
501 * cleared and the work data contains the cpu number it was last on.
503 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
504 * cwq, cpu or clear work->data. These functions should only be
505 * called while the work is owned - ie. while the PENDING bit is set.
507 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
508 * corresponding to a work. gcwq is available once the work has been
509 * queued anywhere after initialization. cwq is available only from
510 * queueing until execution starts.
512 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
515 BUG_ON(!work_pending(work
));
516 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
519 static void set_work_cwq(struct work_struct
*work
,
520 struct cpu_workqueue_struct
*cwq
,
521 unsigned long extra_flags
)
523 set_work_data(work
, (unsigned long)cwq
,
524 WORK_STRUCT_PENDING
| WORK_STRUCT_CWQ
| extra_flags
);
527 static void set_work_cpu(struct work_struct
*work
, unsigned int cpu
)
529 set_work_data(work
, cpu
<< WORK_STRUCT_FLAG_BITS
, WORK_STRUCT_PENDING
);
532 static void clear_work_data(struct work_struct
*work
)
534 set_work_data(work
, WORK_STRUCT_NO_CPU
, 0);
537 static struct cpu_workqueue_struct
*get_work_cwq(struct work_struct
*work
)
539 unsigned long data
= atomic_long_read(&work
->data
);
541 if (data
& WORK_STRUCT_CWQ
)
542 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
547 static struct global_cwq
*get_work_gcwq(struct work_struct
*work
)
549 unsigned long data
= atomic_long_read(&work
->data
);
552 if (data
& WORK_STRUCT_CWQ
)
553 return ((struct cpu_workqueue_struct
*)
554 (data
& WORK_STRUCT_WQ_DATA_MASK
))->gcwq
;
556 cpu
= data
>> WORK_STRUCT_FLAG_BITS
;
557 if (cpu
== WORK_CPU_NONE
)
560 BUG_ON(cpu
>= nr_cpu_ids
&& cpu
!= WORK_CPU_UNBOUND
);
561 return get_gcwq(cpu
);
565 * Policy functions. These define the policies on how the global
566 * worker pool is managed. Unless noted otherwise, these functions
567 * assume that they're being called with gcwq->lock held.
570 static bool __need_more_worker(struct global_cwq
*gcwq
)
572 return !atomic_read(get_gcwq_nr_running(gcwq
->cpu
)) ||
573 gcwq
->flags
& GCWQ_HIGHPRI_PENDING
;
577 * Need to wake up a worker? Called from anything but currently
580 static bool need_more_worker(struct global_cwq
*gcwq
)
582 return !list_empty(&gcwq
->worklist
) && __need_more_worker(gcwq
);
585 /* Can I start working? Called from busy but !running workers. */
586 static bool may_start_working(struct global_cwq
*gcwq
)
588 return gcwq
->nr_idle
;
591 /* Do I need to keep working? Called from currently running workers. */
592 static bool keep_working(struct global_cwq
*gcwq
)
594 atomic_t
*nr_running
= get_gcwq_nr_running(gcwq
->cpu
);
596 return !list_empty(&gcwq
->worklist
) &&
597 (atomic_read(nr_running
) <= 1 ||
598 gcwq
->flags
& GCWQ_HIGHPRI_PENDING
);
601 /* Do we need a new worker? Called from manager. */
602 static bool need_to_create_worker(struct global_cwq
*gcwq
)
604 return need_more_worker(gcwq
) && !may_start_working(gcwq
);
607 /* Do I need to be the manager? */
608 static bool need_to_manage_workers(struct global_cwq
*gcwq
)
610 return need_to_create_worker(gcwq
) || gcwq
->flags
& GCWQ_MANAGE_WORKERS
;
613 /* Do we have too many workers and should some go away? */
614 static bool too_many_workers(struct global_cwq
*gcwq
)
616 bool managing
= gcwq
->flags
& GCWQ_MANAGING_WORKERS
;
617 int nr_idle
= gcwq
->nr_idle
+ managing
; /* manager is considered idle */
618 int nr_busy
= gcwq
->nr_workers
- nr_idle
;
620 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
627 /* Return the first worker. Safe with preemption disabled */
628 static struct worker
*first_worker(struct global_cwq
*gcwq
)
630 if (unlikely(list_empty(&gcwq
->idle_list
)))
633 return list_first_entry(&gcwq
->idle_list
, struct worker
, entry
);
637 * wake_up_worker - wake up an idle worker
638 * @gcwq: gcwq to wake worker for
640 * Wake up the first idle worker of @gcwq.
643 * spin_lock_irq(gcwq->lock).
645 static void wake_up_worker(struct global_cwq
*gcwq
)
647 struct worker
*worker
= first_worker(gcwq
);
650 wake_up_process(worker
->task
);
654 * wq_worker_waking_up - a worker is waking up
655 * @task: task waking up
656 * @cpu: CPU @task is waking up to
658 * This function is called during try_to_wake_up() when a worker is
662 * spin_lock_irq(rq->lock)
664 void wq_worker_waking_up(struct task_struct
*task
, unsigned int cpu
)
666 struct worker
*worker
= kthread_data(task
);
668 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
669 atomic_inc(get_gcwq_nr_running(cpu
));
673 * wq_worker_sleeping - a worker is going to sleep
674 * @task: task going to sleep
675 * @cpu: CPU in question, must be the current CPU number
677 * This function is called during schedule() when a busy worker is
678 * going to sleep. Worker on the same cpu can be woken up by
679 * returning pointer to its task.
682 * spin_lock_irq(rq->lock)
685 * Worker task on @cpu to wake up, %NULL if none.
687 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
,
690 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
691 struct global_cwq
*gcwq
= get_gcwq(cpu
);
692 atomic_t
*nr_running
= get_gcwq_nr_running(cpu
);
694 if (worker
->flags
& WORKER_NOT_RUNNING
)
697 /* this can only happen on the local cpu */
698 BUG_ON(cpu
!= raw_smp_processor_id());
701 * The counterpart of the following dec_and_test, implied mb,
702 * worklist not empty test sequence is in insert_work().
703 * Please read comment there.
705 * NOT_RUNNING is clear. This means that trustee is not in
706 * charge and we're running on the local cpu w/ rq lock held
707 * and preemption disabled, which in turn means that none else
708 * could be manipulating idle_list, so dereferencing idle_list
709 * without gcwq lock is safe.
711 if (atomic_dec_and_test(nr_running
) && !list_empty(&gcwq
->worklist
))
712 to_wakeup
= first_worker(gcwq
);
713 return to_wakeup
? to_wakeup
->task
: NULL
;
717 * worker_set_flags - set worker flags and adjust nr_running accordingly
719 * @flags: flags to set
720 * @wakeup: wakeup an idle worker if necessary
722 * Set @flags in @worker->flags and adjust nr_running accordingly. If
723 * nr_running becomes zero and @wakeup is %true, an idle worker is
727 * spin_lock_irq(gcwq->lock)
729 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
732 struct global_cwq
*gcwq
= worker
->gcwq
;
734 WARN_ON_ONCE(worker
->task
!= current
);
737 * If transitioning into NOT_RUNNING, adjust nr_running and
738 * wake up an idle worker as necessary if requested by
741 if ((flags
& WORKER_NOT_RUNNING
) &&
742 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
743 atomic_t
*nr_running
= get_gcwq_nr_running(gcwq
->cpu
);
746 if (atomic_dec_and_test(nr_running
) &&
747 !list_empty(&gcwq
->worklist
))
748 wake_up_worker(gcwq
);
750 atomic_dec(nr_running
);
753 worker
->flags
|= flags
;
757 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
759 * @flags: flags to clear
761 * Clear @flags in @worker->flags and adjust nr_running accordingly.
764 * spin_lock_irq(gcwq->lock)
766 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
768 struct global_cwq
*gcwq
= worker
->gcwq
;
769 unsigned int oflags
= worker
->flags
;
771 WARN_ON_ONCE(worker
->task
!= current
);
773 worker
->flags
&= ~flags
;
776 * If transitioning out of NOT_RUNNING, increment nr_running. Note
777 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
778 * of multiple flags, not a single flag.
780 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
781 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
782 atomic_inc(get_gcwq_nr_running(gcwq
->cpu
));
786 * busy_worker_head - return the busy hash head for a work
787 * @gcwq: gcwq of interest
788 * @work: work to be hashed
790 * Return hash head of @gcwq for @work.
793 * spin_lock_irq(gcwq->lock).
796 * Pointer to the hash head.
798 static struct hlist_head
*busy_worker_head(struct global_cwq
*gcwq
,
799 struct work_struct
*work
)
801 const int base_shift
= ilog2(sizeof(struct work_struct
));
802 unsigned long v
= (unsigned long)work
;
804 /* simple shift and fold hash, do we need something better? */
806 v
+= v
>> BUSY_WORKER_HASH_ORDER
;
807 v
&= BUSY_WORKER_HASH_MASK
;
809 return &gcwq
->busy_hash
[v
];
813 * __find_worker_executing_work - find worker which is executing a work
814 * @gcwq: gcwq of interest
815 * @bwh: hash head as returned by busy_worker_head()
816 * @work: work to find worker for
818 * Find a worker which is executing @work on @gcwq. @bwh should be
819 * the hash head obtained by calling busy_worker_head() with the same
823 * spin_lock_irq(gcwq->lock).
826 * Pointer to worker which is executing @work if found, NULL
829 static struct worker
*__find_worker_executing_work(struct global_cwq
*gcwq
,
830 struct hlist_head
*bwh
,
831 struct work_struct
*work
)
833 struct worker
*worker
;
834 struct hlist_node
*tmp
;
836 hlist_for_each_entry(worker
, tmp
, bwh
, hentry
)
837 if (worker
->current_work
== work
)
843 * find_worker_executing_work - find worker which is executing a work
844 * @gcwq: gcwq of interest
845 * @work: work to find worker for
847 * Find a worker which is executing @work on @gcwq. This function is
848 * identical to __find_worker_executing_work() except that this
849 * function calculates @bwh itself.
852 * spin_lock_irq(gcwq->lock).
855 * Pointer to worker which is executing @work if found, NULL
858 static struct worker
*find_worker_executing_work(struct global_cwq
*gcwq
,
859 struct work_struct
*work
)
861 return __find_worker_executing_work(gcwq
, busy_worker_head(gcwq
, work
),
866 * gcwq_determine_ins_pos - find insertion position
867 * @gcwq: gcwq of interest
868 * @cwq: cwq a work is being queued for
870 * A work for @cwq is about to be queued on @gcwq, determine insertion
871 * position for the work. If @cwq is for HIGHPRI wq, the work is
872 * queued at the head of the queue but in FIFO order with respect to
873 * other HIGHPRI works; otherwise, at the end of the queue. This
874 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
875 * there are HIGHPRI works pending.
878 * spin_lock_irq(gcwq->lock).
881 * Pointer to inserstion position.
883 static inline struct list_head
*gcwq_determine_ins_pos(struct global_cwq
*gcwq
,
884 struct cpu_workqueue_struct
*cwq
)
886 struct work_struct
*twork
;
888 if (likely(!(cwq
->wq
->flags
& WQ_HIGHPRI
)))
889 return &gcwq
->worklist
;
891 list_for_each_entry(twork
, &gcwq
->worklist
, entry
) {
892 struct cpu_workqueue_struct
*tcwq
= get_work_cwq(twork
);
894 if (!(tcwq
->wq
->flags
& WQ_HIGHPRI
))
898 gcwq
->flags
|= GCWQ_HIGHPRI_PENDING
;
899 return &twork
->entry
;
903 * insert_work - insert a work into gcwq
904 * @cwq: cwq @work belongs to
905 * @work: work to insert
906 * @head: insertion point
907 * @extra_flags: extra WORK_STRUCT_* flags to set
909 * Insert @work which belongs to @cwq into @gcwq after @head.
910 * @extra_flags is or'd to work_struct flags.
913 * spin_lock_irq(gcwq->lock).
915 static void insert_work(struct cpu_workqueue_struct
*cwq
,
916 struct work_struct
*work
, struct list_head
*head
,
917 unsigned int extra_flags
)
919 struct global_cwq
*gcwq
= cwq
->gcwq
;
921 /* we own @work, set data and link */
922 set_work_cwq(work
, cwq
, extra_flags
);
925 * Ensure that we get the right work->data if we see the
926 * result of list_add() below, see try_to_grab_pending().
930 list_add_tail(&work
->entry
, head
);
933 * Ensure either worker_sched_deactivated() sees the above
934 * list_add_tail() or we see zero nr_running to avoid workers
935 * lying around lazily while there are works to be processed.
939 if (__need_more_worker(gcwq
))
940 wake_up_worker(gcwq
);
944 * Test whether @work is being queued from another work executing on the
945 * same workqueue. This is rather expensive and should only be used from
948 static bool is_chained_work(struct workqueue_struct
*wq
)
953 for_each_gcwq_cpu(cpu
) {
954 struct global_cwq
*gcwq
= get_gcwq(cpu
);
955 struct worker
*worker
;
956 struct hlist_node
*pos
;
959 spin_lock_irqsave(&gcwq
->lock
, flags
);
960 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
961 if (worker
->task
!= current
)
963 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
965 * I'm @worker, no locking necessary. See if @work
966 * is headed to the same workqueue.
968 return worker
->current_cwq
->wq
== wq
;
970 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
975 static void __queue_work(unsigned int cpu
, struct workqueue_struct
*wq
,
976 struct work_struct
*work
)
978 struct global_cwq
*gcwq
;
979 struct cpu_workqueue_struct
*cwq
;
980 struct list_head
*worklist
;
981 unsigned int work_flags
;
984 debug_work_activate(work
);
986 /* if dying, only works from the same workqueue are allowed */
987 if (unlikely(wq
->flags
& WQ_DYING
) &&
988 WARN_ON_ONCE(!is_chained_work(wq
)))
991 /* determine gcwq to use */
992 if (!(wq
->flags
& WQ_UNBOUND
)) {
993 struct global_cwq
*last_gcwq
;
995 if (unlikely(cpu
== WORK_CPU_UNBOUND
))
996 cpu
= raw_smp_processor_id();
999 * It's multi cpu. If @wq is non-reentrant and @work
1000 * was previously on a different cpu, it might still
1001 * be running there, in which case the work needs to
1002 * be queued on that cpu to guarantee non-reentrance.
1004 gcwq
= get_gcwq(cpu
);
1005 if (wq
->flags
& WQ_NON_REENTRANT
&&
1006 (last_gcwq
= get_work_gcwq(work
)) && last_gcwq
!= gcwq
) {
1007 struct worker
*worker
;
1009 spin_lock_irqsave(&last_gcwq
->lock
, flags
);
1011 worker
= find_worker_executing_work(last_gcwq
, work
);
1013 if (worker
&& worker
->current_cwq
->wq
== wq
)
1016 /* meh... not running there, queue here */
1017 spin_unlock_irqrestore(&last_gcwq
->lock
, flags
);
1018 spin_lock_irqsave(&gcwq
->lock
, flags
);
1021 spin_lock_irqsave(&gcwq
->lock
, flags
);
1023 gcwq
= get_gcwq(WORK_CPU_UNBOUND
);
1024 spin_lock_irqsave(&gcwq
->lock
, flags
);
1027 /* gcwq determined, get cwq and queue */
1028 cwq
= get_cwq(gcwq
->cpu
, wq
);
1029 trace_workqueue_queue_work(cpu
, cwq
, work
);
1031 BUG_ON(!list_empty(&work
->entry
));
1033 cwq
->nr_in_flight
[cwq
->work_color
]++;
1034 work_flags
= work_color_to_flags(cwq
->work_color
);
1036 if (likely(cwq
->nr_active
< cwq
->max_active
)) {
1037 trace_workqueue_activate_work(work
);
1039 worklist
= gcwq_determine_ins_pos(gcwq
, cwq
);
1041 work_flags
|= WORK_STRUCT_DELAYED
;
1042 worklist
= &cwq
->delayed_works
;
1045 insert_work(cwq
, work
, worklist
, work_flags
);
1047 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1051 * queue_work - queue work on a workqueue
1052 * @wq: workqueue to use
1053 * @work: work to queue
1055 * Returns 0 if @work was already on a queue, non-zero otherwise.
1057 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1058 * it can be processed by another CPU.
1060 int queue_work(struct workqueue_struct
*wq
, struct work_struct
*work
)
1064 ret
= queue_work_on(get_cpu(), wq
, work
);
1069 EXPORT_SYMBOL_GPL(queue_work
);
1072 * queue_work_on - queue work on specific cpu
1073 * @cpu: CPU number to execute work on
1074 * @wq: workqueue to use
1075 * @work: work to queue
1077 * Returns 0 if @work was already on a queue, non-zero otherwise.
1079 * We queue the work to a specific CPU, the caller must ensure it
1083 queue_work_on(int cpu
, struct workqueue_struct
*wq
, struct work_struct
*work
)
1087 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1088 __queue_work(cpu
, wq
, work
);
1093 EXPORT_SYMBOL_GPL(queue_work_on
);
1095 static void delayed_work_timer_fn(unsigned long __data
)
1097 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1098 struct cpu_workqueue_struct
*cwq
= get_work_cwq(&dwork
->work
);
1100 __queue_work(smp_processor_id(), cwq
->wq
, &dwork
->work
);
1104 * queue_delayed_work - queue work on a workqueue after delay
1105 * @wq: workqueue to use
1106 * @dwork: delayable work to queue
1107 * @delay: number of jiffies to wait before queueing
1109 * Returns 0 if @work was already on a queue, non-zero otherwise.
1111 int queue_delayed_work(struct workqueue_struct
*wq
,
1112 struct delayed_work
*dwork
, unsigned long delay
)
1115 return queue_work(wq
, &dwork
->work
);
1117 return queue_delayed_work_on(-1, wq
, dwork
, delay
);
1119 EXPORT_SYMBOL_GPL(queue_delayed_work
);
1122 * queue_delayed_work_on - queue work on specific CPU after delay
1123 * @cpu: CPU number to execute work on
1124 * @wq: workqueue to use
1125 * @dwork: work to queue
1126 * @delay: number of jiffies to wait before queueing
1128 * Returns 0 if @work was already on a queue, non-zero otherwise.
1130 int queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1131 struct delayed_work
*dwork
, unsigned long delay
)
1134 struct timer_list
*timer
= &dwork
->timer
;
1135 struct work_struct
*work
= &dwork
->work
;
1137 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1140 BUG_ON(timer_pending(timer
));
1141 BUG_ON(!list_empty(&work
->entry
));
1143 timer_stats_timer_set_start_info(&dwork
->timer
);
1146 * This stores cwq for the moment, for the timer_fn.
1147 * Note that the work's gcwq is preserved to allow
1148 * reentrance detection for delayed works.
1150 if (!(wq
->flags
& WQ_UNBOUND
)) {
1151 struct global_cwq
*gcwq
= get_work_gcwq(work
);
1153 if (gcwq
&& gcwq
->cpu
!= WORK_CPU_UNBOUND
)
1156 lcpu
= raw_smp_processor_id();
1158 lcpu
= WORK_CPU_UNBOUND
;
1160 set_work_cwq(work
, get_cwq(lcpu
, wq
), 0);
1162 timer
->expires
= jiffies
+ delay
;
1163 timer
->data
= (unsigned long)dwork
;
1164 timer
->function
= delayed_work_timer_fn
;
1166 if (unlikely(cpu
>= 0))
1167 add_timer_on(timer
, cpu
);
1174 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1177 * worker_enter_idle - enter idle state
1178 * @worker: worker which is entering idle state
1180 * @worker is entering idle state. Update stats and idle timer if
1184 * spin_lock_irq(gcwq->lock).
1186 static void worker_enter_idle(struct worker
*worker
)
1188 struct global_cwq
*gcwq
= worker
->gcwq
;
1190 BUG_ON(worker
->flags
& WORKER_IDLE
);
1191 BUG_ON(!list_empty(&worker
->entry
) &&
1192 (worker
->hentry
.next
|| worker
->hentry
.pprev
));
1194 /* can't use worker_set_flags(), also called from start_worker() */
1195 worker
->flags
|= WORKER_IDLE
;
1197 worker
->last_active
= jiffies
;
1199 /* idle_list is LIFO */
1200 list_add(&worker
->entry
, &gcwq
->idle_list
);
1202 if (likely(!(worker
->flags
& WORKER_ROGUE
))) {
1203 if (too_many_workers(gcwq
) && !timer_pending(&gcwq
->idle_timer
))
1204 mod_timer(&gcwq
->idle_timer
,
1205 jiffies
+ IDLE_WORKER_TIMEOUT
);
1207 wake_up_all(&gcwq
->trustee_wait
);
1209 /* sanity check nr_running */
1210 WARN_ON_ONCE(gcwq
->nr_workers
== gcwq
->nr_idle
&&
1211 atomic_read(get_gcwq_nr_running(gcwq
->cpu
)));
1215 * worker_leave_idle - leave idle state
1216 * @worker: worker which is leaving idle state
1218 * @worker is leaving idle state. Update stats.
1221 * spin_lock_irq(gcwq->lock).
1223 static void worker_leave_idle(struct worker
*worker
)
1225 struct global_cwq
*gcwq
= worker
->gcwq
;
1227 BUG_ON(!(worker
->flags
& WORKER_IDLE
));
1228 worker_clr_flags(worker
, WORKER_IDLE
);
1230 list_del_init(&worker
->entry
);
1234 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1237 * Works which are scheduled while the cpu is online must at least be
1238 * scheduled to a worker which is bound to the cpu so that if they are
1239 * flushed from cpu callbacks while cpu is going down, they are
1240 * guaranteed to execute on the cpu.
1242 * This function is to be used by rogue workers and rescuers to bind
1243 * themselves to the target cpu and may race with cpu going down or
1244 * coming online. kthread_bind() can't be used because it may put the
1245 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1246 * verbatim as it's best effort and blocking and gcwq may be
1247 * [dis]associated in the meantime.
1249 * This function tries set_cpus_allowed() and locks gcwq and verifies
1250 * the binding against GCWQ_DISASSOCIATED which is set during
1251 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1252 * idle state or fetches works without dropping lock, it can guarantee
1253 * the scheduling requirement described in the first paragraph.
1256 * Might sleep. Called without any lock but returns with gcwq->lock
1260 * %true if the associated gcwq is online (@worker is successfully
1261 * bound), %false if offline.
1263 static bool worker_maybe_bind_and_lock(struct worker
*worker
)
1264 __acquires(&gcwq
->lock
)
1266 struct global_cwq
*gcwq
= worker
->gcwq
;
1267 struct task_struct
*task
= worker
->task
;
1271 * The following call may fail, succeed or succeed
1272 * without actually migrating the task to the cpu if
1273 * it races with cpu hotunplug operation. Verify
1274 * against GCWQ_DISASSOCIATED.
1276 if (!(gcwq
->flags
& GCWQ_DISASSOCIATED
))
1277 set_cpus_allowed_ptr(task
, get_cpu_mask(gcwq
->cpu
));
1279 spin_lock_irq(&gcwq
->lock
);
1280 if (gcwq
->flags
& GCWQ_DISASSOCIATED
)
1282 if (task_cpu(task
) == gcwq
->cpu
&&
1283 cpumask_equal(¤t
->cpus_allowed
,
1284 get_cpu_mask(gcwq
->cpu
)))
1286 spin_unlock_irq(&gcwq
->lock
);
1288 /* CPU has come up inbetween, retry migration */
1294 * Function for worker->rebind_work used to rebind rogue busy workers
1295 * to the associated cpu which is coming back online. This is
1296 * scheduled by cpu up but can race with other cpu hotplug operations
1297 * and may be executed twice without intervening cpu down.
1299 static void worker_rebind_fn(struct work_struct
*work
)
1301 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1302 struct global_cwq
*gcwq
= worker
->gcwq
;
1304 if (worker_maybe_bind_and_lock(worker
))
1305 worker_clr_flags(worker
, WORKER_REBIND
);
1307 spin_unlock_irq(&gcwq
->lock
);
1310 static struct worker
*alloc_worker(void)
1312 struct worker
*worker
;
1314 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1316 INIT_LIST_HEAD(&worker
->entry
);
1317 INIT_LIST_HEAD(&worker
->scheduled
);
1318 INIT_WORK(&worker
->rebind_work
, worker_rebind_fn
);
1319 /* on creation a worker is in !idle && prep state */
1320 worker
->flags
= WORKER_PREP
;
1326 * create_worker - create a new workqueue worker
1327 * @gcwq: gcwq the new worker will belong to
1328 * @bind: whether to set affinity to @cpu or not
1330 * Create a new worker which is bound to @gcwq. The returned worker
1331 * can be started by calling start_worker() or destroyed using
1335 * Might sleep. Does GFP_KERNEL allocations.
1338 * Pointer to the newly created worker.
1340 static struct worker
*create_worker(struct global_cwq
*gcwq
, bool bind
)
1342 bool on_unbound_cpu
= gcwq
->cpu
== WORK_CPU_UNBOUND
;
1343 struct worker
*worker
= NULL
;
1346 spin_lock_irq(&gcwq
->lock
);
1347 while (ida_get_new(&gcwq
->worker_ida
, &id
)) {
1348 spin_unlock_irq(&gcwq
->lock
);
1349 if (!ida_pre_get(&gcwq
->worker_ida
, GFP_KERNEL
))
1351 spin_lock_irq(&gcwq
->lock
);
1353 spin_unlock_irq(&gcwq
->lock
);
1355 worker
= alloc_worker();
1359 worker
->gcwq
= gcwq
;
1362 if (!on_unbound_cpu
)
1363 worker
->task
= kthread_create(worker_thread
, worker
,
1364 "kworker/%u:%d", gcwq
->cpu
, id
);
1366 worker
->task
= kthread_create(worker_thread
, worker
,
1367 "kworker/u:%d", id
);
1368 if (IS_ERR(worker
->task
))
1372 * A rogue worker will become a regular one if CPU comes
1373 * online later on. Make sure every worker has
1374 * PF_THREAD_BOUND set.
1376 if (bind
&& !on_unbound_cpu
)
1377 kthread_bind(worker
->task
, gcwq
->cpu
);
1379 worker
->task
->flags
|= PF_THREAD_BOUND
;
1381 worker
->flags
|= WORKER_UNBOUND
;
1387 spin_lock_irq(&gcwq
->lock
);
1388 ida_remove(&gcwq
->worker_ida
, id
);
1389 spin_unlock_irq(&gcwq
->lock
);
1396 * start_worker - start a newly created worker
1397 * @worker: worker to start
1399 * Make the gcwq aware of @worker and start it.
1402 * spin_lock_irq(gcwq->lock).
1404 static void start_worker(struct worker
*worker
)
1406 worker
->flags
|= WORKER_STARTED
;
1407 worker
->gcwq
->nr_workers
++;
1408 worker_enter_idle(worker
);
1409 wake_up_process(worker
->task
);
1413 * destroy_worker - destroy a workqueue worker
1414 * @worker: worker to be destroyed
1416 * Destroy @worker and adjust @gcwq stats accordingly.
1419 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1421 static void destroy_worker(struct worker
*worker
)
1423 struct global_cwq
*gcwq
= worker
->gcwq
;
1424 int id
= worker
->id
;
1426 /* sanity check frenzy */
1427 BUG_ON(worker
->current_work
);
1428 BUG_ON(!list_empty(&worker
->scheduled
));
1430 if (worker
->flags
& WORKER_STARTED
)
1432 if (worker
->flags
& WORKER_IDLE
)
1435 list_del_init(&worker
->entry
);
1436 worker
->flags
|= WORKER_DIE
;
1438 spin_unlock_irq(&gcwq
->lock
);
1440 kthread_stop(worker
->task
);
1443 spin_lock_irq(&gcwq
->lock
);
1444 ida_remove(&gcwq
->worker_ida
, id
);
1447 static void idle_worker_timeout(unsigned long __gcwq
)
1449 struct global_cwq
*gcwq
= (void *)__gcwq
;
1451 spin_lock_irq(&gcwq
->lock
);
1453 if (too_many_workers(gcwq
)) {
1454 struct worker
*worker
;
1455 unsigned long expires
;
1457 /* idle_list is kept in LIFO order, check the last one */
1458 worker
= list_entry(gcwq
->idle_list
.prev
, struct worker
, entry
);
1459 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1461 if (time_before(jiffies
, expires
))
1462 mod_timer(&gcwq
->idle_timer
, expires
);
1464 /* it's been idle for too long, wake up manager */
1465 gcwq
->flags
|= GCWQ_MANAGE_WORKERS
;
1466 wake_up_worker(gcwq
);
1470 spin_unlock_irq(&gcwq
->lock
);
1473 static bool send_mayday(struct work_struct
*work
)
1475 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1476 struct workqueue_struct
*wq
= cwq
->wq
;
1479 if (!(wq
->flags
& WQ_RESCUER
))
1482 /* mayday mayday mayday */
1483 cpu
= cwq
->gcwq
->cpu
;
1484 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1485 if (cpu
== WORK_CPU_UNBOUND
)
1487 if (!mayday_test_and_set_cpu(cpu
, wq
->mayday_mask
))
1488 wake_up_process(wq
->rescuer
->task
);
1492 static void gcwq_mayday_timeout(unsigned long __gcwq
)
1494 struct global_cwq
*gcwq
= (void *)__gcwq
;
1495 struct work_struct
*work
;
1497 spin_lock_irq(&gcwq
->lock
);
1499 if (need_to_create_worker(gcwq
)) {
1501 * We've been trying to create a new worker but
1502 * haven't been successful. We might be hitting an
1503 * allocation deadlock. Send distress signals to
1506 list_for_each_entry(work
, &gcwq
->worklist
, entry
)
1510 spin_unlock_irq(&gcwq
->lock
);
1512 mod_timer(&gcwq
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1516 * maybe_create_worker - create a new worker if necessary
1517 * @gcwq: gcwq to create a new worker for
1519 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1520 * have at least one idle worker on return from this function. If
1521 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1522 * sent to all rescuers with works scheduled on @gcwq to resolve
1523 * possible allocation deadlock.
1525 * On return, need_to_create_worker() is guaranteed to be false and
1526 * may_start_working() true.
1529 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1530 * multiple times. Does GFP_KERNEL allocations. Called only from
1534 * false if no action was taken and gcwq->lock stayed locked, true
1537 static bool maybe_create_worker(struct global_cwq
*gcwq
)
1538 __releases(&gcwq
->lock
)
1539 __acquires(&gcwq
->lock
)
1541 if (!need_to_create_worker(gcwq
))
1544 spin_unlock_irq(&gcwq
->lock
);
1546 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1547 mod_timer(&gcwq
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1550 struct worker
*worker
;
1552 worker
= create_worker(gcwq
, true);
1554 del_timer_sync(&gcwq
->mayday_timer
);
1555 spin_lock_irq(&gcwq
->lock
);
1556 start_worker(worker
);
1557 BUG_ON(need_to_create_worker(gcwq
));
1561 if (!need_to_create_worker(gcwq
))
1564 __set_current_state(TASK_INTERRUPTIBLE
);
1565 schedule_timeout(CREATE_COOLDOWN
);
1567 if (!need_to_create_worker(gcwq
))
1571 del_timer_sync(&gcwq
->mayday_timer
);
1572 spin_lock_irq(&gcwq
->lock
);
1573 if (need_to_create_worker(gcwq
))
1579 * maybe_destroy_worker - destroy workers which have been idle for a while
1580 * @gcwq: gcwq to destroy workers for
1582 * Destroy @gcwq workers which have been idle for longer than
1583 * IDLE_WORKER_TIMEOUT.
1586 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1587 * multiple times. Called only from manager.
1590 * false if no action was taken and gcwq->lock stayed locked, true
1593 static bool maybe_destroy_workers(struct global_cwq
*gcwq
)
1597 while (too_many_workers(gcwq
)) {
1598 struct worker
*worker
;
1599 unsigned long expires
;
1601 worker
= list_entry(gcwq
->idle_list
.prev
, struct worker
, entry
);
1602 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1604 if (time_before(jiffies
, expires
)) {
1605 mod_timer(&gcwq
->idle_timer
, expires
);
1609 destroy_worker(worker
);
1617 * manage_workers - manage worker pool
1620 * Assume the manager role and manage gcwq worker pool @worker belongs
1621 * to. At any given time, there can be only zero or one manager per
1622 * gcwq. The exclusion is handled automatically by this function.
1624 * The caller can safely start processing works on false return. On
1625 * true return, it's guaranteed that need_to_create_worker() is false
1626 * and may_start_working() is true.
1629 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1630 * multiple times. Does GFP_KERNEL allocations.
1633 * false if no action was taken and gcwq->lock stayed locked, true if
1634 * some action was taken.
1636 static bool manage_workers(struct worker
*worker
)
1638 struct global_cwq
*gcwq
= worker
->gcwq
;
1641 if (gcwq
->flags
& GCWQ_MANAGING_WORKERS
)
1644 gcwq
->flags
&= ~GCWQ_MANAGE_WORKERS
;
1645 gcwq
->flags
|= GCWQ_MANAGING_WORKERS
;
1648 * Destroy and then create so that may_start_working() is true
1651 ret
|= maybe_destroy_workers(gcwq
);
1652 ret
|= maybe_create_worker(gcwq
);
1654 gcwq
->flags
&= ~GCWQ_MANAGING_WORKERS
;
1657 * The trustee might be waiting to take over the manager
1658 * position, tell it we're done.
1660 if (unlikely(gcwq
->trustee
))
1661 wake_up_all(&gcwq
->trustee_wait
);
1667 * move_linked_works - move linked works to a list
1668 * @work: start of series of works to be scheduled
1669 * @head: target list to append @work to
1670 * @nextp: out paramter for nested worklist walking
1672 * Schedule linked works starting from @work to @head. Work series to
1673 * be scheduled starts at @work and includes any consecutive work with
1674 * WORK_STRUCT_LINKED set in its predecessor.
1676 * If @nextp is not NULL, it's updated to point to the next work of
1677 * the last scheduled work. This allows move_linked_works() to be
1678 * nested inside outer list_for_each_entry_safe().
1681 * spin_lock_irq(gcwq->lock).
1683 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1684 struct work_struct
**nextp
)
1686 struct work_struct
*n
;
1689 * Linked worklist will always end before the end of the list,
1690 * use NULL for list head.
1692 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1693 list_move_tail(&work
->entry
, head
);
1694 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1699 * If we're already inside safe list traversal and have moved
1700 * multiple works to the scheduled queue, the next position
1701 * needs to be updated.
1707 static void cwq_activate_first_delayed(struct cpu_workqueue_struct
*cwq
)
1709 struct work_struct
*work
= list_first_entry(&cwq
->delayed_works
,
1710 struct work_struct
, entry
);
1711 struct list_head
*pos
= gcwq_determine_ins_pos(cwq
->gcwq
, cwq
);
1713 trace_workqueue_activate_work(work
);
1714 move_linked_works(work
, pos
, NULL
);
1715 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1720 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1721 * @cwq: cwq of interest
1722 * @color: color of work which left the queue
1723 * @delayed: for a delayed work
1725 * A work either has completed or is removed from pending queue,
1726 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1729 * spin_lock_irq(gcwq->lock).
1731 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct
*cwq
, int color
,
1734 /* ignore uncolored works */
1735 if (color
== WORK_NO_COLOR
)
1738 cwq
->nr_in_flight
[color
]--;
1742 if (!list_empty(&cwq
->delayed_works
)) {
1743 /* one down, submit a delayed one */
1744 if (cwq
->nr_active
< cwq
->max_active
)
1745 cwq_activate_first_delayed(cwq
);
1749 /* is flush in progress and are we at the flushing tip? */
1750 if (likely(cwq
->flush_color
!= color
))
1753 /* are there still in-flight works? */
1754 if (cwq
->nr_in_flight
[color
])
1757 /* this cwq is done, clear flush_color */
1758 cwq
->flush_color
= -1;
1761 * If this was the last cwq, wake up the first flusher. It
1762 * will handle the rest.
1764 if (atomic_dec_and_test(&cwq
->wq
->nr_cwqs_to_flush
))
1765 complete(&cwq
->wq
->first_flusher
->done
);
1769 * process_one_work - process single work
1771 * @work: work to process
1773 * Process @work. This function contains all the logics necessary to
1774 * process a single work including synchronization against and
1775 * interaction with other workers on the same cpu, queueing and
1776 * flushing. As long as context requirement is met, any worker can
1777 * call this function to process a work.
1780 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1782 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
1783 __releases(&gcwq
->lock
)
1784 __acquires(&gcwq
->lock
)
1786 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1787 struct global_cwq
*gcwq
= cwq
->gcwq
;
1788 struct hlist_head
*bwh
= busy_worker_head(gcwq
, work
);
1789 bool cpu_intensive
= cwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
1790 work_func_t f
= work
->func
;
1792 struct worker
*collision
;
1793 #ifdef CONFIG_LOCKDEP
1795 * It is permissible to free the struct work_struct from
1796 * inside the function that is called from it, this we need to
1797 * take into account for lockdep too. To avoid bogus "held
1798 * lock freed" warnings as well as problems when looking into
1799 * work->lockdep_map, make a copy and use that here.
1801 struct lockdep_map lockdep_map
= work
->lockdep_map
;
1804 * A single work shouldn't be executed concurrently by
1805 * multiple workers on a single cpu. Check whether anyone is
1806 * already processing the work. If so, defer the work to the
1807 * currently executing one.
1809 collision
= __find_worker_executing_work(gcwq
, bwh
, work
);
1810 if (unlikely(collision
)) {
1811 move_linked_works(work
, &collision
->scheduled
, NULL
);
1815 /* claim and process */
1816 debug_work_deactivate(work
);
1817 hlist_add_head(&worker
->hentry
, bwh
);
1818 worker
->current_work
= work
;
1819 worker
->current_cwq
= cwq
;
1820 work_color
= get_work_color(work
);
1822 /* record the current cpu number in the work data and dequeue */
1823 set_work_cpu(work
, gcwq
->cpu
);
1824 list_del_init(&work
->entry
);
1827 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1828 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1830 if (unlikely(gcwq
->flags
& GCWQ_HIGHPRI_PENDING
)) {
1831 struct work_struct
*nwork
= list_first_entry(&gcwq
->worklist
,
1832 struct work_struct
, entry
);
1834 if (!list_empty(&gcwq
->worklist
) &&
1835 get_work_cwq(nwork
)->wq
->flags
& WQ_HIGHPRI
)
1836 wake_up_worker(gcwq
);
1838 gcwq
->flags
&= ~GCWQ_HIGHPRI_PENDING
;
1842 * CPU intensive works don't participate in concurrency
1843 * management. They're the scheduler's responsibility.
1845 if (unlikely(cpu_intensive
))
1846 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
1848 spin_unlock_irq(&gcwq
->lock
);
1850 work_clear_pending(work
);
1851 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
1852 lock_map_acquire(&lockdep_map
);
1853 trace_workqueue_execute_start(work
);
1856 * While we must be careful to not use "work" after this, the trace
1857 * point will only record its address.
1859 trace_workqueue_execute_end(work
);
1860 lock_map_release(&lockdep_map
);
1861 lock_map_release(&cwq
->wq
->lockdep_map
);
1863 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
1864 printk(KERN_ERR
"BUG: workqueue leaked lock or atomic: "
1866 current
->comm
, preempt_count(), task_pid_nr(current
));
1867 printk(KERN_ERR
" last function: ");
1868 print_symbol("%s\n", (unsigned long)f
);
1869 debug_show_held_locks(current
);
1873 spin_lock_irq(&gcwq
->lock
);
1875 /* clear cpu intensive status */
1876 if (unlikely(cpu_intensive
))
1877 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
1879 /* we're done with it, release */
1880 hlist_del_init(&worker
->hentry
);
1881 worker
->current_work
= NULL
;
1882 worker
->current_cwq
= NULL
;
1883 cwq_dec_nr_in_flight(cwq
, work_color
, false);
1887 * process_scheduled_works - process scheduled works
1890 * Process all scheduled works. Please note that the scheduled list
1891 * may change while processing a work, so this function repeatedly
1892 * fetches a work from the top and executes it.
1895 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1898 static void process_scheduled_works(struct worker
*worker
)
1900 while (!list_empty(&worker
->scheduled
)) {
1901 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
1902 struct work_struct
, entry
);
1903 process_one_work(worker
, work
);
1908 * worker_thread - the worker thread function
1911 * The gcwq worker thread function. There's a single dynamic pool of
1912 * these per each cpu. These workers process all works regardless of
1913 * their specific target workqueue. The only exception is works which
1914 * belong to workqueues with a rescuer which will be explained in
1917 static int worker_thread(void *__worker
)
1919 struct worker
*worker
= __worker
;
1920 struct global_cwq
*gcwq
= worker
->gcwq
;
1922 /* tell the scheduler that this is a workqueue worker */
1923 worker
->task
->flags
|= PF_WQ_WORKER
;
1925 spin_lock_irq(&gcwq
->lock
);
1927 /* DIE can be set only while we're idle, checking here is enough */
1928 if (worker
->flags
& WORKER_DIE
) {
1929 spin_unlock_irq(&gcwq
->lock
);
1930 worker
->task
->flags
&= ~PF_WQ_WORKER
;
1934 worker_leave_idle(worker
);
1936 /* no more worker necessary? */
1937 if (!need_more_worker(gcwq
))
1940 /* do we need to manage? */
1941 if (unlikely(!may_start_working(gcwq
)) && manage_workers(worker
))
1945 * ->scheduled list can only be filled while a worker is
1946 * preparing to process a work or actually processing it.
1947 * Make sure nobody diddled with it while I was sleeping.
1949 BUG_ON(!list_empty(&worker
->scheduled
));
1952 * When control reaches this point, we're guaranteed to have
1953 * at least one idle worker or that someone else has already
1954 * assumed the manager role.
1956 worker_clr_flags(worker
, WORKER_PREP
);
1959 struct work_struct
*work
=
1960 list_first_entry(&gcwq
->worklist
,
1961 struct work_struct
, entry
);
1963 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
1964 /* optimization path, not strictly necessary */
1965 process_one_work(worker
, work
);
1966 if (unlikely(!list_empty(&worker
->scheduled
)))
1967 process_scheduled_works(worker
);
1969 move_linked_works(work
, &worker
->scheduled
, NULL
);
1970 process_scheduled_works(worker
);
1972 } while (keep_working(gcwq
));
1974 worker_set_flags(worker
, WORKER_PREP
, false);
1976 if (unlikely(need_to_manage_workers(gcwq
)) && manage_workers(worker
))
1980 * gcwq->lock is held and there's no work to process and no
1981 * need to manage, sleep. Workers are woken up only while
1982 * holding gcwq->lock or from local cpu, so setting the
1983 * current state before releasing gcwq->lock is enough to
1984 * prevent losing any event.
1986 worker_enter_idle(worker
);
1987 __set_current_state(TASK_INTERRUPTIBLE
);
1988 spin_unlock_irq(&gcwq
->lock
);
1994 * rescuer_thread - the rescuer thread function
1995 * @__wq: the associated workqueue
1997 * Workqueue rescuer thread function. There's one rescuer for each
1998 * workqueue which has WQ_RESCUER set.
2000 * Regular work processing on a gcwq may block trying to create a new
2001 * worker which uses GFP_KERNEL allocation which has slight chance of
2002 * developing into deadlock if some works currently on the same queue
2003 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2004 * the problem rescuer solves.
2006 * When such condition is possible, the gcwq summons rescuers of all
2007 * workqueues which have works queued on the gcwq and let them process
2008 * those works so that forward progress can be guaranteed.
2010 * This should happen rarely.
2012 static int rescuer_thread(void *__wq
)
2014 struct workqueue_struct
*wq
= __wq
;
2015 struct worker
*rescuer
= wq
->rescuer
;
2016 struct list_head
*scheduled
= &rescuer
->scheduled
;
2017 bool is_unbound
= wq
->flags
& WQ_UNBOUND
;
2020 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2022 set_current_state(TASK_INTERRUPTIBLE
);
2024 if (kthread_should_stop())
2028 * See whether any cpu is asking for help. Unbounded
2029 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2031 for_each_mayday_cpu(cpu
, wq
->mayday_mask
) {
2032 unsigned int tcpu
= is_unbound
? WORK_CPU_UNBOUND
: cpu
;
2033 struct cpu_workqueue_struct
*cwq
= get_cwq(tcpu
, wq
);
2034 struct global_cwq
*gcwq
= cwq
->gcwq
;
2035 struct work_struct
*work
, *n
;
2037 __set_current_state(TASK_RUNNING
);
2038 mayday_clear_cpu(cpu
, wq
->mayday_mask
);
2040 /* migrate to the target cpu if possible */
2041 rescuer
->gcwq
= gcwq
;
2042 worker_maybe_bind_and_lock(rescuer
);
2045 * Slurp in all works issued via this workqueue and
2048 BUG_ON(!list_empty(&rescuer
->scheduled
));
2049 list_for_each_entry_safe(work
, n
, &gcwq
->worklist
, entry
)
2050 if (get_work_cwq(work
) == cwq
)
2051 move_linked_works(work
, scheduled
, &n
);
2053 process_scheduled_works(rescuer
);
2056 * Leave this gcwq. If keep_working() is %true, notify a
2057 * regular worker; otherwise, we end up with 0 concurrency
2058 * and stalling the execution.
2060 if (keep_working(gcwq
))
2061 wake_up_worker(gcwq
);
2063 spin_unlock_irq(&gcwq
->lock
);
2071 struct work_struct work
;
2072 struct completion done
;
2075 static void wq_barrier_func(struct work_struct
*work
)
2077 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2078 complete(&barr
->done
);
2082 * insert_wq_barrier - insert a barrier work
2083 * @cwq: cwq to insert barrier into
2084 * @barr: wq_barrier to insert
2085 * @target: target work to attach @barr to
2086 * @worker: worker currently executing @target, NULL if @target is not executing
2088 * @barr is linked to @target such that @barr is completed only after
2089 * @target finishes execution. Please note that the ordering
2090 * guarantee is observed only with respect to @target and on the local
2093 * Currently, a queued barrier can't be canceled. This is because
2094 * try_to_grab_pending() can't determine whether the work to be
2095 * grabbed is at the head of the queue and thus can't clear LINKED
2096 * flag of the previous work while there must be a valid next work
2097 * after a work with LINKED flag set.
2099 * Note that when @worker is non-NULL, @target may be modified
2100 * underneath us, so we can't reliably determine cwq from @target.
2103 * spin_lock_irq(gcwq->lock).
2105 static void insert_wq_barrier(struct cpu_workqueue_struct
*cwq
,
2106 struct wq_barrier
*barr
,
2107 struct work_struct
*target
, struct worker
*worker
)
2109 struct list_head
*head
;
2110 unsigned int linked
= 0;
2113 * debugobject calls are safe here even with gcwq->lock locked
2114 * as we know for sure that this will not trigger any of the
2115 * checks and call back into the fixup functions where we
2118 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2119 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2120 init_completion(&barr
->done
);
2123 * If @target is currently being executed, schedule the
2124 * barrier to the worker; otherwise, put it after @target.
2127 head
= worker
->scheduled
.next
;
2129 unsigned long *bits
= work_data_bits(target
);
2131 head
= target
->entry
.next
;
2132 /* there can already be other linked works, inherit and set */
2133 linked
= *bits
& WORK_STRUCT_LINKED
;
2134 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2137 debug_work_activate(&barr
->work
);
2138 insert_work(cwq
, &barr
->work
, head
,
2139 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2143 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2144 * @wq: workqueue being flushed
2145 * @flush_color: new flush color, < 0 for no-op
2146 * @work_color: new work color, < 0 for no-op
2148 * Prepare cwqs for workqueue flushing.
2150 * If @flush_color is non-negative, flush_color on all cwqs should be
2151 * -1. If no cwq has in-flight commands at the specified color, all
2152 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2153 * has in flight commands, its cwq->flush_color is set to
2154 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2155 * wakeup logic is armed and %true is returned.
2157 * The caller should have initialized @wq->first_flusher prior to
2158 * calling this function with non-negative @flush_color. If
2159 * @flush_color is negative, no flush color update is done and %false
2162 * If @work_color is non-negative, all cwqs should have the same
2163 * work_color which is previous to @work_color and all will be
2164 * advanced to @work_color.
2167 * mutex_lock(wq->flush_mutex).
2170 * %true if @flush_color >= 0 and there's something to flush. %false
2173 static bool flush_workqueue_prep_cwqs(struct workqueue_struct
*wq
,
2174 int flush_color
, int work_color
)
2179 if (flush_color
>= 0) {
2180 BUG_ON(atomic_read(&wq
->nr_cwqs_to_flush
));
2181 atomic_set(&wq
->nr_cwqs_to_flush
, 1);
2184 for_each_cwq_cpu(cpu
, wq
) {
2185 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2186 struct global_cwq
*gcwq
= cwq
->gcwq
;
2188 spin_lock_irq(&gcwq
->lock
);
2190 if (flush_color
>= 0) {
2191 BUG_ON(cwq
->flush_color
!= -1);
2193 if (cwq
->nr_in_flight
[flush_color
]) {
2194 cwq
->flush_color
= flush_color
;
2195 atomic_inc(&wq
->nr_cwqs_to_flush
);
2200 if (work_color
>= 0) {
2201 BUG_ON(work_color
!= work_next_color(cwq
->work_color
));
2202 cwq
->work_color
= work_color
;
2205 spin_unlock_irq(&gcwq
->lock
);
2208 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_cwqs_to_flush
))
2209 complete(&wq
->first_flusher
->done
);
2215 * flush_workqueue - ensure that any scheduled work has run to completion.
2216 * @wq: workqueue to flush
2218 * Forces execution of the workqueue and blocks until its completion.
2219 * This is typically used in driver shutdown handlers.
2221 * We sleep until all works which were queued on entry have been handled,
2222 * but we are not livelocked by new incoming ones.
2224 void flush_workqueue(struct workqueue_struct
*wq
)
2226 struct wq_flusher this_flusher
= {
2227 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2229 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2233 lock_map_acquire(&wq
->lockdep_map
);
2234 lock_map_release(&wq
->lockdep_map
);
2236 mutex_lock(&wq
->flush_mutex
);
2239 * Start-to-wait phase
2241 next_color
= work_next_color(wq
->work_color
);
2243 if (next_color
!= wq
->flush_color
) {
2245 * Color space is not full. The current work_color
2246 * becomes our flush_color and work_color is advanced
2249 BUG_ON(!list_empty(&wq
->flusher_overflow
));
2250 this_flusher
.flush_color
= wq
->work_color
;
2251 wq
->work_color
= next_color
;
2253 if (!wq
->first_flusher
) {
2254 /* no flush in progress, become the first flusher */
2255 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2257 wq
->first_flusher
= &this_flusher
;
2259 if (!flush_workqueue_prep_cwqs(wq
, wq
->flush_color
,
2261 /* nothing to flush, done */
2262 wq
->flush_color
= next_color
;
2263 wq
->first_flusher
= NULL
;
2268 BUG_ON(wq
->flush_color
== this_flusher
.flush_color
);
2269 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2270 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2274 * Oops, color space is full, wait on overflow queue.
2275 * The next flush completion will assign us
2276 * flush_color and transfer to flusher_queue.
2278 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2281 mutex_unlock(&wq
->flush_mutex
);
2283 wait_for_completion(&this_flusher
.done
);
2286 * Wake-up-and-cascade phase
2288 * First flushers are responsible for cascading flushes and
2289 * handling overflow. Non-first flushers can simply return.
2291 if (wq
->first_flusher
!= &this_flusher
)
2294 mutex_lock(&wq
->flush_mutex
);
2296 /* we might have raced, check again with mutex held */
2297 if (wq
->first_flusher
!= &this_flusher
)
2300 wq
->first_flusher
= NULL
;
2302 BUG_ON(!list_empty(&this_flusher
.list
));
2303 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2306 struct wq_flusher
*next
, *tmp
;
2308 /* complete all the flushers sharing the current flush color */
2309 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2310 if (next
->flush_color
!= wq
->flush_color
)
2312 list_del_init(&next
->list
);
2313 complete(&next
->done
);
2316 BUG_ON(!list_empty(&wq
->flusher_overflow
) &&
2317 wq
->flush_color
!= work_next_color(wq
->work_color
));
2319 /* this flush_color is finished, advance by one */
2320 wq
->flush_color
= work_next_color(wq
->flush_color
);
2322 /* one color has been freed, handle overflow queue */
2323 if (!list_empty(&wq
->flusher_overflow
)) {
2325 * Assign the same color to all overflowed
2326 * flushers, advance work_color and append to
2327 * flusher_queue. This is the start-to-wait
2328 * phase for these overflowed flushers.
2330 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2331 tmp
->flush_color
= wq
->work_color
;
2333 wq
->work_color
= work_next_color(wq
->work_color
);
2335 list_splice_tail_init(&wq
->flusher_overflow
,
2336 &wq
->flusher_queue
);
2337 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2340 if (list_empty(&wq
->flusher_queue
)) {
2341 BUG_ON(wq
->flush_color
!= wq
->work_color
);
2346 * Need to flush more colors. Make the next flusher
2347 * the new first flusher and arm cwqs.
2349 BUG_ON(wq
->flush_color
== wq
->work_color
);
2350 BUG_ON(wq
->flush_color
!= next
->flush_color
);
2352 list_del_init(&next
->list
);
2353 wq
->first_flusher
= next
;
2355 if (flush_workqueue_prep_cwqs(wq
, wq
->flush_color
, -1))
2359 * Meh... this color is already done, clear first
2360 * flusher and repeat cascading.
2362 wq
->first_flusher
= NULL
;
2366 mutex_unlock(&wq
->flush_mutex
);
2368 EXPORT_SYMBOL_GPL(flush_workqueue
);
2370 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
,
2371 bool wait_executing
)
2373 struct worker
*worker
= NULL
;
2374 struct global_cwq
*gcwq
;
2375 struct cpu_workqueue_struct
*cwq
;
2378 gcwq
= get_work_gcwq(work
);
2382 spin_lock_irq(&gcwq
->lock
);
2383 if (!list_empty(&work
->entry
)) {
2385 * See the comment near try_to_grab_pending()->smp_rmb().
2386 * If it was re-queued to a different gcwq under us, we
2387 * are not going to wait.
2390 cwq
= get_work_cwq(work
);
2391 if (unlikely(!cwq
|| gcwq
!= cwq
->gcwq
))
2393 } else if (wait_executing
) {
2394 worker
= find_worker_executing_work(gcwq
, work
);
2397 cwq
= worker
->current_cwq
;
2401 insert_wq_barrier(cwq
, barr
, work
, worker
);
2402 spin_unlock_irq(&gcwq
->lock
);
2405 * If @max_active is 1 or rescuer is in use, flushing another work
2406 * item on the same workqueue may lead to deadlock. Make sure the
2407 * flusher is not running on the same workqueue by verifying write
2410 if (cwq
->wq
->saved_max_active
== 1 || cwq
->wq
->flags
& WQ_RESCUER
)
2411 lock_map_acquire(&cwq
->wq
->lockdep_map
);
2413 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2414 lock_map_release(&cwq
->wq
->lockdep_map
);
2418 spin_unlock_irq(&gcwq
->lock
);
2423 * flush_work - wait for a work to finish executing the last queueing instance
2424 * @work: the work to flush
2426 * Wait until @work has finished execution. This function considers
2427 * only the last queueing instance of @work. If @work has been
2428 * enqueued across different CPUs on a non-reentrant workqueue or on
2429 * multiple workqueues, @work might still be executing on return on
2430 * some of the CPUs from earlier queueing.
2432 * If @work was queued only on a non-reentrant, ordered or unbound
2433 * workqueue, @work is guaranteed to be idle on return if it hasn't
2434 * been requeued since flush started.
2437 * %true if flush_work() waited for the work to finish execution,
2438 * %false if it was already idle.
2440 bool flush_work(struct work_struct
*work
)
2442 struct wq_barrier barr
;
2444 if (start_flush_work(work
, &barr
, true)) {
2445 wait_for_completion(&barr
.done
);
2446 destroy_work_on_stack(&barr
.work
);
2451 EXPORT_SYMBOL_GPL(flush_work
);
2453 static bool wait_on_cpu_work(struct global_cwq
*gcwq
, struct work_struct
*work
)
2455 struct wq_barrier barr
;
2456 struct worker
*worker
;
2458 spin_lock_irq(&gcwq
->lock
);
2460 worker
= find_worker_executing_work(gcwq
, work
);
2461 if (unlikely(worker
))
2462 insert_wq_barrier(worker
->current_cwq
, &barr
, work
, worker
);
2464 spin_unlock_irq(&gcwq
->lock
);
2466 if (unlikely(worker
)) {
2467 wait_for_completion(&barr
.done
);
2468 destroy_work_on_stack(&barr
.work
);
2474 static bool wait_on_work(struct work_struct
*work
)
2481 lock_map_acquire(&work
->lockdep_map
);
2482 lock_map_release(&work
->lockdep_map
);
2484 for_each_gcwq_cpu(cpu
)
2485 ret
|= wait_on_cpu_work(get_gcwq(cpu
), work
);
2490 * flush_work_sync - wait until a work has finished execution
2491 * @work: the work to flush
2493 * Wait until @work has finished execution. On return, it's
2494 * guaranteed that all queueing instances of @work which happened
2495 * before this function is called are finished. In other words, if
2496 * @work hasn't been requeued since this function was called, @work is
2497 * guaranteed to be idle on return.
2500 * %true if flush_work_sync() waited for the work to finish execution,
2501 * %false if it was already idle.
2503 bool flush_work_sync(struct work_struct
*work
)
2505 struct wq_barrier barr
;
2506 bool pending
, waited
;
2508 /* we'll wait for executions separately, queue barr only if pending */
2509 pending
= start_flush_work(work
, &barr
, false);
2511 /* wait for executions to finish */
2512 waited
= wait_on_work(work
);
2514 /* wait for the pending one */
2516 wait_for_completion(&barr
.done
);
2517 destroy_work_on_stack(&barr
.work
);
2520 return pending
|| waited
;
2522 EXPORT_SYMBOL_GPL(flush_work_sync
);
2525 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2526 * so this work can't be re-armed in any way.
2528 static int try_to_grab_pending(struct work_struct
*work
)
2530 struct global_cwq
*gcwq
;
2533 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
2537 * The queueing is in progress, or it is already queued. Try to
2538 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2540 gcwq
= get_work_gcwq(work
);
2544 spin_lock_irq(&gcwq
->lock
);
2545 if (!list_empty(&work
->entry
)) {
2547 * This work is queued, but perhaps we locked the wrong gcwq.
2548 * In that case we must see the new value after rmb(), see
2549 * insert_work()->wmb().
2552 if (gcwq
== get_work_gcwq(work
)) {
2553 debug_work_deactivate(work
);
2554 list_del_init(&work
->entry
);
2555 cwq_dec_nr_in_flight(get_work_cwq(work
),
2556 get_work_color(work
),
2557 *work_data_bits(work
) & WORK_STRUCT_DELAYED
);
2561 spin_unlock_irq(&gcwq
->lock
);
2566 static bool __cancel_work_timer(struct work_struct
*work
,
2567 struct timer_list
* timer
)
2572 ret
= (timer
&& likely(del_timer(timer
)));
2574 ret
= try_to_grab_pending(work
);
2576 } while (unlikely(ret
< 0));
2578 clear_work_data(work
);
2583 * cancel_work_sync - cancel a work and wait for it to finish
2584 * @work: the work to cancel
2586 * Cancel @work and wait for its execution to finish. This function
2587 * can be used even if the work re-queues itself or migrates to
2588 * another workqueue. On return from this function, @work is
2589 * guaranteed to be not pending or executing on any CPU.
2591 * cancel_work_sync(&delayed_work->work) must not be used for
2592 * delayed_work's. Use cancel_delayed_work_sync() instead.
2594 * The caller must ensure that the workqueue on which @work was last
2595 * queued can't be destroyed before this function returns.
2598 * %true if @work was pending, %false otherwise.
2600 bool cancel_work_sync(struct work_struct
*work
)
2602 return __cancel_work_timer(work
, NULL
);
2604 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2607 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2608 * @dwork: the delayed work to flush
2610 * Delayed timer is cancelled and the pending work is queued for
2611 * immediate execution. Like flush_work(), this function only
2612 * considers the last queueing instance of @dwork.
2615 * %true if flush_work() waited for the work to finish execution,
2616 * %false if it was already idle.
2618 bool flush_delayed_work(struct delayed_work
*dwork
)
2620 if (del_timer_sync(&dwork
->timer
))
2621 __queue_work(raw_smp_processor_id(),
2622 get_work_cwq(&dwork
->work
)->wq
, &dwork
->work
);
2623 return flush_work(&dwork
->work
);
2625 EXPORT_SYMBOL(flush_delayed_work
);
2628 * flush_delayed_work_sync - wait for a dwork to finish
2629 * @dwork: the delayed work to flush
2631 * Delayed timer is cancelled and the pending work is queued for
2632 * execution immediately. Other than timer handling, its behavior
2633 * is identical to flush_work_sync().
2636 * %true if flush_work_sync() waited for the work to finish execution,
2637 * %false if it was already idle.
2639 bool flush_delayed_work_sync(struct delayed_work
*dwork
)
2641 if (del_timer_sync(&dwork
->timer
))
2642 __queue_work(raw_smp_processor_id(),
2643 get_work_cwq(&dwork
->work
)->wq
, &dwork
->work
);
2644 return flush_work_sync(&dwork
->work
);
2646 EXPORT_SYMBOL(flush_delayed_work_sync
);
2649 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2650 * @dwork: the delayed work cancel
2652 * This is cancel_work_sync() for delayed works.
2655 * %true if @dwork was pending, %false otherwise.
2657 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2659 return __cancel_work_timer(&dwork
->work
, &dwork
->timer
);
2661 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2664 * schedule_work - put work task in global workqueue
2665 * @work: job to be done
2667 * Returns zero if @work was already on the kernel-global workqueue and
2668 * non-zero otherwise.
2670 * This puts a job in the kernel-global workqueue if it was not already
2671 * queued and leaves it in the same position on the kernel-global
2672 * workqueue otherwise.
2674 int schedule_work(struct work_struct
*work
)
2676 return queue_work(system_wq
, work
);
2678 EXPORT_SYMBOL(schedule_work
);
2681 * schedule_work_on - put work task on a specific cpu
2682 * @cpu: cpu to put the work task on
2683 * @work: job to be done
2685 * This puts a job on a specific cpu
2687 int schedule_work_on(int cpu
, struct work_struct
*work
)
2689 return queue_work_on(cpu
, system_wq
, work
);
2691 EXPORT_SYMBOL(schedule_work_on
);
2694 * schedule_delayed_work - put work task in global workqueue after delay
2695 * @dwork: job to be done
2696 * @delay: number of jiffies to wait or 0 for immediate execution
2698 * After waiting for a given time this puts a job in the kernel-global
2701 int schedule_delayed_work(struct delayed_work
*dwork
,
2702 unsigned long delay
)
2704 return queue_delayed_work(system_wq
, dwork
, delay
);
2706 EXPORT_SYMBOL(schedule_delayed_work
);
2709 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2711 * @dwork: job to be done
2712 * @delay: number of jiffies to wait
2714 * After waiting for a given time this puts a job in the kernel-global
2715 * workqueue on the specified CPU.
2717 int schedule_delayed_work_on(int cpu
,
2718 struct delayed_work
*dwork
, unsigned long delay
)
2720 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
2722 EXPORT_SYMBOL(schedule_delayed_work_on
);
2725 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2726 * @func: the function to call
2728 * schedule_on_each_cpu() executes @func on each online CPU using the
2729 * system workqueue and blocks until all CPUs have completed.
2730 * schedule_on_each_cpu() is very slow.
2733 * 0 on success, -errno on failure.
2735 int schedule_on_each_cpu(work_func_t func
)
2738 struct work_struct __percpu
*works
;
2740 works
= alloc_percpu(struct work_struct
);
2746 for_each_online_cpu(cpu
) {
2747 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
2749 INIT_WORK(work
, func
);
2750 schedule_work_on(cpu
, work
);
2753 for_each_online_cpu(cpu
)
2754 flush_work(per_cpu_ptr(works
, cpu
));
2762 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2764 * Forces execution of the kernel-global workqueue and blocks until its
2767 * Think twice before calling this function! It's very easy to get into
2768 * trouble if you don't take great care. Either of the following situations
2769 * will lead to deadlock:
2771 * One of the work items currently on the workqueue needs to acquire
2772 * a lock held by your code or its caller.
2774 * Your code is running in the context of a work routine.
2776 * They will be detected by lockdep when they occur, but the first might not
2777 * occur very often. It depends on what work items are on the workqueue and
2778 * what locks they need, which you have no control over.
2780 * In most situations flushing the entire workqueue is overkill; you merely
2781 * need to know that a particular work item isn't queued and isn't running.
2782 * In such cases you should use cancel_delayed_work_sync() or
2783 * cancel_work_sync() instead.
2785 void flush_scheduled_work(void)
2787 flush_workqueue(system_wq
);
2789 EXPORT_SYMBOL(flush_scheduled_work
);
2792 * execute_in_process_context - reliably execute the routine with user context
2793 * @fn: the function to execute
2794 * @ew: guaranteed storage for the execute work structure (must
2795 * be available when the work executes)
2797 * Executes the function immediately if process context is available,
2798 * otherwise schedules the function for delayed execution.
2800 * Returns: 0 - function was executed
2801 * 1 - function was scheduled for execution
2803 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
2805 if (!in_interrupt()) {
2810 INIT_WORK(&ew
->work
, fn
);
2811 schedule_work(&ew
->work
);
2815 EXPORT_SYMBOL_GPL(execute_in_process_context
);
2817 int keventd_up(void)
2819 return system_wq
!= NULL
;
2822 static int alloc_cwqs(struct workqueue_struct
*wq
)
2825 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2826 * Make sure that the alignment isn't lower than that of
2827 * unsigned long long.
2829 const size_t size
= sizeof(struct cpu_workqueue_struct
);
2830 const size_t align
= max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS
,
2831 __alignof__(unsigned long long));
2833 bool percpu
= !(wq
->flags
& WQ_UNBOUND
);
2835 bool percpu
= false;
2839 wq
->cpu_wq
.pcpu
= __alloc_percpu(size
, align
);
2844 * Allocate enough room to align cwq and put an extra
2845 * pointer at the end pointing back to the originally
2846 * allocated pointer which will be used for free.
2848 ptr
= kzalloc(size
+ align
+ sizeof(void *), GFP_KERNEL
);
2850 wq
->cpu_wq
.single
= PTR_ALIGN(ptr
, align
);
2851 *(void **)(wq
->cpu_wq
.single
+ 1) = ptr
;
2855 /* just in case, make sure it's actually aligned
2856 * - this is affected by PERCPU() alignment in vmlinux.lds.S
2858 BUG_ON(!IS_ALIGNED(wq
->cpu_wq
.v
, align
));
2859 return wq
->cpu_wq
.v
? 0 : -ENOMEM
;
2862 static void free_cwqs(struct workqueue_struct
*wq
)
2865 bool percpu
= !(wq
->flags
& WQ_UNBOUND
);
2867 bool percpu
= false;
2871 free_percpu(wq
->cpu_wq
.pcpu
);
2872 else if (wq
->cpu_wq
.single
) {
2873 /* the pointer to free is stored right after the cwq */
2874 kfree(*(void **)(wq
->cpu_wq
.single
+ 1));
2878 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
2881 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
2883 if (max_active
< 1 || max_active
> lim
)
2884 printk(KERN_WARNING
"workqueue: max_active %d requested for %s "
2885 "is out of range, clamping between %d and %d\n",
2886 max_active
, name
, 1, lim
);
2888 return clamp_val(max_active
, 1, lim
);
2891 struct workqueue_struct
*__alloc_workqueue_key(const char *name
,
2894 struct lock_class_key
*key
,
2895 const char *lock_name
)
2897 struct workqueue_struct
*wq
;
2901 * Workqueues which may be used during memory reclaim should
2902 * have a rescuer to guarantee forward progress.
2904 if (flags
& WQ_MEM_RECLAIM
)
2905 flags
|= WQ_RESCUER
;
2908 * Unbound workqueues aren't concurrency managed and should be
2909 * dispatched to workers immediately.
2911 if (flags
& WQ_UNBOUND
)
2912 flags
|= WQ_HIGHPRI
;
2914 max_active
= max_active
?: WQ_DFL_ACTIVE
;
2915 max_active
= wq_clamp_max_active(max_active
, flags
, name
);
2917 wq
= kzalloc(sizeof(*wq
), GFP_KERNEL
);
2922 wq
->saved_max_active
= max_active
;
2923 mutex_init(&wq
->flush_mutex
);
2924 atomic_set(&wq
->nr_cwqs_to_flush
, 0);
2925 INIT_LIST_HEAD(&wq
->flusher_queue
);
2926 INIT_LIST_HEAD(&wq
->flusher_overflow
);
2929 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
2930 INIT_LIST_HEAD(&wq
->list
);
2932 if (alloc_cwqs(wq
) < 0)
2935 for_each_cwq_cpu(cpu
, wq
) {
2936 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2937 struct global_cwq
*gcwq
= get_gcwq(cpu
);
2939 BUG_ON((unsigned long)cwq
& WORK_STRUCT_FLAG_MASK
);
2942 cwq
->flush_color
= -1;
2943 cwq
->max_active
= max_active
;
2944 INIT_LIST_HEAD(&cwq
->delayed_works
);
2947 if (flags
& WQ_RESCUER
) {
2948 struct worker
*rescuer
;
2950 if (!alloc_mayday_mask(&wq
->mayday_mask
, GFP_KERNEL
))
2953 wq
->rescuer
= rescuer
= alloc_worker();
2957 rescuer
->task
= kthread_create(rescuer_thread
, wq
, "%s", name
);
2958 if (IS_ERR(rescuer
->task
))
2961 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
2962 wake_up_process(rescuer
->task
);
2966 * workqueue_lock protects global freeze state and workqueues
2967 * list. Grab it, set max_active accordingly and add the new
2968 * workqueue to workqueues list.
2970 spin_lock(&workqueue_lock
);
2972 if (workqueue_freezing
&& wq
->flags
& WQ_FREEZABLE
)
2973 for_each_cwq_cpu(cpu
, wq
)
2974 get_cwq(cpu
, wq
)->max_active
= 0;
2976 list_add(&wq
->list
, &workqueues
);
2978 spin_unlock(&workqueue_lock
);
2984 free_mayday_mask(wq
->mayday_mask
);
2990 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
2993 * destroy_workqueue - safely terminate a workqueue
2994 * @wq: target workqueue
2996 * Safely destroy a workqueue. All work currently pending will be done first.
2998 void destroy_workqueue(struct workqueue_struct
*wq
)
3000 unsigned int flush_cnt
= 0;
3004 * Mark @wq dying and drain all pending works. Once WQ_DYING is
3005 * set, only chain queueing is allowed. IOW, only currently
3006 * pending or running work items on @wq can queue further work
3007 * items on it. @wq is flushed repeatedly until it becomes empty.
3008 * The number of flushing is detemined by the depth of chaining and
3009 * should be relatively short. Whine if it takes too long.
3011 wq
->flags
|= WQ_DYING
;
3013 flush_workqueue(wq
);
3015 for_each_cwq_cpu(cpu
, wq
) {
3016 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3018 if (!cwq
->nr_active
&& list_empty(&cwq
->delayed_works
))
3021 if (++flush_cnt
== 10 ||
3022 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
3023 printk(KERN_WARNING
"workqueue %s: flush on "
3024 "destruction isn't complete after %u tries\n",
3025 wq
->name
, flush_cnt
);
3030 * wq list is used to freeze wq, remove from list after
3031 * flushing is complete in case freeze races us.
3033 spin_lock(&workqueue_lock
);
3034 list_del(&wq
->list
);
3035 spin_unlock(&workqueue_lock
);
3038 for_each_cwq_cpu(cpu
, wq
) {
3039 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3042 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
3043 BUG_ON(cwq
->nr_in_flight
[i
]);
3044 BUG_ON(cwq
->nr_active
);
3045 BUG_ON(!list_empty(&cwq
->delayed_works
));
3048 if (wq
->flags
& WQ_RESCUER
) {
3049 kthread_stop(wq
->rescuer
->task
);
3050 free_mayday_mask(wq
->mayday_mask
);
3057 EXPORT_SYMBOL_GPL(destroy_workqueue
);
3060 * workqueue_set_max_active - adjust max_active of a workqueue
3061 * @wq: target workqueue
3062 * @max_active: new max_active value.
3064 * Set max_active of @wq to @max_active.
3067 * Don't call from IRQ context.
3069 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
3073 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
3075 spin_lock(&workqueue_lock
);
3077 wq
->saved_max_active
= max_active
;
3079 for_each_cwq_cpu(cpu
, wq
) {
3080 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3082 spin_lock_irq(&gcwq
->lock
);
3084 if (!(wq
->flags
& WQ_FREEZABLE
) ||
3085 !(gcwq
->flags
& GCWQ_FREEZING
))
3086 get_cwq(gcwq
->cpu
, wq
)->max_active
= max_active
;
3088 spin_unlock_irq(&gcwq
->lock
);
3091 spin_unlock(&workqueue_lock
);
3093 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
3096 * workqueue_congested - test whether a workqueue is congested
3097 * @cpu: CPU in question
3098 * @wq: target workqueue
3100 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3101 * no synchronization around this function and the test result is
3102 * unreliable and only useful as advisory hints or for debugging.
3105 * %true if congested, %false otherwise.
3107 bool workqueue_congested(unsigned int cpu
, struct workqueue_struct
*wq
)
3109 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3111 return !list_empty(&cwq
->delayed_works
);
3113 EXPORT_SYMBOL_GPL(workqueue_congested
);
3116 * work_cpu - return the last known associated cpu for @work
3117 * @work: the work of interest
3120 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3122 unsigned int work_cpu(struct work_struct
*work
)
3124 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3126 return gcwq
? gcwq
->cpu
: WORK_CPU_NONE
;
3128 EXPORT_SYMBOL_GPL(work_cpu
);
3131 * work_busy - test whether a work is currently pending or running
3132 * @work: the work to be tested
3134 * Test whether @work is currently pending or running. There is no
3135 * synchronization around this function and the test result is
3136 * unreliable and only useful as advisory hints or for debugging.
3137 * Especially for reentrant wqs, the pending state might hide the
3141 * OR'd bitmask of WORK_BUSY_* bits.
3143 unsigned int work_busy(struct work_struct
*work
)
3145 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3146 unsigned long flags
;
3147 unsigned int ret
= 0;
3152 spin_lock_irqsave(&gcwq
->lock
, flags
);
3154 if (work_pending(work
))
3155 ret
|= WORK_BUSY_PENDING
;
3156 if (find_worker_executing_work(gcwq
, work
))
3157 ret
|= WORK_BUSY_RUNNING
;
3159 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
3163 EXPORT_SYMBOL_GPL(work_busy
);
3168 * There are two challenges in supporting CPU hotplug. Firstly, there
3169 * are a lot of assumptions on strong associations among work, cwq and
3170 * gcwq which make migrating pending and scheduled works very
3171 * difficult to implement without impacting hot paths. Secondly,
3172 * gcwqs serve mix of short, long and very long running works making
3173 * blocked draining impractical.
3175 * This is solved by allowing a gcwq to be detached from CPU, running
3176 * it with unbound (rogue) workers and allowing it to be reattached
3177 * later if the cpu comes back online. A separate thread is created
3178 * to govern a gcwq in such state and is called the trustee of the
3181 * Trustee states and their descriptions.
3183 * START Command state used on startup. On CPU_DOWN_PREPARE, a
3184 * new trustee is started with this state.
3186 * IN_CHARGE Once started, trustee will enter this state after
3187 * assuming the manager role and making all existing
3188 * workers rogue. DOWN_PREPARE waits for trustee to
3189 * enter this state. After reaching IN_CHARGE, trustee
3190 * tries to execute the pending worklist until it's empty
3191 * and the state is set to BUTCHER, or the state is set
3194 * BUTCHER Command state which is set by the cpu callback after
3195 * the cpu has went down. Once this state is set trustee
3196 * knows that there will be no new works on the worklist
3197 * and once the worklist is empty it can proceed to
3198 * killing idle workers.
3200 * RELEASE Command state which is set by the cpu callback if the
3201 * cpu down has been canceled or it has come online
3202 * again. After recognizing this state, trustee stops
3203 * trying to drain or butcher and clears ROGUE, rebinds
3204 * all remaining workers back to the cpu and releases
3207 * DONE Trustee will enter this state after BUTCHER or RELEASE
3210 * trustee CPU draining
3211 * took over down complete
3212 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3214 * | CPU is back online v return workers |
3215 * ----------------> RELEASE --------------
3219 * trustee_wait_event_timeout - timed event wait for trustee
3220 * @cond: condition to wait for
3221 * @timeout: timeout in jiffies
3223 * wait_event_timeout() for trustee to use. Handles locking and
3224 * checks for RELEASE request.
3227 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3228 * multiple times. To be used by trustee.
3231 * Positive indicating left time if @cond is satisfied, 0 if timed
3232 * out, -1 if canceled.
3234 #define trustee_wait_event_timeout(cond, timeout) ({ \
3235 long __ret = (timeout); \
3236 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3238 spin_unlock_irq(&gcwq->lock); \
3239 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3240 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3242 spin_lock_irq(&gcwq->lock); \
3244 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3248 * trustee_wait_event - event wait for trustee
3249 * @cond: condition to wait for
3251 * wait_event() for trustee to use. Automatically handles locking and
3252 * checks for CANCEL request.
3255 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3256 * multiple times. To be used by trustee.
3259 * 0 if @cond is satisfied, -1 if canceled.
3261 #define trustee_wait_event(cond) ({ \
3263 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3264 __ret1 < 0 ? -1 : 0; \
3267 static int __cpuinit
trustee_thread(void *__gcwq
)
3269 struct global_cwq
*gcwq
= __gcwq
;
3270 struct worker
*worker
;
3271 struct work_struct
*work
;
3272 struct hlist_node
*pos
;
3276 BUG_ON(gcwq
->cpu
!= smp_processor_id());
3278 spin_lock_irq(&gcwq
->lock
);
3280 * Claim the manager position and make all workers rogue.
3281 * Trustee must be bound to the target cpu and can't be
3284 BUG_ON(gcwq
->cpu
!= smp_processor_id());
3285 rc
= trustee_wait_event(!(gcwq
->flags
& GCWQ_MANAGING_WORKERS
));
3288 gcwq
->flags
|= GCWQ_MANAGING_WORKERS
;
3290 list_for_each_entry(worker
, &gcwq
->idle_list
, entry
)
3291 worker
->flags
|= WORKER_ROGUE
;
3293 for_each_busy_worker(worker
, i
, pos
, gcwq
)
3294 worker
->flags
|= WORKER_ROGUE
;
3297 * Call schedule() so that we cross rq->lock and thus can
3298 * guarantee sched callbacks see the rogue flag. This is
3299 * necessary as scheduler callbacks may be invoked from other
3302 spin_unlock_irq(&gcwq
->lock
);
3304 spin_lock_irq(&gcwq
->lock
);
3307 * Sched callbacks are disabled now. Zap nr_running. After
3308 * this, nr_running stays zero and need_more_worker() and
3309 * keep_working() are always true as long as the worklist is
3312 atomic_set(get_gcwq_nr_running(gcwq
->cpu
), 0);
3314 spin_unlock_irq(&gcwq
->lock
);
3315 del_timer_sync(&gcwq
->idle_timer
);
3316 spin_lock_irq(&gcwq
->lock
);
3319 * We're now in charge. Notify and proceed to drain. We need
3320 * to keep the gcwq running during the whole CPU down
3321 * procedure as other cpu hotunplug callbacks may need to
3322 * flush currently running tasks.
3324 gcwq
->trustee_state
= TRUSTEE_IN_CHARGE
;
3325 wake_up_all(&gcwq
->trustee_wait
);
3328 * The original cpu is in the process of dying and may go away
3329 * anytime now. When that happens, we and all workers would
3330 * be migrated to other cpus. Try draining any left work. We
3331 * want to get it over with ASAP - spam rescuers, wake up as
3332 * many idlers as necessary and create new ones till the
3333 * worklist is empty. Note that if the gcwq is frozen, there
3334 * may be frozen works in freezable cwqs. Don't declare
3335 * completion while frozen.
3337 while (gcwq
->nr_workers
!= gcwq
->nr_idle
||
3338 gcwq
->flags
& GCWQ_FREEZING
||
3339 gcwq
->trustee_state
== TRUSTEE_IN_CHARGE
) {
3342 list_for_each_entry(work
, &gcwq
->worklist
, entry
) {
3347 list_for_each_entry(worker
, &gcwq
->idle_list
, entry
) {
3350 wake_up_process(worker
->task
);
3353 if (need_to_create_worker(gcwq
)) {
3354 spin_unlock_irq(&gcwq
->lock
);
3355 worker
= create_worker(gcwq
, false);
3356 spin_lock_irq(&gcwq
->lock
);
3358 worker
->flags
|= WORKER_ROGUE
;
3359 start_worker(worker
);
3363 /* give a breather */
3364 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN
) < 0)
3369 * Either all works have been scheduled and cpu is down, or
3370 * cpu down has already been canceled. Wait for and butcher
3371 * all workers till we're canceled.
3374 rc
= trustee_wait_event(!list_empty(&gcwq
->idle_list
));
3375 while (!list_empty(&gcwq
->idle_list
))
3376 destroy_worker(list_first_entry(&gcwq
->idle_list
,
3377 struct worker
, entry
));
3378 } while (gcwq
->nr_workers
&& rc
>= 0);
3381 * At this point, either draining has completed and no worker
3382 * is left, or cpu down has been canceled or the cpu is being
3383 * brought back up. There shouldn't be any idle one left.
3384 * Tell the remaining busy ones to rebind once it finishes the
3385 * currently scheduled works by scheduling the rebind_work.
3387 WARN_ON(!list_empty(&gcwq
->idle_list
));
3389 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
3390 struct work_struct
*rebind_work
= &worker
->rebind_work
;
3393 * Rebind_work may race with future cpu hotplug
3394 * operations. Use a separate flag to mark that
3395 * rebinding is scheduled.
3397 worker
->flags
|= WORKER_REBIND
;
3398 worker
->flags
&= ~WORKER_ROGUE
;
3400 /* queue rebind_work, wq doesn't matter, use the default one */
3401 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
3402 work_data_bits(rebind_work
)))
3405 debug_work_activate(rebind_work
);
3406 insert_work(get_cwq(gcwq
->cpu
, system_wq
), rebind_work
,
3407 worker
->scheduled
.next
,
3408 work_color_to_flags(WORK_NO_COLOR
));
3411 /* relinquish manager role */
3412 gcwq
->flags
&= ~GCWQ_MANAGING_WORKERS
;
3414 /* notify completion */
3415 gcwq
->trustee
= NULL
;
3416 gcwq
->trustee_state
= TRUSTEE_DONE
;
3417 wake_up_all(&gcwq
->trustee_wait
);
3418 spin_unlock_irq(&gcwq
->lock
);
3423 * wait_trustee_state - wait for trustee to enter the specified state
3424 * @gcwq: gcwq the trustee of interest belongs to
3425 * @state: target state to wait for
3427 * Wait for the trustee to reach @state. DONE is already matched.
3430 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3431 * multiple times. To be used by cpu_callback.
3433 static void __cpuinit
wait_trustee_state(struct global_cwq
*gcwq
, int state
)
3434 __releases(&gcwq
->lock
)
3435 __acquires(&gcwq
->lock
)
3437 if (!(gcwq
->trustee_state
== state
||
3438 gcwq
->trustee_state
== TRUSTEE_DONE
)) {
3439 spin_unlock_irq(&gcwq
->lock
);
3440 __wait_event(gcwq
->trustee_wait
,
3441 gcwq
->trustee_state
== state
||
3442 gcwq
->trustee_state
== TRUSTEE_DONE
);
3443 spin_lock_irq(&gcwq
->lock
);
3447 static int __devinit
workqueue_cpu_callback(struct notifier_block
*nfb
,
3448 unsigned long action
,
3451 unsigned int cpu
= (unsigned long)hcpu
;
3452 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3453 struct task_struct
*new_trustee
= NULL
;
3454 struct worker
*uninitialized_var(new_worker
);
3455 unsigned long flags
;
3457 action
&= ~CPU_TASKS_FROZEN
;
3460 case CPU_DOWN_PREPARE
:
3461 new_trustee
= kthread_create(trustee_thread
, gcwq
,
3462 "workqueue_trustee/%d\n", cpu
);
3463 if (IS_ERR(new_trustee
))
3464 return notifier_from_errno(PTR_ERR(new_trustee
));
3465 kthread_bind(new_trustee
, cpu
);
3467 case CPU_UP_PREPARE
:
3468 BUG_ON(gcwq
->first_idle
);
3469 new_worker
= create_worker(gcwq
, false);
3472 kthread_stop(new_trustee
);
3477 /* some are called w/ irq disabled, don't disturb irq status */
3478 spin_lock_irqsave(&gcwq
->lock
, flags
);
3481 case CPU_DOWN_PREPARE
:
3482 /* initialize trustee and tell it to acquire the gcwq */
3483 BUG_ON(gcwq
->trustee
|| gcwq
->trustee_state
!= TRUSTEE_DONE
);
3484 gcwq
->trustee
= new_trustee
;
3485 gcwq
->trustee_state
= TRUSTEE_START
;
3486 wake_up_process(gcwq
->trustee
);
3487 wait_trustee_state(gcwq
, TRUSTEE_IN_CHARGE
);
3489 case CPU_UP_PREPARE
:
3490 BUG_ON(gcwq
->first_idle
);
3491 gcwq
->first_idle
= new_worker
;
3496 * Before this, the trustee and all workers except for
3497 * the ones which are still executing works from
3498 * before the last CPU down must be on the cpu. After
3499 * this, they'll all be diasporas.
3501 gcwq
->flags
|= GCWQ_DISASSOCIATED
;
3505 gcwq
->trustee_state
= TRUSTEE_BUTCHER
;
3507 case CPU_UP_CANCELED
:
3508 destroy_worker(gcwq
->first_idle
);
3509 gcwq
->first_idle
= NULL
;
3512 case CPU_DOWN_FAILED
:
3514 gcwq
->flags
&= ~GCWQ_DISASSOCIATED
;
3515 if (gcwq
->trustee_state
!= TRUSTEE_DONE
) {
3516 gcwq
->trustee_state
= TRUSTEE_RELEASE
;
3517 wake_up_process(gcwq
->trustee
);
3518 wait_trustee_state(gcwq
, TRUSTEE_DONE
);
3522 * Trustee is done and there might be no worker left.
3523 * Put the first_idle in and request a real manager to
3526 spin_unlock_irq(&gcwq
->lock
);
3527 kthread_bind(gcwq
->first_idle
->task
, cpu
);
3528 spin_lock_irq(&gcwq
->lock
);
3529 gcwq
->flags
|= GCWQ_MANAGE_WORKERS
;
3530 start_worker(gcwq
->first_idle
);
3531 gcwq
->first_idle
= NULL
;
3535 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
3537 return notifier_from_errno(0);
3542 struct work_for_cpu
{
3543 struct completion completion
;
3549 static int do_work_for_cpu(void *_wfc
)
3551 struct work_for_cpu
*wfc
= _wfc
;
3552 wfc
->ret
= wfc
->fn(wfc
->arg
);
3553 complete(&wfc
->completion
);
3558 * work_on_cpu - run a function in user context on a particular cpu
3559 * @cpu: the cpu to run on
3560 * @fn: the function to run
3561 * @arg: the function arg
3563 * This will return the value @fn returns.
3564 * It is up to the caller to ensure that the cpu doesn't go offline.
3565 * The caller must not hold any locks which would prevent @fn from completing.
3567 long work_on_cpu(unsigned int cpu
, long (*fn
)(void *), void *arg
)
3569 struct task_struct
*sub_thread
;
3570 struct work_for_cpu wfc
= {
3571 .completion
= COMPLETION_INITIALIZER_ONSTACK(wfc
.completion
),
3576 sub_thread
= kthread_create(do_work_for_cpu
, &wfc
, "work_for_cpu");
3577 if (IS_ERR(sub_thread
))
3578 return PTR_ERR(sub_thread
);
3579 kthread_bind(sub_thread
, cpu
);
3580 wake_up_process(sub_thread
);
3581 wait_for_completion(&wfc
.completion
);
3584 EXPORT_SYMBOL_GPL(work_on_cpu
);
3585 #endif /* CONFIG_SMP */
3587 #ifdef CONFIG_FREEZER
3590 * freeze_workqueues_begin - begin freezing workqueues
3592 * Start freezing workqueues. After this function returns, all freezable
3593 * workqueues will queue new works to their frozen_works list instead of
3597 * Grabs and releases workqueue_lock and gcwq->lock's.
3599 void freeze_workqueues_begin(void)
3603 spin_lock(&workqueue_lock
);
3605 BUG_ON(workqueue_freezing
);
3606 workqueue_freezing
= true;
3608 for_each_gcwq_cpu(cpu
) {
3609 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3610 struct workqueue_struct
*wq
;
3612 spin_lock_irq(&gcwq
->lock
);
3614 BUG_ON(gcwq
->flags
& GCWQ_FREEZING
);
3615 gcwq
->flags
|= GCWQ_FREEZING
;
3617 list_for_each_entry(wq
, &workqueues
, list
) {
3618 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3620 if (cwq
&& wq
->flags
& WQ_FREEZABLE
)
3621 cwq
->max_active
= 0;
3624 spin_unlock_irq(&gcwq
->lock
);
3627 spin_unlock(&workqueue_lock
);
3631 * freeze_workqueues_busy - are freezable workqueues still busy?
3633 * Check whether freezing is complete. This function must be called
3634 * between freeze_workqueues_begin() and thaw_workqueues().
3637 * Grabs and releases workqueue_lock.
3640 * %true if some freezable workqueues are still busy. %false if freezing
3643 bool freeze_workqueues_busy(void)
3648 spin_lock(&workqueue_lock
);
3650 BUG_ON(!workqueue_freezing
);
3652 for_each_gcwq_cpu(cpu
) {
3653 struct workqueue_struct
*wq
;
3655 * nr_active is monotonically decreasing. It's safe
3656 * to peek without lock.
3658 list_for_each_entry(wq
, &workqueues
, list
) {
3659 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3661 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3664 BUG_ON(cwq
->nr_active
< 0);
3665 if (cwq
->nr_active
) {
3672 spin_unlock(&workqueue_lock
);
3677 * thaw_workqueues - thaw workqueues
3679 * Thaw workqueues. Normal queueing is restored and all collected
3680 * frozen works are transferred to their respective gcwq worklists.
3683 * Grabs and releases workqueue_lock and gcwq->lock's.
3685 void thaw_workqueues(void)
3689 spin_lock(&workqueue_lock
);
3691 if (!workqueue_freezing
)
3694 for_each_gcwq_cpu(cpu
) {
3695 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3696 struct workqueue_struct
*wq
;
3698 spin_lock_irq(&gcwq
->lock
);
3700 BUG_ON(!(gcwq
->flags
& GCWQ_FREEZING
));
3701 gcwq
->flags
&= ~GCWQ_FREEZING
;
3703 list_for_each_entry(wq
, &workqueues
, list
) {
3704 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3706 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3709 /* restore max_active and repopulate worklist */
3710 cwq
->max_active
= wq
->saved_max_active
;
3712 while (!list_empty(&cwq
->delayed_works
) &&
3713 cwq
->nr_active
< cwq
->max_active
)
3714 cwq_activate_first_delayed(cwq
);
3717 wake_up_worker(gcwq
);
3719 spin_unlock_irq(&gcwq
->lock
);
3722 workqueue_freezing
= false;
3724 spin_unlock(&workqueue_lock
);
3726 #endif /* CONFIG_FREEZER */
3728 static int __init
init_workqueues(void)
3733 cpu_notifier(workqueue_cpu_callback
, CPU_PRI_WORKQUEUE
);
3735 /* initialize gcwqs */
3736 for_each_gcwq_cpu(cpu
) {
3737 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3739 spin_lock_init(&gcwq
->lock
);
3740 INIT_LIST_HEAD(&gcwq
->worklist
);
3742 gcwq
->flags
|= GCWQ_DISASSOCIATED
;
3744 INIT_LIST_HEAD(&gcwq
->idle_list
);
3745 for (i
= 0; i
< BUSY_WORKER_HASH_SIZE
; i
++)
3746 INIT_HLIST_HEAD(&gcwq
->busy_hash
[i
]);
3748 init_timer_deferrable(&gcwq
->idle_timer
);
3749 gcwq
->idle_timer
.function
= idle_worker_timeout
;
3750 gcwq
->idle_timer
.data
= (unsigned long)gcwq
;
3752 setup_timer(&gcwq
->mayday_timer
, gcwq_mayday_timeout
,
3753 (unsigned long)gcwq
);
3755 ida_init(&gcwq
->worker_ida
);
3757 gcwq
->trustee_state
= TRUSTEE_DONE
;
3758 init_waitqueue_head(&gcwq
->trustee_wait
);
3761 /* create the initial worker */
3762 for_each_online_gcwq_cpu(cpu
) {
3763 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3764 struct worker
*worker
;
3766 if (cpu
!= WORK_CPU_UNBOUND
)
3767 gcwq
->flags
&= ~GCWQ_DISASSOCIATED
;
3768 worker
= create_worker(gcwq
, true);
3770 spin_lock_irq(&gcwq
->lock
);
3771 start_worker(worker
);
3772 spin_unlock_irq(&gcwq
->lock
);
3775 system_wq
= alloc_workqueue("events", 0, 0);
3776 system_long_wq
= alloc_workqueue("events_long", 0, 0);
3777 system_nrt_wq
= alloc_workqueue("events_nrt", WQ_NON_REENTRANT
, 0);
3778 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
3779 WQ_UNBOUND_MAX_ACTIVE
);
3780 system_freezable_wq
= alloc_workqueue("events_freezable",
3782 BUG_ON(!system_wq
|| !system_long_wq
|| !system_nrt_wq
||
3783 !system_unbound_wq
|| !system_freezable_wq
);
3786 early_initcall(init_workqueues
);