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/export.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>
44 #include <linux/hashtable.h>
46 #include "workqueue_internal.h"
52 GCWQ_FREEZING
= 1 << 1, /* freeze in progress */
57 * A bound pool is either associated or disassociated with its CPU.
58 * While associated (!DISASSOCIATED), all workers are bound to the
59 * CPU and none has %WORKER_UNBOUND set and concurrency management
62 * While DISASSOCIATED, the cpu may be offline and all workers have
63 * %WORKER_UNBOUND set and concurrency management disabled, and may
64 * be executing on any CPU. The pool behaves as an unbound one.
66 * Note that DISASSOCIATED can be flipped only while holding
67 * assoc_mutex to avoid changing binding state while
68 * create_worker() is in progress.
70 POOL_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
71 POOL_MANAGING_WORKERS
= 1 << 1, /* managing workers */
72 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
75 WORKER_STARTED
= 1 << 0, /* started */
76 WORKER_DIE
= 1 << 1, /* die die die */
77 WORKER_IDLE
= 1 << 2, /* is idle */
78 WORKER_PREP
= 1 << 3, /* preparing to run works */
79 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
80 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
82 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_UNBOUND
|
85 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
87 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
89 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
92 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
93 /* call for help after 10ms
95 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
96 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give -20.
102 RESCUER_NICE_LEVEL
= -20,
103 HIGHPRI_NICE_LEVEL
= -20,
107 * Structure fields follow one of the following exclusion rules.
109 * I: Modifiable by initialization/destruction paths and read-only for
112 * P: Preemption protected. Disabling preemption is enough and should
113 * only be modified and accessed from the local cpu.
115 * L: gcwq->lock protected. Access with gcwq->lock held.
117 * X: During normal operation, modification requires gcwq->lock and
118 * should be done only from local cpu. Either disabling preemption
119 * on local cpu or grabbing gcwq->lock is enough for read access.
120 * If POOL_DISASSOCIATED is set, it's identical to L.
122 * F: wq->flush_mutex protected.
124 * W: workqueue_lock protected.
127 /* struct worker is defined in workqueue_internal.h */
130 struct global_cwq
*gcwq
; /* I: the owning gcwq */
131 unsigned int flags
; /* X: flags */
133 struct list_head worklist
; /* L: list of pending works */
134 int nr_workers
; /* L: total number of workers */
136 /* nr_idle includes the ones off idle_list for rebinding */
137 int nr_idle
; /* L: currently idle ones */
139 struct list_head idle_list
; /* X: list of idle workers */
140 struct timer_list idle_timer
; /* L: worker idle timeout */
141 struct timer_list mayday_timer
; /* L: SOS timer for workers */
143 struct mutex assoc_mutex
; /* protect POOL_DISASSOCIATED */
144 struct ida worker_ida
; /* L: for worker IDs */
148 * Global per-cpu workqueue. There's one and only one for each cpu
149 * and all works are queued and processed here regardless of their
153 spinlock_t lock
; /* the gcwq lock */
154 unsigned int cpu
; /* I: the associated cpu */
155 unsigned int flags
; /* L: GCWQ_* flags */
157 /* workers are chained either in busy_hash or pool idle_list */
158 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
159 /* L: hash of busy workers */
161 struct worker_pool pools
[NR_STD_WORKER_POOLS
];
162 /* normal and highpri pools */
163 } ____cacheline_aligned_in_smp
;
166 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
167 * work_struct->data are used for flags and thus cwqs need to be
168 * aligned at two's power of the number of flag bits.
170 struct cpu_workqueue_struct
{
171 struct worker_pool
*pool
; /* I: the associated pool */
172 struct workqueue_struct
*wq
; /* I: the owning workqueue */
173 int work_color
; /* L: current color */
174 int flush_color
; /* L: flushing color */
175 int nr_in_flight
[WORK_NR_COLORS
];
176 /* L: nr of in_flight works */
177 int nr_active
; /* L: nr of active works */
178 int max_active
; /* L: max active works */
179 struct list_head delayed_works
; /* L: delayed works */
183 * Structure used to wait for workqueue flush.
186 struct list_head list
; /* F: list of flushers */
187 int flush_color
; /* F: flush color waiting for */
188 struct completion done
; /* flush completion */
192 * All cpumasks are assumed to be always set on UP and thus can't be
193 * used to determine whether there's something to be done.
196 typedef cpumask_var_t mayday_mask_t
;
197 #define mayday_test_and_set_cpu(cpu, mask) \
198 cpumask_test_and_set_cpu((cpu), (mask))
199 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
200 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
201 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
202 #define free_mayday_mask(mask) free_cpumask_var((mask))
204 typedef unsigned long mayday_mask_t
;
205 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
206 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
207 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
208 #define alloc_mayday_mask(maskp, gfp) true
209 #define free_mayday_mask(mask) do { } while (0)
213 * The externally visible workqueue abstraction is an array of
214 * per-CPU workqueues:
216 struct workqueue_struct
{
217 unsigned int flags
; /* W: WQ_* flags */
219 struct cpu_workqueue_struct __percpu
*pcpu
;
220 struct cpu_workqueue_struct
*single
;
222 } cpu_wq
; /* I: cwq's */
223 struct list_head list
; /* W: list of all workqueues */
225 struct mutex flush_mutex
; /* protects wq flushing */
226 int work_color
; /* F: current work color */
227 int flush_color
; /* F: current flush color */
228 atomic_t nr_cwqs_to_flush
; /* flush in progress */
229 struct wq_flusher
*first_flusher
; /* F: first flusher */
230 struct list_head flusher_queue
; /* F: flush waiters */
231 struct list_head flusher_overflow
; /* F: flush overflow list */
233 mayday_mask_t mayday_mask
; /* cpus requesting rescue */
234 struct worker
*rescuer
; /* I: rescue worker */
236 int nr_drainers
; /* W: drain in progress */
237 int saved_max_active
; /* W: saved cwq max_active */
238 #ifdef CONFIG_LOCKDEP
239 struct lockdep_map lockdep_map
;
241 char name
[]; /* I: workqueue name */
244 struct workqueue_struct
*system_wq __read_mostly
;
245 EXPORT_SYMBOL_GPL(system_wq
);
246 struct workqueue_struct
*system_highpri_wq __read_mostly
;
247 EXPORT_SYMBOL_GPL(system_highpri_wq
);
248 struct workqueue_struct
*system_long_wq __read_mostly
;
249 EXPORT_SYMBOL_GPL(system_long_wq
);
250 struct workqueue_struct
*system_unbound_wq __read_mostly
;
251 EXPORT_SYMBOL_GPL(system_unbound_wq
);
252 struct workqueue_struct
*system_freezable_wq __read_mostly
;
253 EXPORT_SYMBOL_GPL(system_freezable_wq
);
255 #define CREATE_TRACE_POINTS
256 #include <trace/events/workqueue.h>
258 #define for_each_worker_pool(pool, gcwq) \
259 for ((pool) = &(gcwq)->pools[0]; \
260 (pool) < &(gcwq)->pools[NR_STD_WORKER_POOLS]; (pool)++)
262 #define for_each_busy_worker(worker, i, pos, gcwq) \
263 hash_for_each(gcwq->busy_hash, i, pos, worker, hentry)
265 static inline int __next_gcwq_cpu(int cpu
, const struct cpumask
*mask
,
268 if (cpu
< nr_cpu_ids
) {
270 cpu
= cpumask_next(cpu
, mask
);
271 if (cpu
< nr_cpu_ids
)
275 return WORK_CPU_UNBOUND
;
277 return WORK_CPU_NONE
;
280 static inline int __next_wq_cpu(int cpu
, const struct cpumask
*mask
,
281 struct workqueue_struct
*wq
)
283 return __next_gcwq_cpu(cpu
, mask
, !(wq
->flags
& WQ_UNBOUND
) ? 1 : 2);
289 * An extra gcwq is defined for an invalid cpu number
290 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
291 * specific CPU. The following iterators are similar to
292 * for_each_*_cpu() iterators but also considers the unbound gcwq.
294 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
295 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
296 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
297 * WORK_CPU_UNBOUND for unbound workqueues
299 #define for_each_gcwq_cpu(cpu) \
300 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
301 (cpu) < WORK_CPU_NONE; \
302 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
304 #define for_each_online_gcwq_cpu(cpu) \
305 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
306 (cpu) < WORK_CPU_NONE; \
307 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
309 #define for_each_cwq_cpu(cpu, wq) \
310 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
311 (cpu) < WORK_CPU_NONE; \
312 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
314 #ifdef CONFIG_DEBUG_OBJECTS_WORK
316 static struct debug_obj_descr work_debug_descr
;
318 static void *work_debug_hint(void *addr
)
320 return ((struct work_struct
*) addr
)->func
;
324 * fixup_init is called when:
325 * - an active object is initialized
327 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
329 struct work_struct
*work
= addr
;
332 case ODEBUG_STATE_ACTIVE
:
333 cancel_work_sync(work
);
334 debug_object_init(work
, &work_debug_descr
);
342 * fixup_activate is called when:
343 * - an active object is activated
344 * - an unknown object is activated (might be a statically initialized object)
346 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
348 struct work_struct
*work
= addr
;
352 case ODEBUG_STATE_NOTAVAILABLE
:
354 * This is not really a fixup. The work struct was
355 * statically initialized. We just make sure that it
356 * is tracked in the object tracker.
358 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
359 debug_object_init(work
, &work_debug_descr
);
360 debug_object_activate(work
, &work_debug_descr
);
366 case ODEBUG_STATE_ACTIVE
:
375 * fixup_free is called when:
376 * - an active object is freed
378 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
380 struct work_struct
*work
= addr
;
383 case ODEBUG_STATE_ACTIVE
:
384 cancel_work_sync(work
);
385 debug_object_free(work
, &work_debug_descr
);
392 static struct debug_obj_descr work_debug_descr
= {
393 .name
= "work_struct",
394 .debug_hint
= work_debug_hint
,
395 .fixup_init
= work_fixup_init
,
396 .fixup_activate
= work_fixup_activate
,
397 .fixup_free
= work_fixup_free
,
400 static inline void debug_work_activate(struct work_struct
*work
)
402 debug_object_activate(work
, &work_debug_descr
);
405 static inline void debug_work_deactivate(struct work_struct
*work
)
407 debug_object_deactivate(work
, &work_debug_descr
);
410 void __init_work(struct work_struct
*work
, int onstack
)
413 debug_object_init_on_stack(work
, &work_debug_descr
);
415 debug_object_init(work
, &work_debug_descr
);
417 EXPORT_SYMBOL_GPL(__init_work
);
419 void destroy_work_on_stack(struct work_struct
*work
)
421 debug_object_free(work
, &work_debug_descr
);
423 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
426 static inline void debug_work_activate(struct work_struct
*work
) { }
427 static inline void debug_work_deactivate(struct work_struct
*work
) { }
430 /* Serializes the accesses to the list of workqueues. */
431 static DEFINE_SPINLOCK(workqueue_lock
);
432 static LIST_HEAD(workqueues
);
433 static bool workqueue_freezing
; /* W: have wqs started freezing? */
436 * The almighty global cpu workqueues. nr_running is the only field
437 * which is expected to be used frequently by other cpus via
438 * try_to_wake_up(). Put it in a separate cacheline.
440 static DEFINE_PER_CPU(struct global_cwq
, global_cwq
);
441 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t
, pool_nr_running
[NR_STD_WORKER_POOLS
]);
444 * Global cpu workqueue and nr_running counter for unbound gcwq. The pools
445 * for online CPUs have POOL_DISASSOCIATED set, and all their workers have
446 * WORKER_UNBOUND set.
448 static struct global_cwq unbound_global_cwq
;
449 static atomic_t unbound_pool_nr_running
[NR_STD_WORKER_POOLS
] = {
450 [0 ... NR_STD_WORKER_POOLS
- 1] = ATOMIC_INIT(0), /* always 0 */
453 static int worker_thread(void *__worker
);
454 static unsigned int work_cpu(struct work_struct
*work
);
456 static int std_worker_pool_pri(struct worker_pool
*pool
)
458 return pool
- pool
->gcwq
->pools
;
461 static struct global_cwq
*get_gcwq(unsigned int cpu
)
463 if (cpu
!= WORK_CPU_UNBOUND
)
464 return &per_cpu(global_cwq
, cpu
);
466 return &unbound_global_cwq
;
469 static atomic_t
*get_pool_nr_running(struct worker_pool
*pool
)
471 int cpu
= pool
->gcwq
->cpu
;
472 int idx
= std_worker_pool_pri(pool
);
474 if (cpu
!= WORK_CPU_UNBOUND
)
475 return &per_cpu(pool_nr_running
, cpu
)[idx
];
477 return &unbound_pool_nr_running
[idx
];
480 static struct cpu_workqueue_struct
*get_cwq(unsigned int cpu
,
481 struct workqueue_struct
*wq
)
483 if (!(wq
->flags
& WQ_UNBOUND
)) {
484 if (likely(cpu
< nr_cpu_ids
))
485 return per_cpu_ptr(wq
->cpu_wq
.pcpu
, cpu
);
486 } else if (likely(cpu
== WORK_CPU_UNBOUND
))
487 return wq
->cpu_wq
.single
;
491 static unsigned int work_color_to_flags(int color
)
493 return color
<< WORK_STRUCT_COLOR_SHIFT
;
496 static int get_work_color(struct work_struct
*work
)
498 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
499 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
502 static int work_next_color(int color
)
504 return (color
+ 1) % WORK_NR_COLORS
;
508 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
509 * contain the pointer to the queued cwq. Once execution starts, the flag
510 * is cleared and the high bits contain OFFQ flags and CPU number.
512 * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling()
513 * and clear_work_data() can be used to set the cwq, cpu or clear
514 * work->data. These functions should only be called while the work is
515 * owned - ie. while the PENDING bit is set.
517 * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to
518 * a work. gcwq is available once the work has been queued anywhere after
519 * initialization until it is sync canceled. cwq is available only while
520 * the work item is queued.
522 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
523 * canceled. While being canceled, a work item may have its PENDING set
524 * but stay off timer and worklist for arbitrarily long and nobody should
525 * try to steal the PENDING bit.
527 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
530 BUG_ON(!work_pending(work
));
531 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
534 static void set_work_cwq(struct work_struct
*work
,
535 struct cpu_workqueue_struct
*cwq
,
536 unsigned long extra_flags
)
538 set_work_data(work
, (unsigned long)cwq
,
539 WORK_STRUCT_PENDING
| WORK_STRUCT_CWQ
| extra_flags
);
542 static void set_work_cpu_and_clear_pending(struct work_struct
*work
,
546 * The following wmb is paired with the implied mb in
547 * test_and_set_bit(PENDING) and ensures all updates to @work made
548 * here are visible to and precede any updates by the next PENDING
552 set_work_data(work
, (unsigned long)cpu
<< WORK_OFFQ_CPU_SHIFT
, 0);
555 static void clear_work_data(struct work_struct
*work
)
557 smp_wmb(); /* see set_work_cpu_and_clear_pending() */
558 set_work_data(work
, WORK_STRUCT_NO_CPU
, 0);
561 static struct cpu_workqueue_struct
*get_work_cwq(struct work_struct
*work
)
563 unsigned long data
= atomic_long_read(&work
->data
);
565 if (data
& WORK_STRUCT_CWQ
)
566 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
571 static struct global_cwq
*get_work_gcwq(struct work_struct
*work
)
573 unsigned long data
= atomic_long_read(&work
->data
);
576 if (data
& WORK_STRUCT_CWQ
)
577 return ((struct cpu_workqueue_struct
*)
578 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->gcwq
;
580 cpu
= data
>> WORK_OFFQ_CPU_SHIFT
;
581 if (cpu
== WORK_CPU_NONE
)
584 BUG_ON(cpu
>= nr_cpu_ids
&& cpu
!= WORK_CPU_UNBOUND
);
585 return get_gcwq(cpu
);
588 static void mark_work_canceling(struct work_struct
*work
)
590 struct global_cwq
*gcwq
= get_work_gcwq(work
);
591 unsigned long cpu
= gcwq
? gcwq
->cpu
: WORK_CPU_NONE
;
593 set_work_data(work
, (cpu
<< WORK_OFFQ_CPU_SHIFT
) | WORK_OFFQ_CANCELING
,
594 WORK_STRUCT_PENDING
);
597 static bool work_is_canceling(struct work_struct
*work
)
599 unsigned long data
= atomic_long_read(&work
->data
);
601 return !(data
& WORK_STRUCT_CWQ
) && (data
& WORK_OFFQ_CANCELING
);
605 * Policy functions. These define the policies on how the global worker
606 * pools are managed. Unless noted otherwise, these functions assume that
607 * they're being called with gcwq->lock held.
610 static bool __need_more_worker(struct worker_pool
*pool
)
612 return !atomic_read(get_pool_nr_running(pool
));
616 * Need to wake up a worker? Called from anything but currently
619 * Note that, because unbound workers never contribute to nr_running, this
620 * function will always return %true for unbound gcwq as long as the
621 * worklist isn't empty.
623 static bool need_more_worker(struct worker_pool
*pool
)
625 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
628 /* Can I start working? Called from busy but !running workers. */
629 static bool may_start_working(struct worker_pool
*pool
)
631 return pool
->nr_idle
;
634 /* Do I need to keep working? Called from currently running workers. */
635 static bool keep_working(struct worker_pool
*pool
)
637 atomic_t
*nr_running
= get_pool_nr_running(pool
);
639 return !list_empty(&pool
->worklist
) && atomic_read(nr_running
) <= 1;
642 /* Do we need a new worker? Called from manager. */
643 static bool need_to_create_worker(struct worker_pool
*pool
)
645 return need_more_worker(pool
) && !may_start_working(pool
);
648 /* Do I need to be the manager? */
649 static bool need_to_manage_workers(struct worker_pool
*pool
)
651 return need_to_create_worker(pool
) ||
652 (pool
->flags
& POOL_MANAGE_WORKERS
);
655 /* Do we have too many workers and should some go away? */
656 static bool too_many_workers(struct worker_pool
*pool
)
658 bool managing
= pool
->flags
& POOL_MANAGING_WORKERS
;
659 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
660 int nr_busy
= pool
->nr_workers
- nr_idle
;
663 * nr_idle and idle_list may disagree if idle rebinding is in
664 * progress. Never return %true if idle_list is empty.
666 if (list_empty(&pool
->idle_list
))
669 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
676 /* Return the first worker. Safe with preemption disabled */
677 static struct worker
*first_worker(struct worker_pool
*pool
)
679 if (unlikely(list_empty(&pool
->idle_list
)))
682 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
686 * wake_up_worker - wake up an idle worker
687 * @pool: worker pool to wake worker from
689 * Wake up the first idle worker of @pool.
692 * spin_lock_irq(gcwq->lock).
694 static void wake_up_worker(struct worker_pool
*pool
)
696 struct worker
*worker
= first_worker(pool
);
699 wake_up_process(worker
->task
);
703 * wq_worker_waking_up - a worker is waking up
704 * @task: task waking up
705 * @cpu: CPU @task is waking up to
707 * This function is called during try_to_wake_up() when a worker is
711 * spin_lock_irq(rq->lock)
713 void wq_worker_waking_up(struct task_struct
*task
, unsigned int cpu
)
715 struct worker
*worker
= kthread_data(task
);
717 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
718 WARN_ON_ONCE(worker
->pool
->gcwq
->cpu
!= cpu
);
719 atomic_inc(get_pool_nr_running(worker
->pool
));
724 * wq_worker_sleeping - a worker is going to sleep
725 * @task: task going to sleep
726 * @cpu: CPU in question, must be the current CPU number
728 * This function is called during schedule() when a busy worker is
729 * going to sleep. Worker on the same cpu can be woken up by
730 * returning pointer to its task.
733 * spin_lock_irq(rq->lock)
736 * Worker task on @cpu to wake up, %NULL if none.
738 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
,
741 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
742 struct worker_pool
*pool
;
743 atomic_t
*nr_running
;
746 * Rescuers, which may not have all the fields set up like normal
747 * workers, also reach here, let's not access anything before
748 * checking NOT_RUNNING.
750 if (worker
->flags
& WORKER_NOT_RUNNING
)
754 nr_running
= get_pool_nr_running(pool
);
756 /* this can only happen on the local cpu */
757 BUG_ON(cpu
!= raw_smp_processor_id());
760 * The counterpart of the following dec_and_test, implied mb,
761 * worklist not empty test sequence is in insert_work().
762 * Please read comment there.
764 * NOT_RUNNING is clear. This means that we're bound to and
765 * running on the local cpu w/ rq lock held and preemption
766 * disabled, which in turn means that none else could be
767 * manipulating idle_list, so dereferencing idle_list without gcwq
770 if (atomic_dec_and_test(nr_running
) && !list_empty(&pool
->worklist
))
771 to_wakeup
= first_worker(pool
);
772 return to_wakeup
? to_wakeup
->task
: NULL
;
776 * worker_set_flags - set worker flags and adjust nr_running accordingly
778 * @flags: flags to set
779 * @wakeup: wakeup an idle worker if necessary
781 * Set @flags in @worker->flags and adjust nr_running accordingly. If
782 * nr_running becomes zero and @wakeup is %true, an idle worker is
786 * spin_lock_irq(gcwq->lock)
788 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
791 struct worker_pool
*pool
= worker
->pool
;
793 WARN_ON_ONCE(worker
->task
!= current
);
796 * If transitioning into NOT_RUNNING, adjust nr_running and
797 * wake up an idle worker as necessary if requested by
800 if ((flags
& WORKER_NOT_RUNNING
) &&
801 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
802 atomic_t
*nr_running
= get_pool_nr_running(pool
);
805 if (atomic_dec_and_test(nr_running
) &&
806 !list_empty(&pool
->worklist
))
807 wake_up_worker(pool
);
809 atomic_dec(nr_running
);
812 worker
->flags
|= flags
;
816 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
818 * @flags: flags to clear
820 * Clear @flags in @worker->flags and adjust nr_running accordingly.
823 * spin_lock_irq(gcwq->lock)
825 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
827 struct worker_pool
*pool
= worker
->pool
;
828 unsigned int oflags
= worker
->flags
;
830 WARN_ON_ONCE(worker
->task
!= current
);
832 worker
->flags
&= ~flags
;
835 * If transitioning out of NOT_RUNNING, increment nr_running. Note
836 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
837 * of multiple flags, not a single flag.
839 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
840 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
841 atomic_inc(get_pool_nr_running(pool
));
845 * find_worker_executing_work - find worker which is executing a work
846 * @gcwq: gcwq of interest
847 * @work: work to find worker for
849 * Find a worker which is executing @work on @gcwq by searching
850 * @gcwq->busy_hash which is keyed by the address of @work. For a worker
851 * to match, its current execution should match the address of @work and
852 * its work function. This is to avoid unwanted dependency between
853 * unrelated work executions through a work item being recycled while still
856 * This is a bit tricky. A work item may be freed once its execution
857 * starts and nothing prevents the freed area from being recycled for
858 * another work item. If the same work item address ends up being reused
859 * before the original execution finishes, workqueue will identify the
860 * recycled work item as currently executing and make it wait until the
861 * current execution finishes, introducing an unwanted dependency.
863 * This function checks the work item address, work function and workqueue
864 * to avoid false positives. Note that this isn't complete as one may
865 * construct a work function which can introduce dependency onto itself
866 * through a recycled work item. Well, if somebody wants to shoot oneself
867 * in the foot that badly, there's only so much we can do, and if such
868 * deadlock actually occurs, it should be easy to locate the culprit work
872 * spin_lock_irq(gcwq->lock).
875 * Pointer to worker which is executing @work if found, NULL
878 static struct worker
*find_worker_executing_work(struct global_cwq
*gcwq
,
879 struct work_struct
*work
)
881 struct worker
*worker
;
882 struct hlist_node
*tmp
;
884 hash_for_each_possible(gcwq
->busy_hash
, worker
, tmp
, hentry
,
886 if (worker
->current_work
== work
&&
887 worker
->current_func
== work
->func
)
894 * move_linked_works - move linked works to a list
895 * @work: start of series of works to be scheduled
896 * @head: target list to append @work to
897 * @nextp: out paramter for nested worklist walking
899 * Schedule linked works starting from @work to @head. Work series to
900 * be scheduled starts at @work and includes any consecutive work with
901 * WORK_STRUCT_LINKED set in its predecessor.
903 * If @nextp is not NULL, it's updated to point to the next work of
904 * the last scheduled work. This allows move_linked_works() to be
905 * nested inside outer list_for_each_entry_safe().
908 * spin_lock_irq(gcwq->lock).
910 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
911 struct work_struct
**nextp
)
913 struct work_struct
*n
;
916 * Linked worklist will always end before the end of the list,
917 * use NULL for list head.
919 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
920 list_move_tail(&work
->entry
, head
);
921 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
926 * If we're already inside safe list traversal and have moved
927 * multiple works to the scheduled queue, the next position
928 * needs to be updated.
934 static void cwq_activate_delayed_work(struct work_struct
*work
)
936 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
938 trace_workqueue_activate_work(work
);
939 move_linked_works(work
, &cwq
->pool
->worklist
, NULL
);
940 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
944 static void cwq_activate_first_delayed(struct cpu_workqueue_struct
*cwq
)
946 struct work_struct
*work
= list_first_entry(&cwq
->delayed_works
,
947 struct work_struct
, entry
);
949 cwq_activate_delayed_work(work
);
953 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
954 * @cwq: cwq of interest
955 * @color: color of work which left the queue
957 * A work either has completed or is removed from pending queue,
958 * decrement nr_in_flight of its cwq and handle workqueue flushing.
961 * spin_lock_irq(gcwq->lock).
963 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct
*cwq
, int color
)
965 /* ignore uncolored works */
966 if (color
== WORK_NO_COLOR
)
969 cwq
->nr_in_flight
[color
]--;
972 if (!list_empty(&cwq
->delayed_works
)) {
973 /* one down, submit a delayed one */
974 if (cwq
->nr_active
< cwq
->max_active
)
975 cwq_activate_first_delayed(cwq
);
978 /* is flush in progress and are we at the flushing tip? */
979 if (likely(cwq
->flush_color
!= color
))
982 /* are there still in-flight works? */
983 if (cwq
->nr_in_flight
[color
])
986 /* this cwq is done, clear flush_color */
987 cwq
->flush_color
= -1;
990 * If this was the last cwq, wake up the first flusher. It
991 * will handle the rest.
993 if (atomic_dec_and_test(&cwq
->wq
->nr_cwqs_to_flush
))
994 complete(&cwq
->wq
->first_flusher
->done
);
998 * try_to_grab_pending - steal work item from worklist and disable irq
999 * @work: work item to steal
1000 * @is_dwork: @work is a delayed_work
1001 * @flags: place to store irq state
1003 * Try to grab PENDING bit of @work. This function can handle @work in any
1004 * stable state - idle, on timer or on worklist. Return values are
1006 * 1 if @work was pending and we successfully stole PENDING
1007 * 0 if @work was idle and we claimed PENDING
1008 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1009 * -ENOENT if someone else is canceling @work, this state may persist
1010 * for arbitrarily long
1012 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1013 * interrupted while holding PENDING and @work off queue, irq must be
1014 * disabled on entry. This, combined with delayed_work->timer being
1015 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1017 * On successful return, >= 0, irq is disabled and the caller is
1018 * responsible for releasing it using local_irq_restore(*@flags).
1020 * This function is safe to call from any context including IRQ handler.
1022 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1023 unsigned long *flags
)
1025 struct global_cwq
*gcwq
;
1027 local_irq_save(*flags
);
1029 /* try to steal the timer if it exists */
1031 struct delayed_work
*dwork
= to_delayed_work(work
);
1034 * dwork->timer is irqsafe. If del_timer() fails, it's
1035 * guaranteed that the timer is not queued anywhere and not
1036 * running on the local CPU.
1038 if (likely(del_timer(&dwork
->timer
)))
1042 /* try to claim PENDING the normal way */
1043 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1047 * The queueing is in progress, or it is already queued. Try to
1048 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1050 gcwq
= get_work_gcwq(work
);
1054 spin_lock(&gcwq
->lock
);
1055 if (!list_empty(&work
->entry
)) {
1057 * This work is queued, but perhaps we locked the wrong gcwq.
1058 * In that case we must see the new value after rmb(), see
1059 * insert_work()->wmb().
1062 if (gcwq
== get_work_gcwq(work
)) {
1063 debug_work_deactivate(work
);
1066 * A delayed work item cannot be grabbed directly
1067 * because it might have linked NO_COLOR work items
1068 * which, if left on the delayed_list, will confuse
1069 * cwq->nr_active management later on and cause
1070 * stall. Make sure the work item is activated
1073 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1074 cwq_activate_delayed_work(work
);
1076 list_del_init(&work
->entry
);
1077 cwq_dec_nr_in_flight(get_work_cwq(work
),
1078 get_work_color(work
));
1080 spin_unlock(&gcwq
->lock
);
1084 spin_unlock(&gcwq
->lock
);
1086 local_irq_restore(*flags
);
1087 if (work_is_canceling(work
))
1094 * insert_work - insert a work into gcwq
1095 * @cwq: cwq @work belongs to
1096 * @work: work to insert
1097 * @head: insertion point
1098 * @extra_flags: extra WORK_STRUCT_* flags to set
1100 * Insert @work which belongs to @cwq into @gcwq after @head.
1101 * @extra_flags is or'd to work_struct flags.
1104 * spin_lock_irq(gcwq->lock).
1106 static void insert_work(struct cpu_workqueue_struct
*cwq
,
1107 struct work_struct
*work
, struct list_head
*head
,
1108 unsigned int extra_flags
)
1110 struct worker_pool
*pool
= cwq
->pool
;
1112 /* we own @work, set data and link */
1113 set_work_cwq(work
, cwq
, extra_flags
);
1116 * Ensure that we get the right work->data if we see the
1117 * result of list_add() below, see try_to_grab_pending().
1121 list_add_tail(&work
->entry
, head
);
1124 * Ensure either worker_sched_deactivated() sees the above
1125 * list_add_tail() or we see zero nr_running to avoid workers
1126 * lying around lazily while there are works to be processed.
1130 if (__need_more_worker(pool
))
1131 wake_up_worker(pool
);
1135 * Test whether @work is being queued from another work executing on the
1136 * same workqueue. This is rather expensive and should only be used from
1139 static bool is_chained_work(struct workqueue_struct
*wq
)
1141 unsigned long flags
;
1144 for_each_gcwq_cpu(cpu
) {
1145 struct global_cwq
*gcwq
= get_gcwq(cpu
);
1146 struct worker
*worker
;
1147 struct hlist_node
*pos
;
1150 spin_lock_irqsave(&gcwq
->lock
, flags
);
1151 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
1152 if (worker
->task
!= current
)
1154 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1156 * I'm @worker, no locking necessary. See if @work
1157 * is headed to the same workqueue.
1159 return worker
->current_cwq
->wq
== wq
;
1161 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1166 static void __queue_work(unsigned int cpu
, struct workqueue_struct
*wq
,
1167 struct work_struct
*work
)
1169 struct global_cwq
*gcwq
;
1170 struct cpu_workqueue_struct
*cwq
;
1171 struct list_head
*worklist
;
1172 unsigned int work_flags
;
1173 unsigned int req_cpu
= cpu
;
1176 * While a work item is PENDING && off queue, a task trying to
1177 * steal the PENDING will busy-loop waiting for it to either get
1178 * queued or lose PENDING. Grabbing PENDING and queueing should
1179 * happen with IRQ disabled.
1181 WARN_ON_ONCE(!irqs_disabled());
1183 debug_work_activate(work
);
1185 /* if dying, only works from the same workqueue are allowed */
1186 if (unlikely(wq
->flags
& WQ_DRAINING
) &&
1187 WARN_ON_ONCE(!is_chained_work(wq
)))
1190 /* determine gcwq to use */
1191 if (!(wq
->flags
& WQ_UNBOUND
)) {
1192 struct global_cwq
*last_gcwq
;
1194 if (cpu
== WORK_CPU_UNBOUND
)
1195 cpu
= raw_smp_processor_id();
1198 * It's multi cpu. If @work was previously on a different
1199 * cpu, it might still be running there, in which case the
1200 * work needs to be queued on that cpu to guarantee
1203 gcwq
= get_gcwq(cpu
);
1204 last_gcwq
= get_work_gcwq(work
);
1206 if (last_gcwq
&& last_gcwq
!= gcwq
) {
1207 struct worker
*worker
;
1209 spin_lock(&last_gcwq
->lock
);
1211 worker
= find_worker_executing_work(last_gcwq
, work
);
1213 if (worker
&& worker
->current_cwq
->wq
== wq
)
1216 /* meh... not running there, queue here */
1217 spin_unlock(&last_gcwq
->lock
);
1218 spin_lock(&gcwq
->lock
);
1221 spin_lock(&gcwq
->lock
);
1224 gcwq
= get_gcwq(WORK_CPU_UNBOUND
);
1225 spin_lock(&gcwq
->lock
);
1228 /* gcwq determined, get cwq and queue */
1229 cwq
= get_cwq(gcwq
->cpu
, wq
);
1230 trace_workqueue_queue_work(req_cpu
, cwq
, work
);
1232 if (WARN_ON(!list_empty(&work
->entry
))) {
1233 spin_unlock(&gcwq
->lock
);
1237 cwq
->nr_in_flight
[cwq
->work_color
]++;
1238 work_flags
= work_color_to_flags(cwq
->work_color
);
1240 if (likely(cwq
->nr_active
< cwq
->max_active
)) {
1241 trace_workqueue_activate_work(work
);
1243 worklist
= &cwq
->pool
->worklist
;
1245 work_flags
|= WORK_STRUCT_DELAYED
;
1246 worklist
= &cwq
->delayed_works
;
1249 insert_work(cwq
, work
, worklist
, work_flags
);
1251 spin_unlock(&gcwq
->lock
);
1255 * queue_work_on - queue work on specific cpu
1256 * @cpu: CPU number to execute work on
1257 * @wq: workqueue to use
1258 * @work: work to queue
1260 * Returns %false if @work was already on a queue, %true otherwise.
1262 * We queue the work to a specific CPU, the caller must ensure it
1265 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1266 struct work_struct
*work
)
1269 unsigned long flags
;
1271 local_irq_save(flags
);
1273 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1274 __queue_work(cpu
, wq
, work
);
1278 local_irq_restore(flags
);
1281 EXPORT_SYMBOL_GPL(queue_work_on
);
1284 * queue_work - queue work on a workqueue
1285 * @wq: workqueue to use
1286 * @work: work to queue
1288 * Returns %false if @work was already on a queue, %true otherwise.
1290 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1291 * it can be processed by another CPU.
1293 bool queue_work(struct workqueue_struct
*wq
, struct work_struct
*work
)
1295 return queue_work_on(WORK_CPU_UNBOUND
, wq
, work
);
1297 EXPORT_SYMBOL_GPL(queue_work
);
1299 void delayed_work_timer_fn(unsigned long __data
)
1301 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1302 struct cpu_workqueue_struct
*cwq
= get_work_cwq(&dwork
->work
);
1304 /* should have been called from irqsafe timer with irq already off */
1305 __queue_work(dwork
->cpu
, cwq
->wq
, &dwork
->work
);
1307 EXPORT_SYMBOL_GPL(delayed_work_timer_fn
);
1309 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1310 struct delayed_work
*dwork
, unsigned long delay
)
1312 struct timer_list
*timer
= &dwork
->timer
;
1313 struct work_struct
*work
= &dwork
->work
;
1316 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1317 timer
->data
!= (unsigned long)dwork
);
1318 WARN_ON_ONCE(timer_pending(timer
));
1319 WARN_ON_ONCE(!list_empty(&work
->entry
));
1322 * If @delay is 0, queue @dwork->work immediately. This is for
1323 * both optimization and correctness. The earliest @timer can
1324 * expire is on the closest next tick and delayed_work users depend
1325 * on that there's no such delay when @delay is 0.
1328 __queue_work(cpu
, wq
, &dwork
->work
);
1332 timer_stats_timer_set_start_info(&dwork
->timer
);
1335 * This stores cwq for the moment, for the timer_fn. Note that the
1336 * work's gcwq is preserved to allow reentrance detection for
1339 if (!(wq
->flags
& WQ_UNBOUND
)) {
1340 struct global_cwq
*gcwq
= get_work_gcwq(work
);
1343 * If we cannot get the last gcwq from @work directly,
1344 * select the last CPU such that it avoids unnecessarily
1345 * triggering non-reentrancy check in __queue_work().
1350 if (lcpu
== WORK_CPU_UNBOUND
)
1351 lcpu
= raw_smp_processor_id();
1353 lcpu
= WORK_CPU_UNBOUND
;
1356 set_work_cwq(work
, get_cwq(lcpu
, wq
), 0);
1359 timer
->expires
= jiffies
+ delay
;
1361 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1362 add_timer_on(timer
, cpu
);
1368 * queue_delayed_work_on - queue work on specific CPU after delay
1369 * @cpu: CPU number to execute work on
1370 * @wq: workqueue to use
1371 * @dwork: work to queue
1372 * @delay: number of jiffies to wait before queueing
1374 * Returns %false if @work was already on a queue, %true otherwise. If
1375 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1378 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1379 struct delayed_work
*dwork
, unsigned long delay
)
1381 struct work_struct
*work
= &dwork
->work
;
1383 unsigned long flags
;
1385 /* read the comment in __queue_work() */
1386 local_irq_save(flags
);
1388 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1389 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1393 local_irq_restore(flags
);
1396 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1399 * queue_delayed_work - queue work on a workqueue after delay
1400 * @wq: workqueue to use
1401 * @dwork: delayable work to queue
1402 * @delay: number of jiffies to wait before queueing
1404 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1406 bool queue_delayed_work(struct workqueue_struct
*wq
,
1407 struct delayed_work
*dwork
, unsigned long delay
)
1409 return queue_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1411 EXPORT_SYMBOL_GPL(queue_delayed_work
);
1414 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1415 * @cpu: CPU number to execute work on
1416 * @wq: workqueue to use
1417 * @dwork: work to queue
1418 * @delay: number of jiffies to wait before queueing
1420 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1421 * modify @dwork's timer so that it expires after @delay. If @delay is
1422 * zero, @work is guaranteed to be scheduled immediately regardless of its
1425 * Returns %false if @dwork was idle and queued, %true if @dwork was
1426 * pending and its timer was modified.
1428 * This function is safe to call from any context including IRQ handler.
1429 * See try_to_grab_pending() for details.
1431 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1432 struct delayed_work
*dwork
, unsigned long delay
)
1434 unsigned long flags
;
1438 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1439 } while (unlikely(ret
== -EAGAIN
));
1441 if (likely(ret
>= 0)) {
1442 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1443 local_irq_restore(flags
);
1446 /* -ENOENT from try_to_grab_pending() becomes %true */
1449 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1452 * mod_delayed_work - modify delay of or queue a delayed work
1453 * @wq: workqueue to use
1454 * @dwork: work to queue
1455 * @delay: number of jiffies to wait before queueing
1457 * mod_delayed_work_on() on local CPU.
1459 bool mod_delayed_work(struct workqueue_struct
*wq
, struct delayed_work
*dwork
,
1460 unsigned long delay
)
1462 return mod_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1464 EXPORT_SYMBOL_GPL(mod_delayed_work
);
1467 * worker_enter_idle - enter idle state
1468 * @worker: worker which is entering idle state
1470 * @worker is entering idle state. Update stats and idle timer if
1474 * spin_lock_irq(gcwq->lock).
1476 static void worker_enter_idle(struct worker
*worker
)
1478 struct worker_pool
*pool
= worker
->pool
;
1480 BUG_ON(worker
->flags
& WORKER_IDLE
);
1481 BUG_ON(!list_empty(&worker
->entry
) &&
1482 (worker
->hentry
.next
|| worker
->hentry
.pprev
));
1484 /* can't use worker_set_flags(), also called from start_worker() */
1485 worker
->flags
|= WORKER_IDLE
;
1487 worker
->last_active
= jiffies
;
1489 /* idle_list is LIFO */
1490 list_add(&worker
->entry
, &pool
->idle_list
);
1492 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1493 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1496 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1497 * gcwq->lock between setting %WORKER_UNBOUND and zapping
1498 * nr_running, the warning may trigger spuriously. Check iff
1499 * unbind is not in progress.
1501 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1502 pool
->nr_workers
== pool
->nr_idle
&&
1503 atomic_read(get_pool_nr_running(pool
)));
1507 * worker_leave_idle - leave idle state
1508 * @worker: worker which is leaving idle state
1510 * @worker is leaving idle state. Update stats.
1513 * spin_lock_irq(gcwq->lock).
1515 static void worker_leave_idle(struct worker
*worker
)
1517 struct worker_pool
*pool
= worker
->pool
;
1519 BUG_ON(!(worker
->flags
& WORKER_IDLE
));
1520 worker_clr_flags(worker
, WORKER_IDLE
);
1522 list_del_init(&worker
->entry
);
1526 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1529 * Works which are scheduled while the cpu is online must at least be
1530 * scheduled to a worker which is bound to the cpu so that if they are
1531 * flushed from cpu callbacks while cpu is going down, they are
1532 * guaranteed to execute on the cpu.
1534 * This function is to be used by rogue workers and rescuers to bind
1535 * themselves to the target cpu and may race with cpu going down or
1536 * coming online. kthread_bind() can't be used because it may put the
1537 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1538 * verbatim as it's best effort and blocking and gcwq may be
1539 * [dis]associated in the meantime.
1541 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1542 * binding against %POOL_DISASSOCIATED which is set during
1543 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1544 * enters idle state or fetches works without dropping lock, it can
1545 * guarantee the scheduling requirement described in the first paragraph.
1548 * Might sleep. Called without any lock but returns with gcwq->lock
1552 * %true if the associated gcwq is online (@worker is successfully
1553 * bound), %false if offline.
1555 static bool worker_maybe_bind_and_lock(struct worker
*worker
)
1556 __acquires(&gcwq
->lock
)
1558 struct worker_pool
*pool
= worker
->pool
;
1559 struct global_cwq
*gcwq
= pool
->gcwq
;
1560 struct task_struct
*task
= worker
->task
;
1564 * The following call may fail, succeed or succeed
1565 * without actually migrating the task to the cpu if
1566 * it races with cpu hotunplug operation. Verify
1567 * against POOL_DISASSOCIATED.
1569 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1570 set_cpus_allowed_ptr(task
, get_cpu_mask(gcwq
->cpu
));
1572 spin_lock_irq(&gcwq
->lock
);
1573 if (pool
->flags
& POOL_DISASSOCIATED
)
1575 if (task_cpu(task
) == gcwq
->cpu
&&
1576 cpumask_equal(¤t
->cpus_allowed
,
1577 get_cpu_mask(gcwq
->cpu
)))
1579 spin_unlock_irq(&gcwq
->lock
);
1582 * We've raced with CPU hot[un]plug. Give it a breather
1583 * and retry migration. cond_resched() is required here;
1584 * otherwise, we might deadlock against cpu_stop trying to
1585 * bring down the CPU on non-preemptive kernel.
1593 * Rebind an idle @worker to its CPU. worker_thread() will test
1594 * list_empty(@worker->entry) before leaving idle and call this function.
1596 static void idle_worker_rebind(struct worker
*worker
)
1598 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1600 /* CPU may go down again inbetween, clear UNBOUND only on success */
1601 if (worker_maybe_bind_and_lock(worker
))
1602 worker_clr_flags(worker
, WORKER_UNBOUND
);
1604 /* rebind complete, become available again */
1605 list_add(&worker
->entry
, &worker
->pool
->idle_list
);
1606 spin_unlock_irq(&gcwq
->lock
);
1610 * Function for @worker->rebind.work used to rebind unbound busy workers to
1611 * the associated cpu which is coming back online. This is scheduled by
1612 * cpu up but can race with other cpu hotplug operations and may be
1613 * executed twice without intervening cpu down.
1615 static void busy_worker_rebind_fn(struct work_struct
*work
)
1617 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1618 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1620 if (worker_maybe_bind_and_lock(worker
))
1621 worker_clr_flags(worker
, WORKER_UNBOUND
);
1623 spin_unlock_irq(&gcwq
->lock
);
1627 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1628 * @gcwq: gcwq of interest
1630 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1631 * is different for idle and busy ones.
1633 * Idle ones will be removed from the idle_list and woken up. They will
1634 * add themselves back after completing rebind. This ensures that the
1635 * idle_list doesn't contain any unbound workers when re-bound busy workers
1636 * try to perform local wake-ups for concurrency management.
1638 * Busy workers can rebind after they finish their current work items.
1639 * Queueing the rebind work item at the head of the scheduled list is
1640 * enough. Note that nr_running will be properly bumped as busy workers
1643 * On return, all non-manager workers are scheduled for rebind - see
1644 * manage_workers() for the manager special case. Any idle worker
1645 * including the manager will not appear on @idle_list until rebind is
1646 * complete, making local wake-ups safe.
1648 static void rebind_workers(struct global_cwq
*gcwq
)
1650 struct worker_pool
*pool
;
1651 struct worker
*worker
, *n
;
1652 struct hlist_node
*pos
;
1655 lockdep_assert_held(&gcwq
->lock
);
1657 for_each_worker_pool(pool
, gcwq
)
1658 lockdep_assert_held(&pool
->assoc_mutex
);
1660 /* dequeue and kick idle ones */
1661 for_each_worker_pool(pool
, gcwq
) {
1662 list_for_each_entry_safe(worker
, n
, &pool
->idle_list
, entry
) {
1664 * idle workers should be off @pool->idle_list
1665 * until rebind is complete to avoid receiving
1666 * premature local wake-ups.
1668 list_del_init(&worker
->entry
);
1671 * worker_thread() will see the above dequeuing
1672 * and call idle_worker_rebind().
1674 wake_up_process(worker
->task
);
1678 /* rebind busy workers */
1679 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
1680 struct work_struct
*rebind_work
= &worker
->rebind_work
;
1681 struct workqueue_struct
*wq
;
1683 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
1684 work_data_bits(rebind_work
)))
1687 debug_work_activate(rebind_work
);
1690 * wq doesn't really matter but let's keep @worker->pool
1691 * and @cwq->pool consistent for sanity.
1693 if (std_worker_pool_pri(worker
->pool
))
1694 wq
= system_highpri_wq
;
1698 insert_work(get_cwq(gcwq
->cpu
, wq
), rebind_work
,
1699 worker
->scheduled
.next
,
1700 work_color_to_flags(WORK_NO_COLOR
));
1704 static struct worker
*alloc_worker(void)
1706 struct worker
*worker
;
1708 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1710 INIT_LIST_HEAD(&worker
->entry
);
1711 INIT_LIST_HEAD(&worker
->scheduled
);
1712 INIT_WORK(&worker
->rebind_work
, busy_worker_rebind_fn
);
1713 /* on creation a worker is in !idle && prep state */
1714 worker
->flags
= WORKER_PREP
;
1720 * create_worker - create a new workqueue worker
1721 * @pool: pool the new worker will belong to
1723 * Create a new worker which is bound to @pool. The returned worker
1724 * can be started by calling start_worker() or destroyed using
1728 * Might sleep. Does GFP_KERNEL allocations.
1731 * Pointer to the newly created worker.
1733 static struct worker
*create_worker(struct worker_pool
*pool
)
1735 struct global_cwq
*gcwq
= pool
->gcwq
;
1736 const char *pri
= std_worker_pool_pri(pool
) ? "H" : "";
1737 struct worker
*worker
= NULL
;
1740 spin_lock_irq(&gcwq
->lock
);
1741 while (ida_get_new(&pool
->worker_ida
, &id
)) {
1742 spin_unlock_irq(&gcwq
->lock
);
1743 if (!ida_pre_get(&pool
->worker_ida
, GFP_KERNEL
))
1745 spin_lock_irq(&gcwq
->lock
);
1747 spin_unlock_irq(&gcwq
->lock
);
1749 worker
= alloc_worker();
1753 worker
->pool
= pool
;
1756 if (gcwq
->cpu
!= WORK_CPU_UNBOUND
)
1757 worker
->task
= kthread_create_on_node(worker_thread
,
1758 worker
, cpu_to_node(gcwq
->cpu
),
1759 "kworker/%u:%d%s", gcwq
->cpu
, id
, pri
);
1761 worker
->task
= kthread_create(worker_thread
, worker
,
1762 "kworker/u:%d%s", id
, pri
);
1763 if (IS_ERR(worker
->task
))
1766 if (std_worker_pool_pri(pool
))
1767 set_user_nice(worker
->task
, HIGHPRI_NICE_LEVEL
);
1770 * Determine CPU binding of the new worker depending on
1771 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1772 * flag remains stable across this function. See the comments
1773 * above the flag definition for details.
1775 * As an unbound worker may later become a regular one if CPU comes
1776 * online, make sure every worker has %PF_THREAD_BOUND set.
1778 if (!(pool
->flags
& POOL_DISASSOCIATED
)) {
1779 kthread_bind(worker
->task
, gcwq
->cpu
);
1781 worker
->task
->flags
|= PF_THREAD_BOUND
;
1782 worker
->flags
|= WORKER_UNBOUND
;
1788 spin_lock_irq(&gcwq
->lock
);
1789 ida_remove(&pool
->worker_ida
, id
);
1790 spin_unlock_irq(&gcwq
->lock
);
1797 * start_worker - start a newly created worker
1798 * @worker: worker to start
1800 * Make the gcwq aware of @worker and start it.
1803 * spin_lock_irq(gcwq->lock).
1805 static void start_worker(struct worker
*worker
)
1807 worker
->flags
|= WORKER_STARTED
;
1808 worker
->pool
->nr_workers
++;
1809 worker_enter_idle(worker
);
1810 wake_up_process(worker
->task
);
1814 * destroy_worker - destroy a workqueue worker
1815 * @worker: worker to be destroyed
1817 * Destroy @worker and adjust @gcwq stats accordingly.
1820 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1822 static void destroy_worker(struct worker
*worker
)
1824 struct worker_pool
*pool
= worker
->pool
;
1825 struct global_cwq
*gcwq
= pool
->gcwq
;
1826 int id
= worker
->id
;
1828 /* sanity check frenzy */
1829 BUG_ON(worker
->current_work
);
1830 BUG_ON(!list_empty(&worker
->scheduled
));
1832 if (worker
->flags
& WORKER_STARTED
)
1834 if (worker
->flags
& WORKER_IDLE
)
1837 list_del_init(&worker
->entry
);
1838 worker
->flags
|= WORKER_DIE
;
1840 spin_unlock_irq(&gcwq
->lock
);
1842 kthread_stop(worker
->task
);
1845 spin_lock_irq(&gcwq
->lock
);
1846 ida_remove(&pool
->worker_ida
, id
);
1849 static void idle_worker_timeout(unsigned long __pool
)
1851 struct worker_pool
*pool
= (void *)__pool
;
1852 struct global_cwq
*gcwq
= pool
->gcwq
;
1854 spin_lock_irq(&gcwq
->lock
);
1856 if (too_many_workers(pool
)) {
1857 struct worker
*worker
;
1858 unsigned long expires
;
1860 /* idle_list is kept in LIFO order, check the last one */
1861 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1862 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1864 if (time_before(jiffies
, expires
))
1865 mod_timer(&pool
->idle_timer
, expires
);
1867 /* it's been idle for too long, wake up manager */
1868 pool
->flags
|= POOL_MANAGE_WORKERS
;
1869 wake_up_worker(pool
);
1873 spin_unlock_irq(&gcwq
->lock
);
1876 static bool send_mayday(struct work_struct
*work
)
1878 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1879 struct workqueue_struct
*wq
= cwq
->wq
;
1882 if (!(wq
->flags
& WQ_RESCUER
))
1885 /* mayday mayday mayday */
1886 cpu
= cwq
->pool
->gcwq
->cpu
;
1887 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1888 if (cpu
== WORK_CPU_UNBOUND
)
1890 if (!mayday_test_and_set_cpu(cpu
, wq
->mayday_mask
))
1891 wake_up_process(wq
->rescuer
->task
);
1895 static void gcwq_mayday_timeout(unsigned long __pool
)
1897 struct worker_pool
*pool
= (void *)__pool
;
1898 struct global_cwq
*gcwq
= pool
->gcwq
;
1899 struct work_struct
*work
;
1901 spin_lock_irq(&gcwq
->lock
);
1903 if (need_to_create_worker(pool
)) {
1905 * We've been trying to create a new worker but
1906 * haven't been successful. We might be hitting an
1907 * allocation deadlock. Send distress signals to
1910 list_for_each_entry(work
, &pool
->worklist
, entry
)
1914 spin_unlock_irq(&gcwq
->lock
);
1916 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1920 * maybe_create_worker - create a new worker if necessary
1921 * @pool: pool to create a new worker for
1923 * Create a new worker for @pool if necessary. @pool is guaranteed to
1924 * have at least one idle worker on return from this function. If
1925 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1926 * sent to all rescuers with works scheduled on @pool to resolve
1927 * possible allocation deadlock.
1929 * On return, need_to_create_worker() is guaranteed to be false and
1930 * may_start_working() true.
1933 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1934 * multiple times. Does GFP_KERNEL allocations. Called only from
1938 * false if no action was taken and gcwq->lock stayed locked, true
1941 static bool maybe_create_worker(struct worker_pool
*pool
)
1942 __releases(&gcwq
->lock
)
1943 __acquires(&gcwq
->lock
)
1945 struct global_cwq
*gcwq
= pool
->gcwq
;
1947 if (!need_to_create_worker(pool
))
1950 spin_unlock_irq(&gcwq
->lock
);
1952 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1953 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1956 struct worker
*worker
;
1958 worker
= create_worker(pool
);
1960 del_timer_sync(&pool
->mayday_timer
);
1961 spin_lock_irq(&gcwq
->lock
);
1962 start_worker(worker
);
1963 BUG_ON(need_to_create_worker(pool
));
1967 if (!need_to_create_worker(pool
))
1970 __set_current_state(TASK_INTERRUPTIBLE
);
1971 schedule_timeout(CREATE_COOLDOWN
);
1973 if (!need_to_create_worker(pool
))
1977 del_timer_sync(&pool
->mayday_timer
);
1978 spin_lock_irq(&gcwq
->lock
);
1979 if (need_to_create_worker(pool
))
1985 * maybe_destroy_worker - destroy workers which have been idle for a while
1986 * @pool: pool to destroy workers for
1988 * Destroy @pool workers which have been idle for longer than
1989 * IDLE_WORKER_TIMEOUT.
1992 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1993 * multiple times. Called only from manager.
1996 * false if no action was taken and gcwq->lock stayed locked, true
1999 static bool maybe_destroy_workers(struct worker_pool
*pool
)
2003 while (too_many_workers(pool
)) {
2004 struct worker
*worker
;
2005 unsigned long expires
;
2007 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2008 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2010 if (time_before(jiffies
, expires
)) {
2011 mod_timer(&pool
->idle_timer
, expires
);
2015 destroy_worker(worker
);
2023 * manage_workers - manage worker pool
2026 * Assume the manager role and manage gcwq worker pool @worker belongs
2027 * to. At any given time, there can be only zero or one manager per
2028 * gcwq. The exclusion is handled automatically by this function.
2030 * The caller can safely start processing works on false return. On
2031 * true return, it's guaranteed that need_to_create_worker() is false
2032 * and may_start_working() is true.
2035 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2036 * multiple times. Does GFP_KERNEL allocations.
2039 * false if no action was taken and gcwq->lock stayed locked, true if
2040 * some action was taken.
2042 static bool manage_workers(struct worker
*worker
)
2044 struct worker_pool
*pool
= worker
->pool
;
2047 if (pool
->flags
& POOL_MANAGING_WORKERS
)
2050 pool
->flags
|= POOL_MANAGING_WORKERS
;
2053 * To simplify both worker management and CPU hotplug, hold off
2054 * management while hotplug is in progress. CPU hotplug path can't
2055 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2056 * lead to idle worker depletion (all become busy thinking someone
2057 * else is managing) which in turn can result in deadlock under
2058 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2059 * manager against CPU hotplug.
2061 * assoc_mutex would always be free unless CPU hotplug is in
2062 * progress. trylock first without dropping @gcwq->lock.
2064 if (unlikely(!mutex_trylock(&pool
->assoc_mutex
))) {
2065 spin_unlock_irq(&pool
->gcwq
->lock
);
2066 mutex_lock(&pool
->assoc_mutex
);
2068 * CPU hotplug could have happened while we were waiting
2069 * for assoc_mutex. Hotplug itself can't handle us
2070 * because manager isn't either on idle or busy list, and
2071 * @gcwq's state and ours could have deviated.
2073 * As hotplug is now excluded via assoc_mutex, we can
2074 * simply try to bind. It will succeed or fail depending
2075 * on @gcwq's current state. Try it and adjust
2076 * %WORKER_UNBOUND accordingly.
2078 if (worker_maybe_bind_and_lock(worker
))
2079 worker
->flags
&= ~WORKER_UNBOUND
;
2081 worker
->flags
|= WORKER_UNBOUND
;
2086 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2089 * Destroy and then create so that may_start_working() is true
2092 ret
|= maybe_destroy_workers(pool
);
2093 ret
|= maybe_create_worker(pool
);
2095 pool
->flags
&= ~POOL_MANAGING_WORKERS
;
2096 mutex_unlock(&pool
->assoc_mutex
);
2101 * process_one_work - process single work
2103 * @work: work to process
2105 * Process @work. This function contains all the logics necessary to
2106 * process a single work including synchronization against and
2107 * interaction with other workers on the same cpu, queueing and
2108 * flushing. As long as context requirement is met, any worker can
2109 * call this function to process a work.
2112 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2114 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2115 __releases(&gcwq
->lock
)
2116 __acquires(&gcwq
->lock
)
2118 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
2119 struct worker_pool
*pool
= worker
->pool
;
2120 struct global_cwq
*gcwq
= pool
->gcwq
;
2121 bool cpu_intensive
= cwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2123 struct worker
*collision
;
2124 #ifdef CONFIG_LOCKDEP
2126 * It is permissible to free the struct work_struct from
2127 * inside the function that is called from it, this we need to
2128 * take into account for lockdep too. To avoid bogus "held
2129 * lock freed" warnings as well as problems when looking into
2130 * work->lockdep_map, make a copy and use that here.
2132 struct lockdep_map lockdep_map
;
2134 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2137 * Ensure we're on the correct CPU. DISASSOCIATED test is
2138 * necessary to avoid spurious warnings from rescuers servicing the
2139 * unbound or a disassociated pool.
2141 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2142 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2143 raw_smp_processor_id() != gcwq
->cpu
);
2146 * A single work shouldn't be executed concurrently by
2147 * multiple workers on a single cpu. Check whether anyone is
2148 * already processing the work. If so, defer the work to the
2149 * currently executing one.
2151 collision
= find_worker_executing_work(gcwq
, work
);
2152 if (unlikely(collision
)) {
2153 move_linked_works(work
, &collision
->scheduled
, NULL
);
2157 /* claim and dequeue */
2158 debug_work_deactivate(work
);
2159 hash_add(gcwq
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2160 worker
->current_work
= work
;
2161 worker
->current_func
= work
->func
;
2162 worker
->current_cwq
= cwq
;
2163 work_color
= get_work_color(work
);
2165 list_del_init(&work
->entry
);
2168 * CPU intensive works don't participate in concurrency
2169 * management. They're the scheduler's responsibility.
2171 if (unlikely(cpu_intensive
))
2172 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2175 * Unbound gcwq isn't concurrency managed and work items should be
2176 * executed ASAP. Wake up another worker if necessary.
2178 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2179 wake_up_worker(pool
);
2182 * Record the last CPU and clear PENDING which should be the last
2183 * update to @work. Also, do this inside @gcwq->lock so that
2184 * PENDING and queued state changes happen together while IRQ is
2187 set_work_cpu_and_clear_pending(work
, gcwq
->cpu
);
2189 spin_unlock_irq(&gcwq
->lock
);
2191 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2192 lock_map_acquire(&lockdep_map
);
2193 trace_workqueue_execute_start(work
);
2194 worker
->current_func(work
);
2196 * While we must be careful to not use "work" after this, the trace
2197 * point will only record its address.
2199 trace_workqueue_execute_end(work
);
2200 lock_map_release(&lockdep_map
);
2201 lock_map_release(&cwq
->wq
->lockdep_map
);
2203 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2204 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2205 " last function: %pf\n",
2206 current
->comm
, preempt_count(), task_pid_nr(current
),
2207 worker
->current_func
);
2208 debug_show_held_locks(current
);
2212 spin_lock_irq(&gcwq
->lock
);
2214 /* clear cpu intensive status */
2215 if (unlikely(cpu_intensive
))
2216 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2218 /* we're done with it, release */
2219 hash_del(&worker
->hentry
);
2220 worker
->current_work
= NULL
;
2221 worker
->current_func
= NULL
;
2222 worker
->current_cwq
= NULL
;
2223 cwq_dec_nr_in_flight(cwq
, work_color
);
2227 * process_scheduled_works - process scheduled works
2230 * Process all scheduled works. Please note that the scheduled list
2231 * may change while processing a work, so this function repeatedly
2232 * fetches a work from the top and executes it.
2235 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2238 static void process_scheduled_works(struct worker
*worker
)
2240 while (!list_empty(&worker
->scheduled
)) {
2241 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2242 struct work_struct
, entry
);
2243 process_one_work(worker
, work
);
2248 * worker_thread - the worker thread function
2251 * The gcwq worker thread function. There's a single dynamic pool of
2252 * these per each cpu. These workers process all works regardless of
2253 * their specific target workqueue. The only exception is works which
2254 * belong to workqueues with a rescuer which will be explained in
2257 static int worker_thread(void *__worker
)
2259 struct worker
*worker
= __worker
;
2260 struct worker_pool
*pool
= worker
->pool
;
2261 struct global_cwq
*gcwq
= pool
->gcwq
;
2263 /* tell the scheduler that this is a workqueue worker */
2264 worker
->task
->flags
|= PF_WQ_WORKER
;
2266 spin_lock_irq(&gcwq
->lock
);
2268 /* we are off idle list if destruction or rebind is requested */
2269 if (unlikely(list_empty(&worker
->entry
))) {
2270 spin_unlock_irq(&gcwq
->lock
);
2272 /* if DIE is set, destruction is requested */
2273 if (worker
->flags
& WORKER_DIE
) {
2274 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2278 /* otherwise, rebind */
2279 idle_worker_rebind(worker
);
2283 worker_leave_idle(worker
);
2285 /* no more worker necessary? */
2286 if (!need_more_worker(pool
))
2289 /* do we need to manage? */
2290 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2294 * ->scheduled list can only be filled while a worker is
2295 * preparing to process a work or actually processing it.
2296 * Make sure nobody diddled with it while I was sleeping.
2298 BUG_ON(!list_empty(&worker
->scheduled
));
2301 * When control reaches this point, we're guaranteed to have
2302 * at least one idle worker or that someone else has already
2303 * assumed the manager role.
2305 worker_clr_flags(worker
, WORKER_PREP
);
2308 struct work_struct
*work
=
2309 list_first_entry(&pool
->worklist
,
2310 struct work_struct
, entry
);
2312 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2313 /* optimization path, not strictly necessary */
2314 process_one_work(worker
, work
);
2315 if (unlikely(!list_empty(&worker
->scheduled
)))
2316 process_scheduled_works(worker
);
2318 move_linked_works(work
, &worker
->scheduled
, NULL
);
2319 process_scheduled_works(worker
);
2321 } while (keep_working(pool
));
2323 worker_set_flags(worker
, WORKER_PREP
, false);
2325 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2329 * gcwq->lock is held and there's no work to process and no
2330 * need to manage, sleep. Workers are woken up only while
2331 * holding gcwq->lock or from local cpu, so setting the
2332 * current state before releasing gcwq->lock is enough to
2333 * prevent losing any event.
2335 worker_enter_idle(worker
);
2336 __set_current_state(TASK_INTERRUPTIBLE
);
2337 spin_unlock_irq(&gcwq
->lock
);
2343 * rescuer_thread - the rescuer thread function
2346 * Workqueue rescuer thread function. There's one rescuer for each
2347 * workqueue which has WQ_RESCUER set.
2349 * Regular work processing on a gcwq may block trying to create a new
2350 * worker which uses GFP_KERNEL allocation which has slight chance of
2351 * developing into deadlock if some works currently on the same queue
2352 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2353 * the problem rescuer solves.
2355 * When such condition is possible, the gcwq summons rescuers of all
2356 * workqueues which have works queued on the gcwq and let them process
2357 * those works so that forward progress can be guaranteed.
2359 * This should happen rarely.
2361 static int rescuer_thread(void *__rescuer
)
2363 struct worker
*rescuer
= __rescuer
;
2364 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2365 struct list_head
*scheduled
= &rescuer
->scheduled
;
2366 bool is_unbound
= wq
->flags
& WQ_UNBOUND
;
2369 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2372 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2373 * doesn't participate in concurrency management.
2375 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2377 set_current_state(TASK_INTERRUPTIBLE
);
2379 if (kthread_should_stop()) {
2380 __set_current_state(TASK_RUNNING
);
2381 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2386 * See whether any cpu is asking for help. Unbounded
2387 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2389 for_each_mayday_cpu(cpu
, wq
->mayday_mask
) {
2390 unsigned int tcpu
= is_unbound
? WORK_CPU_UNBOUND
: cpu
;
2391 struct cpu_workqueue_struct
*cwq
= get_cwq(tcpu
, wq
);
2392 struct worker_pool
*pool
= cwq
->pool
;
2393 struct global_cwq
*gcwq
= pool
->gcwq
;
2394 struct work_struct
*work
, *n
;
2396 __set_current_state(TASK_RUNNING
);
2397 mayday_clear_cpu(cpu
, wq
->mayday_mask
);
2399 /* migrate to the target cpu if possible */
2400 rescuer
->pool
= pool
;
2401 worker_maybe_bind_and_lock(rescuer
);
2404 * Slurp in all works issued via this workqueue and
2407 BUG_ON(!list_empty(&rescuer
->scheduled
));
2408 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2409 if (get_work_cwq(work
) == cwq
)
2410 move_linked_works(work
, scheduled
, &n
);
2412 process_scheduled_works(rescuer
);
2415 * Leave this gcwq. If keep_working() is %true, notify a
2416 * regular worker; otherwise, we end up with 0 concurrency
2417 * and stalling the execution.
2419 if (keep_working(pool
))
2420 wake_up_worker(pool
);
2422 spin_unlock_irq(&gcwq
->lock
);
2425 /* rescuers should never participate in concurrency management */
2426 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2432 struct work_struct work
;
2433 struct completion done
;
2436 static void wq_barrier_func(struct work_struct
*work
)
2438 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2439 complete(&barr
->done
);
2443 * insert_wq_barrier - insert a barrier work
2444 * @cwq: cwq to insert barrier into
2445 * @barr: wq_barrier to insert
2446 * @target: target work to attach @barr to
2447 * @worker: worker currently executing @target, NULL if @target is not executing
2449 * @barr is linked to @target such that @barr is completed only after
2450 * @target finishes execution. Please note that the ordering
2451 * guarantee is observed only with respect to @target and on the local
2454 * Currently, a queued barrier can't be canceled. This is because
2455 * try_to_grab_pending() can't determine whether the work to be
2456 * grabbed is at the head of the queue and thus can't clear LINKED
2457 * flag of the previous work while there must be a valid next work
2458 * after a work with LINKED flag set.
2460 * Note that when @worker is non-NULL, @target may be modified
2461 * underneath us, so we can't reliably determine cwq from @target.
2464 * spin_lock_irq(gcwq->lock).
2466 static void insert_wq_barrier(struct cpu_workqueue_struct
*cwq
,
2467 struct wq_barrier
*barr
,
2468 struct work_struct
*target
, struct worker
*worker
)
2470 struct list_head
*head
;
2471 unsigned int linked
= 0;
2474 * debugobject calls are safe here even with gcwq->lock locked
2475 * as we know for sure that this will not trigger any of the
2476 * checks and call back into the fixup functions where we
2479 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2480 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2481 init_completion(&barr
->done
);
2484 * If @target is currently being executed, schedule the
2485 * barrier to the worker; otherwise, put it after @target.
2488 head
= worker
->scheduled
.next
;
2490 unsigned long *bits
= work_data_bits(target
);
2492 head
= target
->entry
.next
;
2493 /* there can already be other linked works, inherit and set */
2494 linked
= *bits
& WORK_STRUCT_LINKED
;
2495 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2498 debug_work_activate(&barr
->work
);
2499 insert_work(cwq
, &barr
->work
, head
,
2500 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2504 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2505 * @wq: workqueue being flushed
2506 * @flush_color: new flush color, < 0 for no-op
2507 * @work_color: new work color, < 0 for no-op
2509 * Prepare cwqs for workqueue flushing.
2511 * If @flush_color is non-negative, flush_color on all cwqs should be
2512 * -1. If no cwq has in-flight commands at the specified color, all
2513 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2514 * has in flight commands, its cwq->flush_color is set to
2515 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2516 * wakeup logic is armed and %true is returned.
2518 * The caller should have initialized @wq->first_flusher prior to
2519 * calling this function with non-negative @flush_color. If
2520 * @flush_color is negative, no flush color update is done and %false
2523 * If @work_color is non-negative, all cwqs should have the same
2524 * work_color which is previous to @work_color and all will be
2525 * advanced to @work_color.
2528 * mutex_lock(wq->flush_mutex).
2531 * %true if @flush_color >= 0 and there's something to flush. %false
2534 static bool flush_workqueue_prep_cwqs(struct workqueue_struct
*wq
,
2535 int flush_color
, int work_color
)
2540 if (flush_color
>= 0) {
2541 BUG_ON(atomic_read(&wq
->nr_cwqs_to_flush
));
2542 atomic_set(&wq
->nr_cwqs_to_flush
, 1);
2545 for_each_cwq_cpu(cpu
, wq
) {
2546 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2547 struct global_cwq
*gcwq
= cwq
->pool
->gcwq
;
2549 spin_lock_irq(&gcwq
->lock
);
2551 if (flush_color
>= 0) {
2552 BUG_ON(cwq
->flush_color
!= -1);
2554 if (cwq
->nr_in_flight
[flush_color
]) {
2555 cwq
->flush_color
= flush_color
;
2556 atomic_inc(&wq
->nr_cwqs_to_flush
);
2561 if (work_color
>= 0) {
2562 BUG_ON(work_color
!= work_next_color(cwq
->work_color
));
2563 cwq
->work_color
= work_color
;
2566 spin_unlock_irq(&gcwq
->lock
);
2569 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_cwqs_to_flush
))
2570 complete(&wq
->first_flusher
->done
);
2576 * flush_workqueue - ensure that any scheduled work has run to completion.
2577 * @wq: workqueue to flush
2579 * Forces execution of the workqueue and blocks until its completion.
2580 * This is typically used in driver shutdown handlers.
2582 * We sleep until all works which were queued on entry have been handled,
2583 * but we are not livelocked by new incoming ones.
2585 void flush_workqueue(struct workqueue_struct
*wq
)
2587 struct wq_flusher this_flusher
= {
2588 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2590 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2594 lock_map_acquire(&wq
->lockdep_map
);
2595 lock_map_release(&wq
->lockdep_map
);
2597 mutex_lock(&wq
->flush_mutex
);
2600 * Start-to-wait phase
2602 next_color
= work_next_color(wq
->work_color
);
2604 if (next_color
!= wq
->flush_color
) {
2606 * Color space is not full. The current work_color
2607 * becomes our flush_color and work_color is advanced
2610 BUG_ON(!list_empty(&wq
->flusher_overflow
));
2611 this_flusher
.flush_color
= wq
->work_color
;
2612 wq
->work_color
= next_color
;
2614 if (!wq
->first_flusher
) {
2615 /* no flush in progress, become the first flusher */
2616 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2618 wq
->first_flusher
= &this_flusher
;
2620 if (!flush_workqueue_prep_cwqs(wq
, wq
->flush_color
,
2622 /* nothing to flush, done */
2623 wq
->flush_color
= next_color
;
2624 wq
->first_flusher
= NULL
;
2629 BUG_ON(wq
->flush_color
== this_flusher
.flush_color
);
2630 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2631 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2635 * Oops, color space is full, wait on overflow queue.
2636 * The next flush completion will assign us
2637 * flush_color and transfer to flusher_queue.
2639 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2642 mutex_unlock(&wq
->flush_mutex
);
2644 wait_for_completion(&this_flusher
.done
);
2647 * Wake-up-and-cascade phase
2649 * First flushers are responsible for cascading flushes and
2650 * handling overflow. Non-first flushers can simply return.
2652 if (wq
->first_flusher
!= &this_flusher
)
2655 mutex_lock(&wq
->flush_mutex
);
2657 /* we might have raced, check again with mutex held */
2658 if (wq
->first_flusher
!= &this_flusher
)
2661 wq
->first_flusher
= NULL
;
2663 BUG_ON(!list_empty(&this_flusher
.list
));
2664 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2667 struct wq_flusher
*next
, *tmp
;
2669 /* complete all the flushers sharing the current flush color */
2670 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2671 if (next
->flush_color
!= wq
->flush_color
)
2673 list_del_init(&next
->list
);
2674 complete(&next
->done
);
2677 BUG_ON(!list_empty(&wq
->flusher_overflow
) &&
2678 wq
->flush_color
!= work_next_color(wq
->work_color
));
2680 /* this flush_color is finished, advance by one */
2681 wq
->flush_color
= work_next_color(wq
->flush_color
);
2683 /* one color has been freed, handle overflow queue */
2684 if (!list_empty(&wq
->flusher_overflow
)) {
2686 * Assign the same color to all overflowed
2687 * flushers, advance work_color and append to
2688 * flusher_queue. This is the start-to-wait
2689 * phase for these overflowed flushers.
2691 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2692 tmp
->flush_color
= wq
->work_color
;
2694 wq
->work_color
= work_next_color(wq
->work_color
);
2696 list_splice_tail_init(&wq
->flusher_overflow
,
2697 &wq
->flusher_queue
);
2698 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2701 if (list_empty(&wq
->flusher_queue
)) {
2702 BUG_ON(wq
->flush_color
!= wq
->work_color
);
2707 * Need to flush more colors. Make the next flusher
2708 * the new first flusher and arm cwqs.
2710 BUG_ON(wq
->flush_color
== wq
->work_color
);
2711 BUG_ON(wq
->flush_color
!= next
->flush_color
);
2713 list_del_init(&next
->list
);
2714 wq
->first_flusher
= next
;
2716 if (flush_workqueue_prep_cwqs(wq
, wq
->flush_color
, -1))
2720 * Meh... this color is already done, clear first
2721 * flusher and repeat cascading.
2723 wq
->first_flusher
= NULL
;
2727 mutex_unlock(&wq
->flush_mutex
);
2729 EXPORT_SYMBOL_GPL(flush_workqueue
);
2732 * drain_workqueue - drain a workqueue
2733 * @wq: workqueue to drain
2735 * Wait until the workqueue becomes empty. While draining is in progress,
2736 * only chain queueing is allowed. IOW, only currently pending or running
2737 * work items on @wq can queue further work items on it. @wq is flushed
2738 * repeatedly until it becomes empty. The number of flushing is detemined
2739 * by the depth of chaining and should be relatively short. Whine if it
2742 void drain_workqueue(struct workqueue_struct
*wq
)
2744 unsigned int flush_cnt
= 0;
2748 * __queue_work() needs to test whether there are drainers, is much
2749 * hotter than drain_workqueue() and already looks at @wq->flags.
2750 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2752 spin_lock(&workqueue_lock
);
2753 if (!wq
->nr_drainers
++)
2754 wq
->flags
|= WQ_DRAINING
;
2755 spin_unlock(&workqueue_lock
);
2757 flush_workqueue(wq
);
2759 for_each_cwq_cpu(cpu
, wq
) {
2760 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2763 spin_lock_irq(&cwq
->pool
->gcwq
->lock
);
2764 drained
= !cwq
->nr_active
&& list_empty(&cwq
->delayed_works
);
2765 spin_unlock_irq(&cwq
->pool
->gcwq
->lock
);
2770 if (++flush_cnt
== 10 ||
2771 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2772 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2773 wq
->name
, flush_cnt
);
2777 spin_lock(&workqueue_lock
);
2778 if (!--wq
->nr_drainers
)
2779 wq
->flags
&= ~WQ_DRAINING
;
2780 spin_unlock(&workqueue_lock
);
2782 EXPORT_SYMBOL_GPL(drain_workqueue
);
2784 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2786 struct worker
*worker
= NULL
;
2787 struct global_cwq
*gcwq
;
2788 struct cpu_workqueue_struct
*cwq
;
2791 gcwq
= get_work_gcwq(work
);
2795 spin_lock_irq(&gcwq
->lock
);
2796 if (!list_empty(&work
->entry
)) {
2798 * See the comment near try_to_grab_pending()->smp_rmb().
2799 * If it was re-queued to a different gcwq under us, we
2800 * are not going to wait.
2803 cwq
= get_work_cwq(work
);
2804 if (unlikely(!cwq
|| gcwq
!= cwq
->pool
->gcwq
))
2807 worker
= find_worker_executing_work(gcwq
, work
);
2810 cwq
= worker
->current_cwq
;
2813 insert_wq_barrier(cwq
, barr
, work
, worker
);
2814 spin_unlock_irq(&gcwq
->lock
);
2817 * If @max_active is 1 or rescuer is in use, flushing another work
2818 * item on the same workqueue may lead to deadlock. Make sure the
2819 * flusher is not running on the same workqueue by verifying write
2822 if (cwq
->wq
->saved_max_active
== 1 || cwq
->wq
->flags
& WQ_RESCUER
)
2823 lock_map_acquire(&cwq
->wq
->lockdep_map
);
2825 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2826 lock_map_release(&cwq
->wq
->lockdep_map
);
2830 spin_unlock_irq(&gcwq
->lock
);
2835 * flush_work - wait for a work to finish executing the last queueing instance
2836 * @work: the work to flush
2838 * Wait until @work has finished execution. @work is guaranteed to be idle
2839 * on return if it hasn't been requeued since flush started.
2842 * %true if flush_work() waited for the work to finish execution,
2843 * %false if it was already idle.
2845 bool flush_work(struct work_struct
*work
)
2847 struct wq_barrier barr
;
2849 lock_map_acquire(&work
->lockdep_map
);
2850 lock_map_release(&work
->lockdep_map
);
2852 if (start_flush_work(work
, &barr
)) {
2853 wait_for_completion(&barr
.done
);
2854 destroy_work_on_stack(&barr
.work
);
2860 EXPORT_SYMBOL_GPL(flush_work
);
2862 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2864 unsigned long flags
;
2868 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2870 * If someone else is canceling, wait for the same event it
2871 * would be waiting for before retrying.
2873 if (unlikely(ret
== -ENOENT
))
2875 } while (unlikely(ret
< 0));
2877 /* tell other tasks trying to grab @work to back off */
2878 mark_work_canceling(work
);
2879 local_irq_restore(flags
);
2882 clear_work_data(work
);
2887 * cancel_work_sync - cancel a work and wait for it to finish
2888 * @work: the work to cancel
2890 * Cancel @work and wait for its execution to finish. This function
2891 * can be used even if the work re-queues itself or migrates to
2892 * another workqueue. On return from this function, @work is
2893 * guaranteed to be not pending or executing on any CPU.
2895 * cancel_work_sync(&delayed_work->work) must not be used for
2896 * delayed_work's. Use cancel_delayed_work_sync() instead.
2898 * The caller must ensure that the workqueue on which @work was last
2899 * queued can't be destroyed before this function returns.
2902 * %true if @work was pending, %false otherwise.
2904 bool cancel_work_sync(struct work_struct
*work
)
2906 return __cancel_work_timer(work
, false);
2908 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2911 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2912 * @dwork: the delayed work to flush
2914 * Delayed timer is cancelled and the pending work is queued for
2915 * immediate execution. Like flush_work(), this function only
2916 * considers the last queueing instance of @dwork.
2919 * %true if flush_work() waited for the work to finish execution,
2920 * %false if it was already idle.
2922 bool flush_delayed_work(struct delayed_work
*dwork
)
2924 local_irq_disable();
2925 if (del_timer_sync(&dwork
->timer
))
2926 __queue_work(dwork
->cpu
,
2927 get_work_cwq(&dwork
->work
)->wq
, &dwork
->work
);
2929 return flush_work(&dwork
->work
);
2931 EXPORT_SYMBOL(flush_delayed_work
);
2934 * cancel_delayed_work - cancel a delayed work
2935 * @dwork: delayed_work to cancel
2937 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2938 * and canceled; %false if wasn't pending. Note that the work callback
2939 * function may still be running on return, unless it returns %true and the
2940 * work doesn't re-arm itself. Explicitly flush or use
2941 * cancel_delayed_work_sync() to wait on it.
2943 * This function is safe to call from any context including IRQ handler.
2945 bool cancel_delayed_work(struct delayed_work
*dwork
)
2947 unsigned long flags
;
2951 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2952 } while (unlikely(ret
== -EAGAIN
));
2954 if (unlikely(ret
< 0))
2957 set_work_cpu_and_clear_pending(&dwork
->work
, work_cpu(&dwork
->work
));
2958 local_irq_restore(flags
);
2961 EXPORT_SYMBOL(cancel_delayed_work
);
2964 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2965 * @dwork: the delayed work cancel
2967 * This is cancel_work_sync() for delayed works.
2970 * %true if @dwork was pending, %false otherwise.
2972 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2974 return __cancel_work_timer(&dwork
->work
, true);
2976 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2979 * schedule_work_on - put work task on a specific cpu
2980 * @cpu: cpu to put the work task on
2981 * @work: job to be done
2983 * This puts a job on a specific cpu
2985 bool schedule_work_on(int cpu
, struct work_struct
*work
)
2987 return queue_work_on(cpu
, system_wq
, work
);
2989 EXPORT_SYMBOL(schedule_work_on
);
2992 * schedule_work - put work task in global workqueue
2993 * @work: job to be done
2995 * Returns %false if @work was already on the kernel-global workqueue and
2998 * This puts a job in the kernel-global workqueue if it was not already
2999 * queued and leaves it in the same position on the kernel-global
3000 * workqueue otherwise.
3002 bool schedule_work(struct work_struct
*work
)
3004 return queue_work(system_wq
, work
);
3006 EXPORT_SYMBOL(schedule_work
);
3009 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3011 * @dwork: job to be done
3012 * @delay: number of jiffies to wait
3014 * After waiting for a given time this puts a job in the kernel-global
3015 * workqueue on the specified CPU.
3017 bool schedule_delayed_work_on(int cpu
, struct delayed_work
*dwork
,
3018 unsigned long delay
)
3020 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
3022 EXPORT_SYMBOL(schedule_delayed_work_on
);
3025 * schedule_delayed_work - put work task in global workqueue after delay
3026 * @dwork: job to be done
3027 * @delay: number of jiffies to wait or 0 for immediate execution
3029 * After waiting for a given time this puts a job in the kernel-global
3032 bool schedule_delayed_work(struct delayed_work
*dwork
, unsigned long delay
)
3034 return queue_delayed_work(system_wq
, dwork
, delay
);
3036 EXPORT_SYMBOL(schedule_delayed_work
);
3039 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3040 * @func: the function to call
3042 * schedule_on_each_cpu() executes @func on each online CPU using the
3043 * system workqueue and blocks until all CPUs have completed.
3044 * schedule_on_each_cpu() is very slow.
3047 * 0 on success, -errno on failure.
3049 int schedule_on_each_cpu(work_func_t func
)
3052 struct work_struct __percpu
*works
;
3054 works
= alloc_percpu(struct work_struct
);
3060 for_each_online_cpu(cpu
) {
3061 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3063 INIT_WORK(work
, func
);
3064 schedule_work_on(cpu
, work
);
3067 for_each_online_cpu(cpu
)
3068 flush_work(per_cpu_ptr(works
, cpu
));
3076 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3078 * Forces execution of the kernel-global workqueue and blocks until its
3081 * Think twice before calling this function! It's very easy to get into
3082 * trouble if you don't take great care. Either of the following situations
3083 * will lead to deadlock:
3085 * One of the work items currently on the workqueue needs to acquire
3086 * a lock held by your code or its caller.
3088 * Your code is running in the context of a work routine.
3090 * They will be detected by lockdep when they occur, but the first might not
3091 * occur very often. It depends on what work items are on the workqueue and
3092 * what locks they need, which you have no control over.
3094 * In most situations flushing the entire workqueue is overkill; you merely
3095 * need to know that a particular work item isn't queued and isn't running.
3096 * In such cases you should use cancel_delayed_work_sync() or
3097 * cancel_work_sync() instead.
3099 void flush_scheduled_work(void)
3101 flush_workqueue(system_wq
);
3103 EXPORT_SYMBOL(flush_scheduled_work
);
3106 * execute_in_process_context - reliably execute the routine with user context
3107 * @fn: the function to execute
3108 * @ew: guaranteed storage for the execute work structure (must
3109 * be available when the work executes)
3111 * Executes the function immediately if process context is available,
3112 * otherwise schedules the function for delayed execution.
3114 * Returns: 0 - function was executed
3115 * 1 - function was scheduled for execution
3117 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3119 if (!in_interrupt()) {
3124 INIT_WORK(&ew
->work
, fn
);
3125 schedule_work(&ew
->work
);
3129 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3131 int keventd_up(void)
3133 return system_wq
!= NULL
;
3136 static int alloc_cwqs(struct workqueue_struct
*wq
)
3139 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3140 * Make sure that the alignment isn't lower than that of
3141 * unsigned long long.
3143 const size_t size
= sizeof(struct cpu_workqueue_struct
);
3144 const size_t align
= max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS
,
3145 __alignof__(unsigned long long));
3147 if (!(wq
->flags
& WQ_UNBOUND
))
3148 wq
->cpu_wq
.pcpu
= __alloc_percpu(size
, align
);
3153 * Allocate enough room to align cwq and put an extra
3154 * pointer at the end pointing back to the originally
3155 * allocated pointer which will be used for free.
3157 ptr
= kzalloc(size
+ align
+ sizeof(void *), GFP_KERNEL
);
3159 wq
->cpu_wq
.single
= PTR_ALIGN(ptr
, align
);
3160 *(void **)(wq
->cpu_wq
.single
+ 1) = ptr
;
3164 /* just in case, make sure it's actually aligned */
3165 BUG_ON(!IS_ALIGNED(wq
->cpu_wq
.v
, align
));
3166 return wq
->cpu_wq
.v
? 0 : -ENOMEM
;
3169 static void free_cwqs(struct workqueue_struct
*wq
)
3171 if (!(wq
->flags
& WQ_UNBOUND
))
3172 free_percpu(wq
->cpu_wq
.pcpu
);
3173 else if (wq
->cpu_wq
.single
) {
3174 /* the pointer to free is stored right after the cwq */
3175 kfree(*(void **)(wq
->cpu_wq
.single
+ 1));
3179 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3182 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3184 if (max_active
< 1 || max_active
> lim
)
3185 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3186 max_active
, name
, 1, lim
);
3188 return clamp_val(max_active
, 1, lim
);
3191 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3194 struct lock_class_key
*key
,
3195 const char *lock_name
, ...)
3197 va_list args
, args1
;
3198 struct workqueue_struct
*wq
;
3202 /* determine namelen, allocate wq and format name */
3203 va_start(args
, lock_name
);
3204 va_copy(args1
, args
);
3205 namelen
= vsnprintf(NULL
, 0, fmt
, args
) + 1;
3207 wq
= kzalloc(sizeof(*wq
) + namelen
, GFP_KERNEL
);
3211 vsnprintf(wq
->name
, namelen
, fmt
, args1
);
3216 * Workqueues which may be used during memory reclaim should
3217 * have a rescuer to guarantee forward progress.
3219 if (flags
& WQ_MEM_RECLAIM
)
3220 flags
|= WQ_RESCUER
;
3222 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3223 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3227 wq
->saved_max_active
= max_active
;
3228 mutex_init(&wq
->flush_mutex
);
3229 atomic_set(&wq
->nr_cwqs_to_flush
, 0);
3230 INIT_LIST_HEAD(&wq
->flusher_queue
);
3231 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3233 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3234 INIT_LIST_HEAD(&wq
->list
);
3236 if (alloc_cwqs(wq
) < 0)
3239 for_each_cwq_cpu(cpu
, wq
) {
3240 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3241 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3242 int pool_idx
= (bool)(flags
& WQ_HIGHPRI
);
3244 BUG_ON((unsigned long)cwq
& WORK_STRUCT_FLAG_MASK
);
3245 cwq
->pool
= &gcwq
->pools
[pool_idx
];
3247 cwq
->flush_color
= -1;
3248 cwq
->max_active
= max_active
;
3249 INIT_LIST_HEAD(&cwq
->delayed_works
);
3252 if (flags
& WQ_RESCUER
) {
3253 struct worker
*rescuer
;
3255 if (!alloc_mayday_mask(&wq
->mayday_mask
, GFP_KERNEL
))
3258 wq
->rescuer
= rescuer
= alloc_worker();
3262 rescuer
->rescue_wq
= wq
;
3263 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
3265 if (IS_ERR(rescuer
->task
))
3268 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
3269 wake_up_process(rescuer
->task
);
3273 * workqueue_lock protects global freeze state and workqueues
3274 * list. Grab it, set max_active accordingly and add the new
3275 * workqueue to workqueues list.
3277 spin_lock(&workqueue_lock
);
3279 if (workqueue_freezing
&& wq
->flags
& WQ_FREEZABLE
)
3280 for_each_cwq_cpu(cpu
, wq
)
3281 get_cwq(cpu
, wq
)->max_active
= 0;
3283 list_add(&wq
->list
, &workqueues
);
3285 spin_unlock(&workqueue_lock
);
3291 free_mayday_mask(wq
->mayday_mask
);
3297 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3300 * destroy_workqueue - safely terminate a workqueue
3301 * @wq: target workqueue
3303 * Safely destroy a workqueue. All work currently pending will be done first.
3305 void destroy_workqueue(struct workqueue_struct
*wq
)
3309 /* drain it before proceeding with destruction */
3310 drain_workqueue(wq
);
3313 * wq list is used to freeze wq, remove from list after
3314 * flushing is complete in case freeze races us.
3316 spin_lock(&workqueue_lock
);
3317 list_del(&wq
->list
);
3318 spin_unlock(&workqueue_lock
);
3321 for_each_cwq_cpu(cpu
, wq
) {
3322 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3325 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
3326 BUG_ON(cwq
->nr_in_flight
[i
]);
3327 BUG_ON(cwq
->nr_active
);
3328 BUG_ON(!list_empty(&cwq
->delayed_works
));
3331 if (wq
->flags
& WQ_RESCUER
) {
3332 kthread_stop(wq
->rescuer
->task
);
3333 free_mayday_mask(wq
->mayday_mask
);
3340 EXPORT_SYMBOL_GPL(destroy_workqueue
);
3343 * cwq_set_max_active - adjust max_active of a cwq
3344 * @cwq: target cpu_workqueue_struct
3345 * @max_active: new max_active value.
3347 * Set @cwq->max_active to @max_active and activate delayed works if
3351 * spin_lock_irq(gcwq->lock).
3353 static void cwq_set_max_active(struct cpu_workqueue_struct
*cwq
, int max_active
)
3355 cwq
->max_active
= max_active
;
3357 while (!list_empty(&cwq
->delayed_works
) &&
3358 cwq
->nr_active
< cwq
->max_active
)
3359 cwq_activate_first_delayed(cwq
);
3363 * workqueue_set_max_active - adjust max_active of a workqueue
3364 * @wq: target workqueue
3365 * @max_active: new max_active value.
3367 * Set max_active of @wq to @max_active.
3370 * Don't call from IRQ context.
3372 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
3376 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
3378 spin_lock(&workqueue_lock
);
3380 wq
->saved_max_active
= max_active
;
3382 for_each_cwq_cpu(cpu
, wq
) {
3383 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3385 spin_lock_irq(&gcwq
->lock
);
3387 if (!(wq
->flags
& WQ_FREEZABLE
) ||
3388 !(gcwq
->flags
& GCWQ_FREEZING
))
3389 cwq_set_max_active(get_cwq(gcwq
->cpu
, wq
), max_active
);
3391 spin_unlock_irq(&gcwq
->lock
);
3394 spin_unlock(&workqueue_lock
);
3396 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
3399 * workqueue_congested - test whether a workqueue is congested
3400 * @cpu: CPU in question
3401 * @wq: target workqueue
3403 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3404 * no synchronization around this function and the test result is
3405 * unreliable and only useful as advisory hints or for debugging.
3408 * %true if congested, %false otherwise.
3410 bool workqueue_congested(unsigned int cpu
, struct workqueue_struct
*wq
)
3412 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3414 return !list_empty(&cwq
->delayed_works
);
3416 EXPORT_SYMBOL_GPL(workqueue_congested
);
3419 * work_cpu - return the last known associated cpu for @work
3420 * @work: the work of interest
3423 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3425 static unsigned int work_cpu(struct work_struct
*work
)
3427 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3429 return gcwq
? gcwq
->cpu
: WORK_CPU_NONE
;
3433 * work_busy - test whether a work is currently pending or running
3434 * @work: the work to be tested
3436 * Test whether @work is currently pending or running. There is no
3437 * synchronization around this function and the test result is
3438 * unreliable and only useful as advisory hints or for debugging.
3439 * Especially for reentrant wqs, the pending state might hide the
3443 * OR'd bitmask of WORK_BUSY_* bits.
3445 unsigned int work_busy(struct work_struct
*work
)
3447 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3448 unsigned long flags
;
3449 unsigned int ret
= 0;
3454 spin_lock_irqsave(&gcwq
->lock
, flags
);
3456 if (work_pending(work
))
3457 ret
|= WORK_BUSY_PENDING
;
3458 if (find_worker_executing_work(gcwq
, work
))
3459 ret
|= WORK_BUSY_RUNNING
;
3461 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
3465 EXPORT_SYMBOL_GPL(work_busy
);
3470 * There are two challenges in supporting CPU hotplug. Firstly, there
3471 * are a lot of assumptions on strong associations among work, cwq and
3472 * gcwq which make migrating pending and scheduled works very
3473 * difficult to implement without impacting hot paths. Secondly,
3474 * gcwqs serve mix of short, long and very long running works making
3475 * blocked draining impractical.
3477 * This is solved by allowing the pools to be disassociated from the CPU
3478 * running as an unbound one and allowing it to be reattached later if the
3479 * cpu comes back online.
3482 /* claim manager positions of all pools */
3483 static void gcwq_claim_assoc_and_lock(struct global_cwq
*gcwq
)
3485 struct worker_pool
*pool
;
3487 for_each_worker_pool(pool
, gcwq
)
3488 mutex_lock_nested(&pool
->assoc_mutex
, pool
- gcwq
->pools
);
3489 spin_lock_irq(&gcwq
->lock
);
3492 /* release manager positions */
3493 static void gcwq_release_assoc_and_unlock(struct global_cwq
*gcwq
)
3495 struct worker_pool
*pool
;
3497 spin_unlock_irq(&gcwq
->lock
);
3498 for_each_worker_pool(pool
, gcwq
)
3499 mutex_unlock(&pool
->assoc_mutex
);
3502 static void gcwq_unbind_fn(struct work_struct
*work
)
3504 struct global_cwq
*gcwq
= get_gcwq(smp_processor_id());
3505 struct worker_pool
*pool
;
3506 struct worker
*worker
;
3507 struct hlist_node
*pos
;
3510 BUG_ON(gcwq
->cpu
!= smp_processor_id());
3512 gcwq_claim_assoc_and_lock(gcwq
);
3515 * We've claimed all manager positions. Make all workers unbound
3516 * and set DISASSOCIATED. Before this, all workers except for the
3517 * ones which are still executing works from before the last CPU
3518 * down must be on the cpu. After this, they may become diasporas.
3520 for_each_worker_pool(pool
, gcwq
)
3521 list_for_each_entry(worker
, &pool
->idle_list
, entry
)
3522 worker
->flags
|= WORKER_UNBOUND
;
3524 for_each_busy_worker(worker
, i
, pos
, gcwq
)
3525 worker
->flags
|= WORKER_UNBOUND
;
3527 for_each_worker_pool(pool
, gcwq
)
3528 pool
->flags
|= POOL_DISASSOCIATED
;
3530 gcwq_release_assoc_and_unlock(gcwq
);
3533 * Call schedule() so that we cross rq->lock and thus can guarantee
3534 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3535 * as scheduler callbacks may be invoked from other cpus.
3540 * Sched callbacks are disabled now. Zap nr_running. After this,
3541 * nr_running stays zero and need_more_worker() and keep_working()
3542 * are always true as long as the worklist is not empty. @gcwq now
3543 * behaves as unbound (in terms of concurrency management) gcwq
3544 * which is served by workers tied to the CPU.
3546 * On return from this function, the current worker would trigger
3547 * unbound chain execution of pending work items if other workers
3550 for_each_worker_pool(pool
, gcwq
)
3551 atomic_set(get_pool_nr_running(pool
), 0);
3555 * Workqueues should be brought up before normal priority CPU notifiers.
3556 * This will be registered high priority CPU notifier.
3558 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
3559 unsigned long action
,
3562 unsigned int cpu
= (unsigned long)hcpu
;
3563 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3564 struct worker_pool
*pool
;
3566 switch (action
& ~CPU_TASKS_FROZEN
) {
3567 case CPU_UP_PREPARE
:
3568 for_each_worker_pool(pool
, gcwq
) {
3569 struct worker
*worker
;
3571 if (pool
->nr_workers
)
3574 worker
= create_worker(pool
);
3578 spin_lock_irq(&gcwq
->lock
);
3579 start_worker(worker
);
3580 spin_unlock_irq(&gcwq
->lock
);
3584 case CPU_DOWN_FAILED
:
3586 gcwq_claim_assoc_and_lock(gcwq
);
3587 for_each_worker_pool(pool
, gcwq
)
3588 pool
->flags
&= ~POOL_DISASSOCIATED
;
3589 rebind_workers(gcwq
);
3590 gcwq_release_assoc_and_unlock(gcwq
);
3597 * Workqueues should be brought down after normal priority CPU notifiers.
3598 * This will be registered as low priority CPU notifier.
3600 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
3601 unsigned long action
,
3604 unsigned int cpu
= (unsigned long)hcpu
;
3605 struct work_struct unbind_work
;
3607 switch (action
& ~CPU_TASKS_FROZEN
) {
3608 case CPU_DOWN_PREPARE
:
3609 /* unbinding should happen on the local CPU */
3610 INIT_WORK_ONSTACK(&unbind_work
, gcwq_unbind_fn
);
3611 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
3612 flush_work(&unbind_work
);
3620 struct work_for_cpu
{
3621 struct work_struct work
;
3627 static void work_for_cpu_fn(struct work_struct
*work
)
3629 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
3631 wfc
->ret
= wfc
->fn(wfc
->arg
);
3635 * work_on_cpu - run a function in user context on a particular cpu
3636 * @cpu: the cpu to run on
3637 * @fn: the function to run
3638 * @arg: the function arg
3640 * This will return the value @fn returns.
3641 * It is up to the caller to ensure that the cpu doesn't go offline.
3642 * The caller must not hold any locks which would prevent @fn from completing.
3644 long work_on_cpu(unsigned int cpu
, long (*fn
)(void *), void *arg
)
3646 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
3648 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
3649 schedule_work_on(cpu
, &wfc
.work
);
3650 flush_work(&wfc
.work
);
3653 EXPORT_SYMBOL_GPL(work_on_cpu
);
3654 #endif /* CONFIG_SMP */
3656 #ifdef CONFIG_FREEZER
3659 * freeze_workqueues_begin - begin freezing workqueues
3661 * Start freezing workqueues. After this function returns, all freezable
3662 * workqueues will queue new works to their frozen_works list instead of
3666 * Grabs and releases workqueue_lock and gcwq->lock's.
3668 void freeze_workqueues_begin(void)
3672 spin_lock(&workqueue_lock
);
3674 BUG_ON(workqueue_freezing
);
3675 workqueue_freezing
= true;
3677 for_each_gcwq_cpu(cpu
) {
3678 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3679 struct workqueue_struct
*wq
;
3681 spin_lock_irq(&gcwq
->lock
);
3683 BUG_ON(gcwq
->flags
& GCWQ_FREEZING
);
3684 gcwq
->flags
|= GCWQ_FREEZING
;
3686 list_for_each_entry(wq
, &workqueues
, list
) {
3687 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3689 if (cwq
&& wq
->flags
& WQ_FREEZABLE
)
3690 cwq
->max_active
= 0;
3693 spin_unlock_irq(&gcwq
->lock
);
3696 spin_unlock(&workqueue_lock
);
3700 * freeze_workqueues_busy - are freezable workqueues still busy?
3702 * Check whether freezing is complete. This function must be called
3703 * between freeze_workqueues_begin() and thaw_workqueues().
3706 * Grabs and releases workqueue_lock.
3709 * %true if some freezable workqueues are still busy. %false if freezing
3712 bool freeze_workqueues_busy(void)
3717 spin_lock(&workqueue_lock
);
3719 BUG_ON(!workqueue_freezing
);
3721 for_each_gcwq_cpu(cpu
) {
3722 struct workqueue_struct
*wq
;
3724 * nr_active is monotonically decreasing. It's safe
3725 * to peek without lock.
3727 list_for_each_entry(wq
, &workqueues
, list
) {
3728 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3730 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3733 BUG_ON(cwq
->nr_active
< 0);
3734 if (cwq
->nr_active
) {
3741 spin_unlock(&workqueue_lock
);
3746 * thaw_workqueues - thaw workqueues
3748 * Thaw workqueues. Normal queueing is restored and all collected
3749 * frozen works are transferred to their respective gcwq worklists.
3752 * Grabs and releases workqueue_lock and gcwq->lock's.
3754 void thaw_workqueues(void)
3758 spin_lock(&workqueue_lock
);
3760 if (!workqueue_freezing
)
3763 for_each_gcwq_cpu(cpu
) {
3764 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3765 struct worker_pool
*pool
;
3766 struct workqueue_struct
*wq
;
3768 spin_lock_irq(&gcwq
->lock
);
3770 BUG_ON(!(gcwq
->flags
& GCWQ_FREEZING
));
3771 gcwq
->flags
&= ~GCWQ_FREEZING
;
3773 list_for_each_entry(wq
, &workqueues
, list
) {
3774 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3776 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3779 /* restore max_active and repopulate worklist */
3780 cwq_set_max_active(cwq
, wq
->saved_max_active
);
3783 for_each_worker_pool(pool
, gcwq
)
3784 wake_up_worker(pool
);
3786 spin_unlock_irq(&gcwq
->lock
);
3789 workqueue_freezing
= false;
3791 spin_unlock(&workqueue_lock
);
3793 #endif /* CONFIG_FREEZER */
3795 static int __init
init_workqueues(void)
3799 /* make sure we have enough bits for OFFQ CPU number */
3800 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_CPU_SHIFT
)) <
3803 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
3804 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
3806 /* initialize gcwqs */
3807 for_each_gcwq_cpu(cpu
) {
3808 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3809 struct worker_pool
*pool
;
3811 spin_lock_init(&gcwq
->lock
);
3814 hash_init(gcwq
->busy_hash
);
3816 for_each_worker_pool(pool
, gcwq
) {
3818 pool
->flags
|= POOL_DISASSOCIATED
;
3819 INIT_LIST_HEAD(&pool
->worklist
);
3820 INIT_LIST_HEAD(&pool
->idle_list
);
3822 init_timer_deferrable(&pool
->idle_timer
);
3823 pool
->idle_timer
.function
= idle_worker_timeout
;
3824 pool
->idle_timer
.data
= (unsigned long)pool
;
3826 setup_timer(&pool
->mayday_timer
, gcwq_mayday_timeout
,
3827 (unsigned long)pool
);
3829 mutex_init(&pool
->assoc_mutex
);
3830 ida_init(&pool
->worker_ida
);
3834 /* create the initial worker */
3835 for_each_online_gcwq_cpu(cpu
) {
3836 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3837 struct worker_pool
*pool
;
3839 for_each_worker_pool(pool
, gcwq
) {
3840 struct worker
*worker
;
3842 if (cpu
!= WORK_CPU_UNBOUND
)
3843 pool
->flags
&= ~POOL_DISASSOCIATED
;
3845 worker
= create_worker(pool
);
3847 spin_lock_irq(&gcwq
->lock
);
3848 start_worker(worker
);
3849 spin_unlock_irq(&gcwq
->lock
);
3853 system_wq
= alloc_workqueue("events", 0, 0);
3854 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
3855 system_long_wq
= alloc_workqueue("events_long", 0, 0);
3856 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
3857 WQ_UNBOUND_MAX_ACTIVE
);
3858 system_freezable_wq
= alloc_workqueue("events_freezable",
3860 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
3861 !system_unbound_wq
|| !system_freezable_wq
);
3864 early_initcall(init_workqueues
);