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 * A bound pool is either associated or disassociated with its CPU.
53 * While associated (!DISASSOCIATED), all workers are bound to the
54 * CPU and none has %WORKER_UNBOUND set and concurrency management
57 * While DISASSOCIATED, the cpu may be offline and all workers have
58 * %WORKER_UNBOUND set and concurrency management disabled, and may
59 * be executing on any CPU. The pool behaves as an unbound one.
61 * Note that DISASSOCIATED can be flipped only while holding
62 * assoc_mutex to avoid changing binding state while
63 * create_worker() is in progress.
65 POOL_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
66 POOL_MANAGING_WORKERS
= 1 << 1, /* managing workers */
67 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
68 POOL_FREEZING
= 1 << 3, /* freeze in progress */
71 WORKER_STARTED
= 1 << 0, /* started */
72 WORKER_DIE
= 1 << 1, /* die die die */
73 WORKER_IDLE
= 1 << 2, /* is idle */
74 WORKER_PREP
= 1 << 3, /* preparing to run works */
75 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
76 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
78 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_UNBOUND
|
81 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
83 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
85 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
86 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
88 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
89 /* call for help after 10ms
91 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
92 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
95 * Rescue workers are used only on emergencies and shared by
98 RESCUER_NICE_LEVEL
= -20,
99 HIGHPRI_NICE_LEVEL
= -20,
103 * Structure fields follow one of the following exclusion rules.
105 * I: Modifiable by initialization/destruction paths and read-only for
108 * P: Preemption protected. Disabling preemption is enough and should
109 * only be modified and accessed from the local cpu.
111 * L: gcwq->lock protected. Access with gcwq->lock held.
113 * X: During normal operation, modification requires gcwq->lock and
114 * should be done only from local cpu. Either disabling preemption
115 * on local cpu or grabbing gcwq->lock is enough for read access.
116 * If POOL_DISASSOCIATED is set, it's identical to L.
118 * F: wq->flush_mutex protected.
120 * W: workqueue_lock protected.
123 /* struct worker is defined in workqueue_internal.h */
126 struct global_cwq
*gcwq
; /* I: the owning gcwq */
127 unsigned int flags
; /* X: flags */
129 struct list_head worklist
; /* L: list of pending works */
130 int nr_workers
; /* L: total number of workers */
132 /* nr_idle includes the ones off idle_list for rebinding */
133 int nr_idle
; /* L: currently idle ones */
135 struct list_head idle_list
; /* X: list of idle workers */
136 struct timer_list idle_timer
; /* L: worker idle timeout */
137 struct timer_list mayday_timer
; /* L: SOS timer for workers */
139 struct mutex assoc_mutex
; /* protect POOL_DISASSOCIATED */
140 struct ida worker_ida
; /* L: for worker IDs */
144 * Global per-cpu workqueue. There's one and only one for each cpu
145 * and all works are queued and processed here regardless of their
149 spinlock_t lock
; /* the gcwq lock */
150 unsigned int cpu
; /* I: the associated cpu */
152 /* workers are chained either in busy_hash or pool idle_list */
153 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
154 /* L: hash of busy workers */
156 struct worker_pool pools
[NR_STD_WORKER_POOLS
];
157 /* normal and highpri pools */
158 } ____cacheline_aligned_in_smp
;
161 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
162 * work_struct->data are used for flags and thus cwqs need to be
163 * aligned at two's power of the number of flag bits.
165 struct cpu_workqueue_struct
{
166 struct worker_pool
*pool
; /* I: the associated pool */
167 struct workqueue_struct
*wq
; /* I: the owning workqueue */
168 int work_color
; /* L: current color */
169 int flush_color
; /* L: flushing color */
170 int nr_in_flight
[WORK_NR_COLORS
];
171 /* L: nr of in_flight works */
172 int nr_active
; /* L: nr of active works */
173 int max_active
; /* L: max active works */
174 struct list_head delayed_works
; /* L: delayed works */
178 * Structure used to wait for workqueue flush.
181 struct list_head list
; /* F: list of flushers */
182 int flush_color
; /* F: flush color waiting for */
183 struct completion done
; /* flush completion */
187 * All cpumasks are assumed to be always set on UP and thus can't be
188 * used to determine whether there's something to be done.
191 typedef cpumask_var_t mayday_mask_t
;
192 #define mayday_test_and_set_cpu(cpu, mask) \
193 cpumask_test_and_set_cpu((cpu), (mask))
194 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
195 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
196 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
197 #define free_mayday_mask(mask) free_cpumask_var((mask))
199 typedef unsigned long mayday_mask_t
;
200 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
201 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
202 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
203 #define alloc_mayday_mask(maskp, gfp) true
204 #define free_mayday_mask(mask) do { } while (0)
208 * The externally visible workqueue abstraction is an array of
209 * per-CPU workqueues:
211 struct workqueue_struct
{
212 unsigned int flags
; /* W: WQ_* flags */
214 struct cpu_workqueue_struct __percpu
*pcpu
;
215 struct cpu_workqueue_struct
*single
;
217 } cpu_wq
; /* I: cwq's */
218 struct list_head list
; /* W: list of all workqueues */
220 struct mutex flush_mutex
; /* protects wq flushing */
221 int work_color
; /* F: current work color */
222 int flush_color
; /* F: current flush color */
223 atomic_t nr_cwqs_to_flush
; /* flush in progress */
224 struct wq_flusher
*first_flusher
; /* F: first flusher */
225 struct list_head flusher_queue
; /* F: flush waiters */
226 struct list_head flusher_overflow
; /* F: flush overflow list */
228 mayday_mask_t mayday_mask
; /* cpus requesting rescue */
229 struct worker
*rescuer
; /* I: rescue worker */
231 int nr_drainers
; /* W: drain in progress */
232 int saved_max_active
; /* W: saved cwq max_active */
233 #ifdef CONFIG_LOCKDEP
234 struct lockdep_map lockdep_map
;
236 char name
[]; /* I: workqueue name */
239 struct workqueue_struct
*system_wq __read_mostly
;
240 EXPORT_SYMBOL_GPL(system_wq
);
241 struct workqueue_struct
*system_highpri_wq __read_mostly
;
242 EXPORT_SYMBOL_GPL(system_highpri_wq
);
243 struct workqueue_struct
*system_long_wq __read_mostly
;
244 EXPORT_SYMBOL_GPL(system_long_wq
);
245 struct workqueue_struct
*system_unbound_wq __read_mostly
;
246 EXPORT_SYMBOL_GPL(system_unbound_wq
);
247 struct workqueue_struct
*system_freezable_wq __read_mostly
;
248 EXPORT_SYMBOL_GPL(system_freezable_wq
);
250 #define CREATE_TRACE_POINTS
251 #include <trace/events/workqueue.h>
253 #define for_each_worker_pool(pool, gcwq) \
254 for ((pool) = &(gcwq)->pools[0]; \
255 (pool) < &(gcwq)->pools[NR_STD_WORKER_POOLS]; (pool)++)
257 #define for_each_busy_worker(worker, i, pos, gcwq) \
258 hash_for_each(gcwq->busy_hash, i, pos, worker, hentry)
260 static inline int __next_gcwq_cpu(int cpu
, const struct cpumask
*mask
,
263 if (cpu
< nr_cpu_ids
) {
265 cpu
= cpumask_next(cpu
, mask
);
266 if (cpu
< nr_cpu_ids
)
270 return WORK_CPU_UNBOUND
;
272 return WORK_CPU_NONE
;
275 static inline int __next_wq_cpu(int cpu
, const struct cpumask
*mask
,
276 struct workqueue_struct
*wq
)
278 return __next_gcwq_cpu(cpu
, mask
, !(wq
->flags
& WQ_UNBOUND
) ? 1 : 2);
284 * An extra gcwq is defined for an invalid cpu number
285 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
286 * specific CPU. The following iterators are similar to
287 * for_each_*_cpu() iterators but also considers the unbound gcwq.
289 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
290 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
291 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
292 * WORK_CPU_UNBOUND for unbound workqueues
294 #define for_each_gcwq_cpu(cpu) \
295 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
296 (cpu) < WORK_CPU_NONE; \
297 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
299 #define for_each_online_gcwq_cpu(cpu) \
300 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
301 (cpu) < WORK_CPU_NONE; \
302 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
304 #define for_each_cwq_cpu(cpu, wq) \
305 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
306 (cpu) < WORK_CPU_NONE; \
307 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
309 #ifdef CONFIG_DEBUG_OBJECTS_WORK
311 static struct debug_obj_descr work_debug_descr
;
313 static void *work_debug_hint(void *addr
)
315 return ((struct work_struct
*) addr
)->func
;
319 * fixup_init is called when:
320 * - an active object is initialized
322 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
324 struct work_struct
*work
= addr
;
327 case ODEBUG_STATE_ACTIVE
:
328 cancel_work_sync(work
);
329 debug_object_init(work
, &work_debug_descr
);
337 * fixup_activate is called when:
338 * - an active object is activated
339 * - an unknown object is activated (might be a statically initialized object)
341 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
343 struct work_struct
*work
= addr
;
347 case ODEBUG_STATE_NOTAVAILABLE
:
349 * This is not really a fixup. The work struct was
350 * statically initialized. We just make sure that it
351 * is tracked in the object tracker.
353 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
354 debug_object_init(work
, &work_debug_descr
);
355 debug_object_activate(work
, &work_debug_descr
);
361 case ODEBUG_STATE_ACTIVE
:
370 * fixup_free is called when:
371 * - an active object is freed
373 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
375 struct work_struct
*work
= addr
;
378 case ODEBUG_STATE_ACTIVE
:
379 cancel_work_sync(work
);
380 debug_object_free(work
, &work_debug_descr
);
387 static struct debug_obj_descr work_debug_descr
= {
388 .name
= "work_struct",
389 .debug_hint
= work_debug_hint
,
390 .fixup_init
= work_fixup_init
,
391 .fixup_activate
= work_fixup_activate
,
392 .fixup_free
= work_fixup_free
,
395 static inline void debug_work_activate(struct work_struct
*work
)
397 debug_object_activate(work
, &work_debug_descr
);
400 static inline void debug_work_deactivate(struct work_struct
*work
)
402 debug_object_deactivate(work
, &work_debug_descr
);
405 void __init_work(struct work_struct
*work
, int onstack
)
408 debug_object_init_on_stack(work
, &work_debug_descr
);
410 debug_object_init(work
, &work_debug_descr
);
412 EXPORT_SYMBOL_GPL(__init_work
);
414 void destroy_work_on_stack(struct work_struct
*work
)
416 debug_object_free(work
, &work_debug_descr
);
418 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
421 static inline void debug_work_activate(struct work_struct
*work
) { }
422 static inline void debug_work_deactivate(struct work_struct
*work
) { }
425 /* Serializes the accesses to the list of workqueues. */
426 static DEFINE_SPINLOCK(workqueue_lock
);
427 static LIST_HEAD(workqueues
);
428 static bool workqueue_freezing
; /* W: have wqs started freezing? */
431 * The almighty global cpu workqueues. nr_running is the only field
432 * which is expected to be used frequently by other cpus via
433 * try_to_wake_up(). Put it in a separate cacheline.
435 static DEFINE_PER_CPU(struct global_cwq
, global_cwq
);
436 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t
, pool_nr_running
[NR_STD_WORKER_POOLS
]);
439 * Global cpu workqueue and nr_running counter for unbound gcwq. The pools
440 * for online CPUs have POOL_DISASSOCIATED set, and all their workers have
441 * WORKER_UNBOUND set.
443 static struct global_cwq unbound_global_cwq
;
444 static atomic_t unbound_pool_nr_running
[NR_STD_WORKER_POOLS
] = {
445 [0 ... NR_STD_WORKER_POOLS
- 1] = ATOMIC_INIT(0), /* always 0 */
448 static int worker_thread(void *__worker
);
449 static unsigned int work_cpu(struct work_struct
*work
);
451 static int std_worker_pool_pri(struct worker_pool
*pool
)
453 return pool
- pool
->gcwq
->pools
;
456 static struct global_cwq
*get_gcwq(unsigned int cpu
)
458 if (cpu
!= WORK_CPU_UNBOUND
)
459 return &per_cpu(global_cwq
, cpu
);
461 return &unbound_global_cwq
;
464 static atomic_t
*get_pool_nr_running(struct worker_pool
*pool
)
466 int cpu
= pool
->gcwq
->cpu
;
467 int idx
= std_worker_pool_pri(pool
);
469 if (cpu
!= WORK_CPU_UNBOUND
)
470 return &per_cpu(pool_nr_running
, cpu
)[idx
];
472 return &unbound_pool_nr_running
[idx
];
475 static struct cpu_workqueue_struct
*get_cwq(unsigned int cpu
,
476 struct workqueue_struct
*wq
)
478 if (!(wq
->flags
& WQ_UNBOUND
)) {
479 if (likely(cpu
< nr_cpu_ids
))
480 return per_cpu_ptr(wq
->cpu_wq
.pcpu
, cpu
);
481 } else if (likely(cpu
== WORK_CPU_UNBOUND
))
482 return wq
->cpu_wq
.single
;
486 static unsigned int work_color_to_flags(int color
)
488 return color
<< WORK_STRUCT_COLOR_SHIFT
;
491 static int get_work_color(struct work_struct
*work
)
493 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
494 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
497 static int work_next_color(int color
)
499 return (color
+ 1) % WORK_NR_COLORS
;
503 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
504 * contain the pointer to the queued cwq. Once execution starts, the flag
505 * is cleared and the high bits contain OFFQ flags and CPU number.
507 * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling()
508 * and clear_work_data() can be used to set the cwq, cpu or clear
509 * work->data. These functions should only be called while the work is
510 * owned - ie. while the PENDING bit is set.
512 * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to
513 * a work. gcwq is available once the work has been queued anywhere after
514 * initialization until it is sync canceled. cwq is available only while
515 * the work item is queued.
517 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
518 * canceled. While being canceled, a work item may have its PENDING set
519 * but stay off timer and worklist for arbitrarily long and nobody should
520 * try to steal the PENDING bit.
522 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
525 BUG_ON(!work_pending(work
));
526 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
529 static void set_work_cwq(struct work_struct
*work
,
530 struct cpu_workqueue_struct
*cwq
,
531 unsigned long extra_flags
)
533 set_work_data(work
, (unsigned long)cwq
,
534 WORK_STRUCT_PENDING
| WORK_STRUCT_CWQ
| extra_flags
);
537 static void set_work_cpu_and_clear_pending(struct work_struct
*work
,
541 * The following wmb is paired with the implied mb in
542 * test_and_set_bit(PENDING) and ensures all updates to @work made
543 * here are visible to and precede any updates by the next PENDING
547 set_work_data(work
, (unsigned long)cpu
<< WORK_OFFQ_CPU_SHIFT
, 0);
550 static void clear_work_data(struct work_struct
*work
)
552 smp_wmb(); /* see set_work_cpu_and_clear_pending() */
553 set_work_data(work
, WORK_STRUCT_NO_CPU
, 0);
556 static struct cpu_workqueue_struct
*get_work_cwq(struct work_struct
*work
)
558 unsigned long data
= atomic_long_read(&work
->data
);
560 if (data
& WORK_STRUCT_CWQ
)
561 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
566 static struct global_cwq
*get_work_gcwq(struct work_struct
*work
)
568 unsigned long data
= atomic_long_read(&work
->data
);
571 if (data
& WORK_STRUCT_CWQ
)
572 return ((struct cpu_workqueue_struct
*)
573 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->gcwq
;
575 cpu
= data
>> WORK_OFFQ_CPU_SHIFT
;
576 if (cpu
== WORK_CPU_NONE
)
579 BUG_ON(cpu
>= nr_cpu_ids
&& cpu
!= WORK_CPU_UNBOUND
);
580 return get_gcwq(cpu
);
583 static void mark_work_canceling(struct work_struct
*work
)
585 struct global_cwq
*gcwq
= get_work_gcwq(work
);
586 unsigned long cpu
= gcwq
? gcwq
->cpu
: WORK_CPU_NONE
;
588 set_work_data(work
, (cpu
<< WORK_OFFQ_CPU_SHIFT
) | WORK_OFFQ_CANCELING
,
589 WORK_STRUCT_PENDING
);
592 static bool work_is_canceling(struct work_struct
*work
)
594 unsigned long data
= atomic_long_read(&work
->data
);
596 return !(data
& WORK_STRUCT_CWQ
) && (data
& WORK_OFFQ_CANCELING
);
600 * Policy functions. These define the policies on how the global worker
601 * pools are managed. Unless noted otherwise, these functions assume that
602 * they're being called with gcwq->lock held.
605 static bool __need_more_worker(struct worker_pool
*pool
)
607 return !atomic_read(get_pool_nr_running(pool
));
611 * Need to wake up a worker? Called from anything but currently
614 * Note that, because unbound workers never contribute to nr_running, this
615 * function will always return %true for unbound gcwq as long as the
616 * worklist isn't empty.
618 static bool need_more_worker(struct worker_pool
*pool
)
620 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
623 /* Can I start working? Called from busy but !running workers. */
624 static bool may_start_working(struct worker_pool
*pool
)
626 return pool
->nr_idle
;
629 /* Do I need to keep working? Called from currently running workers. */
630 static bool keep_working(struct worker_pool
*pool
)
632 atomic_t
*nr_running
= get_pool_nr_running(pool
);
634 return !list_empty(&pool
->worklist
) && atomic_read(nr_running
) <= 1;
637 /* Do we need a new worker? Called from manager. */
638 static bool need_to_create_worker(struct worker_pool
*pool
)
640 return need_more_worker(pool
) && !may_start_working(pool
);
643 /* Do I need to be the manager? */
644 static bool need_to_manage_workers(struct worker_pool
*pool
)
646 return need_to_create_worker(pool
) ||
647 (pool
->flags
& POOL_MANAGE_WORKERS
);
650 /* Do we have too many workers and should some go away? */
651 static bool too_many_workers(struct worker_pool
*pool
)
653 bool managing
= pool
->flags
& POOL_MANAGING_WORKERS
;
654 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
655 int nr_busy
= pool
->nr_workers
- nr_idle
;
658 * nr_idle and idle_list may disagree if idle rebinding is in
659 * progress. Never return %true if idle_list is empty.
661 if (list_empty(&pool
->idle_list
))
664 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
671 /* Return the first worker. Safe with preemption disabled */
672 static struct worker
*first_worker(struct worker_pool
*pool
)
674 if (unlikely(list_empty(&pool
->idle_list
)))
677 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
681 * wake_up_worker - wake up an idle worker
682 * @pool: worker pool to wake worker from
684 * Wake up the first idle worker of @pool.
687 * spin_lock_irq(gcwq->lock).
689 static void wake_up_worker(struct worker_pool
*pool
)
691 struct worker
*worker
= first_worker(pool
);
694 wake_up_process(worker
->task
);
698 * wq_worker_waking_up - a worker is waking up
699 * @task: task waking up
700 * @cpu: CPU @task is waking up to
702 * This function is called during try_to_wake_up() when a worker is
706 * spin_lock_irq(rq->lock)
708 void wq_worker_waking_up(struct task_struct
*task
, unsigned int cpu
)
710 struct worker
*worker
= kthread_data(task
);
712 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
713 WARN_ON_ONCE(worker
->pool
->gcwq
->cpu
!= cpu
);
714 atomic_inc(get_pool_nr_running(worker
->pool
));
719 * wq_worker_sleeping - a worker is going to sleep
720 * @task: task going to sleep
721 * @cpu: CPU in question, must be the current CPU number
723 * This function is called during schedule() when a busy worker is
724 * going to sleep. Worker on the same cpu can be woken up by
725 * returning pointer to its task.
728 * spin_lock_irq(rq->lock)
731 * Worker task on @cpu to wake up, %NULL if none.
733 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
,
736 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
737 struct worker_pool
*pool
;
738 atomic_t
*nr_running
;
741 * Rescuers, which may not have all the fields set up like normal
742 * workers, also reach here, let's not access anything before
743 * checking NOT_RUNNING.
745 if (worker
->flags
& WORKER_NOT_RUNNING
)
749 nr_running
= get_pool_nr_running(pool
);
751 /* this can only happen on the local cpu */
752 BUG_ON(cpu
!= raw_smp_processor_id());
755 * The counterpart of the following dec_and_test, implied mb,
756 * worklist not empty test sequence is in insert_work().
757 * Please read comment there.
759 * NOT_RUNNING is clear. This means that we're bound to and
760 * running on the local cpu w/ rq lock held and preemption
761 * disabled, which in turn means that none else could be
762 * manipulating idle_list, so dereferencing idle_list without gcwq
765 if (atomic_dec_and_test(nr_running
) && !list_empty(&pool
->worklist
))
766 to_wakeup
= first_worker(pool
);
767 return to_wakeup
? to_wakeup
->task
: NULL
;
771 * worker_set_flags - set worker flags and adjust nr_running accordingly
773 * @flags: flags to set
774 * @wakeup: wakeup an idle worker if necessary
776 * Set @flags in @worker->flags and adjust nr_running accordingly. If
777 * nr_running becomes zero and @wakeup is %true, an idle worker is
781 * spin_lock_irq(gcwq->lock)
783 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
786 struct worker_pool
*pool
= worker
->pool
;
788 WARN_ON_ONCE(worker
->task
!= current
);
791 * If transitioning into NOT_RUNNING, adjust nr_running and
792 * wake up an idle worker as necessary if requested by
795 if ((flags
& WORKER_NOT_RUNNING
) &&
796 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
797 atomic_t
*nr_running
= get_pool_nr_running(pool
);
800 if (atomic_dec_and_test(nr_running
) &&
801 !list_empty(&pool
->worklist
))
802 wake_up_worker(pool
);
804 atomic_dec(nr_running
);
807 worker
->flags
|= flags
;
811 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
813 * @flags: flags to clear
815 * Clear @flags in @worker->flags and adjust nr_running accordingly.
818 * spin_lock_irq(gcwq->lock)
820 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
822 struct worker_pool
*pool
= worker
->pool
;
823 unsigned int oflags
= worker
->flags
;
825 WARN_ON_ONCE(worker
->task
!= current
);
827 worker
->flags
&= ~flags
;
830 * If transitioning out of NOT_RUNNING, increment nr_running. Note
831 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
832 * of multiple flags, not a single flag.
834 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
835 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
836 atomic_inc(get_pool_nr_running(pool
));
840 * find_worker_executing_work - find worker which is executing a work
841 * @gcwq: gcwq of interest
842 * @work: work to find worker for
844 * Find a worker which is executing @work on @gcwq by searching
845 * @gcwq->busy_hash which is keyed by the address of @work. For a worker
846 * to match, its current execution should match the address of @work and
847 * its work function. This is to avoid unwanted dependency between
848 * unrelated work executions through a work item being recycled while still
851 * This is a bit tricky. A work item may be freed once its execution
852 * starts and nothing prevents the freed area from being recycled for
853 * another work item. If the same work item address ends up being reused
854 * before the original execution finishes, workqueue will identify the
855 * recycled work item as currently executing and make it wait until the
856 * current execution finishes, introducing an unwanted dependency.
858 * This function checks the work item address, work function and workqueue
859 * to avoid false positives. Note that this isn't complete as one may
860 * construct a work function which can introduce dependency onto itself
861 * through a recycled work item. Well, if somebody wants to shoot oneself
862 * in the foot that badly, there's only so much we can do, and if such
863 * deadlock actually occurs, it should be easy to locate the culprit work
867 * spin_lock_irq(gcwq->lock).
870 * Pointer to worker which is executing @work if found, NULL
873 static struct worker
*find_worker_executing_work(struct global_cwq
*gcwq
,
874 struct work_struct
*work
)
876 struct worker
*worker
;
877 struct hlist_node
*tmp
;
879 hash_for_each_possible(gcwq
->busy_hash
, worker
, tmp
, hentry
,
881 if (worker
->current_work
== work
&&
882 worker
->current_func
== work
->func
)
889 * move_linked_works - move linked works to a list
890 * @work: start of series of works to be scheduled
891 * @head: target list to append @work to
892 * @nextp: out paramter for nested worklist walking
894 * Schedule linked works starting from @work to @head. Work series to
895 * be scheduled starts at @work and includes any consecutive work with
896 * WORK_STRUCT_LINKED set in its predecessor.
898 * If @nextp is not NULL, it's updated to point to the next work of
899 * the last scheduled work. This allows move_linked_works() to be
900 * nested inside outer list_for_each_entry_safe().
903 * spin_lock_irq(gcwq->lock).
905 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
906 struct work_struct
**nextp
)
908 struct work_struct
*n
;
911 * Linked worklist will always end before the end of the list,
912 * use NULL for list head.
914 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
915 list_move_tail(&work
->entry
, head
);
916 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
921 * If we're already inside safe list traversal and have moved
922 * multiple works to the scheduled queue, the next position
923 * needs to be updated.
929 static void cwq_activate_delayed_work(struct work_struct
*work
)
931 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
933 trace_workqueue_activate_work(work
);
934 move_linked_works(work
, &cwq
->pool
->worklist
, NULL
);
935 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
939 static void cwq_activate_first_delayed(struct cpu_workqueue_struct
*cwq
)
941 struct work_struct
*work
= list_first_entry(&cwq
->delayed_works
,
942 struct work_struct
, entry
);
944 cwq_activate_delayed_work(work
);
948 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
949 * @cwq: cwq of interest
950 * @color: color of work which left the queue
952 * A work either has completed or is removed from pending queue,
953 * decrement nr_in_flight of its cwq and handle workqueue flushing.
956 * spin_lock_irq(gcwq->lock).
958 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct
*cwq
, int color
)
960 /* ignore uncolored works */
961 if (color
== WORK_NO_COLOR
)
964 cwq
->nr_in_flight
[color
]--;
967 if (!list_empty(&cwq
->delayed_works
)) {
968 /* one down, submit a delayed one */
969 if (cwq
->nr_active
< cwq
->max_active
)
970 cwq_activate_first_delayed(cwq
);
973 /* is flush in progress and are we at the flushing tip? */
974 if (likely(cwq
->flush_color
!= color
))
977 /* are there still in-flight works? */
978 if (cwq
->nr_in_flight
[color
])
981 /* this cwq is done, clear flush_color */
982 cwq
->flush_color
= -1;
985 * If this was the last cwq, wake up the first flusher. It
986 * will handle the rest.
988 if (atomic_dec_and_test(&cwq
->wq
->nr_cwqs_to_flush
))
989 complete(&cwq
->wq
->first_flusher
->done
);
993 * try_to_grab_pending - steal work item from worklist and disable irq
994 * @work: work item to steal
995 * @is_dwork: @work is a delayed_work
996 * @flags: place to store irq state
998 * Try to grab PENDING bit of @work. This function can handle @work in any
999 * stable state - idle, on timer or on worklist. Return values are
1001 * 1 if @work was pending and we successfully stole PENDING
1002 * 0 if @work was idle and we claimed PENDING
1003 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1004 * -ENOENT if someone else is canceling @work, this state may persist
1005 * for arbitrarily long
1007 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1008 * interrupted while holding PENDING and @work off queue, irq must be
1009 * disabled on entry. This, combined with delayed_work->timer being
1010 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1012 * On successful return, >= 0, irq is disabled and the caller is
1013 * responsible for releasing it using local_irq_restore(*@flags).
1015 * This function is safe to call from any context including IRQ handler.
1017 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1018 unsigned long *flags
)
1020 struct global_cwq
*gcwq
;
1022 local_irq_save(*flags
);
1024 /* try to steal the timer if it exists */
1026 struct delayed_work
*dwork
= to_delayed_work(work
);
1029 * dwork->timer is irqsafe. If del_timer() fails, it's
1030 * guaranteed that the timer is not queued anywhere and not
1031 * running on the local CPU.
1033 if (likely(del_timer(&dwork
->timer
)))
1037 /* try to claim PENDING the normal way */
1038 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1042 * The queueing is in progress, or it is already queued. Try to
1043 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1045 gcwq
= get_work_gcwq(work
);
1049 spin_lock(&gcwq
->lock
);
1050 if (!list_empty(&work
->entry
)) {
1052 * This work is queued, but perhaps we locked the wrong gcwq.
1053 * In that case we must see the new value after rmb(), see
1054 * insert_work()->wmb().
1057 if (gcwq
== get_work_gcwq(work
)) {
1058 debug_work_deactivate(work
);
1061 * A delayed work item cannot be grabbed directly
1062 * because it might have linked NO_COLOR work items
1063 * which, if left on the delayed_list, will confuse
1064 * cwq->nr_active management later on and cause
1065 * stall. Make sure the work item is activated
1068 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1069 cwq_activate_delayed_work(work
);
1071 list_del_init(&work
->entry
);
1072 cwq_dec_nr_in_flight(get_work_cwq(work
),
1073 get_work_color(work
));
1075 spin_unlock(&gcwq
->lock
);
1079 spin_unlock(&gcwq
->lock
);
1081 local_irq_restore(*flags
);
1082 if (work_is_canceling(work
))
1089 * insert_work - insert a work into gcwq
1090 * @cwq: cwq @work belongs to
1091 * @work: work to insert
1092 * @head: insertion point
1093 * @extra_flags: extra WORK_STRUCT_* flags to set
1095 * Insert @work which belongs to @cwq into @gcwq after @head.
1096 * @extra_flags is or'd to work_struct flags.
1099 * spin_lock_irq(gcwq->lock).
1101 static void insert_work(struct cpu_workqueue_struct
*cwq
,
1102 struct work_struct
*work
, struct list_head
*head
,
1103 unsigned int extra_flags
)
1105 struct worker_pool
*pool
= cwq
->pool
;
1107 /* we own @work, set data and link */
1108 set_work_cwq(work
, cwq
, extra_flags
);
1111 * Ensure that we get the right work->data if we see the
1112 * result of list_add() below, see try_to_grab_pending().
1116 list_add_tail(&work
->entry
, head
);
1119 * Ensure either worker_sched_deactivated() sees the above
1120 * list_add_tail() or we see zero nr_running to avoid workers
1121 * lying around lazily while there are works to be processed.
1125 if (__need_more_worker(pool
))
1126 wake_up_worker(pool
);
1130 * Test whether @work is being queued from another work executing on the
1131 * same workqueue. This is rather expensive and should only be used from
1134 static bool is_chained_work(struct workqueue_struct
*wq
)
1136 unsigned long flags
;
1139 for_each_gcwq_cpu(cpu
) {
1140 struct global_cwq
*gcwq
= get_gcwq(cpu
);
1141 struct worker
*worker
;
1142 struct hlist_node
*pos
;
1145 spin_lock_irqsave(&gcwq
->lock
, flags
);
1146 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
1147 if (worker
->task
!= current
)
1149 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1151 * I'm @worker, no locking necessary. See if @work
1152 * is headed to the same workqueue.
1154 return worker
->current_cwq
->wq
== wq
;
1156 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1161 static void __queue_work(unsigned int cpu
, struct workqueue_struct
*wq
,
1162 struct work_struct
*work
)
1164 struct global_cwq
*gcwq
;
1165 struct cpu_workqueue_struct
*cwq
;
1166 struct list_head
*worklist
;
1167 unsigned int work_flags
;
1168 unsigned int req_cpu
= cpu
;
1171 * While a work item is PENDING && off queue, a task trying to
1172 * steal the PENDING will busy-loop waiting for it to either get
1173 * queued or lose PENDING. Grabbing PENDING and queueing should
1174 * happen with IRQ disabled.
1176 WARN_ON_ONCE(!irqs_disabled());
1178 debug_work_activate(work
);
1180 /* if dying, only works from the same workqueue are allowed */
1181 if (unlikely(wq
->flags
& WQ_DRAINING
) &&
1182 WARN_ON_ONCE(!is_chained_work(wq
)))
1185 /* determine gcwq to use */
1186 if (!(wq
->flags
& WQ_UNBOUND
)) {
1187 struct global_cwq
*last_gcwq
;
1189 if (cpu
== WORK_CPU_UNBOUND
)
1190 cpu
= raw_smp_processor_id();
1193 * It's multi cpu. If @work was previously on a different
1194 * cpu, it might still be running there, in which case the
1195 * work needs to be queued on that cpu to guarantee
1198 gcwq
= get_gcwq(cpu
);
1199 last_gcwq
= get_work_gcwq(work
);
1201 if (last_gcwq
&& last_gcwq
!= gcwq
) {
1202 struct worker
*worker
;
1204 spin_lock(&last_gcwq
->lock
);
1206 worker
= find_worker_executing_work(last_gcwq
, work
);
1208 if (worker
&& worker
->current_cwq
->wq
== wq
)
1211 /* meh... not running there, queue here */
1212 spin_unlock(&last_gcwq
->lock
);
1213 spin_lock(&gcwq
->lock
);
1216 spin_lock(&gcwq
->lock
);
1219 gcwq
= get_gcwq(WORK_CPU_UNBOUND
);
1220 spin_lock(&gcwq
->lock
);
1223 /* gcwq determined, get cwq and queue */
1224 cwq
= get_cwq(gcwq
->cpu
, wq
);
1225 trace_workqueue_queue_work(req_cpu
, cwq
, work
);
1227 if (WARN_ON(!list_empty(&work
->entry
))) {
1228 spin_unlock(&gcwq
->lock
);
1232 cwq
->nr_in_flight
[cwq
->work_color
]++;
1233 work_flags
= work_color_to_flags(cwq
->work_color
);
1235 if (likely(cwq
->nr_active
< cwq
->max_active
)) {
1236 trace_workqueue_activate_work(work
);
1238 worklist
= &cwq
->pool
->worklist
;
1240 work_flags
|= WORK_STRUCT_DELAYED
;
1241 worklist
= &cwq
->delayed_works
;
1244 insert_work(cwq
, work
, worklist
, work_flags
);
1246 spin_unlock(&gcwq
->lock
);
1250 * queue_work_on - queue work on specific cpu
1251 * @cpu: CPU number to execute work on
1252 * @wq: workqueue to use
1253 * @work: work to queue
1255 * Returns %false if @work was already on a queue, %true otherwise.
1257 * We queue the work to a specific CPU, the caller must ensure it
1260 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1261 struct work_struct
*work
)
1264 unsigned long flags
;
1266 local_irq_save(flags
);
1268 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1269 __queue_work(cpu
, wq
, work
);
1273 local_irq_restore(flags
);
1276 EXPORT_SYMBOL_GPL(queue_work_on
);
1279 * queue_work - queue work on a workqueue
1280 * @wq: workqueue to use
1281 * @work: work to queue
1283 * Returns %false if @work was already on a queue, %true otherwise.
1285 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1286 * it can be processed by another CPU.
1288 bool queue_work(struct workqueue_struct
*wq
, struct work_struct
*work
)
1290 return queue_work_on(WORK_CPU_UNBOUND
, wq
, work
);
1292 EXPORT_SYMBOL_GPL(queue_work
);
1294 void delayed_work_timer_fn(unsigned long __data
)
1296 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1297 struct cpu_workqueue_struct
*cwq
= get_work_cwq(&dwork
->work
);
1299 /* should have been called from irqsafe timer with irq already off */
1300 __queue_work(dwork
->cpu
, cwq
->wq
, &dwork
->work
);
1302 EXPORT_SYMBOL_GPL(delayed_work_timer_fn
);
1304 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1305 struct delayed_work
*dwork
, unsigned long delay
)
1307 struct timer_list
*timer
= &dwork
->timer
;
1308 struct work_struct
*work
= &dwork
->work
;
1311 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1312 timer
->data
!= (unsigned long)dwork
);
1313 WARN_ON_ONCE(timer_pending(timer
));
1314 WARN_ON_ONCE(!list_empty(&work
->entry
));
1317 * If @delay is 0, queue @dwork->work immediately. This is for
1318 * both optimization and correctness. The earliest @timer can
1319 * expire is on the closest next tick and delayed_work users depend
1320 * on that there's no such delay when @delay is 0.
1323 __queue_work(cpu
, wq
, &dwork
->work
);
1327 timer_stats_timer_set_start_info(&dwork
->timer
);
1330 * This stores cwq for the moment, for the timer_fn. Note that the
1331 * work's gcwq is preserved to allow reentrance detection for
1334 if (!(wq
->flags
& WQ_UNBOUND
)) {
1335 struct global_cwq
*gcwq
= get_work_gcwq(work
);
1338 * If we cannot get the last gcwq from @work directly,
1339 * select the last CPU such that it avoids unnecessarily
1340 * triggering non-reentrancy check in __queue_work().
1345 if (lcpu
== WORK_CPU_UNBOUND
)
1346 lcpu
= raw_smp_processor_id();
1348 lcpu
= WORK_CPU_UNBOUND
;
1351 set_work_cwq(work
, get_cwq(lcpu
, wq
), 0);
1354 timer
->expires
= jiffies
+ delay
;
1356 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1357 add_timer_on(timer
, cpu
);
1363 * queue_delayed_work_on - queue work on specific CPU after delay
1364 * @cpu: CPU number to execute work on
1365 * @wq: workqueue to use
1366 * @dwork: work to queue
1367 * @delay: number of jiffies to wait before queueing
1369 * Returns %false if @work was already on a queue, %true otherwise. If
1370 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1373 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1374 struct delayed_work
*dwork
, unsigned long delay
)
1376 struct work_struct
*work
= &dwork
->work
;
1378 unsigned long flags
;
1380 /* read the comment in __queue_work() */
1381 local_irq_save(flags
);
1383 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1384 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1388 local_irq_restore(flags
);
1391 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1394 * queue_delayed_work - queue work on a workqueue after delay
1395 * @wq: workqueue to use
1396 * @dwork: delayable work to queue
1397 * @delay: number of jiffies to wait before queueing
1399 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1401 bool queue_delayed_work(struct workqueue_struct
*wq
,
1402 struct delayed_work
*dwork
, unsigned long delay
)
1404 return queue_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1406 EXPORT_SYMBOL_GPL(queue_delayed_work
);
1409 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1410 * @cpu: CPU number to execute work on
1411 * @wq: workqueue to use
1412 * @dwork: work to queue
1413 * @delay: number of jiffies to wait before queueing
1415 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1416 * modify @dwork's timer so that it expires after @delay. If @delay is
1417 * zero, @work is guaranteed to be scheduled immediately regardless of its
1420 * Returns %false if @dwork was idle and queued, %true if @dwork was
1421 * pending and its timer was modified.
1423 * This function is safe to call from any context including IRQ handler.
1424 * See try_to_grab_pending() for details.
1426 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1427 struct delayed_work
*dwork
, unsigned long delay
)
1429 unsigned long flags
;
1433 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1434 } while (unlikely(ret
== -EAGAIN
));
1436 if (likely(ret
>= 0)) {
1437 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1438 local_irq_restore(flags
);
1441 /* -ENOENT from try_to_grab_pending() becomes %true */
1444 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1447 * mod_delayed_work - modify delay of or queue a delayed work
1448 * @wq: workqueue to use
1449 * @dwork: work to queue
1450 * @delay: number of jiffies to wait before queueing
1452 * mod_delayed_work_on() on local CPU.
1454 bool mod_delayed_work(struct workqueue_struct
*wq
, struct delayed_work
*dwork
,
1455 unsigned long delay
)
1457 return mod_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1459 EXPORT_SYMBOL_GPL(mod_delayed_work
);
1462 * worker_enter_idle - enter idle state
1463 * @worker: worker which is entering idle state
1465 * @worker is entering idle state. Update stats and idle timer if
1469 * spin_lock_irq(gcwq->lock).
1471 static void worker_enter_idle(struct worker
*worker
)
1473 struct worker_pool
*pool
= worker
->pool
;
1475 BUG_ON(worker
->flags
& WORKER_IDLE
);
1476 BUG_ON(!list_empty(&worker
->entry
) &&
1477 (worker
->hentry
.next
|| worker
->hentry
.pprev
));
1479 /* can't use worker_set_flags(), also called from start_worker() */
1480 worker
->flags
|= WORKER_IDLE
;
1482 worker
->last_active
= jiffies
;
1484 /* idle_list is LIFO */
1485 list_add(&worker
->entry
, &pool
->idle_list
);
1487 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1488 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1491 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1492 * gcwq->lock between setting %WORKER_UNBOUND and zapping
1493 * nr_running, the warning may trigger spuriously. Check iff
1494 * unbind is not in progress.
1496 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1497 pool
->nr_workers
== pool
->nr_idle
&&
1498 atomic_read(get_pool_nr_running(pool
)));
1502 * worker_leave_idle - leave idle state
1503 * @worker: worker which is leaving idle state
1505 * @worker is leaving idle state. Update stats.
1508 * spin_lock_irq(gcwq->lock).
1510 static void worker_leave_idle(struct worker
*worker
)
1512 struct worker_pool
*pool
= worker
->pool
;
1514 BUG_ON(!(worker
->flags
& WORKER_IDLE
));
1515 worker_clr_flags(worker
, WORKER_IDLE
);
1517 list_del_init(&worker
->entry
);
1521 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1524 * Works which are scheduled while the cpu is online must at least be
1525 * scheduled to a worker which is bound to the cpu so that if they are
1526 * flushed from cpu callbacks while cpu is going down, they are
1527 * guaranteed to execute on the cpu.
1529 * This function is to be used by rogue workers and rescuers to bind
1530 * themselves to the target cpu and may race with cpu going down or
1531 * coming online. kthread_bind() can't be used because it may put the
1532 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1533 * verbatim as it's best effort and blocking and gcwq may be
1534 * [dis]associated in the meantime.
1536 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1537 * binding against %POOL_DISASSOCIATED which is set during
1538 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1539 * enters idle state or fetches works without dropping lock, it can
1540 * guarantee the scheduling requirement described in the first paragraph.
1543 * Might sleep. Called without any lock but returns with gcwq->lock
1547 * %true if the associated gcwq is online (@worker is successfully
1548 * bound), %false if offline.
1550 static bool worker_maybe_bind_and_lock(struct worker
*worker
)
1551 __acquires(&gcwq
->lock
)
1553 struct worker_pool
*pool
= worker
->pool
;
1554 struct global_cwq
*gcwq
= pool
->gcwq
;
1555 struct task_struct
*task
= worker
->task
;
1559 * The following call may fail, succeed or succeed
1560 * without actually migrating the task to the cpu if
1561 * it races with cpu hotunplug operation. Verify
1562 * against POOL_DISASSOCIATED.
1564 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1565 set_cpus_allowed_ptr(task
, get_cpu_mask(gcwq
->cpu
));
1567 spin_lock_irq(&gcwq
->lock
);
1568 if (pool
->flags
& POOL_DISASSOCIATED
)
1570 if (task_cpu(task
) == gcwq
->cpu
&&
1571 cpumask_equal(¤t
->cpus_allowed
,
1572 get_cpu_mask(gcwq
->cpu
)))
1574 spin_unlock_irq(&gcwq
->lock
);
1577 * We've raced with CPU hot[un]plug. Give it a breather
1578 * and retry migration. cond_resched() is required here;
1579 * otherwise, we might deadlock against cpu_stop trying to
1580 * bring down the CPU on non-preemptive kernel.
1588 * Rebind an idle @worker to its CPU. worker_thread() will test
1589 * list_empty(@worker->entry) before leaving idle and call this function.
1591 static void idle_worker_rebind(struct worker
*worker
)
1593 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1595 /* CPU may go down again inbetween, clear UNBOUND only on success */
1596 if (worker_maybe_bind_and_lock(worker
))
1597 worker_clr_flags(worker
, WORKER_UNBOUND
);
1599 /* rebind complete, become available again */
1600 list_add(&worker
->entry
, &worker
->pool
->idle_list
);
1601 spin_unlock_irq(&gcwq
->lock
);
1605 * Function for @worker->rebind.work used to rebind unbound busy workers to
1606 * the associated cpu which is coming back online. This is scheduled by
1607 * cpu up but can race with other cpu hotplug operations and may be
1608 * executed twice without intervening cpu down.
1610 static void busy_worker_rebind_fn(struct work_struct
*work
)
1612 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1613 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1615 if (worker_maybe_bind_and_lock(worker
))
1616 worker_clr_flags(worker
, WORKER_UNBOUND
);
1618 spin_unlock_irq(&gcwq
->lock
);
1622 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1623 * @gcwq: gcwq of interest
1625 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1626 * is different for idle and busy ones.
1628 * Idle ones will be removed from the idle_list and woken up. They will
1629 * add themselves back after completing rebind. This ensures that the
1630 * idle_list doesn't contain any unbound workers when re-bound busy workers
1631 * try to perform local wake-ups for concurrency management.
1633 * Busy workers can rebind after they finish their current work items.
1634 * Queueing the rebind work item at the head of the scheduled list is
1635 * enough. Note that nr_running will be properly bumped as busy workers
1638 * On return, all non-manager workers are scheduled for rebind - see
1639 * manage_workers() for the manager special case. Any idle worker
1640 * including the manager will not appear on @idle_list until rebind is
1641 * complete, making local wake-ups safe.
1643 static void rebind_workers(struct global_cwq
*gcwq
)
1645 struct worker_pool
*pool
;
1646 struct worker
*worker
, *n
;
1647 struct hlist_node
*pos
;
1650 lockdep_assert_held(&gcwq
->lock
);
1652 for_each_worker_pool(pool
, gcwq
)
1653 lockdep_assert_held(&pool
->assoc_mutex
);
1655 /* dequeue and kick idle ones */
1656 for_each_worker_pool(pool
, gcwq
) {
1657 list_for_each_entry_safe(worker
, n
, &pool
->idle_list
, entry
) {
1659 * idle workers should be off @pool->idle_list
1660 * until rebind is complete to avoid receiving
1661 * premature local wake-ups.
1663 list_del_init(&worker
->entry
);
1666 * worker_thread() will see the above dequeuing
1667 * and call idle_worker_rebind().
1669 wake_up_process(worker
->task
);
1673 /* rebind busy workers */
1674 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
1675 struct work_struct
*rebind_work
= &worker
->rebind_work
;
1676 struct workqueue_struct
*wq
;
1678 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
1679 work_data_bits(rebind_work
)))
1682 debug_work_activate(rebind_work
);
1685 * wq doesn't really matter but let's keep @worker->pool
1686 * and @cwq->pool consistent for sanity.
1688 if (std_worker_pool_pri(worker
->pool
))
1689 wq
= system_highpri_wq
;
1693 insert_work(get_cwq(gcwq
->cpu
, wq
), rebind_work
,
1694 worker
->scheduled
.next
,
1695 work_color_to_flags(WORK_NO_COLOR
));
1699 static struct worker
*alloc_worker(void)
1701 struct worker
*worker
;
1703 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1705 INIT_LIST_HEAD(&worker
->entry
);
1706 INIT_LIST_HEAD(&worker
->scheduled
);
1707 INIT_WORK(&worker
->rebind_work
, busy_worker_rebind_fn
);
1708 /* on creation a worker is in !idle && prep state */
1709 worker
->flags
= WORKER_PREP
;
1715 * create_worker - create a new workqueue worker
1716 * @pool: pool the new worker will belong to
1718 * Create a new worker which is bound to @pool. The returned worker
1719 * can be started by calling start_worker() or destroyed using
1723 * Might sleep. Does GFP_KERNEL allocations.
1726 * Pointer to the newly created worker.
1728 static struct worker
*create_worker(struct worker_pool
*pool
)
1730 struct global_cwq
*gcwq
= pool
->gcwq
;
1731 const char *pri
= std_worker_pool_pri(pool
) ? "H" : "";
1732 struct worker
*worker
= NULL
;
1735 spin_lock_irq(&gcwq
->lock
);
1736 while (ida_get_new(&pool
->worker_ida
, &id
)) {
1737 spin_unlock_irq(&gcwq
->lock
);
1738 if (!ida_pre_get(&pool
->worker_ida
, GFP_KERNEL
))
1740 spin_lock_irq(&gcwq
->lock
);
1742 spin_unlock_irq(&gcwq
->lock
);
1744 worker
= alloc_worker();
1748 worker
->pool
= pool
;
1751 if (gcwq
->cpu
!= WORK_CPU_UNBOUND
)
1752 worker
->task
= kthread_create_on_node(worker_thread
,
1753 worker
, cpu_to_node(gcwq
->cpu
),
1754 "kworker/%u:%d%s", gcwq
->cpu
, id
, pri
);
1756 worker
->task
= kthread_create(worker_thread
, worker
,
1757 "kworker/u:%d%s", id
, pri
);
1758 if (IS_ERR(worker
->task
))
1761 if (std_worker_pool_pri(pool
))
1762 set_user_nice(worker
->task
, HIGHPRI_NICE_LEVEL
);
1765 * Determine CPU binding of the new worker depending on
1766 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1767 * flag remains stable across this function. See the comments
1768 * above the flag definition for details.
1770 * As an unbound worker may later become a regular one if CPU comes
1771 * online, make sure every worker has %PF_THREAD_BOUND set.
1773 if (!(pool
->flags
& POOL_DISASSOCIATED
)) {
1774 kthread_bind(worker
->task
, gcwq
->cpu
);
1776 worker
->task
->flags
|= PF_THREAD_BOUND
;
1777 worker
->flags
|= WORKER_UNBOUND
;
1783 spin_lock_irq(&gcwq
->lock
);
1784 ida_remove(&pool
->worker_ida
, id
);
1785 spin_unlock_irq(&gcwq
->lock
);
1792 * start_worker - start a newly created worker
1793 * @worker: worker to start
1795 * Make the gcwq aware of @worker and start it.
1798 * spin_lock_irq(gcwq->lock).
1800 static void start_worker(struct worker
*worker
)
1802 worker
->flags
|= WORKER_STARTED
;
1803 worker
->pool
->nr_workers
++;
1804 worker_enter_idle(worker
);
1805 wake_up_process(worker
->task
);
1809 * destroy_worker - destroy a workqueue worker
1810 * @worker: worker to be destroyed
1812 * Destroy @worker and adjust @gcwq stats accordingly.
1815 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1817 static void destroy_worker(struct worker
*worker
)
1819 struct worker_pool
*pool
= worker
->pool
;
1820 struct global_cwq
*gcwq
= pool
->gcwq
;
1821 int id
= worker
->id
;
1823 /* sanity check frenzy */
1824 BUG_ON(worker
->current_work
);
1825 BUG_ON(!list_empty(&worker
->scheduled
));
1827 if (worker
->flags
& WORKER_STARTED
)
1829 if (worker
->flags
& WORKER_IDLE
)
1832 list_del_init(&worker
->entry
);
1833 worker
->flags
|= WORKER_DIE
;
1835 spin_unlock_irq(&gcwq
->lock
);
1837 kthread_stop(worker
->task
);
1840 spin_lock_irq(&gcwq
->lock
);
1841 ida_remove(&pool
->worker_ida
, id
);
1844 static void idle_worker_timeout(unsigned long __pool
)
1846 struct worker_pool
*pool
= (void *)__pool
;
1847 struct global_cwq
*gcwq
= pool
->gcwq
;
1849 spin_lock_irq(&gcwq
->lock
);
1851 if (too_many_workers(pool
)) {
1852 struct worker
*worker
;
1853 unsigned long expires
;
1855 /* idle_list is kept in LIFO order, check the last one */
1856 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1857 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1859 if (time_before(jiffies
, expires
))
1860 mod_timer(&pool
->idle_timer
, expires
);
1862 /* it's been idle for too long, wake up manager */
1863 pool
->flags
|= POOL_MANAGE_WORKERS
;
1864 wake_up_worker(pool
);
1868 spin_unlock_irq(&gcwq
->lock
);
1871 static bool send_mayday(struct work_struct
*work
)
1873 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1874 struct workqueue_struct
*wq
= cwq
->wq
;
1877 if (!(wq
->flags
& WQ_RESCUER
))
1880 /* mayday mayday mayday */
1881 cpu
= cwq
->pool
->gcwq
->cpu
;
1882 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1883 if (cpu
== WORK_CPU_UNBOUND
)
1885 if (!mayday_test_and_set_cpu(cpu
, wq
->mayday_mask
))
1886 wake_up_process(wq
->rescuer
->task
);
1890 static void gcwq_mayday_timeout(unsigned long __pool
)
1892 struct worker_pool
*pool
= (void *)__pool
;
1893 struct global_cwq
*gcwq
= pool
->gcwq
;
1894 struct work_struct
*work
;
1896 spin_lock_irq(&gcwq
->lock
);
1898 if (need_to_create_worker(pool
)) {
1900 * We've been trying to create a new worker but
1901 * haven't been successful. We might be hitting an
1902 * allocation deadlock. Send distress signals to
1905 list_for_each_entry(work
, &pool
->worklist
, entry
)
1909 spin_unlock_irq(&gcwq
->lock
);
1911 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1915 * maybe_create_worker - create a new worker if necessary
1916 * @pool: pool to create a new worker for
1918 * Create a new worker for @pool if necessary. @pool is guaranteed to
1919 * have at least one idle worker on return from this function. If
1920 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1921 * sent to all rescuers with works scheduled on @pool to resolve
1922 * possible allocation deadlock.
1924 * On return, need_to_create_worker() is guaranteed to be false and
1925 * may_start_working() true.
1928 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1929 * multiple times. Does GFP_KERNEL allocations. Called only from
1933 * false if no action was taken and gcwq->lock stayed locked, true
1936 static bool maybe_create_worker(struct worker_pool
*pool
)
1937 __releases(&gcwq
->lock
)
1938 __acquires(&gcwq
->lock
)
1940 struct global_cwq
*gcwq
= pool
->gcwq
;
1942 if (!need_to_create_worker(pool
))
1945 spin_unlock_irq(&gcwq
->lock
);
1947 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1948 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1951 struct worker
*worker
;
1953 worker
= create_worker(pool
);
1955 del_timer_sync(&pool
->mayday_timer
);
1956 spin_lock_irq(&gcwq
->lock
);
1957 start_worker(worker
);
1958 BUG_ON(need_to_create_worker(pool
));
1962 if (!need_to_create_worker(pool
))
1965 __set_current_state(TASK_INTERRUPTIBLE
);
1966 schedule_timeout(CREATE_COOLDOWN
);
1968 if (!need_to_create_worker(pool
))
1972 del_timer_sync(&pool
->mayday_timer
);
1973 spin_lock_irq(&gcwq
->lock
);
1974 if (need_to_create_worker(pool
))
1980 * maybe_destroy_worker - destroy workers which have been idle for a while
1981 * @pool: pool to destroy workers for
1983 * Destroy @pool workers which have been idle for longer than
1984 * IDLE_WORKER_TIMEOUT.
1987 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1988 * multiple times. Called only from manager.
1991 * false if no action was taken and gcwq->lock stayed locked, true
1994 static bool maybe_destroy_workers(struct worker_pool
*pool
)
1998 while (too_many_workers(pool
)) {
1999 struct worker
*worker
;
2000 unsigned long expires
;
2002 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2003 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2005 if (time_before(jiffies
, expires
)) {
2006 mod_timer(&pool
->idle_timer
, expires
);
2010 destroy_worker(worker
);
2018 * manage_workers - manage worker pool
2021 * Assume the manager role and manage gcwq worker pool @worker belongs
2022 * to. At any given time, there can be only zero or one manager per
2023 * gcwq. The exclusion is handled automatically by this function.
2025 * The caller can safely start processing works on false return. On
2026 * true return, it's guaranteed that need_to_create_worker() is false
2027 * and may_start_working() is true.
2030 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2031 * multiple times. Does GFP_KERNEL allocations.
2034 * false if no action was taken and gcwq->lock stayed locked, true if
2035 * some action was taken.
2037 static bool manage_workers(struct worker
*worker
)
2039 struct worker_pool
*pool
= worker
->pool
;
2042 if (pool
->flags
& POOL_MANAGING_WORKERS
)
2045 pool
->flags
|= POOL_MANAGING_WORKERS
;
2048 * To simplify both worker management and CPU hotplug, hold off
2049 * management while hotplug is in progress. CPU hotplug path can't
2050 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2051 * lead to idle worker depletion (all become busy thinking someone
2052 * else is managing) which in turn can result in deadlock under
2053 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2054 * manager against CPU hotplug.
2056 * assoc_mutex would always be free unless CPU hotplug is in
2057 * progress. trylock first without dropping @gcwq->lock.
2059 if (unlikely(!mutex_trylock(&pool
->assoc_mutex
))) {
2060 spin_unlock_irq(&pool
->gcwq
->lock
);
2061 mutex_lock(&pool
->assoc_mutex
);
2063 * CPU hotplug could have happened while we were waiting
2064 * for assoc_mutex. Hotplug itself can't handle us
2065 * because manager isn't either on idle or busy list, and
2066 * @gcwq's state and ours could have deviated.
2068 * As hotplug is now excluded via assoc_mutex, we can
2069 * simply try to bind. It will succeed or fail depending
2070 * on @gcwq's current state. Try it and adjust
2071 * %WORKER_UNBOUND accordingly.
2073 if (worker_maybe_bind_and_lock(worker
))
2074 worker
->flags
&= ~WORKER_UNBOUND
;
2076 worker
->flags
|= WORKER_UNBOUND
;
2081 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2084 * Destroy and then create so that may_start_working() is true
2087 ret
|= maybe_destroy_workers(pool
);
2088 ret
|= maybe_create_worker(pool
);
2090 pool
->flags
&= ~POOL_MANAGING_WORKERS
;
2091 mutex_unlock(&pool
->assoc_mutex
);
2096 * process_one_work - process single work
2098 * @work: work to process
2100 * Process @work. This function contains all the logics necessary to
2101 * process a single work including synchronization against and
2102 * interaction with other workers on the same cpu, queueing and
2103 * flushing. As long as context requirement is met, any worker can
2104 * call this function to process a work.
2107 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2109 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2110 __releases(&gcwq
->lock
)
2111 __acquires(&gcwq
->lock
)
2113 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
2114 struct worker_pool
*pool
= worker
->pool
;
2115 struct global_cwq
*gcwq
= pool
->gcwq
;
2116 bool cpu_intensive
= cwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2118 struct worker
*collision
;
2119 #ifdef CONFIG_LOCKDEP
2121 * It is permissible to free the struct work_struct from
2122 * inside the function that is called from it, this we need to
2123 * take into account for lockdep too. To avoid bogus "held
2124 * lock freed" warnings as well as problems when looking into
2125 * work->lockdep_map, make a copy and use that here.
2127 struct lockdep_map lockdep_map
;
2129 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2132 * Ensure we're on the correct CPU. DISASSOCIATED test is
2133 * necessary to avoid spurious warnings from rescuers servicing the
2134 * unbound or a disassociated pool.
2136 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2137 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2138 raw_smp_processor_id() != gcwq
->cpu
);
2141 * A single work shouldn't be executed concurrently by
2142 * multiple workers on a single cpu. Check whether anyone is
2143 * already processing the work. If so, defer the work to the
2144 * currently executing one.
2146 collision
= find_worker_executing_work(gcwq
, work
);
2147 if (unlikely(collision
)) {
2148 move_linked_works(work
, &collision
->scheduled
, NULL
);
2152 /* claim and dequeue */
2153 debug_work_deactivate(work
);
2154 hash_add(gcwq
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2155 worker
->current_work
= work
;
2156 worker
->current_func
= work
->func
;
2157 worker
->current_cwq
= cwq
;
2158 work_color
= get_work_color(work
);
2160 list_del_init(&work
->entry
);
2163 * CPU intensive works don't participate in concurrency
2164 * management. They're the scheduler's responsibility.
2166 if (unlikely(cpu_intensive
))
2167 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2170 * Unbound gcwq isn't concurrency managed and work items should be
2171 * executed ASAP. Wake up another worker if necessary.
2173 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2174 wake_up_worker(pool
);
2177 * Record the last CPU and clear PENDING which should be the last
2178 * update to @work. Also, do this inside @gcwq->lock so that
2179 * PENDING and queued state changes happen together while IRQ is
2182 set_work_cpu_and_clear_pending(work
, gcwq
->cpu
);
2184 spin_unlock_irq(&gcwq
->lock
);
2186 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2187 lock_map_acquire(&lockdep_map
);
2188 trace_workqueue_execute_start(work
);
2189 worker
->current_func(work
);
2191 * While we must be careful to not use "work" after this, the trace
2192 * point will only record its address.
2194 trace_workqueue_execute_end(work
);
2195 lock_map_release(&lockdep_map
);
2196 lock_map_release(&cwq
->wq
->lockdep_map
);
2198 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2199 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2200 " last function: %pf\n",
2201 current
->comm
, preempt_count(), task_pid_nr(current
),
2202 worker
->current_func
);
2203 debug_show_held_locks(current
);
2207 spin_lock_irq(&gcwq
->lock
);
2209 /* clear cpu intensive status */
2210 if (unlikely(cpu_intensive
))
2211 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2213 /* we're done with it, release */
2214 hash_del(&worker
->hentry
);
2215 worker
->current_work
= NULL
;
2216 worker
->current_func
= NULL
;
2217 worker
->current_cwq
= NULL
;
2218 cwq_dec_nr_in_flight(cwq
, work_color
);
2222 * process_scheduled_works - process scheduled works
2225 * Process all scheduled works. Please note that the scheduled list
2226 * may change while processing a work, so this function repeatedly
2227 * fetches a work from the top and executes it.
2230 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2233 static void process_scheduled_works(struct worker
*worker
)
2235 while (!list_empty(&worker
->scheduled
)) {
2236 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2237 struct work_struct
, entry
);
2238 process_one_work(worker
, work
);
2243 * worker_thread - the worker thread function
2246 * The gcwq worker thread function. There's a single dynamic pool of
2247 * these per each cpu. These workers process all works regardless of
2248 * their specific target workqueue. The only exception is works which
2249 * belong to workqueues with a rescuer which will be explained in
2252 static int worker_thread(void *__worker
)
2254 struct worker
*worker
= __worker
;
2255 struct worker_pool
*pool
= worker
->pool
;
2256 struct global_cwq
*gcwq
= pool
->gcwq
;
2258 /* tell the scheduler that this is a workqueue worker */
2259 worker
->task
->flags
|= PF_WQ_WORKER
;
2261 spin_lock_irq(&gcwq
->lock
);
2263 /* we are off idle list if destruction or rebind is requested */
2264 if (unlikely(list_empty(&worker
->entry
))) {
2265 spin_unlock_irq(&gcwq
->lock
);
2267 /* if DIE is set, destruction is requested */
2268 if (worker
->flags
& WORKER_DIE
) {
2269 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2273 /* otherwise, rebind */
2274 idle_worker_rebind(worker
);
2278 worker_leave_idle(worker
);
2280 /* no more worker necessary? */
2281 if (!need_more_worker(pool
))
2284 /* do we need to manage? */
2285 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2289 * ->scheduled list can only be filled while a worker is
2290 * preparing to process a work or actually processing it.
2291 * Make sure nobody diddled with it while I was sleeping.
2293 BUG_ON(!list_empty(&worker
->scheduled
));
2296 * When control reaches this point, we're guaranteed to have
2297 * at least one idle worker or that someone else has already
2298 * assumed the manager role.
2300 worker_clr_flags(worker
, WORKER_PREP
);
2303 struct work_struct
*work
=
2304 list_first_entry(&pool
->worklist
,
2305 struct work_struct
, entry
);
2307 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2308 /* optimization path, not strictly necessary */
2309 process_one_work(worker
, work
);
2310 if (unlikely(!list_empty(&worker
->scheduled
)))
2311 process_scheduled_works(worker
);
2313 move_linked_works(work
, &worker
->scheduled
, NULL
);
2314 process_scheduled_works(worker
);
2316 } while (keep_working(pool
));
2318 worker_set_flags(worker
, WORKER_PREP
, false);
2320 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2324 * gcwq->lock is held and there's no work to process and no
2325 * need to manage, sleep. Workers are woken up only while
2326 * holding gcwq->lock or from local cpu, so setting the
2327 * current state before releasing gcwq->lock is enough to
2328 * prevent losing any event.
2330 worker_enter_idle(worker
);
2331 __set_current_state(TASK_INTERRUPTIBLE
);
2332 spin_unlock_irq(&gcwq
->lock
);
2338 * rescuer_thread - the rescuer thread function
2341 * Workqueue rescuer thread function. There's one rescuer for each
2342 * workqueue which has WQ_RESCUER set.
2344 * Regular work processing on a gcwq may block trying to create a new
2345 * worker which uses GFP_KERNEL allocation which has slight chance of
2346 * developing into deadlock if some works currently on the same queue
2347 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2348 * the problem rescuer solves.
2350 * When such condition is possible, the gcwq summons rescuers of all
2351 * workqueues which have works queued on the gcwq and let them process
2352 * those works so that forward progress can be guaranteed.
2354 * This should happen rarely.
2356 static int rescuer_thread(void *__rescuer
)
2358 struct worker
*rescuer
= __rescuer
;
2359 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2360 struct list_head
*scheduled
= &rescuer
->scheduled
;
2361 bool is_unbound
= wq
->flags
& WQ_UNBOUND
;
2364 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2367 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2368 * doesn't participate in concurrency management.
2370 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2372 set_current_state(TASK_INTERRUPTIBLE
);
2374 if (kthread_should_stop()) {
2375 __set_current_state(TASK_RUNNING
);
2376 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2381 * See whether any cpu is asking for help. Unbounded
2382 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2384 for_each_mayday_cpu(cpu
, wq
->mayday_mask
) {
2385 unsigned int tcpu
= is_unbound
? WORK_CPU_UNBOUND
: cpu
;
2386 struct cpu_workqueue_struct
*cwq
= get_cwq(tcpu
, wq
);
2387 struct worker_pool
*pool
= cwq
->pool
;
2388 struct global_cwq
*gcwq
= pool
->gcwq
;
2389 struct work_struct
*work
, *n
;
2391 __set_current_state(TASK_RUNNING
);
2392 mayday_clear_cpu(cpu
, wq
->mayday_mask
);
2394 /* migrate to the target cpu if possible */
2395 rescuer
->pool
= pool
;
2396 worker_maybe_bind_and_lock(rescuer
);
2399 * Slurp in all works issued via this workqueue and
2402 BUG_ON(!list_empty(&rescuer
->scheduled
));
2403 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2404 if (get_work_cwq(work
) == cwq
)
2405 move_linked_works(work
, scheduled
, &n
);
2407 process_scheduled_works(rescuer
);
2410 * Leave this gcwq. If keep_working() is %true, notify a
2411 * regular worker; otherwise, we end up with 0 concurrency
2412 * and stalling the execution.
2414 if (keep_working(pool
))
2415 wake_up_worker(pool
);
2417 spin_unlock_irq(&gcwq
->lock
);
2420 /* rescuers should never participate in concurrency management */
2421 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2427 struct work_struct work
;
2428 struct completion done
;
2431 static void wq_barrier_func(struct work_struct
*work
)
2433 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2434 complete(&barr
->done
);
2438 * insert_wq_barrier - insert a barrier work
2439 * @cwq: cwq to insert barrier into
2440 * @barr: wq_barrier to insert
2441 * @target: target work to attach @barr to
2442 * @worker: worker currently executing @target, NULL if @target is not executing
2444 * @barr is linked to @target such that @barr is completed only after
2445 * @target finishes execution. Please note that the ordering
2446 * guarantee is observed only with respect to @target and on the local
2449 * Currently, a queued barrier can't be canceled. This is because
2450 * try_to_grab_pending() can't determine whether the work to be
2451 * grabbed is at the head of the queue and thus can't clear LINKED
2452 * flag of the previous work while there must be a valid next work
2453 * after a work with LINKED flag set.
2455 * Note that when @worker is non-NULL, @target may be modified
2456 * underneath us, so we can't reliably determine cwq from @target.
2459 * spin_lock_irq(gcwq->lock).
2461 static void insert_wq_barrier(struct cpu_workqueue_struct
*cwq
,
2462 struct wq_barrier
*barr
,
2463 struct work_struct
*target
, struct worker
*worker
)
2465 struct list_head
*head
;
2466 unsigned int linked
= 0;
2469 * debugobject calls are safe here even with gcwq->lock locked
2470 * as we know for sure that this will not trigger any of the
2471 * checks and call back into the fixup functions where we
2474 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2475 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2476 init_completion(&barr
->done
);
2479 * If @target is currently being executed, schedule the
2480 * barrier to the worker; otherwise, put it after @target.
2483 head
= worker
->scheduled
.next
;
2485 unsigned long *bits
= work_data_bits(target
);
2487 head
= target
->entry
.next
;
2488 /* there can already be other linked works, inherit and set */
2489 linked
= *bits
& WORK_STRUCT_LINKED
;
2490 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2493 debug_work_activate(&barr
->work
);
2494 insert_work(cwq
, &barr
->work
, head
,
2495 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2499 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2500 * @wq: workqueue being flushed
2501 * @flush_color: new flush color, < 0 for no-op
2502 * @work_color: new work color, < 0 for no-op
2504 * Prepare cwqs for workqueue flushing.
2506 * If @flush_color is non-negative, flush_color on all cwqs should be
2507 * -1. If no cwq has in-flight commands at the specified color, all
2508 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2509 * has in flight commands, its cwq->flush_color is set to
2510 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2511 * wakeup logic is armed and %true is returned.
2513 * The caller should have initialized @wq->first_flusher prior to
2514 * calling this function with non-negative @flush_color. If
2515 * @flush_color is negative, no flush color update is done and %false
2518 * If @work_color is non-negative, all cwqs should have the same
2519 * work_color which is previous to @work_color and all will be
2520 * advanced to @work_color.
2523 * mutex_lock(wq->flush_mutex).
2526 * %true if @flush_color >= 0 and there's something to flush. %false
2529 static bool flush_workqueue_prep_cwqs(struct workqueue_struct
*wq
,
2530 int flush_color
, int work_color
)
2535 if (flush_color
>= 0) {
2536 BUG_ON(atomic_read(&wq
->nr_cwqs_to_flush
));
2537 atomic_set(&wq
->nr_cwqs_to_flush
, 1);
2540 for_each_cwq_cpu(cpu
, wq
) {
2541 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2542 struct global_cwq
*gcwq
= cwq
->pool
->gcwq
;
2544 spin_lock_irq(&gcwq
->lock
);
2546 if (flush_color
>= 0) {
2547 BUG_ON(cwq
->flush_color
!= -1);
2549 if (cwq
->nr_in_flight
[flush_color
]) {
2550 cwq
->flush_color
= flush_color
;
2551 atomic_inc(&wq
->nr_cwqs_to_flush
);
2556 if (work_color
>= 0) {
2557 BUG_ON(work_color
!= work_next_color(cwq
->work_color
));
2558 cwq
->work_color
= work_color
;
2561 spin_unlock_irq(&gcwq
->lock
);
2564 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_cwqs_to_flush
))
2565 complete(&wq
->first_flusher
->done
);
2571 * flush_workqueue - ensure that any scheduled work has run to completion.
2572 * @wq: workqueue to flush
2574 * Forces execution of the workqueue and blocks until its completion.
2575 * This is typically used in driver shutdown handlers.
2577 * We sleep until all works which were queued on entry have been handled,
2578 * but we are not livelocked by new incoming ones.
2580 void flush_workqueue(struct workqueue_struct
*wq
)
2582 struct wq_flusher this_flusher
= {
2583 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2585 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2589 lock_map_acquire(&wq
->lockdep_map
);
2590 lock_map_release(&wq
->lockdep_map
);
2592 mutex_lock(&wq
->flush_mutex
);
2595 * Start-to-wait phase
2597 next_color
= work_next_color(wq
->work_color
);
2599 if (next_color
!= wq
->flush_color
) {
2601 * Color space is not full. The current work_color
2602 * becomes our flush_color and work_color is advanced
2605 BUG_ON(!list_empty(&wq
->flusher_overflow
));
2606 this_flusher
.flush_color
= wq
->work_color
;
2607 wq
->work_color
= next_color
;
2609 if (!wq
->first_flusher
) {
2610 /* no flush in progress, become the first flusher */
2611 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2613 wq
->first_flusher
= &this_flusher
;
2615 if (!flush_workqueue_prep_cwqs(wq
, wq
->flush_color
,
2617 /* nothing to flush, done */
2618 wq
->flush_color
= next_color
;
2619 wq
->first_flusher
= NULL
;
2624 BUG_ON(wq
->flush_color
== this_flusher
.flush_color
);
2625 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2626 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2630 * Oops, color space is full, wait on overflow queue.
2631 * The next flush completion will assign us
2632 * flush_color and transfer to flusher_queue.
2634 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2637 mutex_unlock(&wq
->flush_mutex
);
2639 wait_for_completion(&this_flusher
.done
);
2642 * Wake-up-and-cascade phase
2644 * First flushers are responsible for cascading flushes and
2645 * handling overflow. Non-first flushers can simply return.
2647 if (wq
->first_flusher
!= &this_flusher
)
2650 mutex_lock(&wq
->flush_mutex
);
2652 /* we might have raced, check again with mutex held */
2653 if (wq
->first_flusher
!= &this_flusher
)
2656 wq
->first_flusher
= NULL
;
2658 BUG_ON(!list_empty(&this_flusher
.list
));
2659 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2662 struct wq_flusher
*next
, *tmp
;
2664 /* complete all the flushers sharing the current flush color */
2665 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2666 if (next
->flush_color
!= wq
->flush_color
)
2668 list_del_init(&next
->list
);
2669 complete(&next
->done
);
2672 BUG_ON(!list_empty(&wq
->flusher_overflow
) &&
2673 wq
->flush_color
!= work_next_color(wq
->work_color
));
2675 /* this flush_color is finished, advance by one */
2676 wq
->flush_color
= work_next_color(wq
->flush_color
);
2678 /* one color has been freed, handle overflow queue */
2679 if (!list_empty(&wq
->flusher_overflow
)) {
2681 * Assign the same color to all overflowed
2682 * flushers, advance work_color and append to
2683 * flusher_queue. This is the start-to-wait
2684 * phase for these overflowed flushers.
2686 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2687 tmp
->flush_color
= wq
->work_color
;
2689 wq
->work_color
= work_next_color(wq
->work_color
);
2691 list_splice_tail_init(&wq
->flusher_overflow
,
2692 &wq
->flusher_queue
);
2693 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2696 if (list_empty(&wq
->flusher_queue
)) {
2697 BUG_ON(wq
->flush_color
!= wq
->work_color
);
2702 * Need to flush more colors. Make the next flusher
2703 * the new first flusher and arm cwqs.
2705 BUG_ON(wq
->flush_color
== wq
->work_color
);
2706 BUG_ON(wq
->flush_color
!= next
->flush_color
);
2708 list_del_init(&next
->list
);
2709 wq
->first_flusher
= next
;
2711 if (flush_workqueue_prep_cwqs(wq
, wq
->flush_color
, -1))
2715 * Meh... this color is already done, clear first
2716 * flusher and repeat cascading.
2718 wq
->first_flusher
= NULL
;
2722 mutex_unlock(&wq
->flush_mutex
);
2724 EXPORT_SYMBOL_GPL(flush_workqueue
);
2727 * drain_workqueue - drain a workqueue
2728 * @wq: workqueue to drain
2730 * Wait until the workqueue becomes empty. While draining is in progress,
2731 * only chain queueing is allowed. IOW, only currently pending or running
2732 * work items on @wq can queue further work items on it. @wq is flushed
2733 * repeatedly until it becomes empty. The number of flushing is detemined
2734 * by the depth of chaining and should be relatively short. Whine if it
2737 void drain_workqueue(struct workqueue_struct
*wq
)
2739 unsigned int flush_cnt
= 0;
2743 * __queue_work() needs to test whether there are drainers, is much
2744 * hotter than drain_workqueue() and already looks at @wq->flags.
2745 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2747 spin_lock(&workqueue_lock
);
2748 if (!wq
->nr_drainers
++)
2749 wq
->flags
|= WQ_DRAINING
;
2750 spin_unlock(&workqueue_lock
);
2752 flush_workqueue(wq
);
2754 for_each_cwq_cpu(cpu
, wq
) {
2755 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2758 spin_lock_irq(&cwq
->pool
->gcwq
->lock
);
2759 drained
= !cwq
->nr_active
&& list_empty(&cwq
->delayed_works
);
2760 spin_unlock_irq(&cwq
->pool
->gcwq
->lock
);
2765 if (++flush_cnt
== 10 ||
2766 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2767 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2768 wq
->name
, flush_cnt
);
2772 spin_lock(&workqueue_lock
);
2773 if (!--wq
->nr_drainers
)
2774 wq
->flags
&= ~WQ_DRAINING
;
2775 spin_unlock(&workqueue_lock
);
2777 EXPORT_SYMBOL_GPL(drain_workqueue
);
2779 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2781 struct worker
*worker
= NULL
;
2782 struct global_cwq
*gcwq
;
2783 struct cpu_workqueue_struct
*cwq
;
2786 gcwq
= get_work_gcwq(work
);
2790 spin_lock_irq(&gcwq
->lock
);
2791 if (!list_empty(&work
->entry
)) {
2793 * See the comment near try_to_grab_pending()->smp_rmb().
2794 * If it was re-queued to a different gcwq under us, we
2795 * are not going to wait.
2798 cwq
= get_work_cwq(work
);
2799 if (unlikely(!cwq
|| gcwq
!= cwq
->pool
->gcwq
))
2802 worker
= find_worker_executing_work(gcwq
, work
);
2805 cwq
= worker
->current_cwq
;
2808 insert_wq_barrier(cwq
, barr
, work
, worker
);
2809 spin_unlock_irq(&gcwq
->lock
);
2812 * If @max_active is 1 or rescuer is in use, flushing another work
2813 * item on the same workqueue may lead to deadlock. Make sure the
2814 * flusher is not running on the same workqueue by verifying write
2817 if (cwq
->wq
->saved_max_active
== 1 || cwq
->wq
->flags
& WQ_RESCUER
)
2818 lock_map_acquire(&cwq
->wq
->lockdep_map
);
2820 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2821 lock_map_release(&cwq
->wq
->lockdep_map
);
2825 spin_unlock_irq(&gcwq
->lock
);
2830 * flush_work - wait for a work to finish executing the last queueing instance
2831 * @work: the work to flush
2833 * Wait until @work has finished execution. @work is guaranteed to be idle
2834 * on return if it hasn't been requeued since flush started.
2837 * %true if flush_work() waited for the work to finish execution,
2838 * %false if it was already idle.
2840 bool flush_work(struct work_struct
*work
)
2842 struct wq_barrier barr
;
2844 lock_map_acquire(&work
->lockdep_map
);
2845 lock_map_release(&work
->lockdep_map
);
2847 if (start_flush_work(work
, &barr
)) {
2848 wait_for_completion(&barr
.done
);
2849 destroy_work_on_stack(&barr
.work
);
2855 EXPORT_SYMBOL_GPL(flush_work
);
2857 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2859 unsigned long flags
;
2863 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2865 * If someone else is canceling, wait for the same event it
2866 * would be waiting for before retrying.
2868 if (unlikely(ret
== -ENOENT
))
2870 } while (unlikely(ret
< 0));
2872 /* tell other tasks trying to grab @work to back off */
2873 mark_work_canceling(work
);
2874 local_irq_restore(flags
);
2877 clear_work_data(work
);
2882 * cancel_work_sync - cancel a work and wait for it to finish
2883 * @work: the work to cancel
2885 * Cancel @work and wait for its execution to finish. This function
2886 * can be used even if the work re-queues itself or migrates to
2887 * another workqueue. On return from this function, @work is
2888 * guaranteed to be not pending or executing on any CPU.
2890 * cancel_work_sync(&delayed_work->work) must not be used for
2891 * delayed_work's. Use cancel_delayed_work_sync() instead.
2893 * The caller must ensure that the workqueue on which @work was last
2894 * queued can't be destroyed before this function returns.
2897 * %true if @work was pending, %false otherwise.
2899 bool cancel_work_sync(struct work_struct
*work
)
2901 return __cancel_work_timer(work
, false);
2903 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2906 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2907 * @dwork: the delayed work to flush
2909 * Delayed timer is cancelled and the pending work is queued for
2910 * immediate execution. Like flush_work(), this function only
2911 * considers the last queueing instance of @dwork.
2914 * %true if flush_work() waited for the work to finish execution,
2915 * %false if it was already idle.
2917 bool flush_delayed_work(struct delayed_work
*dwork
)
2919 local_irq_disable();
2920 if (del_timer_sync(&dwork
->timer
))
2921 __queue_work(dwork
->cpu
,
2922 get_work_cwq(&dwork
->work
)->wq
, &dwork
->work
);
2924 return flush_work(&dwork
->work
);
2926 EXPORT_SYMBOL(flush_delayed_work
);
2929 * cancel_delayed_work - cancel a delayed work
2930 * @dwork: delayed_work to cancel
2932 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2933 * and canceled; %false if wasn't pending. Note that the work callback
2934 * function may still be running on return, unless it returns %true and the
2935 * work doesn't re-arm itself. Explicitly flush or use
2936 * cancel_delayed_work_sync() to wait on it.
2938 * This function is safe to call from any context including IRQ handler.
2940 bool cancel_delayed_work(struct delayed_work
*dwork
)
2942 unsigned long flags
;
2946 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2947 } while (unlikely(ret
== -EAGAIN
));
2949 if (unlikely(ret
< 0))
2952 set_work_cpu_and_clear_pending(&dwork
->work
, work_cpu(&dwork
->work
));
2953 local_irq_restore(flags
);
2956 EXPORT_SYMBOL(cancel_delayed_work
);
2959 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2960 * @dwork: the delayed work cancel
2962 * This is cancel_work_sync() for delayed works.
2965 * %true if @dwork was pending, %false otherwise.
2967 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2969 return __cancel_work_timer(&dwork
->work
, true);
2971 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2974 * schedule_work_on - put work task on a specific cpu
2975 * @cpu: cpu to put the work task on
2976 * @work: job to be done
2978 * This puts a job on a specific cpu
2980 bool schedule_work_on(int cpu
, struct work_struct
*work
)
2982 return queue_work_on(cpu
, system_wq
, work
);
2984 EXPORT_SYMBOL(schedule_work_on
);
2987 * schedule_work - put work task in global workqueue
2988 * @work: job to be done
2990 * Returns %false if @work was already on the kernel-global workqueue and
2993 * This puts a job in the kernel-global workqueue if it was not already
2994 * queued and leaves it in the same position on the kernel-global
2995 * workqueue otherwise.
2997 bool schedule_work(struct work_struct
*work
)
2999 return queue_work(system_wq
, work
);
3001 EXPORT_SYMBOL(schedule_work
);
3004 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3006 * @dwork: job to be done
3007 * @delay: number of jiffies to wait
3009 * After waiting for a given time this puts a job in the kernel-global
3010 * workqueue on the specified CPU.
3012 bool schedule_delayed_work_on(int cpu
, struct delayed_work
*dwork
,
3013 unsigned long delay
)
3015 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
3017 EXPORT_SYMBOL(schedule_delayed_work_on
);
3020 * schedule_delayed_work - put work task in global workqueue after delay
3021 * @dwork: job to be done
3022 * @delay: number of jiffies to wait or 0 for immediate execution
3024 * After waiting for a given time this puts a job in the kernel-global
3027 bool schedule_delayed_work(struct delayed_work
*dwork
, unsigned long delay
)
3029 return queue_delayed_work(system_wq
, dwork
, delay
);
3031 EXPORT_SYMBOL(schedule_delayed_work
);
3034 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3035 * @func: the function to call
3037 * schedule_on_each_cpu() executes @func on each online CPU using the
3038 * system workqueue and blocks until all CPUs have completed.
3039 * schedule_on_each_cpu() is very slow.
3042 * 0 on success, -errno on failure.
3044 int schedule_on_each_cpu(work_func_t func
)
3047 struct work_struct __percpu
*works
;
3049 works
= alloc_percpu(struct work_struct
);
3055 for_each_online_cpu(cpu
) {
3056 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3058 INIT_WORK(work
, func
);
3059 schedule_work_on(cpu
, work
);
3062 for_each_online_cpu(cpu
)
3063 flush_work(per_cpu_ptr(works
, cpu
));
3071 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3073 * Forces execution of the kernel-global workqueue and blocks until its
3076 * Think twice before calling this function! It's very easy to get into
3077 * trouble if you don't take great care. Either of the following situations
3078 * will lead to deadlock:
3080 * One of the work items currently on the workqueue needs to acquire
3081 * a lock held by your code or its caller.
3083 * Your code is running in the context of a work routine.
3085 * They will be detected by lockdep when they occur, but the first might not
3086 * occur very often. It depends on what work items are on the workqueue and
3087 * what locks they need, which you have no control over.
3089 * In most situations flushing the entire workqueue is overkill; you merely
3090 * need to know that a particular work item isn't queued and isn't running.
3091 * In such cases you should use cancel_delayed_work_sync() or
3092 * cancel_work_sync() instead.
3094 void flush_scheduled_work(void)
3096 flush_workqueue(system_wq
);
3098 EXPORT_SYMBOL(flush_scheduled_work
);
3101 * execute_in_process_context - reliably execute the routine with user context
3102 * @fn: the function to execute
3103 * @ew: guaranteed storage for the execute work structure (must
3104 * be available when the work executes)
3106 * Executes the function immediately if process context is available,
3107 * otherwise schedules the function for delayed execution.
3109 * Returns: 0 - function was executed
3110 * 1 - function was scheduled for execution
3112 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3114 if (!in_interrupt()) {
3119 INIT_WORK(&ew
->work
, fn
);
3120 schedule_work(&ew
->work
);
3124 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3126 int keventd_up(void)
3128 return system_wq
!= NULL
;
3131 static int alloc_cwqs(struct workqueue_struct
*wq
)
3134 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3135 * Make sure that the alignment isn't lower than that of
3136 * unsigned long long.
3138 const size_t size
= sizeof(struct cpu_workqueue_struct
);
3139 const size_t align
= max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS
,
3140 __alignof__(unsigned long long));
3142 if (!(wq
->flags
& WQ_UNBOUND
))
3143 wq
->cpu_wq
.pcpu
= __alloc_percpu(size
, align
);
3148 * Allocate enough room to align cwq and put an extra
3149 * pointer at the end pointing back to the originally
3150 * allocated pointer which will be used for free.
3152 ptr
= kzalloc(size
+ align
+ sizeof(void *), GFP_KERNEL
);
3154 wq
->cpu_wq
.single
= PTR_ALIGN(ptr
, align
);
3155 *(void **)(wq
->cpu_wq
.single
+ 1) = ptr
;
3159 /* just in case, make sure it's actually aligned */
3160 BUG_ON(!IS_ALIGNED(wq
->cpu_wq
.v
, align
));
3161 return wq
->cpu_wq
.v
? 0 : -ENOMEM
;
3164 static void free_cwqs(struct workqueue_struct
*wq
)
3166 if (!(wq
->flags
& WQ_UNBOUND
))
3167 free_percpu(wq
->cpu_wq
.pcpu
);
3168 else if (wq
->cpu_wq
.single
) {
3169 /* the pointer to free is stored right after the cwq */
3170 kfree(*(void **)(wq
->cpu_wq
.single
+ 1));
3174 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3177 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3179 if (max_active
< 1 || max_active
> lim
)
3180 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3181 max_active
, name
, 1, lim
);
3183 return clamp_val(max_active
, 1, lim
);
3186 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3189 struct lock_class_key
*key
,
3190 const char *lock_name
, ...)
3192 va_list args
, args1
;
3193 struct workqueue_struct
*wq
;
3197 /* determine namelen, allocate wq and format name */
3198 va_start(args
, lock_name
);
3199 va_copy(args1
, args
);
3200 namelen
= vsnprintf(NULL
, 0, fmt
, args
) + 1;
3202 wq
= kzalloc(sizeof(*wq
) + namelen
, GFP_KERNEL
);
3206 vsnprintf(wq
->name
, namelen
, fmt
, args1
);
3211 * Workqueues which may be used during memory reclaim should
3212 * have a rescuer to guarantee forward progress.
3214 if (flags
& WQ_MEM_RECLAIM
)
3215 flags
|= WQ_RESCUER
;
3217 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3218 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3222 wq
->saved_max_active
= max_active
;
3223 mutex_init(&wq
->flush_mutex
);
3224 atomic_set(&wq
->nr_cwqs_to_flush
, 0);
3225 INIT_LIST_HEAD(&wq
->flusher_queue
);
3226 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3228 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3229 INIT_LIST_HEAD(&wq
->list
);
3231 if (alloc_cwqs(wq
) < 0)
3234 for_each_cwq_cpu(cpu
, wq
) {
3235 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3236 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3237 int pool_idx
= (bool)(flags
& WQ_HIGHPRI
);
3239 BUG_ON((unsigned long)cwq
& WORK_STRUCT_FLAG_MASK
);
3240 cwq
->pool
= &gcwq
->pools
[pool_idx
];
3242 cwq
->flush_color
= -1;
3243 cwq
->max_active
= max_active
;
3244 INIT_LIST_HEAD(&cwq
->delayed_works
);
3247 if (flags
& WQ_RESCUER
) {
3248 struct worker
*rescuer
;
3250 if (!alloc_mayday_mask(&wq
->mayday_mask
, GFP_KERNEL
))
3253 wq
->rescuer
= rescuer
= alloc_worker();
3257 rescuer
->rescue_wq
= wq
;
3258 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
3260 if (IS_ERR(rescuer
->task
))
3263 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
3264 wake_up_process(rescuer
->task
);
3268 * workqueue_lock protects global freeze state and workqueues
3269 * list. Grab it, set max_active accordingly and add the new
3270 * workqueue to workqueues list.
3272 spin_lock(&workqueue_lock
);
3274 if (workqueue_freezing
&& wq
->flags
& WQ_FREEZABLE
)
3275 for_each_cwq_cpu(cpu
, wq
)
3276 get_cwq(cpu
, wq
)->max_active
= 0;
3278 list_add(&wq
->list
, &workqueues
);
3280 spin_unlock(&workqueue_lock
);
3286 free_mayday_mask(wq
->mayday_mask
);
3292 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3295 * destroy_workqueue - safely terminate a workqueue
3296 * @wq: target workqueue
3298 * Safely destroy a workqueue. All work currently pending will be done first.
3300 void destroy_workqueue(struct workqueue_struct
*wq
)
3304 /* drain it before proceeding with destruction */
3305 drain_workqueue(wq
);
3308 * wq list is used to freeze wq, remove from list after
3309 * flushing is complete in case freeze races us.
3311 spin_lock(&workqueue_lock
);
3312 list_del(&wq
->list
);
3313 spin_unlock(&workqueue_lock
);
3316 for_each_cwq_cpu(cpu
, wq
) {
3317 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3320 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
3321 BUG_ON(cwq
->nr_in_flight
[i
]);
3322 BUG_ON(cwq
->nr_active
);
3323 BUG_ON(!list_empty(&cwq
->delayed_works
));
3326 if (wq
->flags
& WQ_RESCUER
) {
3327 kthread_stop(wq
->rescuer
->task
);
3328 free_mayday_mask(wq
->mayday_mask
);
3335 EXPORT_SYMBOL_GPL(destroy_workqueue
);
3338 * cwq_set_max_active - adjust max_active of a cwq
3339 * @cwq: target cpu_workqueue_struct
3340 * @max_active: new max_active value.
3342 * Set @cwq->max_active to @max_active and activate delayed works if
3346 * spin_lock_irq(gcwq->lock).
3348 static void cwq_set_max_active(struct cpu_workqueue_struct
*cwq
, int max_active
)
3350 cwq
->max_active
= max_active
;
3352 while (!list_empty(&cwq
->delayed_works
) &&
3353 cwq
->nr_active
< cwq
->max_active
)
3354 cwq_activate_first_delayed(cwq
);
3358 * workqueue_set_max_active - adjust max_active of a workqueue
3359 * @wq: target workqueue
3360 * @max_active: new max_active value.
3362 * Set max_active of @wq to @max_active.
3365 * Don't call from IRQ context.
3367 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
3371 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
3373 spin_lock(&workqueue_lock
);
3375 wq
->saved_max_active
= max_active
;
3377 for_each_cwq_cpu(cpu
, wq
) {
3378 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3379 struct worker_pool
*pool
= cwq
->pool
;
3380 struct global_cwq
*gcwq
= pool
->gcwq
;
3382 spin_lock_irq(&gcwq
->lock
);
3384 if (!(wq
->flags
& WQ_FREEZABLE
) ||
3385 !(pool
->flags
& POOL_FREEZING
))
3386 cwq_set_max_active(cwq
, max_active
);
3388 spin_unlock_irq(&gcwq
->lock
);
3391 spin_unlock(&workqueue_lock
);
3393 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
3396 * workqueue_congested - test whether a workqueue is congested
3397 * @cpu: CPU in question
3398 * @wq: target workqueue
3400 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3401 * no synchronization around this function and the test result is
3402 * unreliable and only useful as advisory hints or for debugging.
3405 * %true if congested, %false otherwise.
3407 bool workqueue_congested(unsigned int cpu
, struct workqueue_struct
*wq
)
3409 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3411 return !list_empty(&cwq
->delayed_works
);
3413 EXPORT_SYMBOL_GPL(workqueue_congested
);
3416 * work_cpu - return the last known associated cpu for @work
3417 * @work: the work of interest
3420 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3422 static unsigned int work_cpu(struct work_struct
*work
)
3424 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3426 return gcwq
? gcwq
->cpu
: WORK_CPU_NONE
;
3430 * work_busy - test whether a work is currently pending or running
3431 * @work: the work to be tested
3433 * Test whether @work is currently pending or running. There is no
3434 * synchronization around this function and the test result is
3435 * unreliable and only useful as advisory hints or for debugging.
3436 * Especially for reentrant wqs, the pending state might hide the
3440 * OR'd bitmask of WORK_BUSY_* bits.
3442 unsigned int work_busy(struct work_struct
*work
)
3444 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3445 unsigned long flags
;
3446 unsigned int ret
= 0;
3451 spin_lock_irqsave(&gcwq
->lock
, flags
);
3453 if (work_pending(work
))
3454 ret
|= WORK_BUSY_PENDING
;
3455 if (find_worker_executing_work(gcwq
, work
))
3456 ret
|= WORK_BUSY_RUNNING
;
3458 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
3462 EXPORT_SYMBOL_GPL(work_busy
);
3467 * There are two challenges in supporting CPU hotplug. Firstly, there
3468 * are a lot of assumptions on strong associations among work, cwq and
3469 * gcwq which make migrating pending and scheduled works very
3470 * difficult to implement without impacting hot paths. Secondly,
3471 * gcwqs serve mix of short, long and very long running works making
3472 * blocked draining impractical.
3474 * This is solved by allowing the pools to be disassociated from the CPU
3475 * running as an unbound one and allowing it to be reattached later if the
3476 * cpu comes back online.
3479 /* claim manager positions of all pools */
3480 static void gcwq_claim_assoc_and_lock(struct global_cwq
*gcwq
)
3482 struct worker_pool
*pool
;
3484 for_each_worker_pool(pool
, gcwq
)
3485 mutex_lock_nested(&pool
->assoc_mutex
, pool
- gcwq
->pools
);
3486 spin_lock_irq(&gcwq
->lock
);
3489 /* release manager positions */
3490 static void gcwq_release_assoc_and_unlock(struct global_cwq
*gcwq
)
3492 struct worker_pool
*pool
;
3494 spin_unlock_irq(&gcwq
->lock
);
3495 for_each_worker_pool(pool
, gcwq
)
3496 mutex_unlock(&pool
->assoc_mutex
);
3499 static void gcwq_unbind_fn(struct work_struct
*work
)
3501 struct global_cwq
*gcwq
= get_gcwq(smp_processor_id());
3502 struct worker_pool
*pool
;
3503 struct worker
*worker
;
3504 struct hlist_node
*pos
;
3507 BUG_ON(gcwq
->cpu
!= smp_processor_id());
3509 gcwq_claim_assoc_and_lock(gcwq
);
3512 * We've claimed all manager positions. Make all workers unbound
3513 * and set DISASSOCIATED. Before this, all workers except for the
3514 * ones which are still executing works from before the last CPU
3515 * down must be on the cpu. After this, they may become diasporas.
3517 for_each_worker_pool(pool
, gcwq
)
3518 list_for_each_entry(worker
, &pool
->idle_list
, entry
)
3519 worker
->flags
|= WORKER_UNBOUND
;
3521 for_each_busy_worker(worker
, i
, pos
, gcwq
)
3522 worker
->flags
|= WORKER_UNBOUND
;
3524 for_each_worker_pool(pool
, gcwq
)
3525 pool
->flags
|= POOL_DISASSOCIATED
;
3527 gcwq_release_assoc_and_unlock(gcwq
);
3530 * Call schedule() so that we cross rq->lock and thus can guarantee
3531 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3532 * as scheduler callbacks may be invoked from other cpus.
3537 * Sched callbacks are disabled now. Zap nr_running. After this,
3538 * nr_running stays zero and need_more_worker() and keep_working()
3539 * are always true as long as the worklist is not empty. @gcwq now
3540 * behaves as unbound (in terms of concurrency management) gcwq
3541 * which is served by workers tied to the CPU.
3543 * On return from this function, the current worker would trigger
3544 * unbound chain execution of pending work items if other workers
3547 for_each_worker_pool(pool
, gcwq
)
3548 atomic_set(get_pool_nr_running(pool
), 0);
3552 * Workqueues should be brought up before normal priority CPU notifiers.
3553 * This will be registered high priority CPU notifier.
3555 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
3556 unsigned long action
,
3559 unsigned int cpu
= (unsigned long)hcpu
;
3560 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3561 struct worker_pool
*pool
;
3563 switch (action
& ~CPU_TASKS_FROZEN
) {
3564 case CPU_UP_PREPARE
:
3565 for_each_worker_pool(pool
, gcwq
) {
3566 struct worker
*worker
;
3568 if (pool
->nr_workers
)
3571 worker
= create_worker(pool
);
3575 spin_lock_irq(&gcwq
->lock
);
3576 start_worker(worker
);
3577 spin_unlock_irq(&gcwq
->lock
);
3581 case CPU_DOWN_FAILED
:
3583 gcwq_claim_assoc_and_lock(gcwq
);
3584 for_each_worker_pool(pool
, gcwq
)
3585 pool
->flags
&= ~POOL_DISASSOCIATED
;
3586 rebind_workers(gcwq
);
3587 gcwq_release_assoc_and_unlock(gcwq
);
3594 * Workqueues should be brought down after normal priority CPU notifiers.
3595 * This will be registered as low priority CPU notifier.
3597 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
3598 unsigned long action
,
3601 unsigned int cpu
= (unsigned long)hcpu
;
3602 struct work_struct unbind_work
;
3604 switch (action
& ~CPU_TASKS_FROZEN
) {
3605 case CPU_DOWN_PREPARE
:
3606 /* unbinding should happen on the local CPU */
3607 INIT_WORK_ONSTACK(&unbind_work
, gcwq_unbind_fn
);
3608 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
3609 flush_work(&unbind_work
);
3617 struct work_for_cpu
{
3618 struct work_struct work
;
3624 static void work_for_cpu_fn(struct work_struct
*work
)
3626 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
3628 wfc
->ret
= wfc
->fn(wfc
->arg
);
3632 * work_on_cpu - run a function in user context on a particular cpu
3633 * @cpu: the cpu to run on
3634 * @fn: the function to run
3635 * @arg: the function arg
3637 * This will return the value @fn returns.
3638 * It is up to the caller to ensure that the cpu doesn't go offline.
3639 * The caller must not hold any locks which would prevent @fn from completing.
3641 long work_on_cpu(unsigned int cpu
, long (*fn
)(void *), void *arg
)
3643 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
3645 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
3646 schedule_work_on(cpu
, &wfc
.work
);
3647 flush_work(&wfc
.work
);
3650 EXPORT_SYMBOL_GPL(work_on_cpu
);
3651 #endif /* CONFIG_SMP */
3653 #ifdef CONFIG_FREEZER
3656 * freeze_workqueues_begin - begin freezing workqueues
3658 * Start freezing workqueues. After this function returns, all freezable
3659 * workqueues will queue new works to their frozen_works list instead of
3663 * Grabs and releases workqueue_lock and gcwq->lock's.
3665 void freeze_workqueues_begin(void)
3669 spin_lock(&workqueue_lock
);
3671 BUG_ON(workqueue_freezing
);
3672 workqueue_freezing
= true;
3674 for_each_gcwq_cpu(cpu
) {
3675 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3676 struct worker_pool
*pool
;
3677 struct workqueue_struct
*wq
;
3679 spin_lock_irq(&gcwq
->lock
);
3681 for_each_worker_pool(pool
, gcwq
) {
3682 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
3683 pool
->flags
|= POOL_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 for_each_worker_pool(pool
, gcwq
) {
3771 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
3772 pool
->flags
&= ~POOL_FREEZING
;
3775 list_for_each_entry(wq
, &workqueues
, list
) {
3776 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3778 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3781 /* restore max_active and repopulate worklist */
3782 cwq_set_max_active(cwq
, wq
->saved_max_active
);
3785 for_each_worker_pool(pool
, gcwq
)
3786 wake_up_worker(pool
);
3788 spin_unlock_irq(&gcwq
->lock
);
3791 workqueue_freezing
= false;
3793 spin_unlock(&workqueue_lock
);
3795 #endif /* CONFIG_FREEZER */
3797 static int __init
init_workqueues(void)
3801 /* make sure we have enough bits for OFFQ CPU number */
3802 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_CPU_SHIFT
)) <
3805 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
3806 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
3808 /* initialize gcwqs */
3809 for_each_gcwq_cpu(cpu
) {
3810 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3811 struct worker_pool
*pool
;
3813 spin_lock_init(&gcwq
->lock
);
3816 hash_init(gcwq
->busy_hash
);
3818 for_each_worker_pool(pool
, gcwq
) {
3820 pool
->flags
|= POOL_DISASSOCIATED
;
3821 INIT_LIST_HEAD(&pool
->worklist
);
3822 INIT_LIST_HEAD(&pool
->idle_list
);
3824 init_timer_deferrable(&pool
->idle_timer
);
3825 pool
->idle_timer
.function
= idle_worker_timeout
;
3826 pool
->idle_timer
.data
= (unsigned long)pool
;
3828 setup_timer(&pool
->mayday_timer
, gcwq_mayday_timeout
,
3829 (unsigned long)pool
);
3831 mutex_init(&pool
->assoc_mutex
);
3832 ida_init(&pool
->worker_ida
);
3836 /* create the initial worker */
3837 for_each_online_gcwq_cpu(cpu
) {
3838 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3839 struct worker_pool
*pool
;
3841 for_each_worker_pool(pool
, gcwq
) {
3842 struct worker
*worker
;
3844 if (cpu
!= WORK_CPU_UNBOUND
)
3845 pool
->flags
&= ~POOL_DISASSOCIATED
;
3847 worker
= create_worker(pool
);
3849 spin_lock_irq(&gcwq
->lock
);
3850 start_worker(worker
);
3851 spin_unlock_irq(&gcwq
->lock
);
3855 system_wq
= alloc_workqueue("events", 0, 0);
3856 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
3857 system_long_wq
= alloc_workqueue("events_long", 0, 0);
3858 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
3859 WQ_UNBOUND_MAX_ACTIVE
);
3860 system_freezable_wq
= alloc_workqueue("events_freezable",
3862 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
3863 !system_unbound_wq
|| !system_freezable_wq
);
3866 early_initcall(init_workqueues
);