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 gcwq 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 gcwq behaves as an unbound one.
61 * Note that DISASSOCIATED can be flipped only while holding
62 * assoc_mutex of all pools on the gcwq to avoid changing binding
63 * state while create_worker() is in progress.
65 GCWQ_DISASSOCIATED
= 1 << 0, /* cpu can't serve workers */
66 GCWQ_FREEZING
= 1 << 1, /* freeze in progress */
69 POOL_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
70 POOL_MANAGING_WORKERS
= 1 << 1, /* managing workers */
73 WORKER_STARTED
= 1 << 0, /* started */
74 WORKER_DIE
= 1 << 1, /* die die die */
75 WORKER_IDLE
= 1 << 2, /* is idle */
76 WORKER_PREP
= 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
80 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_UNBOUND
|
83 NR_WORKER_POOLS
= 2, /* # worker pools per gcwq */
85 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
87 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
88 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
90 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
91 /* call for help after 10ms
93 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
94 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
97 * Rescue workers are used only on emergencies and shared by
100 RESCUER_NICE_LEVEL
= -20,
101 HIGHPRI_NICE_LEVEL
= -20,
105 * Structure fields follow one of the following exclusion rules.
107 * I: Modifiable by initialization/destruction paths and read-only for
110 * P: Preemption protected. Disabling preemption is enough and should
111 * only be modified and accessed from the local cpu.
113 * L: gcwq->lock protected. Access with gcwq->lock held.
115 * X: During normal operation, modification requires gcwq->lock and
116 * should be done only from local cpu. Either disabling preemption
117 * on local cpu or grabbing gcwq->lock is enough for read access.
118 * If GCWQ_DISASSOCIATED is set, it's identical to L.
120 * F: wq->flush_mutex protected.
122 * W: workqueue_lock protected.
125 /* struct worker is defined in workqueue_internal.h */
128 struct global_cwq
*gcwq
; /* I: the owning gcwq */
129 unsigned int flags
; /* X: flags */
131 struct list_head worklist
; /* L: list of pending works */
132 int nr_workers
; /* L: total number of workers */
134 /* nr_idle includes the ones off idle_list for rebinding */
135 int nr_idle
; /* L: currently idle ones */
137 struct list_head idle_list
; /* X: list of idle workers */
138 struct timer_list idle_timer
; /* L: worker idle timeout */
139 struct timer_list mayday_timer
; /* L: SOS timer for workers */
141 struct mutex assoc_mutex
; /* protect GCWQ_DISASSOCIATED */
142 struct ida worker_ida
; /* L: for worker IDs */
146 * Global per-cpu workqueue. There's one and only one for each cpu
147 * and all works are queued and processed here regardless of their
151 spinlock_t lock
; /* the gcwq lock */
152 unsigned int cpu
; /* I: the associated cpu */
153 unsigned int flags
; /* L: GCWQ_* flags */
155 /* workers are chained either in busy_hash or pool idle_list */
156 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
157 /* L: hash of busy workers */
159 struct worker_pool pools
[NR_WORKER_POOLS
];
160 /* normal and highpri pools */
161 } ____cacheline_aligned_in_smp
;
164 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
165 * work_struct->data are used for flags and thus cwqs need to be
166 * aligned at two's power of the number of flag bits.
168 struct cpu_workqueue_struct
{
169 struct worker_pool
*pool
; /* I: the associated pool */
170 struct workqueue_struct
*wq
; /* I: the owning workqueue */
171 int work_color
; /* L: current color */
172 int flush_color
; /* L: flushing color */
173 int nr_in_flight
[WORK_NR_COLORS
];
174 /* L: nr of in_flight works */
175 int nr_active
; /* L: nr of active works */
176 int max_active
; /* L: max active works */
177 struct list_head delayed_works
; /* L: delayed works */
181 * Structure used to wait for workqueue flush.
184 struct list_head list
; /* F: list of flushers */
185 int flush_color
; /* F: flush color waiting for */
186 struct completion done
; /* flush completion */
190 * All cpumasks are assumed to be always set on UP and thus can't be
191 * used to determine whether there's something to be done.
194 typedef cpumask_var_t mayday_mask_t
;
195 #define mayday_test_and_set_cpu(cpu, mask) \
196 cpumask_test_and_set_cpu((cpu), (mask))
197 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
198 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
199 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
200 #define free_mayday_mask(mask) free_cpumask_var((mask))
202 typedef unsigned long mayday_mask_t
;
203 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
204 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
205 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
206 #define alloc_mayday_mask(maskp, gfp) true
207 #define free_mayday_mask(mask) do { } while (0)
211 * The externally visible workqueue abstraction is an array of
212 * per-CPU workqueues:
214 struct workqueue_struct
{
215 unsigned int flags
; /* W: WQ_* flags */
217 struct cpu_workqueue_struct __percpu
*pcpu
;
218 struct cpu_workqueue_struct
*single
;
220 } cpu_wq
; /* I: cwq's */
221 struct list_head list
; /* W: list of all workqueues */
223 struct mutex flush_mutex
; /* protects wq flushing */
224 int work_color
; /* F: current work color */
225 int flush_color
; /* F: current flush color */
226 atomic_t nr_cwqs_to_flush
; /* flush in progress */
227 struct wq_flusher
*first_flusher
; /* F: first flusher */
228 struct list_head flusher_queue
; /* F: flush waiters */
229 struct list_head flusher_overflow
; /* F: flush overflow list */
231 mayday_mask_t mayday_mask
; /* cpus requesting rescue */
232 struct worker
*rescuer
; /* I: rescue worker */
234 int nr_drainers
; /* W: drain in progress */
235 int saved_max_active
; /* W: saved cwq max_active */
236 #ifdef CONFIG_LOCKDEP
237 struct lockdep_map lockdep_map
;
239 char name
[]; /* I: workqueue name */
242 struct workqueue_struct
*system_wq __read_mostly
;
243 EXPORT_SYMBOL_GPL(system_wq
);
244 struct workqueue_struct
*system_highpri_wq __read_mostly
;
245 EXPORT_SYMBOL_GPL(system_highpri_wq
);
246 struct workqueue_struct
*system_long_wq __read_mostly
;
247 EXPORT_SYMBOL_GPL(system_long_wq
);
248 struct workqueue_struct
*system_unbound_wq __read_mostly
;
249 EXPORT_SYMBOL_GPL(system_unbound_wq
);
250 struct workqueue_struct
*system_freezable_wq __read_mostly
;
251 EXPORT_SYMBOL_GPL(system_freezable_wq
);
253 #define CREATE_TRACE_POINTS
254 #include <trace/events/workqueue.h>
256 #define for_each_worker_pool(pool, gcwq) \
257 for ((pool) = &(gcwq)->pools[0]; \
258 (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
260 #define for_each_busy_worker(worker, i, pos, gcwq) \
261 hash_for_each(gcwq->busy_hash, i, pos, worker, hentry)
263 static inline int __next_gcwq_cpu(int cpu
, const struct cpumask
*mask
,
266 if (cpu
< nr_cpu_ids
) {
268 cpu
= cpumask_next(cpu
, mask
);
269 if (cpu
< nr_cpu_ids
)
273 return WORK_CPU_UNBOUND
;
275 return WORK_CPU_NONE
;
278 static inline int __next_wq_cpu(int cpu
, const struct cpumask
*mask
,
279 struct workqueue_struct
*wq
)
281 return __next_gcwq_cpu(cpu
, mask
, !(wq
->flags
& WQ_UNBOUND
) ? 1 : 2);
287 * An extra gcwq is defined for an invalid cpu number
288 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
289 * specific CPU. The following iterators are similar to
290 * for_each_*_cpu() iterators but also considers the unbound gcwq.
292 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
293 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
294 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
295 * WORK_CPU_UNBOUND for unbound workqueues
297 #define for_each_gcwq_cpu(cpu) \
298 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
299 (cpu) < WORK_CPU_NONE; \
300 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
302 #define for_each_online_gcwq_cpu(cpu) \
303 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
304 (cpu) < WORK_CPU_NONE; \
305 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
307 #define for_each_cwq_cpu(cpu, wq) \
308 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
309 (cpu) < WORK_CPU_NONE; \
310 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
312 #ifdef CONFIG_DEBUG_OBJECTS_WORK
314 static struct debug_obj_descr work_debug_descr
;
316 static void *work_debug_hint(void *addr
)
318 return ((struct work_struct
*) addr
)->func
;
322 * fixup_init is called when:
323 * - an active object is initialized
325 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
327 struct work_struct
*work
= addr
;
330 case ODEBUG_STATE_ACTIVE
:
331 cancel_work_sync(work
);
332 debug_object_init(work
, &work_debug_descr
);
340 * fixup_activate is called when:
341 * - an active object is activated
342 * - an unknown object is activated (might be a statically initialized object)
344 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
346 struct work_struct
*work
= addr
;
350 case ODEBUG_STATE_NOTAVAILABLE
:
352 * This is not really a fixup. The work struct was
353 * statically initialized. We just make sure that it
354 * is tracked in the object tracker.
356 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
357 debug_object_init(work
, &work_debug_descr
);
358 debug_object_activate(work
, &work_debug_descr
);
364 case ODEBUG_STATE_ACTIVE
:
373 * fixup_free is called when:
374 * - an active object is freed
376 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
378 struct work_struct
*work
= addr
;
381 case ODEBUG_STATE_ACTIVE
:
382 cancel_work_sync(work
);
383 debug_object_free(work
, &work_debug_descr
);
390 static struct debug_obj_descr work_debug_descr
= {
391 .name
= "work_struct",
392 .debug_hint
= work_debug_hint
,
393 .fixup_init
= work_fixup_init
,
394 .fixup_activate
= work_fixup_activate
,
395 .fixup_free
= work_fixup_free
,
398 static inline void debug_work_activate(struct work_struct
*work
)
400 debug_object_activate(work
, &work_debug_descr
);
403 static inline void debug_work_deactivate(struct work_struct
*work
)
405 debug_object_deactivate(work
, &work_debug_descr
);
408 void __init_work(struct work_struct
*work
, int onstack
)
411 debug_object_init_on_stack(work
, &work_debug_descr
);
413 debug_object_init(work
, &work_debug_descr
);
415 EXPORT_SYMBOL_GPL(__init_work
);
417 void destroy_work_on_stack(struct work_struct
*work
)
419 debug_object_free(work
, &work_debug_descr
);
421 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
424 static inline void debug_work_activate(struct work_struct
*work
) { }
425 static inline void debug_work_deactivate(struct work_struct
*work
) { }
428 /* Serializes the accesses to the list of workqueues. */
429 static DEFINE_SPINLOCK(workqueue_lock
);
430 static LIST_HEAD(workqueues
);
431 static bool workqueue_freezing
; /* W: have wqs started freezing? */
434 * The almighty global cpu workqueues. nr_running is the only field
435 * which is expected to be used frequently by other cpus via
436 * try_to_wake_up(). Put it in a separate cacheline.
438 static DEFINE_PER_CPU(struct global_cwq
, global_cwq
);
439 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t
, pool_nr_running
[NR_WORKER_POOLS
]);
442 * Global cpu workqueue and nr_running counter for unbound gcwq. The
443 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
444 * workers have WORKER_UNBOUND set.
446 static struct global_cwq unbound_global_cwq
;
447 static atomic_t unbound_pool_nr_running
[NR_WORKER_POOLS
] = {
448 [0 ... NR_WORKER_POOLS
- 1] = ATOMIC_INIT(0), /* always 0 */
451 static int worker_thread(void *__worker
);
453 static int worker_pool_pri(struct worker_pool
*pool
)
455 return pool
- pool
->gcwq
->pools
;
458 static struct global_cwq
*get_gcwq(unsigned int cpu
)
460 if (cpu
!= WORK_CPU_UNBOUND
)
461 return &per_cpu(global_cwq
, cpu
);
463 return &unbound_global_cwq
;
466 static atomic_t
*get_pool_nr_running(struct worker_pool
*pool
)
468 int cpu
= pool
->gcwq
->cpu
;
469 int idx
= worker_pool_pri(pool
);
471 if (cpu
!= WORK_CPU_UNBOUND
)
472 return &per_cpu(pool_nr_running
, cpu
)[idx
];
474 return &unbound_pool_nr_running
[idx
];
477 static struct cpu_workqueue_struct
*get_cwq(unsigned int cpu
,
478 struct workqueue_struct
*wq
)
480 if (!(wq
->flags
& WQ_UNBOUND
)) {
481 if (likely(cpu
< nr_cpu_ids
))
482 return per_cpu_ptr(wq
->cpu_wq
.pcpu
, cpu
);
483 } else if (likely(cpu
== WORK_CPU_UNBOUND
))
484 return wq
->cpu_wq
.single
;
488 static unsigned int work_color_to_flags(int color
)
490 return color
<< WORK_STRUCT_COLOR_SHIFT
;
493 static int get_work_color(struct work_struct
*work
)
495 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
496 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
499 static int work_next_color(int color
)
501 return (color
+ 1) % WORK_NR_COLORS
;
505 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
506 * contain the pointer to the queued cwq. Once execution starts, the flag
507 * is cleared and the high bits contain OFFQ flags and CPU number.
509 * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling()
510 * and clear_work_data() can be used to set the cwq, cpu or clear
511 * work->data. These functions should only be called while the work is
512 * owned - ie. while the PENDING bit is set.
514 * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to
515 * a work. gcwq is available once the work has been queued anywhere after
516 * initialization until it is sync canceled. cwq is available only while
517 * the work item is queued.
519 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
520 * canceled. While being canceled, a work item may have its PENDING set
521 * but stay off timer and worklist for arbitrarily long and nobody should
522 * try to steal the PENDING bit.
524 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
527 BUG_ON(!work_pending(work
));
528 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
531 static void set_work_cwq(struct work_struct
*work
,
532 struct cpu_workqueue_struct
*cwq
,
533 unsigned long extra_flags
)
535 set_work_data(work
, (unsigned long)cwq
,
536 WORK_STRUCT_PENDING
| WORK_STRUCT_CWQ
| extra_flags
);
539 static void set_work_cpu_and_clear_pending(struct work_struct
*work
,
543 * The following wmb is paired with the implied mb in
544 * test_and_set_bit(PENDING) and ensures all updates to @work made
545 * here are visible to and precede any updates by the next PENDING
549 set_work_data(work
, (unsigned long)cpu
<< WORK_OFFQ_CPU_SHIFT
, 0);
552 static void clear_work_data(struct work_struct
*work
)
554 smp_wmb(); /* see set_work_cpu_and_clear_pending() */
555 set_work_data(work
, WORK_STRUCT_NO_CPU
, 0);
558 static struct cpu_workqueue_struct
*get_work_cwq(struct work_struct
*work
)
560 unsigned long data
= atomic_long_read(&work
->data
);
562 if (data
& WORK_STRUCT_CWQ
)
563 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
568 static struct global_cwq
*get_work_gcwq(struct work_struct
*work
)
570 unsigned long data
= atomic_long_read(&work
->data
);
573 if (data
& WORK_STRUCT_CWQ
)
574 return ((struct cpu_workqueue_struct
*)
575 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->gcwq
;
577 cpu
= data
>> WORK_OFFQ_CPU_SHIFT
;
578 if (cpu
== WORK_CPU_NONE
)
581 BUG_ON(cpu
>= nr_cpu_ids
&& cpu
!= WORK_CPU_UNBOUND
);
582 return get_gcwq(cpu
);
585 static void mark_work_canceling(struct work_struct
*work
)
587 struct global_cwq
*gcwq
= get_work_gcwq(work
);
588 unsigned long cpu
= gcwq
? gcwq
->cpu
: WORK_CPU_NONE
;
590 set_work_data(work
, (cpu
<< WORK_OFFQ_CPU_SHIFT
) | WORK_OFFQ_CANCELING
,
591 WORK_STRUCT_PENDING
);
594 static bool work_is_canceling(struct work_struct
*work
)
596 unsigned long data
= atomic_long_read(&work
->data
);
598 return !(data
& WORK_STRUCT_CWQ
) && (data
& WORK_OFFQ_CANCELING
);
602 * Policy functions. These define the policies on how the global worker
603 * pools are managed. Unless noted otherwise, these functions assume that
604 * they're being called with gcwq->lock held.
607 static bool __need_more_worker(struct worker_pool
*pool
)
609 return !atomic_read(get_pool_nr_running(pool
));
613 * Need to wake up a worker? Called from anything but currently
616 * Note that, because unbound workers never contribute to nr_running, this
617 * function will always return %true for unbound gcwq as long as the
618 * worklist isn't empty.
620 static bool need_more_worker(struct worker_pool
*pool
)
622 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
625 /* Can I start working? Called from busy but !running workers. */
626 static bool may_start_working(struct worker_pool
*pool
)
628 return pool
->nr_idle
;
631 /* Do I need to keep working? Called from currently running workers. */
632 static bool keep_working(struct worker_pool
*pool
)
634 atomic_t
*nr_running
= get_pool_nr_running(pool
);
636 return !list_empty(&pool
->worklist
) && atomic_read(nr_running
) <= 1;
639 /* Do we need a new worker? Called from manager. */
640 static bool need_to_create_worker(struct worker_pool
*pool
)
642 return need_more_worker(pool
) && !may_start_working(pool
);
645 /* Do I need to be the manager? */
646 static bool need_to_manage_workers(struct worker_pool
*pool
)
648 return need_to_create_worker(pool
) ||
649 (pool
->flags
& POOL_MANAGE_WORKERS
);
652 /* Do we have too many workers and should some go away? */
653 static bool too_many_workers(struct worker_pool
*pool
)
655 bool managing
= pool
->flags
& POOL_MANAGING_WORKERS
;
656 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
657 int nr_busy
= pool
->nr_workers
- nr_idle
;
660 * nr_idle and idle_list may disagree if idle rebinding is in
661 * progress. Never return %true if idle_list is empty.
663 if (list_empty(&pool
->idle_list
))
666 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
673 /* Return the first worker. Safe with preemption disabled */
674 static struct worker
*first_worker(struct worker_pool
*pool
)
676 if (unlikely(list_empty(&pool
->idle_list
)))
679 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
683 * wake_up_worker - wake up an idle worker
684 * @pool: worker pool to wake worker from
686 * Wake up the first idle worker of @pool.
689 * spin_lock_irq(gcwq->lock).
691 static void wake_up_worker(struct worker_pool
*pool
)
693 struct worker
*worker
= first_worker(pool
);
696 wake_up_process(worker
->task
);
700 * wq_worker_waking_up - a worker is waking up
701 * @task: task waking up
702 * @cpu: CPU @task is waking up to
704 * This function is called during try_to_wake_up() when a worker is
708 * spin_lock_irq(rq->lock)
710 void wq_worker_waking_up(struct task_struct
*task
, unsigned int cpu
)
712 struct worker
*worker
= kthread_data(task
);
714 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
715 WARN_ON_ONCE(worker
->pool
->gcwq
->cpu
!= cpu
);
716 atomic_inc(get_pool_nr_running(worker
->pool
));
721 * wq_worker_sleeping - a worker is going to sleep
722 * @task: task going to sleep
723 * @cpu: CPU in question, must be the current CPU number
725 * This function is called during schedule() when a busy worker is
726 * going to sleep. Worker on the same cpu can be woken up by
727 * returning pointer to its task.
730 * spin_lock_irq(rq->lock)
733 * Worker task on @cpu to wake up, %NULL if none.
735 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
,
738 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
739 struct worker_pool
*pool
;
740 atomic_t
*nr_running
;
743 * Rescuers, which may not have all the fields set up like normal
744 * workers, also reach here, let's not access anything before
745 * checking NOT_RUNNING.
747 if (worker
->flags
& WORKER_NOT_RUNNING
)
751 nr_running
= get_pool_nr_running(pool
);
753 /* this can only happen on the local cpu */
754 BUG_ON(cpu
!= raw_smp_processor_id());
757 * The counterpart of the following dec_and_test, implied mb,
758 * worklist not empty test sequence is in insert_work().
759 * Please read comment there.
761 * NOT_RUNNING is clear. This means that we're bound to and
762 * running on the local cpu w/ rq lock held and preemption
763 * disabled, which in turn means that none else could be
764 * manipulating idle_list, so dereferencing idle_list without gcwq
767 if (atomic_dec_and_test(nr_running
) && !list_empty(&pool
->worklist
))
768 to_wakeup
= first_worker(pool
);
769 return to_wakeup
? to_wakeup
->task
: NULL
;
773 * worker_set_flags - set worker flags and adjust nr_running accordingly
775 * @flags: flags to set
776 * @wakeup: wakeup an idle worker if necessary
778 * Set @flags in @worker->flags and adjust nr_running accordingly. If
779 * nr_running becomes zero and @wakeup is %true, an idle worker is
783 * spin_lock_irq(gcwq->lock)
785 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
788 struct worker_pool
*pool
= worker
->pool
;
790 WARN_ON_ONCE(worker
->task
!= current
);
793 * If transitioning into NOT_RUNNING, adjust nr_running and
794 * wake up an idle worker as necessary if requested by
797 if ((flags
& WORKER_NOT_RUNNING
) &&
798 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
799 atomic_t
*nr_running
= get_pool_nr_running(pool
);
802 if (atomic_dec_and_test(nr_running
) &&
803 !list_empty(&pool
->worklist
))
804 wake_up_worker(pool
);
806 atomic_dec(nr_running
);
809 worker
->flags
|= flags
;
813 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
815 * @flags: flags to clear
817 * Clear @flags in @worker->flags and adjust nr_running accordingly.
820 * spin_lock_irq(gcwq->lock)
822 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
824 struct worker_pool
*pool
= worker
->pool
;
825 unsigned int oflags
= worker
->flags
;
827 WARN_ON_ONCE(worker
->task
!= current
);
829 worker
->flags
&= ~flags
;
832 * If transitioning out of NOT_RUNNING, increment nr_running. Note
833 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
834 * of multiple flags, not a single flag.
836 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
837 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
838 atomic_inc(get_pool_nr_running(pool
));
842 * find_worker_executing_work - find worker which is executing a work
843 * @gcwq: gcwq of interest
844 * @work: work to find worker for
846 * Find a worker which is executing @work on @gcwq by searching
847 * @gcwq->busy_hash which is keyed by the address of @work. For a worker
848 * to match, its current execution should match the address of @work and
849 * its work function. This is to avoid unwanted dependency between
850 * unrelated work executions through a work item being recycled while still
853 * This is a bit tricky. A work item may be freed once its execution
854 * starts and nothing prevents the freed area from being recycled for
855 * another work item. If the same work item address ends up being reused
856 * before the original execution finishes, workqueue will identify the
857 * recycled work item as currently executing and make it wait until the
858 * current execution finishes, introducing an unwanted dependency.
860 * This function checks the work item address, work function and workqueue
861 * to avoid false positives. Note that this isn't complete as one may
862 * construct a work function which can introduce dependency onto itself
863 * through a recycled work item. Well, if somebody wants to shoot oneself
864 * in the foot that badly, there's only so much we can do, and if such
865 * deadlock actually occurs, it should be easy to locate the culprit work
869 * spin_lock_irq(gcwq->lock).
872 * Pointer to worker which is executing @work if found, NULL
875 static struct worker
*find_worker_executing_work(struct global_cwq
*gcwq
,
876 struct work_struct
*work
)
878 struct worker
*worker
;
879 struct hlist_node
*tmp
;
881 hash_for_each_possible(gcwq
->busy_hash
, worker
, tmp
, hentry
,
883 if (worker
->current_work
== work
&&
884 worker
->current_func
== work
->func
)
891 * move_linked_works - move linked works to a list
892 * @work: start of series of works to be scheduled
893 * @head: target list to append @work to
894 * @nextp: out paramter for nested worklist walking
896 * Schedule linked works starting from @work to @head. Work series to
897 * be scheduled starts at @work and includes any consecutive work with
898 * WORK_STRUCT_LINKED set in its predecessor.
900 * If @nextp is not NULL, it's updated to point to the next work of
901 * the last scheduled work. This allows move_linked_works() to be
902 * nested inside outer list_for_each_entry_safe().
905 * spin_lock_irq(gcwq->lock).
907 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
908 struct work_struct
**nextp
)
910 struct work_struct
*n
;
913 * Linked worklist will always end before the end of the list,
914 * use NULL for list head.
916 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
917 list_move_tail(&work
->entry
, head
);
918 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
923 * If we're already inside safe list traversal and have moved
924 * multiple works to the scheduled queue, the next position
925 * needs to be updated.
931 static void cwq_activate_delayed_work(struct work_struct
*work
)
933 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
935 trace_workqueue_activate_work(work
);
936 move_linked_works(work
, &cwq
->pool
->worklist
, NULL
);
937 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
941 static void cwq_activate_first_delayed(struct cpu_workqueue_struct
*cwq
)
943 struct work_struct
*work
= list_first_entry(&cwq
->delayed_works
,
944 struct work_struct
, entry
);
946 cwq_activate_delayed_work(work
);
950 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
951 * @cwq: cwq of interest
952 * @color: color of work which left the queue
954 * A work either has completed or is removed from pending queue,
955 * decrement nr_in_flight of its cwq and handle workqueue flushing.
958 * spin_lock_irq(gcwq->lock).
960 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct
*cwq
, int color
)
962 /* ignore uncolored works */
963 if (color
== WORK_NO_COLOR
)
966 cwq
->nr_in_flight
[color
]--;
969 if (!list_empty(&cwq
->delayed_works
)) {
970 /* one down, submit a delayed one */
971 if (cwq
->nr_active
< cwq
->max_active
)
972 cwq_activate_first_delayed(cwq
);
975 /* is flush in progress and are we at the flushing tip? */
976 if (likely(cwq
->flush_color
!= color
))
979 /* are there still in-flight works? */
980 if (cwq
->nr_in_flight
[color
])
983 /* this cwq is done, clear flush_color */
984 cwq
->flush_color
= -1;
987 * If this was the last cwq, wake up the first flusher. It
988 * will handle the rest.
990 if (atomic_dec_and_test(&cwq
->wq
->nr_cwqs_to_flush
))
991 complete(&cwq
->wq
->first_flusher
->done
);
995 * try_to_grab_pending - steal work item from worklist and disable irq
996 * @work: work item to steal
997 * @is_dwork: @work is a delayed_work
998 * @flags: place to store irq state
1000 * Try to grab PENDING bit of @work. This function can handle @work in any
1001 * stable state - idle, on timer or on worklist. Return values are
1003 * 1 if @work was pending and we successfully stole PENDING
1004 * 0 if @work was idle and we claimed PENDING
1005 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1006 * -ENOENT if someone else is canceling @work, this state may persist
1007 * for arbitrarily long
1009 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1010 * interrupted while holding PENDING and @work off queue, irq must be
1011 * disabled on entry. This, combined with delayed_work->timer being
1012 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1014 * On successful return, >= 0, irq is disabled and the caller is
1015 * responsible for releasing it using local_irq_restore(*@flags).
1017 * This function is safe to call from any context including IRQ handler.
1019 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1020 unsigned long *flags
)
1022 struct global_cwq
*gcwq
;
1024 local_irq_save(*flags
);
1026 /* try to steal the timer if it exists */
1028 struct delayed_work
*dwork
= to_delayed_work(work
);
1031 * dwork->timer is irqsafe. If del_timer() fails, it's
1032 * guaranteed that the timer is not queued anywhere and not
1033 * running on the local CPU.
1035 if (likely(del_timer(&dwork
->timer
)))
1039 /* try to claim PENDING the normal way */
1040 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1044 * The queueing is in progress, or it is already queued. Try to
1045 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1047 gcwq
= get_work_gcwq(work
);
1051 spin_lock(&gcwq
->lock
);
1052 if (!list_empty(&work
->entry
)) {
1054 * This work is queued, but perhaps we locked the wrong gcwq.
1055 * In that case we must see the new value after rmb(), see
1056 * insert_work()->wmb().
1059 if (gcwq
== get_work_gcwq(work
)) {
1060 debug_work_deactivate(work
);
1063 * A delayed work item cannot be grabbed directly
1064 * because it might have linked NO_COLOR work items
1065 * which, if left on the delayed_list, will confuse
1066 * cwq->nr_active management later on and cause
1067 * stall. Make sure the work item is activated
1070 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1071 cwq_activate_delayed_work(work
);
1073 list_del_init(&work
->entry
);
1074 cwq_dec_nr_in_flight(get_work_cwq(work
),
1075 get_work_color(work
));
1077 spin_unlock(&gcwq
->lock
);
1081 spin_unlock(&gcwq
->lock
);
1083 local_irq_restore(*flags
);
1084 if (work_is_canceling(work
))
1091 * insert_work - insert a work into gcwq
1092 * @cwq: cwq @work belongs to
1093 * @work: work to insert
1094 * @head: insertion point
1095 * @extra_flags: extra WORK_STRUCT_* flags to set
1097 * Insert @work which belongs to @cwq into @gcwq after @head.
1098 * @extra_flags is or'd to work_struct flags.
1101 * spin_lock_irq(gcwq->lock).
1103 static void insert_work(struct cpu_workqueue_struct
*cwq
,
1104 struct work_struct
*work
, struct list_head
*head
,
1105 unsigned int extra_flags
)
1107 struct worker_pool
*pool
= cwq
->pool
;
1109 /* we own @work, set data and link */
1110 set_work_cwq(work
, cwq
, extra_flags
);
1113 * Ensure that we get the right work->data if we see the
1114 * result of list_add() below, see try_to_grab_pending().
1118 list_add_tail(&work
->entry
, head
);
1121 * Ensure either worker_sched_deactivated() sees the above
1122 * list_add_tail() or we see zero nr_running to avoid workers
1123 * lying around lazily while there are works to be processed.
1127 if (__need_more_worker(pool
))
1128 wake_up_worker(pool
);
1132 * Test whether @work is being queued from another work executing on the
1133 * same workqueue. This is rather expensive and should only be used from
1136 static bool is_chained_work(struct workqueue_struct
*wq
)
1138 unsigned long flags
;
1141 for_each_gcwq_cpu(cpu
) {
1142 struct global_cwq
*gcwq
= get_gcwq(cpu
);
1143 struct worker
*worker
;
1144 struct hlist_node
*pos
;
1147 spin_lock_irqsave(&gcwq
->lock
, flags
);
1148 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
1149 if (worker
->task
!= current
)
1151 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1153 * I'm @worker, no locking necessary. See if @work
1154 * is headed to the same workqueue.
1156 return worker
->current_cwq
->wq
== wq
;
1158 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1163 static void __queue_work(unsigned int cpu
, struct workqueue_struct
*wq
,
1164 struct work_struct
*work
)
1166 struct global_cwq
*gcwq
;
1167 struct cpu_workqueue_struct
*cwq
;
1168 struct list_head
*worklist
;
1169 unsigned int work_flags
;
1170 unsigned int req_cpu
= cpu
;
1173 * While a work item is PENDING && off queue, a task trying to
1174 * steal the PENDING will busy-loop waiting for it to either get
1175 * queued or lose PENDING. Grabbing PENDING and queueing should
1176 * happen with IRQ disabled.
1178 WARN_ON_ONCE(!irqs_disabled());
1180 debug_work_activate(work
);
1182 /* if dying, only works from the same workqueue are allowed */
1183 if (unlikely(wq
->flags
& WQ_DRAINING
) &&
1184 WARN_ON_ONCE(!is_chained_work(wq
)))
1187 /* determine gcwq to use */
1188 if (!(wq
->flags
& WQ_UNBOUND
)) {
1189 struct global_cwq
*last_gcwq
;
1191 if (cpu
== WORK_CPU_UNBOUND
)
1192 cpu
= raw_smp_processor_id();
1195 * It's multi cpu. If @work was previously on a different
1196 * cpu, it might still be running there, in which case the
1197 * work needs to be queued on that cpu to guarantee
1200 gcwq
= get_gcwq(cpu
);
1201 last_gcwq
= get_work_gcwq(work
);
1203 if (last_gcwq
&& last_gcwq
!= gcwq
) {
1204 struct worker
*worker
;
1206 spin_lock(&last_gcwq
->lock
);
1208 worker
= find_worker_executing_work(last_gcwq
, work
);
1210 if (worker
&& worker
->current_cwq
->wq
== wq
)
1213 /* meh... not running there, queue here */
1214 spin_unlock(&last_gcwq
->lock
);
1215 spin_lock(&gcwq
->lock
);
1218 spin_lock(&gcwq
->lock
);
1221 gcwq
= get_gcwq(WORK_CPU_UNBOUND
);
1222 spin_lock(&gcwq
->lock
);
1225 /* gcwq determined, get cwq and queue */
1226 cwq
= get_cwq(gcwq
->cpu
, wq
);
1227 trace_workqueue_queue_work(req_cpu
, cwq
, work
);
1229 if (WARN_ON(!list_empty(&work
->entry
))) {
1230 spin_unlock(&gcwq
->lock
);
1234 cwq
->nr_in_flight
[cwq
->work_color
]++;
1235 work_flags
= work_color_to_flags(cwq
->work_color
);
1237 if (likely(cwq
->nr_active
< cwq
->max_active
)) {
1238 trace_workqueue_activate_work(work
);
1240 worklist
= &cwq
->pool
->worklist
;
1242 work_flags
|= WORK_STRUCT_DELAYED
;
1243 worklist
= &cwq
->delayed_works
;
1246 insert_work(cwq
, work
, worklist
, work_flags
);
1248 spin_unlock(&gcwq
->lock
);
1252 * queue_work_on - queue work on specific cpu
1253 * @cpu: CPU number to execute work on
1254 * @wq: workqueue to use
1255 * @work: work to queue
1257 * Returns %false if @work was already on a queue, %true otherwise.
1259 * We queue the work to a specific CPU, the caller must ensure it
1262 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1263 struct work_struct
*work
)
1266 unsigned long flags
;
1268 local_irq_save(flags
);
1270 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1271 __queue_work(cpu
, wq
, work
);
1275 local_irq_restore(flags
);
1278 EXPORT_SYMBOL_GPL(queue_work_on
);
1281 * queue_work - queue work on a workqueue
1282 * @wq: workqueue to use
1283 * @work: work to queue
1285 * Returns %false if @work was already on a queue, %true otherwise.
1287 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1288 * it can be processed by another CPU.
1290 bool queue_work(struct workqueue_struct
*wq
, struct work_struct
*work
)
1292 return queue_work_on(WORK_CPU_UNBOUND
, wq
, work
);
1294 EXPORT_SYMBOL_GPL(queue_work
);
1296 void delayed_work_timer_fn(unsigned long __data
)
1298 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1299 struct cpu_workqueue_struct
*cwq
= get_work_cwq(&dwork
->work
);
1301 /* should have been called from irqsafe timer with irq already off */
1302 __queue_work(dwork
->cpu
, cwq
->wq
, &dwork
->work
);
1304 EXPORT_SYMBOL_GPL(delayed_work_timer_fn
);
1306 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1307 struct delayed_work
*dwork
, unsigned long delay
)
1309 struct timer_list
*timer
= &dwork
->timer
;
1310 struct work_struct
*work
= &dwork
->work
;
1313 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1314 timer
->data
!= (unsigned long)dwork
);
1315 WARN_ON_ONCE(timer_pending(timer
));
1316 WARN_ON_ONCE(!list_empty(&work
->entry
));
1319 * If @delay is 0, queue @dwork->work immediately. This is for
1320 * both optimization and correctness. The earliest @timer can
1321 * expire is on the closest next tick and delayed_work users depend
1322 * on that there's no such delay when @delay is 0.
1325 __queue_work(cpu
, wq
, &dwork
->work
);
1329 timer_stats_timer_set_start_info(&dwork
->timer
);
1332 * This stores cwq for the moment, for the timer_fn. Note that the
1333 * work's gcwq is preserved to allow reentrance detection for
1336 if (!(wq
->flags
& WQ_UNBOUND
)) {
1337 struct global_cwq
*gcwq
= get_work_gcwq(work
);
1340 * If we cannot get the last gcwq from @work directly,
1341 * select the last CPU such that it avoids unnecessarily
1342 * triggering non-reentrancy check in __queue_work().
1347 if (lcpu
== WORK_CPU_UNBOUND
)
1348 lcpu
= raw_smp_processor_id();
1350 lcpu
= WORK_CPU_UNBOUND
;
1353 set_work_cwq(work
, get_cwq(lcpu
, wq
), 0);
1356 timer
->expires
= jiffies
+ delay
;
1358 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1359 add_timer_on(timer
, cpu
);
1365 * queue_delayed_work_on - queue work on specific CPU after delay
1366 * @cpu: CPU number to execute work on
1367 * @wq: workqueue to use
1368 * @dwork: work to queue
1369 * @delay: number of jiffies to wait before queueing
1371 * Returns %false if @work was already on a queue, %true otherwise. If
1372 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1375 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1376 struct delayed_work
*dwork
, unsigned long delay
)
1378 struct work_struct
*work
= &dwork
->work
;
1380 unsigned long flags
;
1382 /* read the comment in __queue_work() */
1383 local_irq_save(flags
);
1385 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1386 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1390 local_irq_restore(flags
);
1393 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1396 * queue_delayed_work - queue work on a workqueue after delay
1397 * @wq: workqueue to use
1398 * @dwork: delayable work to queue
1399 * @delay: number of jiffies to wait before queueing
1401 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1403 bool queue_delayed_work(struct workqueue_struct
*wq
,
1404 struct delayed_work
*dwork
, unsigned long delay
)
1406 return queue_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1408 EXPORT_SYMBOL_GPL(queue_delayed_work
);
1411 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1412 * @cpu: CPU number to execute work on
1413 * @wq: workqueue to use
1414 * @dwork: work to queue
1415 * @delay: number of jiffies to wait before queueing
1417 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1418 * modify @dwork's timer so that it expires after @delay. If @delay is
1419 * zero, @work is guaranteed to be scheduled immediately regardless of its
1422 * Returns %false if @dwork was idle and queued, %true if @dwork was
1423 * pending and its timer was modified.
1425 * This function is safe to call from any context including IRQ handler.
1426 * See try_to_grab_pending() for details.
1428 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1429 struct delayed_work
*dwork
, unsigned long delay
)
1431 unsigned long flags
;
1435 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1436 } while (unlikely(ret
== -EAGAIN
));
1438 if (likely(ret
>= 0)) {
1439 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1440 local_irq_restore(flags
);
1443 /* -ENOENT from try_to_grab_pending() becomes %true */
1446 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1449 * mod_delayed_work - modify delay of or queue a delayed work
1450 * @wq: workqueue to use
1451 * @dwork: work to queue
1452 * @delay: number of jiffies to wait before queueing
1454 * mod_delayed_work_on() on local CPU.
1456 bool mod_delayed_work(struct workqueue_struct
*wq
, struct delayed_work
*dwork
,
1457 unsigned long delay
)
1459 return mod_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1461 EXPORT_SYMBOL_GPL(mod_delayed_work
);
1464 * worker_enter_idle - enter idle state
1465 * @worker: worker which is entering idle state
1467 * @worker is entering idle state. Update stats and idle timer if
1471 * spin_lock_irq(gcwq->lock).
1473 static void worker_enter_idle(struct worker
*worker
)
1475 struct worker_pool
*pool
= worker
->pool
;
1476 struct global_cwq
*gcwq
= pool
->gcwq
;
1478 BUG_ON(worker
->flags
& WORKER_IDLE
);
1479 BUG_ON(!list_empty(&worker
->entry
) &&
1480 (worker
->hentry
.next
|| worker
->hentry
.pprev
));
1482 /* can't use worker_set_flags(), also called from start_worker() */
1483 worker
->flags
|= WORKER_IDLE
;
1485 worker
->last_active
= jiffies
;
1487 /* idle_list is LIFO */
1488 list_add(&worker
->entry
, &pool
->idle_list
);
1490 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1491 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1494 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1495 * gcwq->lock between setting %WORKER_UNBOUND and zapping
1496 * nr_running, the warning may trigger spuriously. Check iff
1497 * unbind is not in progress.
1499 WARN_ON_ONCE(!(gcwq
->flags
& GCWQ_DISASSOCIATED
) &&
1500 pool
->nr_workers
== pool
->nr_idle
&&
1501 atomic_read(get_pool_nr_running(pool
)));
1505 * worker_leave_idle - leave idle state
1506 * @worker: worker which is leaving idle state
1508 * @worker is leaving idle state. Update stats.
1511 * spin_lock_irq(gcwq->lock).
1513 static void worker_leave_idle(struct worker
*worker
)
1515 struct worker_pool
*pool
= worker
->pool
;
1517 BUG_ON(!(worker
->flags
& WORKER_IDLE
));
1518 worker_clr_flags(worker
, WORKER_IDLE
);
1520 list_del_init(&worker
->entry
);
1524 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1527 * Works which are scheduled while the cpu is online must at least be
1528 * scheduled to a worker which is bound to the cpu so that if they are
1529 * flushed from cpu callbacks while cpu is going down, they are
1530 * guaranteed to execute on the cpu.
1532 * This function is to be used by rogue workers and rescuers to bind
1533 * themselves to the target cpu and may race with cpu going down or
1534 * coming online. kthread_bind() can't be used because it may put the
1535 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1536 * verbatim as it's best effort and blocking and gcwq may be
1537 * [dis]associated in the meantime.
1539 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1540 * binding against %GCWQ_DISASSOCIATED which is set during
1541 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1542 * enters idle state or fetches works without dropping lock, it can
1543 * guarantee the scheduling requirement described in the first paragraph.
1546 * Might sleep. Called without any lock but returns with gcwq->lock
1550 * %true if the associated gcwq is online (@worker is successfully
1551 * bound), %false if offline.
1553 static bool worker_maybe_bind_and_lock(struct worker
*worker
)
1554 __acquires(&gcwq
->lock
)
1556 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1557 struct task_struct
*task
= worker
->task
;
1561 * The following call may fail, succeed or succeed
1562 * without actually migrating the task to the cpu if
1563 * it races with cpu hotunplug operation. Verify
1564 * against GCWQ_DISASSOCIATED.
1566 if (!(gcwq
->flags
& GCWQ_DISASSOCIATED
))
1567 set_cpus_allowed_ptr(task
, get_cpu_mask(gcwq
->cpu
));
1569 spin_lock_irq(&gcwq
->lock
);
1570 if (gcwq
->flags
& GCWQ_DISASSOCIATED
)
1572 if (task_cpu(task
) == gcwq
->cpu
&&
1573 cpumask_equal(¤t
->cpus_allowed
,
1574 get_cpu_mask(gcwq
->cpu
)))
1576 spin_unlock_irq(&gcwq
->lock
);
1579 * We've raced with CPU hot[un]plug. Give it a breather
1580 * and retry migration. cond_resched() is required here;
1581 * otherwise, we might deadlock against cpu_stop trying to
1582 * bring down the CPU on non-preemptive kernel.
1590 * Rebind an idle @worker to its CPU. worker_thread() will test
1591 * list_empty(@worker->entry) before leaving idle and call this function.
1593 static void idle_worker_rebind(struct worker
*worker
)
1595 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1597 /* CPU may go down again inbetween, clear UNBOUND only on success */
1598 if (worker_maybe_bind_and_lock(worker
))
1599 worker_clr_flags(worker
, WORKER_UNBOUND
);
1601 /* rebind complete, become available again */
1602 list_add(&worker
->entry
, &worker
->pool
->idle_list
);
1603 spin_unlock_irq(&gcwq
->lock
);
1607 * Function for @worker->rebind.work used to rebind unbound busy workers to
1608 * the associated cpu which is coming back online. This is scheduled by
1609 * cpu up but can race with other cpu hotplug operations and may be
1610 * executed twice without intervening cpu down.
1612 static void busy_worker_rebind_fn(struct work_struct
*work
)
1614 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1615 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1617 if (worker_maybe_bind_and_lock(worker
))
1618 worker_clr_flags(worker
, WORKER_UNBOUND
);
1620 spin_unlock_irq(&gcwq
->lock
);
1624 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1625 * @gcwq: gcwq of interest
1627 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1628 * is different for idle and busy ones.
1630 * Idle ones will be removed from the idle_list and woken up. They will
1631 * add themselves back after completing rebind. This ensures that the
1632 * idle_list doesn't contain any unbound workers when re-bound busy workers
1633 * try to perform local wake-ups for concurrency management.
1635 * Busy workers can rebind after they finish their current work items.
1636 * Queueing the rebind work item at the head of the scheduled list is
1637 * enough. Note that nr_running will be properly bumped as busy workers
1640 * On return, all non-manager workers are scheduled for rebind - see
1641 * manage_workers() for the manager special case. Any idle worker
1642 * including the manager will not appear on @idle_list until rebind is
1643 * complete, making local wake-ups safe.
1645 static void rebind_workers(struct global_cwq
*gcwq
)
1647 struct worker_pool
*pool
;
1648 struct worker
*worker
, *n
;
1649 struct hlist_node
*pos
;
1652 lockdep_assert_held(&gcwq
->lock
);
1654 for_each_worker_pool(pool
, gcwq
)
1655 lockdep_assert_held(&pool
->assoc_mutex
);
1657 /* dequeue and kick idle ones */
1658 for_each_worker_pool(pool
, gcwq
) {
1659 list_for_each_entry_safe(worker
, n
, &pool
->idle_list
, entry
) {
1661 * idle workers should be off @pool->idle_list
1662 * until rebind is complete to avoid receiving
1663 * premature local wake-ups.
1665 list_del_init(&worker
->entry
);
1668 * worker_thread() will see the above dequeuing
1669 * and call idle_worker_rebind().
1671 wake_up_process(worker
->task
);
1675 /* rebind busy workers */
1676 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
1677 struct work_struct
*rebind_work
= &worker
->rebind_work
;
1678 struct workqueue_struct
*wq
;
1680 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
1681 work_data_bits(rebind_work
)))
1684 debug_work_activate(rebind_work
);
1687 * wq doesn't really matter but let's keep @worker->pool
1688 * and @cwq->pool consistent for sanity.
1690 if (worker_pool_pri(worker
->pool
))
1691 wq
= system_highpri_wq
;
1695 insert_work(get_cwq(gcwq
->cpu
, wq
), rebind_work
,
1696 worker
->scheduled
.next
,
1697 work_color_to_flags(WORK_NO_COLOR
));
1701 static struct worker
*alloc_worker(void)
1703 struct worker
*worker
;
1705 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1707 INIT_LIST_HEAD(&worker
->entry
);
1708 INIT_LIST_HEAD(&worker
->scheduled
);
1709 INIT_WORK(&worker
->rebind_work
, busy_worker_rebind_fn
);
1710 /* on creation a worker is in !idle && prep state */
1711 worker
->flags
= WORKER_PREP
;
1717 * create_worker - create a new workqueue worker
1718 * @pool: pool the new worker will belong to
1720 * Create a new worker which is bound to @pool. The returned worker
1721 * can be started by calling start_worker() or destroyed using
1725 * Might sleep. Does GFP_KERNEL allocations.
1728 * Pointer to the newly created worker.
1730 static struct worker
*create_worker(struct worker_pool
*pool
)
1732 struct global_cwq
*gcwq
= pool
->gcwq
;
1733 const char *pri
= worker_pool_pri(pool
) ? "H" : "";
1734 struct worker
*worker
= NULL
;
1737 spin_lock_irq(&gcwq
->lock
);
1738 while (ida_get_new(&pool
->worker_ida
, &id
)) {
1739 spin_unlock_irq(&gcwq
->lock
);
1740 if (!ida_pre_get(&pool
->worker_ida
, GFP_KERNEL
))
1742 spin_lock_irq(&gcwq
->lock
);
1744 spin_unlock_irq(&gcwq
->lock
);
1746 worker
= alloc_worker();
1750 worker
->pool
= pool
;
1753 if (gcwq
->cpu
!= WORK_CPU_UNBOUND
)
1754 worker
->task
= kthread_create_on_node(worker_thread
,
1755 worker
, cpu_to_node(gcwq
->cpu
),
1756 "kworker/%u:%d%s", gcwq
->cpu
, id
, pri
);
1758 worker
->task
= kthread_create(worker_thread
, worker
,
1759 "kworker/u:%d%s", id
, pri
);
1760 if (IS_ERR(worker
->task
))
1763 if (worker_pool_pri(pool
))
1764 set_user_nice(worker
->task
, HIGHPRI_NICE_LEVEL
);
1767 * Determine CPU binding of the new worker depending on
1768 * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the
1769 * flag remains stable across this function. See the comments
1770 * above the flag definition for details.
1772 * As an unbound worker may later become a regular one if CPU comes
1773 * online, make sure every worker has %PF_THREAD_BOUND set.
1775 if (!(gcwq
->flags
& GCWQ_DISASSOCIATED
)) {
1776 kthread_bind(worker
->task
, gcwq
->cpu
);
1778 worker
->task
->flags
|= PF_THREAD_BOUND
;
1779 worker
->flags
|= WORKER_UNBOUND
;
1785 spin_lock_irq(&gcwq
->lock
);
1786 ida_remove(&pool
->worker_ida
, id
);
1787 spin_unlock_irq(&gcwq
->lock
);
1794 * start_worker - start a newly created worker
1795 * @worker: worker to start
1797 * Make the gcwq aware of @worker and start it.
1800 * spin_lock_irq(gcwq->lock).
1802 static void start_worker(struct worker
*worker
)
1804 worker
->flags
|= WORKER_STARTED
;
1805 worker
->pool
->nr_workers
++;
1806 worker_enter_idle(worker
);
1807 wake_up_process(worker
->task
);
1811 * destroy_worker - destroy a workqueue worker
1812 * @worker: worker to be destroyed
1814 * Destroy @worker and adjust @gcwq stats accordingly.
1817 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1819 static void destroy_worker(struct worker
*worker
)
1821 struct worker_pool
*pool
= worker
->pool
;
1822 struct global_cwq
*gcwq
= pool
->gcwq
;
1823 int id
= worker
->id
;
1825 /* sanity check frenzy */
1826 BUG_ON(worker
->current_work
);
1827 BUG_ON(!list_empty(&worker
->scheduled
));
1829 if (worker
->flags
& WORKER_STARTED
)
1831 if (worker
->flags
& WORKER_IDLE
)
1834 list_del_init(&worker
->entry
);
1835 worker
->flags
|= WORKER_DIE
;
1837 spin_unlock_irq(&gcwq
->lock
);
1839 kthread_stop(worker
->task
);
1842 spin_lock_irq(&gcwq
->lock
);
1843 ida_remove(&pool
->worker_ida
, id
);
1846 static void idle_worker_timeout(unsigned long __pool
)
1848 struct worker_pool
*pool
= (void *)__pool
;
1849 struct global_cwq
*gcwq
= pool
->gcwq
;
1851 spin_lock_irq(&gcwq
->lock
);
1853 if (too_many_workers(pool
)) {
1854 struct worker
*worker
;
1855 unsigned long expires
;
1857 /* idle_list is kept in LIFO order, check the last one */
1858 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1859 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1861 if (time_before(jiffies
, expires
))
1862 mod_timer(&pool
->idle_timer
, expires
);
1864 /* it's been idle for too long, wake up manager */
1865 pool
->flags
|= POOL_MANAGE_WORKERS
;
1866 wake_up_worker(pool
);
1870 spin_unlock_irq(&gcwq
->lock
);
1873 static bool send_mayday(struct work_struct
*work
)
1875 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1876 struct workqueue_struct
*wq
= cwq
->wq
;
1879 if (!(wq
->flags
& WQ_RESCUER
))
1882 /* mayday mayday mayday */
1883 cpu
= cwq
->pool
->gcwq
->cpu
;
1884 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1885 if (cpu
== WORK_CPU_UNBOUND
)
1887 if (!mayday_test_and_set_cpu(cpu
, wq
->mayday_mask
))
1888 wake_up_process(wq
->rescuer
->task
);
1892 static void gcwq_mayday_timeout(unsigned long __pool
)
1894 struct worker_pool
*pool
= (void *)__pool
;
1895 struct global_cwq
*gcwq
= pool
->gcwq
;
1896 struct work_struct
*work
;
1898 spin_lock_irq(&gcwq
->lock
);
1900 if (need_to_create_worker(pool
)) {
1902 * We've been trying to create a new worker but
1903 * haven't been successful. We might be hitting an
1904 * allocation deadlock. Send distress signals to
1907 list_for_each_entry(work
, &pool
->worklist
, entry
)
1911 spin_unlock_irq(&gcwq
->lock
);
1913 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1917 * maybe_create_worker - create a new worker if necessary
1918 * @pool: pool to create a new worker for
1920 * Create a new worker for @pool if necessary. @pool is guaranteed to
1921 * have at least one idle worker on return from this function. If
1922 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1923 * sent to all rescuers with works scheduled on @pool to resolve
1924 * possible allocation deadlock.
1926 * On return, need_to_create_worker() is guaranteed to be false and
1927 * may_start_working() true.
1930 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1931 * multiple times. Does GFP_KERNEL allocations. Called only from
1935 * false if no action was taken and gcwq->lock stayed locked, true
1938 static bool maybe_create_worker(struct worker_pool
*pool
)
1939 __releases(&gcwq
->lock
)
1940 __acquires(&gcwq
->lock
)
1942 struct global_cwq
*gcwq
= pool
->gcwq
;
1944 if (!need_to_create_worker(pool
))
1947 spin_unlock_irq(&gcwq
->lock
);
1949 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1950 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1953 struct worker
*worker
;
1955 worker
= create_worker(pool
);
1957 del_timer_sync(&pool
->mayday_timer
);
1958 spin_lock_irq(&gcwq
->lock
);
1959 start_worker(worker
);
1960 BUG_ON(need_to_create_worker(pool
));
1964 if (!need_to_create_worker(pool
))
1967 __set_current_state(TASK_INTERRUPTIBLE
);
1968 schedule_timeout(CREATE_COOLDOWN
);
1970 if (!need_to_create_worker(pool
))
1974 del_timer_sync(&pool
->mayday_timer
);
1975 spin_lock_irq(&gcwq
->lock
);
1976 if (need_to_create_worker(pool
))
1982 * maybe_destroy_worker - destroy workers which have been idle for a while
1983 * @pool: pool to destroy workers for
1985 * Destroy @pool workers which have been idle for longer than
1986 * IDLE_WORKER_TIMEOUT.
1989 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1990 * multiple times. Called only from manager.
1993 * false if no action was taken and gcwq->lock stayed locked, true
1996 static bool maybe_destroy_workers(struct worker_pool
*pool
)
2000 while (too_many_workers(pool
)) {
2001 struct worker
*worker
;
2002 unsigned long expires
;
2004 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2005 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2007 if (time_before(jiffies
, expires
)) {
2008 mod_timer(&pool
->idle_timer
, expires
);
2012 destroy_worker(worker
);
2020 * manage_workers - manage worker pool
2023 * Assume the manager role and manage gcwq worker pool @worker belongs
2024 * to. At any given time, there can be only zero or one manager per
2025 * gcwq. The exclusion is handled automatically by this function.
2027 * The caller can safely start processing works on false return. On
2028 * true return, it's guaranteed that need_to_create_worker() is false
2029 * and may_start_working() is true.
2032 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2033 * multiple times. Does GFP_KERNEL allocations.
2036 * false if no action was taken and gcwq->lock stayed locked, true if
2037 * some action was taken.
2039 static bool manage_workers(struct worker
*worker
)
2041 struct worker_pool
*pool
= worker
->pool
;
2044 if (pool
->flags
& POOL_MANAGING_WORKERS
)
2047 pool
->flags
|= POOL_MANAGING_WORKERS
;
2050 * To simplify both worker management and CPU hotplug, hold off
2051 * management while hotplug is in progress. CPU hotplug path can't
2052 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2053 * lead to idle worker depletion (all become busy thinking someone
2054 * else is managing) which in turn can result in deadlock under
2055 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2056 * manager against CPU hotplug.
2058 * assoc_mutex would always be free unless CPU hotplug is in
2059 * progress. trylock first without dropping @gcwq->lock.
2061 if (unlikely(!mutex_trylock(&pool
->assoc_mutex
))) {
2062 spin_unlock_irq(&pool
->gcwq
->lock
);
2063 mutex_lock(&pool
->assoc_mutex
);
2065 * CPU hotplug could have happened while we were waiting
2066 * for assoc_mutex. Hotplug itself can't handle us
2067 * because manager isn't either on idle or busy list, and
2068 * @gcwq's state and ours could have deviated.
2070 * As hotplug is now excluded via assoc_mutex, we can
2071 * simply try to bind. It will succeed or fail depending
2072 * on @gcwq's current state. Try it and adjust
2073 * %WORKER_UNBOUND accordingly.
2075 if (worker_maybe_bind_and_lock(worker
))
2076 worker
->flags
&= ~WORKER_UNBOUND
;
2078 worker
->flags
|= WORKER_UNBOUND
;
2083 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2086 * Destroy and then create so that may_start_working() is true
2089 ret
|= maybe_destroy_workers(pool
);
2090 ret
|= maybe_create_worker(pool
);
2092 pool
->flags
&= ~POOL_MANAGING_WORKERS
;
2093 mutex_unlock(&pool
->assoc_mutex
);
2098 * process_one_work - process single work
2100 * @work: work to process
2102 * Process @work. This function contains all the logics necessary to
2103 * process a single work including synchronization against and
2104 * interaction with other workers on the same cpu, queueing and
2105 * flushing. As long as context requirement is met, any worker can
2106 * call this function to process a work.
2109 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2111 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2112 __releases(&gcwq
->lock
)
2113 __acquires(&gcwq
->lock
)
2115 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
2116 struct worker_pool
*pool
= worker
->pool
;
2117 struct global_cwq
*gcwq
= pool
->gcwq
;
2118 bool cpu_intensive
= cwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2120 struct worker
*collision
;
2121 #ifdef CONFIG_LOCKDEP
2123 * It is permissible to free the struct work_struct from
2124 * inside the function that is called from it, this we need to
2125 * take into account for lockdep too. To avoid bogus "held
2126 * lock freed" warnings as well as problems when looking into
2127 * work->lockdep_map, make a copy and use that here.
2129 struct lockdep_map lockdep_map
;
2131 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2134 * Ensure we're on the correct CPU. DISASSOCIATED test is
2135 * necessary to avoid spurious warnings from rescuers servicing the
2136 * unbound or a disassociated gcwq.
2138 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2139 !(gcwq
->flags
& GCWQ_DISASSOCIATED
) &&
2140 raw_smp_processor_id() != gcwq
->cpu
);
2143 * A single work shouldn't be executed concurrently by
2144 * multiple workers on a single cpu. Check whether anyone is
2145 * already processing the work. If so, defer the work to the
2146 * currently executing one.
2148 collision
= find_worker_executing_work(gcwq
, work
);
2149 if (unlikely(collision
)) {
2150 move_linked_works(work
, &collision
->scheduled
, NULL
);
2154 /* claim and dequeue */
2155 debug_work_deactivate(work
);
2156 hash_add(gcwq
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2157 worker
->current_work
= work
;
2158 worker
->current_func
= work
->func
;
2159 worker
->current_cwq
= cwq
;
2160 work_color
= get_work_color(work
);
2162 list_del_init(&work
->entry
);
2165 * CPU intensive works don't participate in concurrency
2166 * management. They're the scheduler's responsibility.
2168 if (unlikely(cpu_intensive
))
2169 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2172 * Unbound gcwq isn't concurrency managed and work items should be
2173 * executed ASAP. Wake up another worker if necessary.
2175 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2176 wake_up_worker(pool
);
2179 * Record the last CPU and clear PENDING which should be the last
2180 * update to @work. Also, do this inside @gcwq->lock so that
2181 * PENDING and queued state changes happen together while IRQ is
2184 set_work_cpu_and_clear_pending(work
, gcwq
->cpu
);
2186 spin_unlock_irq(&gcwq
->lock
);
2188 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2189 lock_map_acquire(&lockdep_map
);
2190 trace_workqueue_execute_start(work
);
2191 worker
->current_func(work
);
2193 * While we must be careful to not use "work" after this, the trace
2194 * point will only record its address.
2196 trace_workqueue_execute_end(work
);
2197 lock_map_release(&lockdep_map
);
2198 lock_map_release(&cwq
->wq
->lockdep_map
);
2200 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2201 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2202 " last function: %pf\n",
2203 current
->comm
, preempt_count(), task_pid_nr(current
),
2204 worker
->current_func
);
2205 debug_show_held_locks(current
);
2209 spin_lock_irq(&gcwq
->lock
);
2211 /* clear cpu intensive status */
2212 if (unlikely(cpu_intensive
))
2213 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2215 /* we're done with it, release */
2216 hash_del(&worker
->hentry
);
2217 worker
->current_work
= NULL
;
2218 worker
->current_func
= NULL
;
2219 worker
->current_cwq
= NULL
;
2220 cwq_dec_nr_in_flight(cwq
, work_color
);
2224 * process_scheduled_works - process scheduled works
2227 * Process all scheduled works. Please note that the scheduled list
2228 * may change while processing a work, so this function repeatedly
2229 * fetches a work from the top and executes it.
2232 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2235 static void process_scheduled_works(struct worker
*worker
)
2237 while (!list_empty(&worker
->scheduled
)) {
2238 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2239 struct work_struct
, entry
);
2240 process_one_work(worker
, work
);
2245 * worker_thread - the worker thread function
2248 * The gcwq worker thread function. There's a single dynamic pool of
2249 * these per each cpu. These workers process all works regardless of
2250 * their specific target workqueue. The only exception is works which
2251 * belong to workqueues with a rescuer which will be explained in
2254 static int worker_thread(void *__worker
)
2256 struct worker
*worker
= __worker
;
2257 struct worker_pool
*pool
= worker
->pool
;
2258 struct global_cwq
*gcwq
= pool
->gcwq
;
2260 /* tell the scheduler that this is a workqueue worker */
2261 worker
->task
->flags
|= PF_WQ_WORKER
;
2263 spin_lock_irq(&gcwq
->lock
);
2265 /* we are off idle list if destruction or rebind is requested */
2266 if (unlikely(list_empty(&worker
->entry
))) {
2267 spin_unlock_irq(&gcwq
->lock
);
2269 /* if DIE is set, destruction is requested */
2270 if (worker
->flags
& WORKER_DIE
) {
2271 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2275 /* otherwise, rebind */
2276 idle_worker_rebind(worker
);
2280 worker_leave_idle(worker
);
2282 /* no more worker necessary? */
2283 if (!need_more_worker(pool
))
2286 /* do we need to manage? */
2287 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2291 * ->scheduled list can only be filled while a worker is
2292 * preparing to process a work or actually processing it.
2293 * Make sure nobody diddled with it while I was sleeping.
2295 BUG_ON(!list_empty(&worker
->scheduled
));
2298 * When control reaches this point, we're guaranteed to have
2299 * at least one idle worker or that someone else has already
2300 * assumed the manager role.
2302 worker_clr_flags(worker
, WORKER_PREP
);
2305 struct work_struct
*work
=
2306 list_first_entry(&pool
->worklist
,
2307 struct work_struct
, entry
);
2309 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2310 /* optimization path, not strictly necessary */
2311 process_one_work(worker
, work
);
2312 if (unlikely(!list_empty(&worker
->scheduled
)))
2313 process_scheduled_works(worker
);
2315 move_linked_works(work
, &worker
->scheduled
, NULL
);
2316 process_scheduled_works(worker
);
2318 } while (keep_working(pool
));
2320 worker_set_flags(worker
, WORKER_PREP
, false);
2322 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2326 * gcwq->lock is held and there's no work to process and no
2327 * need to manage, sleep. Workers are woken up only while
2328 * holding gcwq->lock or from local cpu, so setting the
2329 * current state before releasing gcwq->lock is enough to
2330 * prevent losing any event.
2332 worker_enter_idle(worker
);
2333 __set_current_state(TASK_INTERRUPTIBLE
);
2334 spin_unlock_irq(&gcwq
->lock
);
2340 * rescuer_thread - the rescuer thread function
2343 * Workqueue rescuer thread function. There's one rescuer for each
2344 * workqueue which has WQ_RESCUER set.
2346 * Regular work processing on a gcwq may block trying to create a new
2347 * worker which uses GFP_KERNEL allocation which has slight chance of
2348 * developing into deadlock if some works currently on the same queue
2349 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2350 * the problem rescuer solves.
2352 * When such condition is possible, the gcwq summons rescuers of all
2353 * workqueues which have works queued on the gcwq and let them process
2354 * those works so that forward progress can be guaranteed.
2356 * This should happen rarely.
2358 static int rescuer_thread(void *__rescuer
)
2360 struct worker
*rescuer
= __rescuer
;
2361 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2362 struct list_head
*scheduled
= &rescuer
->scheduled
;
2363 bool is_unbound
= wq
->flags
& WQ_UNBOUND
;
2366 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2369 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2370 * doesn't participate in concurrency management.
2372 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2374 set_current_state(TASK_INTERRUPTIBLE
);
2376 if (kthread_should_stop()) {
2377 __set_current_state(TASK_RUNNING
);
2378 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2383 * See whether any cpu is asking for help. Unbounded
2384 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2386 for_each_mayday_cpu(cpu
, wq
->mayday_mask
) {
2387 unsigned int tcpu
= is_unbound
? WORK_CPU_UNBOUND
: cpu
;
2388 struct cpu_workqueue_struct
*cwq
= get_cwq(tcpu
, wq
);
2389 struct worker_pool
*pool
= cwq
->pool
;
2390 struct global_cwq
*gcwq
= pool
->gcwq
;
2391 struct work_struct
*work
, *n
;
2393 __set_current_state(TASK_RUNNING
);
2394 mayday_clear_cpu(cpu
, wq
->mayday_mask
);
2396 /* migrate to the target cpu if possible */
2397 rescuer
->pool
= pool
;
2398 worker_maybe_bind_and_lock(rescuer
);
2401 * Slurp in all works issued via this workqueue and
2404 BUG_ON(!list_empty(&rescuer
->scheduled
));
2405 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2406 if (get_work_cwq(work
) == cwq
)
2407 move_linked_works(work
, scheduled
, &n
);
2409 process_scheduled_works(rescuer
);
2412 * Leave this gcwq. If keep_working() is %true, notify a
2413 * regular worker; otherwise, we end up with 0 concurrency
2414 * and stalling the execution.
2416 if (keep_working(pool
))
2417 wake_up_worker(pool
);
2419 spin_unlock_irq(&gcwq
->lock
);
2422 /* rescuers should never participate in concurrency management */
2423 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2429 struct work_struct work
;
2430 struct completion done
;
2433 static void wq_barrier_func(struct work_struct
*work
)
2435 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2436 complete(&barr
->done
);
2440 * insert_wq_barrier - insert a barrier work
2441 * @cwq: cwq to insert barrier into
2442 * @barr: wq_barrier to insert
2443 * @target: target work to attach @barr to
2444 * @worker: worker currently executing @target, NULL if @target is not executing
2446 * @barr is linked to @target such that @barr is completed only after
2447 * @target finishes execution. Please note that the ordering
2448 * guarantee is observed only with respect to @target and on the local
2451 * Currently, a queued barrier can't be canceled. This is because
2452 * try_to_grab_pending() can't determine whether the work to be
2453 * grabbed is at the head of the queue and thus can't clear LINKED
2454 * flag of the previous work while there must be a valid next work
2455 * after a work with LINKED flag set.
2457 * Note that when @worker is non-NULL, @target may be modified
2458 * underneath us, so we can't reliably determine cwq from @target.
2461 * spin_lock_irq(gcwq->lock).
2463 static void insert_wq_barrier(struct cpu_workqueue_struct
*cwq
,
2464 struct wq_barrier
*barr
,
2465 struct work_struct
*target
, struct worker
*worker
)
2467 struct list_head
*head
;
2468 unsigned int linked
= 0;
2471 * debugobject calls are safe here even with gcwq->lock locked
2472 * as we know for sure that this will not trigger any of the
2473 * checks and call back into the fixup functions where we
2476 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2477 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2478 init_completion(&barr
->done
);
2481 * If @target is currently being executed, schedule the
2482 * barrier to the worker; otherwise, put it after @target.
2485 head
= worker
->scheduled
.next
;
2487 unsigned long *bits
= work_data_bits(target
);
2489 head
= target
->entry
.next
;
2490 /* there can already be other linked works, inherit and set */
2491 linked
= *bits
& WORK_STRUCT_LINKED
;
2492 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2495 debug_work_activate(&barr
->work
);
2496 insert_work(cwq
, &barr
->work
, head
,
2497 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2501 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2502 * @wq: workqueue being flushed
2503 * @flush_color: new flush color, < 0 for no-op
2504 * @work_color: new work color, < 0 for no-op
2506 * Prepare cwqs for workqueue flushing.
2508 * If @flush_color is non-negative, flush_color on all cwqs should be
2509 * -1. If no cwq has in-flight commands at the specified color, all
2510 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2511 * has in flight commands, its cwq->flush_color is set to
2512 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2513 * wakeup logic is armed and %true is returned.
2515 * The caller should have initialized @wq->first_flusher prior to
2516 * calling this function with non-negative @flush_color. If
2517 * @flush_color is negative, no flush color update is done and %false
2520 * If @work_color is non-negative, all cwqs should have the same
2521 * work_color which is previous to @work_color and all will be
2522 * advanced to @work_color.
2525 * mutex_lock(wq->flush_mutex).
2528 * %true if @flush_color >= 0 and there's something to flush. %false
2531 static bool flush_workqueue_prep_cwqs(struct workqueue_struct
*wq
,
2532 int flush_color
, int work_color
)
2537 if (flush_color
>= 0) {
2538 BUG_ON(atomic_read(&wq
->nr_cwqs_to_flush
));
2539 atomic_set(&wq
->nr_cwqs_to_flush
, 1);
2542 for_each_cwq_cpu(cpu
, wq
) {
2543 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2544 struct global_cwq
*gcwq
= cwq
->pool
->gcwq
;
2546 spin_lock_irq(&gcwq
->lock
);
2548 if (flush_color
>= 0) {
2549 BUG_ON(cwq
->flush_color
!= -1);
2551 if (cwq
->nr_in_flight
[flush_color
]) {
2552 cwq
->flush_color
= flush_color
;
2553 atomic_inc(&wq
->nr_cwqs_to_flush
);
2558 if (work_color
>= 0) {
2559 BUG_ON(work_color
!= work_next_color(cwq
->work_color
));
2560 cwq
->work_color
= work_color
;
2563 spin_unlock_irq(&gcwq
->lock
);
2566 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_cwqs_to_flush
))
2567 complete(&wq
->first_flusher
->done
);
2573 * flush_workqueue - ensure that any scheduled work has run to completion.
2574 * @wq: workqueue to flush
2576 * Forces execution of the workqueue and blocks until its completion.
2577 * This is typically used in driver shutdown handlers.
2579 * We sleep until all works which were queued on entry have been handled,
2580 * but we are not livelocked by new incoming ones.
2582 void flush_workqueue(struct workqueue_struct
*wq
)
2584 struct wq_flusher this_flusher
= {
2585 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2587 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2591 lock_map_acquire(&wq
->lockdep_map
);
2592 lock_map_release(&wq
->lockdep_map
);
2594 mutex_lock(&wq
->flush_mutex
);
2597 * Start-to-wait phase
2599 next_color
= work_next_color(wq
->work_color
);
2601 if (next_color
!= wq
->flush_color
) {
2603 * Color space is not full. The current work_color
2604 * becomes our flush_color and work_color is advanced
2607 BUG_ON(!list_empty(&wq
->flusher_overflow
));
2608 this_flusher
.flush_color
= wq
->work_color
;
2609 wq
->work_color
= next_color
;
2611 if (!wq
->first_flusher
) {
2612 /* no flush in progress, become the first flusher */
2613 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2615 wq
->first_flusher
= &this_flusher
;
2617 if (!flush_workqueue_prep_cwqs(wq
, wq
->flush_color
,
2619 /* nothing to flush, done */
2620 wq
->flush_color
= next_color
;
2621 wq
->first_flusher
= NULL
;
2626 BUG_ON(wq
->flush_color
== this_flusher
.flush_color
);
2627 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2628 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2632 * Oops, color space is full, wait on overflow queue.
2633 * The next flush completion will assign us
2634 * flush_color and transfer to flusher_queue.
2636 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2639 mutex_unlock(&wq
->flush_mutex
);
2641 wait_for_completion(&this_flusher
.done
);
2644 * Wake-up-and-cascade phase
2646 * First flushers are responsible for cascading flushes and
2647 * handling overflow. Non-first flushers can simply return.
2649 if (wq
->first_flusher
!= &this_flusher
)
2652 mutex_lock(&wq
->flush_mutex
);
2654 /* we might have raced, check again with mutex held */
2655 if (wq
->first_flusher
!= &this_flusher
)
2658 wq
->first_flusher
= NULL
;
2660 BUG_ON(!list_empty(&this_flusher
.list
));
2661 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2664 struct wq_flusher
*next
, *tmp
;
2666 /* complete all the flushers sharing the current flush color */
2667 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2668 if (next
->flush_color
!= wq
->flush_color
)
2670 list_del_init(&next
->list
);
2671 complete(&next
->done
);
2674 BUG_ON(!list_empty(&wq
->flusher_overflow
) &&
2675 wq
->flush_color
!= work_next_color(wq
->work_color
));
2677 /* this flush_color is finished, advance by one */
2678 wq
->flush_color
= work_next_color(wq
->flush_color
);
2680 /* one color has been freed, handle overflow queue */
2681 if (!list_empty(&wq
->flusher_overflow
)) {
2683 * Assign the same color to all overflowed
2684 * flushers, advance work_color and append to
2685 * flusher_queue. This is the start-to-wait
2686 * phase for these overflowed flushers.
2688 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2689 tmp
->flush_color
= wq
->work_color
;
2691 wq
->work_color
= work_next_color(wq
->work_color
);
2693 list_splice_tail_init(&wq
->flusher_overflow
,
2694 &wq
->flusher_queue
);
2695 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2698 if (list_empty(&wq
->flusher_queue
)) {
2699 BUG_ON(wq
->flush_color
!= wq
->work_color
);
2704 * Need to flush more colors. Make the next flusher
2705 * the new first flusher and arm cwqs.
2707 BUG_ON(wq
->flush_color
== wq
->work_color
);
2708 BUG_ON(wq
->flush_color
!= next
->flush_color
);
2710 list_del_init(&next
->list
);
2711 wq
->first_flusher
= next
;
2713 if (flush_workqueue_prep_cwqs(wq
, wq
->flush_color
, -1))
2717 * Meh... this color is already done, clear first
2718 * flusher and repeat cascading.
2720 wq
->first_flusher
= NULL
;
2724 mutex_unlock(&wq
->flush_mutex
);
2726 EXPORT_SYMBOL_GPL(flush_workqueue
);
2729 * drain_workqueue - drain a workqueue
2730 * @wq: workqueue to drain
2732 * Wait until the workqueue becomes empty. While draining is in progress,
2733 * only chain queueing is allowed. IOW, only currently pending or running
2734 * work items on @wq can queue further work items on it. @wq is flushed
2735 * repeatedly until it becomes empty. The number of flushing is detemined
2736 * by the depth of chaining and should be relatively short. Whine if it
2739 void drain_workqueue(struct workqueue_struct
*wq
)
2741 unsigned int flush_cnt
= 0;
2745 * __queue_work() needs to test whether there are drainers, is much
2746 * hotter than drain_workqueue() and already looks at @wq->flags.
2747 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2749 spin_lock(&workqueue_lock
);
2750 if (!wq
->nr_drainers
++)
2751 wq
->flags
|= WQ_DRAINING
;
2752 spin_unlock(&workqueue_lock
);
2754 flush_workqueue(wq
);
2756 for_each_cwq_cpu(cpu
, wq
) {
2757 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2760 spin_lock_irq(&cwq
->pool
->gcwq
->lock
);
2761 drained
= !cwq
->nr_active
&& list_empty(&cwq
->delayed_works
);
2762 spin_unlock_irq(&cwq
->pool
->gcwq
->lock
);
2767 if (++flush_cnt
== 10 ||
2768 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2769 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2770 wq
->name
, flush_cnt
);
2774 spin_lock(&workqueue_lock
);
2775 if (!--wq
->nr_drainers
)
2776 wq
->flags
&= ~WQ_DRAINING
;
2777 spin_unlock(&workqueue_lock
);
2779 EXPORT_SYMBOL_GPL(drain_workqueue
);
2781 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2783 struct worker
*worker
= NULL
;
2784 struct global_cwq
*gcwq
;
2785 struct cpu_workqueue_struct
*cwq
;
2788 gcwq
= get_work_gcwq(work
);
2792 spin_lock_irq(&gcwq
->lock
);
2793 if (!list_empty(&work
->entry
)) {
2795 * See the comment near try_to_grab_pending()->smp_rmb().
2796 * If it was re-queued to a different gcwq under us, we
2797 * are not going to wait.
2800 cwq
= get_work_cwq(work
);
2801 if (unlikely(!cwq
|| gcwq
!= cwq
->pool
->gcwq
))
2804 worker
= find_worker_executing_work(gcwq
, work
);
2807 cwq
= worker
->current_cwq
;
2810 insert_wq_barrier(cwq
, barr
, work
, worker
);
2811 spin_unlock_irq(&gcwq
->lock
);
2814 * If @max_active is 1 or rescuer is in use, flushing another work
2815 * item on the same workqueue may lead to deadlock. Make sure the
2816 * flusher is not running on the same workqueue by verifying write
2819 if (cwq
->wq
->saved_max_active
== 1 || cwq
->wq
->flags
& WQ_RESCUER
)
2820 lock_map_acquire(&cwq
->wq
->lockdep_map
);
2822 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2823 lock_map_release(&cwq
->wq
->lockdep_map
);
2827 spin_unlock_irq(&gcwq
->lock
);
2832 * flush_work - wait for a work to finish executing the last queueing instance
2833 * @work: the work to flush
2835 * Wait until @work has finished execution. @work is guaranteed to be idle
2836 * on return if it hasn't been requeued since flush started.
2839 * %true if flush_work() waited for the work to finish execution,
2840 * %false if it was already idle.
2842 bool flush_work(struct work_struct
*work
)
2844 struct wq_barrier barr
;
2846 lock_map_acquire(&work
->lockdep_map
);
2847 lock_map_release(&work
->lockdep_map
);
2849 if (start_flush_work(work
, &barr
)) {
2850 wait_for_completion(&barr
.done
);
2851 destroy_work_on_stack(&barr
.work
);
2857 EXPORT_SYMBOL_GPL(flush_work
);
2859 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2861 unsigned long flags
;
2865 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2867 * If someone else is canceling, wait for the same event it
2868 * would be waiting for before retrying.
2870 if (unlikely(ret
== -ENOENT
))
2872 } while (unlikely(ret
< 0));
2874 /* tell other tasks trying to grab @work to back off */
2875 mark_work_canceling(work
);
2876 local_irq_restore(flags
);
2879 clear_work_data(work
);
2884 * cancel_work_sync - cancel a work and wait for it to finish
2885 * @work: the work to cancel
2887 * Cancel @work and wait for its execution to finish. This function
2888 * can be used even if the work re-queues itself or migrates to
2889 * another workqueue. On return from this function, @work is
2890 * guaranteed to be not pending or executing on any CPU.
2892 * cancel_work_sync(&delayed_work->work) must not be used for
2893 * delayed_work's. Use cancel_delayed_work_sync() instead.
2895 * The caller must ensure that the workqueue on which @work was last
2896 * queued can't be destroyed before this function returns.
2899 * %true if @work was pending, %false otherwise.
2901 bool cancel_work_sync(struct work_struct
*work
)
2903 return __cancel_work_timer(work
, false);
2905 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2908 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2909 * @dwork: the delayed work to flush
2911 * Delayed timer is cancelled and the pending work is queued for
2912 * immediate execution. Like flush_work(), this function only
2913 * considers the last queueing instance of @dwork.
2916 * %true if flush_work() waited for the work to finish execution,
2917 * %false if it was already idle.
2919 bool flush_delayed_work(struct delayed_work
*dwork
)
2921 local_irq_disable();
2922 if (del_timer_sync(&dwork
->timer
))
2923 __queue_work(dwork
->cpu
,
2924 get_work_cwq(&dwork
->work
)->wq
, &dwork
->work
);
2926 return flush_work(&dwork
->work
);
2928 EXPORT_SYMBOL(flush_delayed_work
);
2931 * cancel_delayed_work - cancel a delayed work
2932 * @dwork: delayed_work to cancel
2934 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2935 * and canceled; %false if wasn't pending. Note that the work callback
2936 * function may still be running on return, unless it returns %true and the
2937 * work doesn't re-arm itself. Explicitly flush or use
2938 * cancel_delayed_work_sync() to wait on it.
2940 * This function is safe to call from any context including IRQ handler.
2942 bool cancel_delayed_work(struct delayed_work
*dwork
)
2944 unsigned long flags
;
2948 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2949 } while (unlikely(ret
== -EAGAIN
));
2951 if (unlikely(ret
< 0))
2954 set_work_cpu_and_clear_pending(&dwork
->work
, work_cpu(&dwork
->work
));
2955 local_irq_restore(flags
);
2958 EXPORT_SYMBOL(cancel_delayed_work
);
2961 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2962 * @dwork: the delayed work cancel
2964 * This is cancel_work_sync() for delayed works.
2967 * %true if @dwork was pending, %false otherwise.
2969 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2971 return __cancel_work_timer(&dwork
->work
, true);
2973 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2976 * schedule_work_on - put work task on a specific cpu
2977 * @cpu: cpu to put the work task on
2978 * @work: job to be done
2980 * This puts a job on a specific cpu
2982 bool schedule_work_on(int cpu
, struct work_struct
*work
)
2984 return queue_work_on(cpu
, system_wq
, work
);
2986 EXPORT_SYMBOL(schedule_work_on
);
2989 * schedule_work - put work task in global workqueue
2990 * @work: job to be done
2992 * Returns %false if @work was already on the kernel-global workqueue and
2995 * This puts a job in the kernel-global workqueue if it was not already
2996 * queued and leaves it in the same position on the kernel-global
2997 * workqueue otherwise.
2999 bool schedule_work(struct work_struct
*work
)
3001 return queue_work(system_wq
, work
);
3003 EXPORT_SYMBOL(schedule_work
);
3006 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3008 * @dwork: job to be done
3009 * @delay: number of jiffies to wait
3011 * After waiting for a given time this puts a job in the kernel-global
3012 * workqueue on the specified CPU.
3014 bool schedule_delayed_work_on(int cpu
, struct delayed_work
*dwork
,
3015 unsigned long delay
)
3017 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
3019 EXPORT_SYMBOL(schedule_delayed_work_on
);
3022 * schedule_delayed_work - put work task in global workqueue after delay
3023 * @dwork: job to be done
3024 * @delay: number of jiffies to wait or 0 for immediate execution
3026 * After waiting for a given time this puts a job in the kernel-global
3029 bool schedule_delayed_work(struct delayed_work
*dwork
, unsigned long delay
)
3031 return queue_delayed_work(system_wq
, dwork
, delay
);
3033 EXPORT_SYMBOL(schedule_delayed_work
);
3036 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3037 * @func: the function to call
3039 * schedule_on_each_cpu() executes @func on each online CPU using the
3040 * system workqueue and blocks until all CPUs have completed.
3041 * schedule_on_each_cpu() is very slow.
3044 * 0 on success, -errno on failure.
3046 int schedule_on_each_cpu(work_func_t func
)
3049 struct work_struct __percpu
*works
;
3051 works
= alloc_percpu(struct work_struct
);
3057 for_each_online_cpu(cpu
) {
3058 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3060 INIT_WORK(work
, func
);
3061 schedule_work_on(cpu
, work
);
3064 for_each_online_cpu(cpu
)
3065 flush_work(per_cpu_ptr(works
, cpu
));
3073 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3075 * Forces execution of the kernel-global workqueue and blocks until its
3078 * Think twice before calling this function! It's very easy to get into
3079 * trouble if you don't take great care. Either of the following situations
3080 * will lead to deadlock:
3082 * One of the work items currently on the workqueue needs to acquire
3083 * a lock held by your code or its caller.
3085 * Your code is running in the context of a work routine.
3087 * They will be detected by lockdep when they occur, but the first might not
3088 * occur very often. It depends on what work items are on the workqueue and
3089 * what locks they need, which you have no control over.
3091 * In most situations flushing the entire workqueue is overkill; you merely
3092 * need to know that a particular work item isn't queued and isn't running.
3093 * In such cases you should use cancel_delayed_work_sync() or
3094 * cancel_work_sync() instead.
3096 void flush_scheduled_work(void)
3098 flush_workqueue(system_wq
);
3100 EXPORT_SYMBOL(flush_scheduled_work
);
3103 * execute_in_process_context - reliably execute the routine with user context
3104 * @fn: the function to execute
3105 * @ew: guaranteed storage for the execute work structure (must
3106 * be available when the work executes)
3108 * Executes the function immediately if process context is available,
3109 * otherwise schedules the function for delayed execution.
3111 * Returns: 0 - function was executed
3112 * 1 - function was scheduled for execution
3114 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3116 if (!in_interrupt()) {
3121 INIT_WORK(&ew
->work
, fn
);
3122 schedule_work(&ew
->work
);
3126 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3128 int keventd_up(void)
3130 return system_wq
!= NULL
;
3133 static int alloc_cwqs(struct workqueue_struct
*wq
)
3136 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3137 * Make sure that the alignment isn't lower than that of
3138 * unsigned long long.
3140 const size_t size
= sizeof(struct cpu_workqueue_struct
);
3141 const size_t align
= max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS
,
3142 __alignof__(unsigned long long));
3144 if (!(wq
->flags
& WQ_UNBOUND
))
3145 wq
->cpu_wq
.pcpu
= __alloc_percpu(size
, align
);
3150 * Allocate enough room to align cwq and put an extra
3151 * pointer at the end pointing back to the originally
3152 * allocated pointer which will be used for free.
3154 ptr
= kzalloc(size
+ align
+ sizeof(void *), GFP_KERNEL
);
3156 wq
->cpu_wq
.single
= PTR_ALIGN(ptr
, align
);
3157 *(void **)(wq
->cpu_wq
.single
+ 1) = ptr
;
3161 /* just in case, make sure it's actually aligned */
3162 BUG_ON(!IS_ALIGNED(wq
->cpu_wq
.v
, align
));
3163 return wq
->cpu_wq
.v
? 0 : -ENOMEM
;
3166 static void free_cwqs(struct workqueue_struct
*wq
)
3168 if (!(wq
->flags
& WQ_UNBOUND
))
3169 free_percpu(wq
->cpu_wq
.pcpu
);
3170 else if (wq
->cpu_wq
.single
) {
3171 /* the pointer to free is stored right after the cwq */
3172 kfree(*(void **)(wq
->cpu_wq
.single
+ 1));
3176 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3179 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3181 if (max_active
< 1 || max_active
> lim
)
3182 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3183 max_active
, name
, 1, lim
);
3185 return clamp_val(max_active
, 1, lim
);
3188 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3191 struct lock_class_key
*key
,
3192 const char *lock_name
, ...)
3194 va_list args
, args1
;
3195 struct workqueue_struct
*wq
;
3199 /* determine namelen, allocate wq and format name */
3200 va_start(args
, lock_name
);
3201 va_copy(args1
, args
);
3202 namelen
= vsnprintf(NULL
, 0, fmt
, args
) + 1;
3204 wq
= kzalloc(sizeof(*wq
) + namelen
, GFP_KERNEL
);
3208 vsnprintf(wq
->name
, namelen
, fmt
, args1
);
3213 * Workqueues which may be used during memory reclaim should
3214 * have a rescuer to guarantee forward progress.
3216 if (flags
& WQ_MEM_RECLAIM
)
3217 flags
|= WQ_RESCUER
;
3219 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3220 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3224 wq
->saved_max_active
= max_active
;
3225 mutex_init(&wq
->flush_mutex
);
3226 atomic_set(&wq
->nr_cwqs_to_flush
, 0);
3227 INIT_LIST_HEAD(&wq
->flusher_queue
);
3228 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3230 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3231 INIT_LIST_HEAD(&wq
->list
);
3233 if (alloc_cwqs(wq
) < 0)
3236 for_each_cwq_cpu(cpu
, wq
) {
3237 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3238 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3239 int pool_idx
= (bool)(flags
& WQ_HIGHPRI
);
3241 BUG_ON((unsigned long)cwq
& WORK_STRUCT_FLAG_MASK
);
3242 cwq
->pool
= &gcwq
->pools
[pool_idx
];
3244 cwq
->flush_color
= -1;
3245 cwq
->max_active
= max_active
;
3246 INIT_LIST_HEAD(&cwq
->delayed_works
);
3249 if (flags
& WQ_RESCUER
) {
3250 struct worker
*rescuer
;
3252 if (!alloc_mayday_mask(&wq
->mayday_mask
, GFP_KERNEL
))
3255 wq
->rescuer
= rescuer
= alloc_worker();
3259 rescuer
->rescue_wq
= wq
;
3260 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
3262 if (IS_ERR(rescuer
->task
))
3265 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
3266 wake_up_process(rescuer
->task
);
3270 * workqueue_lock protects global freeze state and workqueues
3271 * list. Grab it, set max_active accordingly and add the new
3272 * workqueue to workqueues list.
3274 spin_lock(&workqueue_lock
);
3276 if (workqueue_freezing
&& wq
->flags
& WQ_FREEZABLE
)
3277 for_each_cwq_cpu(cpu
, wq
)
3278 get_cwq(cpu
, wq
)->max_active
= 0;
3280 list_add(&wq
->list
, &workqueues
);
3282 spin_unlock(&workqueue_lock
);
3288 free_mayday_mask(wq
->mayday_mask
);
3294 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3297 * destroy_workqueue - safely terminate a workqueue
3298 * @wq: target workqueue
3300 * Safely destroy a workqueue. All work currently pending will be done first.
3302 void destroy_workqueue(struct workqueue_struct
*wq
)
3306 /* drain it before proceeding with destruction */
3307 drain_workqueue(wq
);
3310 * wq list is used to freeze wq, remove from list after
3311 * flushing is complete in case freeze races us.
3313 spin_lock(&workqueue_lock
);
3314 list_del(&wq
->list
);
3315 spin_unlock(&workqueue_lock
);
3318 for_each_cwq_cpu(cpu
, wq
) {
3319 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3322 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
3323 BUG_ON(cwq
->nr_in_flight
[i
]);
3324 BUG_ON(cwq
->nr_active
);
3325 BUG_ON(!list_empty(&cwq
->delayed_works
));
3328 if (wq
->flags
& WQ_RESCUER
) {
3329 kthread_stop(wq
->rescuer
->task
);
3330 free_mayday_mask(wq
->mayday_mask
);
3337 EXPORT_SYMBOL_GPL(destroy_workqueue
);
3340 * cwq_set_max_active - adjust max_active of a cwq
3341 * @cwq: target cpu_workqueue_struct
3342 * @max_active: new max_active value.
3344 * Set @cwq->max_active to @max_active and activate delayed works if
3348 * spin_lock_irq(gcwq->lock).
3350 static void cwq_set_max_active(struct cpu_workqueue_struct
*cwq
, int max_active
)
3352 cwq
->max_active
= max_active
;
3354 while (!list_empty(&cwq
->delayed_works
) &&
3355 cwq
->nr_active
< cwq
->max_active
)
3356 cwq_activate_first_delayed(cwq
);
3360 * workqueue_set_max_active - adjust max_active of a workqueue
3361 * @wq: target workqueue
3362 * @max_active: new max_active value.
3364 * Set max_active of @wq to @max_active.
3367 * Don't call from IRQ context.
3369 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
3373 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
3375 spin_lock(&workqueue_lock
);
3377 wq
->saved_max_active
= max_active
;
3379 for_each_cwq_cpu(cpu
, wq
) {
3380 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3382 spin_lock_irq(&gcwq
->lock
);
3384 if (!(wq
->flags
& WQ_FREEZABLE
) ||
3385 !(gcwq
->flags
& GCWQ_FREEZING
))
3386 cwq_set_max_active(get_cwq(gcwq
->cpu
, wq
), 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 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
;
3428 EXPORT_SYMBOL_GPL(work_cpu
);
3431 * work_busy - test whether a work is currently pending or running
3432 * @work: the work to be tested
3434 * Test whether @work is currently pending or running. There is no
3435 * synchronization around this function and the test result is
3436 * unreliable and only useful as advisory hints or for debugging.
3437 * Especially for reentrant wqs, the pending state might hide the
3441 * OR'd bitmask of WORK_BUSY_* bits.
3443 unsigned int work_busy(struct work_struct
*work
)
3445 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3446 unsigned long flags
;
3447 unsigned int ret
= 0;
3452 spin_lock_irqsave(&gcwq
->lock
, flags
);
3454 if (work_pending(work
))
3455 ret
|= WORK_BUSY_PENDING
;
3456 if (find_worker_executing_work(gcwq
, work
))
3457 ret
|= WORK_BUSY_RUNNING
;
3459 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
3463 EXPORT_SYMBOL_GPL(work_busy
);
3468 * There are two challenges in supporting CPU hotplug. Firstly, there
3469 * are a lot of assumptions on strong associations among work, cwq and
3470 * gcwq which make migrating pending and scheduled works very
3471 * difficult to implement without impacting hot paths. Secondly,
3472 * gcwqs serve mix of short, long and very long running works making
3473 * blocked draining impractical.
3475 * This is solved by allowing a gcwq to be disassociated from the CPU
3476 * running as an unbound one and allowing it to be reattached later if the
3477 * cpu comes back online.
3480 /* claim manager positions of all pools */
3481 static void gcwq_claim_assoc_and_lock(struct global_cwq
*gcwq
)
3483 struct worker_pool
*pool
;
3485 for_each_worker_pool(pool
, gcwq
)
3486 mutex_lock_nested(&pool
->assoc_mutex
, pool
- gcwq
->pools
);
3487 spin_lock_irq(&gcwq
->lock
);
3490 /* release manager positions */
3491 static void gcwq_release_assoc_and_unlock(struct global_cwq
*gcwq
)
3493 struct worker_pool
*pool
;
3495 spin_unlock_irq(&gcwq
->lock
);
3496 for_each_worker_pool(pool
, gcwq
)
3497 mutex_unlock(&pool
->assoc_mutex
);
3500 static void gcwq_unbind_fn(struct work_struct
*work
)
3502 struct global_cwq
*gcwq
= get_gcwq(smp_processor_id());
3503 struct worker_pool
*pool
;
3504 struct worker
*worker
;
3505 struct hlist_node
*pos
;
3508 BUG_ON(gcwq
->cpu
!= smp_processor_id());
3510 gcwq_claim_assoc_and_lock(gcwq
);
3513 * We've claimed all manager positions. Make all workers unbound
3514 * and set DISASSOCIATED. Before this, all workers except for the
3515 * ones which are still executing works from before the last CPU
3516 * down must be on the cpu. After this, they may become diasporas.
3518 for_each_worker_pool(pool
, gcwq
)
3519 list_for_each_entry(worker
, &pool
->idle_list
, entry
)
3520 worker
->flags
|= WORKER_UNBOUND
;
3522 for_each_busy_worker(worker
, i
, pos
, gcwq
)
3523 worker
->flags
|= WORKER_UNBOUND
;
3525 gcwq
->flags
|= GCWQ_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 gcwq
->flags
&= ~GCWQ_DISASSOCIATED
;
3585 rebind_workers(gcwq
);
3586 gcwq_release_assoc_and_unlock(gcwq
);
3593 * Workqueues should be brought down after normal priority CPU notifiers.
3594 * This will be registered as low priority CPU notifier.
3596 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
3597 unsigned long action
,
3600 unsigned int cpu
= (unsigned long)hcpu
;
3601 struct work_struct unbind_work
;
3603 switch (action
& ~CPU_TASKS_FROZEN
) {
3604 case CPU_DOWN_PREPARE
:
3605 /* unbinding should happen on the local CPU */
3606 INIT_WORK_ONSTACK(&unbind_work
, gcwq_unbind_fn
);
3607 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
3608 flush_work(&unbind_work
);
3616 struct work_for_cpu
{
3617 struct work_struct work
;
3623 static void work_for_cpu_fn(struct work_struct
*work
)
3625 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
3627 wfc
->ret
= wfc
->fn(wfc
->arg
);
3631 * work_on_cpu - run a function in user context on a particular cpu
3632 * @cpu: the cpu to run on
3633 * @fn: the function to run
3634 * @arg: the function arg
3636 * This will return the value @fn returns.
3637 * It is up to the caller to ensure that the cpu doesn't go offline.
3638 * The caller must not hold any locks which would prevent @fn from completing.
3640 long work_on_cpu(unsigned int cpu
, long (*fn
)(void *), void *arg
)
3642 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
3644 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
3645 schedule_work_on(cpu
, &wfc
.work
);
3646 flush_work(&wfc
.work
);
3649 EXPORT_SYMBOL_GPL(work_on_cpu
);
3650 #endif /* CONFIG_SMP */
3652 #ifdef CONFIG_FREEZER
3655 * freeze_workqueues_begin - begin freezing workqueues
3657 * Start freezing workqueues. After this function returns, all freezable
3658 * workqueues will queue new works to their frozen_works list instead of
3662 * Grabs and releases workqueue_lock and gcwq->lock's.
3664 void freeze_workqueues_begin(void)
3668 spin_lock(&workqueue_lock
);
3670 BUG_ON(workqueue_freezing
);
3671 workqueue_freezing
= true;
3673 for_each_gcwq_cpu(cpu
) {
3674 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3675 struct workqueue_struct
*wq
;
3677 spin_lock_irq(&gcwq
->lock
);
3679 BUG_ON(gcwq
->flags
& GCWQ_FREEZING
);
3680 gcwq
->flags
|= GCWQ_FREEZING
;
3682 list_for_each_entry(wq
, &workqueues
, list
) {
3683 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3685 if (cwq
&& wq
->flags
& WQ_FREEZABLE
)
3686 cwq
->max_active
= 0;
3689 spin_unlock_irq(&gcwq
->lock
);
3692 spin_unlock(&workqueue_lock
);
3696 * freeze_workqueues_busy - are freezable workqueues still busy?
3698 * Check whether freezing is complete. This function must be called
3699 * between freeze_workqueues_begin() and thaw_workqueues().
3702 * Grabs and releases workqueue_lock.
3705 * %true if some freezable workqueues are still busy. %false if freezing
3708 bool freeze_workqueues_busy(void)
3713 spin_lock(&workqueue_lock
);
3715 BUG_ON(!workqueue_freezing
);
3717 for_each_gcwq_cpu(cpu
) {
3718 struct workqueue_struct
*wq
;
3720 * nr_active is monotonically decreasing. It's safe
3721 * to peek without lock.
3723 list_for_each_entry(wq
, &workqueues
, list
) {
3724 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3726 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3729 BUG_ON(cwq
->nr_active
< 0);
3730 if (cwq
->nr_active
) {
3737 spin_unlock(&workqueue_lock
);
3742 * thaw_workqueues - thaw workqueues
3744 * Thaw workqueues. Normal queueing is restored and all collected
3745 * frozen works are transferred to their respective gcwq worklists.
3748 * Grabs and releases workqueue_lock and gcwq->lock's.
3750 void thaw_workqueues(void)
3754 spin_lock(&workqueue_lock
);
3756 if (!workqueue_freezing
)
3759 for_each_gcwq_cpu(cpu
) {
3760 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3761 struct worker_pool
*pool
;
3762 struct workqueue_struct
*wq
;
3764 spin_lock_irq(&gcwq
->lock
);
3766 BUG_ON(!(gcwq
->flags
& GCWQ_FREEZING
));
3767 gcwq
->flags
&= ~GCWQ_FREEZING
;
3769 list_for_each_entry(wq
, &workqueues
, list
) {
3770 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3772 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3775 /* restore max_active and repopulate worklist */
3776 cwq_set_max_active(cwq
, wq
->saved_max_active
);
3779 for_each_worker_pool(pool
, gcwq
)
3780 wake_up_worker(pool
);
3782 spin_unlock_irq(&gcwq
->lock
);
3785 workqueue_freezing
= false;
3787 spin_unlock(&workqueue_lock
);
3789 #endif /* CONFIG_FREEZER */
3791 static int __init
init_workqueues(void)
3795 /* make sure we have enough bits for OFFQ CPU number */
3796 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_CPU_SHIFT
)) <
3799 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
3800 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
3802 /* initialize gcwqs */
3803 for_each_gcwq_cpu(cpu
) {
3804 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3805 struct worker_pool
*pool
;
3807 spin_lock_init(&gcwq
->lock
);
3809 gcwq
->flags
|= GCWQ_DISASSOCIATED
;
3811 hash_init(gcwq
->busy_hash
);
3813 for_each_worker_pool(pool
, gcwq
) {
3815 INIT_LIST_HEAD(&pool
->worklist
);
3816 INIT_LIST_HEAD(&pool
->idle_list
);
3818 init_timer_deferrable(&pool
->idle_timer
);
3819 pool
->idle_timer
.function
= idle_worker_timeout
;
3820 pool
->idle_timer
.data
= (unsigned long)pool
;
3822 setup_timer(&pool
->mayday_timer
, gcwq_mayday_timeout
,
3823 (unsigned long)pool
);
3825 mutex_init(&pool
->assoc_mutex
);
3826 ida_init(&pool
->worker_ida
);
3830 /* create the initial worker */
3831 for_each_online_gcwq_cpu(cpu
) {
3832 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3833 struct worker_pool
*pool
;
3835 if (cpu
!= WORK_CPU_UNBOUND
)
3836 gcwq
->flags
&= ~GCWQ_DISASSOCIATED
;
3838 for_each_worker_pool(pool
, gcwq
) {
3839 struct worker
*worker
;
3841 worker
= create_worker(pool
);
3843 spin_lock_irq(&gcwq
->lock
);
3844 start_worker(worker
);
3845 spin_unlock_irq(&gcwq
->lock
);
3849 system_wq
= alloc_workqueue("events", 0, 0);
3850 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
3851 system_long_wq
= alloc_workqueue("events_long", 0, 0);
3852 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
3853 WQ_UNBOUND_MAX_ACTIVE
);
3854 system_freezable_wq
= alloc_workqueue("events_freezable",
3856 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
3857 !system_unbound_wq
|| !system_freezable_wq
);
3860 early_initcall(init_workqueues
);