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>
45 #include "workqueue_sched.h"
48 /* global_cwq flags */
49 GCWQ_DISASSOCIATED
= 1 << 0, /* cpu can't serve workers */
50 GCWQ_FREEZING
= 1 << 1, /* freeze in progress */
53 POOL_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
56 WORKER_STARTED
= 1 << 0, /* started */
57 WORKER_DIE
= 1 << 1, /* die die die */
58 WORKER_IDLE
= 1 << 2, /* is idle */
59 WORKER_PREP
= 1 << 3, /* preparing to run works */
60 WORKER_REBIND
= 1 << 5, /* mom is home, come back */
61 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
62 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
64 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_REBIND
| WORKER_UNBOUND
|
67 /* gcwq->trustee_state */
68 TRUSTEE_START
= 0, /* start */
69 TRUSTEE_IN_CHARGE
= 1, /* trustee in charge of gcwq */
70 TRUSTEE_BUTCHER
= 2, /* butcher workers */
71 TRUSTEE_RELEASE
= 3, /* release workers */
72 TRUSTEE_DONE
= 4, /* trustee is done */
74 NR_WORKER_POOLS
= 2, /* # worker pools per gcwq */
76 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
77 BUSY_WORKER_HASH_SIZE
= 1 << BUSY_WORKER_HASH_ORDER
,
78 BUSY_WORKER_HASH_MASK
= BUSY_WORKER_HASH_SIZE
- 1,
80 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
81 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
83 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
84 /* call for help after 10ms
86 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
87 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
88 TRUSTEE_COOLDOWN
= HZ
/ 10, /* for trustee draining */
91 * Rescue workers are used only on emergencies and shared by
94 RESCUER_NICE_LEVEL
= -20,
95 HIGHPRI_NICE_LEVEL
= -20,
99 * Structure fields follow one of the following exclusion rules.
101 * I: Modifiable by initialization/destruction paths and read-only for
104 * P: Preemption protected. Disabling preemption is enough and should
105 * only be modified and accessed from the local cpu.
107 * L: gcwq->lock protected. Access with gcwq->lock held.
109 * X: During normal operation, modification requires gcwq->lock and
110 * should be done only from local cpu. Either disabling preemption
111 * on local cpu or grabbing gcwq->lock is enough for read access.
112 * If GCWQ_DISASSOCIATED is set, it's identical to L.
114 * F: wq->flush_mutex protected.
116 * W: workqueue_lock protected.
123 * The poor guys doing the actual heavy lifting. All on-duty workers
124 * are either serving the manager role, on idle list or on busy hash.
127 /* on idle list while idle, on busy hash table while busy */
129 struct list_head entry
; /* L: while idle */
130 struct hlist_node hentry
; /* L: while busy */
133 struct work_struct
*current_work
; /* L: work being processed */
134 struct cpu_workqueue_struct
*current_cwq
; /* L: current_work's cwq */
135 struct list_head scheduled
; /* L: scheduled works */
136 struct task_struct
*task
; /* I: worker task */
137 struct worker_pool
*pool
; /* I: the associated pool */
138 /* 64 bytes boundary on 64bit, 32 on 32bit */
139 unsigned long last_active
; /* L: last active timestamp */
140 unsigned int flags
; /* X: flags */
141 int id
; /* I: worker id */
142 struct work_struct rebind_work
; /* L: rebind worker to cpu */
146 struct global_cwq
*gcwq
; /* I: the owning gcwq */
147 unsigned int flags
; /* X: flags */
149 struct list_head worklist
; /* L: list of pending works */
150 int nr_workers
; /* L: total number of workers */
151 int nr_idle
; /* L: currently idle ones */
153 struct list_head idle_list
; /* X: list of idle workers */
154 struct timer_list idle_timer
; /* L: worker idle timeout */
155 struct timer_list mayday_timer
; /* L: SOS timer for workers */
157 struct mutex manager_mutex
; /* mutex manager should hold */
158 struct ida worker_ida
; /* L: for worker IDs */
159 struct worker
*first_idle
; /* L: first idle worker */
163 * Global per-cpu workqueue. There's one and only one for each cpu
164 * and all works are queued and processed here regardless of their
168 spinlock_t lock
; /* the gcwq lock */
169 unsigned int cpu
; /* I: the associated cpu */
170 unsigned int flags
; /* L: GCWQ_* flags */
172 /* workers are chained either in busy_hash or pool idle_list */
173 struct hlist_head busy_hash
[BUSY_WORKER_HASH_SIZE
];
174 /* L: hash of busy workers */
176 struct worker_pool pools
[2]; /* normal and highpri pools */
178 struct task_struct
*trustee
; /* L: for gcwq shutdown */
179 unsigned int trustee_state
; /* L: trustee state */
180 wait_queue_head_t trustee_wait
; /* trustee wait */
181 } ____cacheline_aligned_in_smp
;
184 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
185 * work_struct->data are used for flags and thus cwqs need to be
186 * aligned at two's power of the number of flag bits.
188 struct cpu_workqueue_struct
{
189 struct worker_pool
*pool
; /* I: the associated pool */
190 struct workqueue_struct
*wq
; /* I: the owning workqueue */
191 int work_color
; /* L: current color */
192 int flush_color
; /* L: flushing color */
193 int nr_in_flight
[WORK_NR_COLORS
];
194 /* L: nr of in_flight works */
195 int nr_active
; /* L: nr of active works */
196 int max_active
; /* L: max active works */
197 struct list_head delayed_works
; /* L: delayed works */
201 * Structure used to wait for workqueue flush.
204 struct list_head list
; /* F: list of flushers */
205 int flush_color
; /* F: flush color waiting for */
206 struct completion done
; /* flush completion */
210 * All cpumasks are assumed to be always set on UP and thus can't be
211 * used to determine whether there's something to be done.
214 typedef cpumask_var_t mayday_mask_t
;
215 #define mayday_test_and_set_cpu(cpu, mask) \
216 cpumask_test_and_set_cpu((cpu), (mask))
217 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
218 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
219 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
220 #define free_mayday_mask(mask) free_cpumask_var((mask))
222 typedef unsigned long mayday_mask_t
;
223 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
224 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
225 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
226 #define alloc_mayday_mask(maskp, gfp) true
227 #define free_mayday_mask(mask) do { } while (0)
231 * The externally visible workqueue abstraction is an array of
232 * per-CPU workqueues:
234 struct workqueue_struct
{
235 unsigned int flags
; /* W: WQ_* flags */
237 struct cpu_workqueue_struct __percpu
*pcpu
;
238 struct cpu_workqueue_struct
*single
;
240 } cpu_wq
; /* I: cwq's */
241 struct list_head list
; /* W: list of all workqueues */
243 struct mutex flush_mutex
; /* protects wq flushing */
244 int work_color
; /* F: current work color */
245 int flush_color
; /* F: current flush color */
246 atomic_t nr_cwqs_to_flush
; /* flush in progress */
247 struct wq_flusher
*first_flusher
; /* F: first flusher */
248 struct list_head flusher_queue
; /* F: flush waiters */
249 struct list_head flusher_overflow
; /* F: flush overflow list */
251 mayday_mask_t mayday_mask
; /* cpus requesting rescue */
252 struct worker
*rescuer
; /* I: rescue worker */
254 int nr_drainers
; /* W: drain in progress */
255 int saved_max_active
; /* W: saved cwq max_active */
256 #ifdef CONFIG_LOCKDEP
257 struct lockdep_map lockdep_map
;
259 char name
[]; /* I: workqueue name */
262 struct workqueue_struct
*system_wq __read_mostly
;
263 struct workqueue_struct
*system_long_wq __read_mostly
;
264 struct workqueue_struct
*system_nrt_wq __read_mostly
;
265 struct workqueue_struct
*system_unbound_wq __read_mostly
;
266 struct workqueue_struct
*system_freezable_wq __read_mostly
;
267 struct workqueue_struct
*system_nrt_freezable_wq __read_mostly
;
268 EXPORT_SYMBOL_GPL(system_wq
);
269 EXPORT_SYMBOL_GPL(system_long_wq
);
270 EXPORT_SYMBOL_GPL(system_nrt_wq
);
271 EXPORT_SYMBOL_GPL(system_unbound_wq
);
272 EXPORT_SYMBOL_GPL(system_freezable_wq
);
273 EXPORT_SYMBOL_GPL(system_nrt_freezable_wq
);
275 #define CREATE_TRACE_POINTS
276 #include <trace/events/workqueue.h>
278 #define for_each_worker_pool(pool, gcwq) \
279 for ((pool) = &(gcwq)->pools[0]; \
280 (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
282 #define for_each_busy_worker(worker, i, pos, gcwq) \
283 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
284 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
286 static inline int __next_gcwq_cpu(int cpu
, const struct cpumask
*mask
,
289 if (cpu
< nr_cpu_ids
) {
291 cpu
= cpumask_next(cpu
, mask
);
292 if (cpu
< nr_cpu_ids
)
296 return WORK_CPU_UNBOUND
;
298 return WORK_CPU_NONE
;
301 static inline int __next_wq_cpu(int cpu
, const struct cpumask
*mask
,
302 struct workqueue_struct
*wq
)
304 return __next_gcwq_cpu(cpu
, mask
, !(wq
->flags
& WQ_UNBOUND
) ? 1 : 2);
310 * An extra gcwq is defined for an invalid cpu number
311 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
312 * specific CPU. The following iterators are similar to
313 * for_each_*_cpu() iterators but also considers the unbound gcwq.
315 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
316 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
317 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
318 * WORK_CPU_UNBOUND for unbound workqueues
320 #define for_each_gcwq_cpu(cpu) \
321 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
322 (cpu) < WORK_CPU_NONE; \
323 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
325 #define for_each_online_gcwq_cpu(cpu) \
326 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
327 (cpu) < WORK_CPU_NONE; \
328 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
330 #define for_each_cwq_cpu(cpu, wq) \
331 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
332 (cpu) < WORK_CPU_NONE; \
333 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
335 #ifdef CONFIG_DEBUG_OBJECTS_WORK
337 static struct debug_obj_descr work_debug_descr
;
339 static void *work_debug_hint(void *addr
)
341 return ((struct work_struct
*) addr
)->func
;
345 * fixup_init is called when:
346 * - an active object is initialized
348 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
350 struct work_struct
*work
= addr
;
353 case ODEBUG_STATE_ACTIVE
:
354 cancel_work_sync(work
);
355 debug_object_init(work
, &work_debug_descr
);
363 * fixup_activate is called when:
364 * - an active object is activated
365 * - an unknown object is activated (might be a statically initialized object)
367 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
369 struct work_struct
*work
= addr
;
373 case ODEBUG_STATE_NOTAVAILABLE
:
375 * This is not really a fixup. The work struct was
376 * statically initialized. We just make sure that it
377 * is tracked in the object tracker.
379 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
380 debug_object_init(work
, &work_debug_descr
);
381 debug_object_activate(work
, &work_debug_descr
);
387 case ODEBUG_STATE_ACTIVE
:
396 * fixup_free is called when:
397 * - an active object is freed
399 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
401 struct work_struct
*work
= addr
;
404 case ODEBUG_STATE_ACTIVE
:
405 cancel_work_sync(work
);
406 debug_object_free(work
, &work_debug_descr
);
413 static struct debug_obj_descr work_debug_descr
= {
414 .name
= "work_struct",
415 .debug_hint
= work_debug_hint
,
416 .fixup_init
= work_fixup_init
,
417 .fixup_activate
= work_fixup_activate
,
418 .fixup_free
= work_fixup_free
,
421 static inline void debug_work_activate(struct work_struct
*work
)
423 debug_object_activate(work
, &work_debug_descr
);
426 static inline void debug_work_deactivate(struct work_struct
*work
)
428 debug_object_deactivate(work
, &work_debug_descr
);
431 void __init_work(struct work_struct
*work
, int onstack
)
434 debug_object_init_on_stack(work
, &work_debug_descr
);
436 debug_object_init(work
, &work_debug_descr
);
438 EXPORT_SYMBOL_GPL(__init_work
);
440 void destroy_work_on_stack(struct work_struct
*work
)
442 debug_object_free(work
, &work_debug_descr
);
444 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
447 static inline void debug_work_activate(struct work_struct
*work
) { }
448 static inline void debug_work_deactivate(struct work_struct
*work
) { }
451 /* Serializes the accesses to the list of workqueues. */
452 static DEFINE_SPINLOCK(workqueue_lock
);
453 static LIST_HEAD(workqueues
);
454 static bool workqueue_freezing
; /* W: have wqs started freezing? */
457 * The almighty global cpu workqueues. nr_running is the only field
458 * which is expected to be used frequently by other cpus via
459 * try_to_wake_up(). Put it in a separate cacheline.
461 static DEFINE_PER_CPU(struct global_cwq
, global_cwq
);
462 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t
, pool_nr_running
[NR_WORKER_POOLS
]);
465 * Global cpu workqueue and nr_running counter for unbound gcwq. The
466 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
467 * workers have WORKER_UNBOUND set.
469 static struct global_cwq unbound_global_cwq
;
470 static atomic_t unbound_pool_nr_running
[NR_WORKER_POOLS
] = {
471 [0 ... NR_WORKER_POOLS
- 1] = ATOMIC_INIT(0), /* always 0 */
474 static int worker_thread(void *__worker
);
476 static int worker_pool_pri(struct worker_pool
*pool
)
478 return pool
- pool
->gcwq
->pools
;
481 static struct global_cwq
*get_gcwq(unsigned int cpu
)
483 if (cpu
!= WORK_CPU_UNBOUND
)
484 return &per_cpu(global_cwq
, cpu
);
486 return &unbound_global_cwq
;
489 static atomic_t
*get_pool_nr_running(struct worker_pool
*pool
)
491 int cpu
= pool
->gcwq
->cpu
;
492 int idx
= worker_pool_pri(pool
);
494 if (cpu
!= WORK_CPU_UNBOUND
)
495 return &per_cpu(pool_nr_running
, cpu
)[idx
];
497 return &unbound_pool_nr_running
[idx
];
500 static struct cpu_workqueue_struct
*get_cwq(unsigned int cpu
,
501 struct workqueue_struct
*wq
)
503 if (!(wq
->flags
& WQ_UNBOUND
)) {
504 if (likely(cpu
< nr_cpu_ids
))
505 return per_cpu_ptr(wq
->cpu_wq
.pcpu
, cpu
);
506 } else if (likely(cpu
== WORK_CPU_UNBOUND
))
507 return wq
->cpu_wq
.single
;
511 static unsigned int work_color_to_flags(int color
)
513 return color
<< WORK_STRUCT_COLOR_SHIFT
;
516 static int get_work_color(struct work_struct
*work
)
518 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
519 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
522 static int work_next_color(int color
)
524 return (color
+ 1) % WORK_NR_COLORS
;
528 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
529 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
530 * cleared and the work data contains the cpu number it was last on.
532 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
533 * cwq, cpu or clear work->data. These functions should only be
534 * called while the work is owned - ie. while the PENDING bit is set.
536 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
537 * corresponding to a work. gcwq is available once the work has been
538 * queued anywhere after initialization. cwq is available only from
539 * queueing until execution starts.
541 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
544 BUG_ON(!work_pending(work
));
545 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
548 static void set_work_cwq(struct work_struct
*work
,
549 struct cpu_workqueue_struct
*cwq
,
550 unsigned long extra_flags
)
552 set_work_data(work
, (unsigned long)cwq
,
553 WORK_STRUCT_PENDING
| WORK_STRUCT_CWQ
| extra_flags
);
556 static void set_work_cpu(struct work_struct
*work
, unsigned int cpu
)
558 set_work_data(work
, cpu
<< WORK_STRUCT_FLAG_BITS
, WORK_STRUCT_PENDING
);
561 static void clear_work_data(struct work_struct
*work
)
563 set_work_data(work
, WORK_STRUCT_NO_CPU
, 0);
566 static struct cpu_workqueue_struct
*get_work_cwq(struct work_struct
*work
)
568 unsigned long data
= atomic_long_read(&work
->data
);
570 if (data
& WORK_STRUCT_CWQ
)
571 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
576 static struct global_cwq
*get_work_gcwq(struct work_struct
*work
)
578 unsigned long data
= atomic_long_read(&work
->data
);
581 if (data
& WORK_STRUCT_CWQ
)
582 return ((struct cpu_workqueue_struct
*)
583 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->gcwq
;
585 cpu
= data
>> WORK_STRUCT_FLAG_BITS
;
586 if (cpu
== WORK_CPU_NONE
)
589 BUG_ON(cpu
>= nr_cpu_ids
&& cpu
!= WORK_CPU_UNBOUND
);
590 return get_gcwq(cpu
);
594 * Policy functions. These define the policies on how the global worker
595 * pools are managed. Unless noted otherwise, these functions assume that
596 * they're being called with gcwq->lock held.
599 static bool __need_more_worker(struct worker_pool
*pool
)
601 return !atomic_read(get_pool_nr_running(pool
));
605 * Need to wake up a worker? Called from anything but currently
608 * Note that, because unbound workers never contribute to nr_running, this
609 * function will always return %true for unbound gcwq as long as the
610 * worklist isn't empty.
612 static bool need_more_worker(struct worker_pool
*pool
)
614 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
617 /* Can I start working? Called from busy but !running workers. */
618 static bool may_start_working(struct worker_pool
*pool
)
620 return pool
->nr_idle
;
623 /* Do I need to keep working? Called from currently running workers. */
624 static bool keep_working(struct worker_pool
*pool
)
626 atomic_t
*nr_running
= get_pool_nr_running(pool
);
628 return !list_empty(&pool
->worklist
) && atomic_read(nr_running
) <= 1;
631 /* Do we need a new worker? Called from manager. */
632 static bool need_to_create_worker(struct worker_pool
*pool
)
634 return need_more_worker(pool
) && !may_start_working(pool
);
637 /* Do I need to be the manager? */
638 static bool need_to_manage_workers(struct worker_pool
*pool
)
640 return need_to_create_worker(pool
) ||
641 (pool
->flags
& POOL_MANAGE_WORKERS
);
644 /* Do we have too many workers and should some go away? */
645 static bool too_many_workers(struct worker_pool
*pool
)
647 bool managing
= mutex_is_locked(&pool
->manager_mutex
);
648 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
649 int nr_busy
= pool
->nr_workers
- nr_idle
;
651 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
658 /* Return the first worker. Safe with preemption disabled */
659 static struct worker
*first_worker(struct worker_pool
*pool
)
661 if (unlikely(list_empty(&pool
->idle_list
)))
664 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
668 * wake_up_worker - wake up an idle worker
669 * @pool: worker pool to wake worker from
671 * Wake up the first idle worker of @pool.
674 * spin_lock_irq(gcwq->lock).
676 static void wake_up_worker(struct worker_pool
*pool
)
678 struct worker
*worker
= first_worker(pool
);
681 wake_up_process(worker
->task
);
685 * wq_worker_waking_up - a worker is waking up
686 * @task: task waking up
687 * @cpu: CPU @task is waking up to
689 * This function is called during try_to_wake_up() when a worker is
693 * spin_lock_irq(rq->lock)
695 void wq_worker_waking_up(struct task_struct
*task
, unsigned int cpu
)
697 struct worker
*worker
= kthread_data(task
);
699 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
700 atomic_inc(get_pool_nr_running(worker
->pool
));
704 * wq_worker_sleeping - a worker is going to sleep
705 * @task: task going to sleep
706 * @cpu: CPU in question, must be the current CPU number
708 * This function is called during schedule() when a busy worker is
709 * going to sleep. Worker on the same cpu can be woken up by
710 * returning pointer to its task.
713 * spin_lock_irq(rq->lock)
716 * Worker task on @cpu to wake up, %NULL if none.
718 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
,
721 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
722 struct worker_pool
*pool
= worker
->pool
;
723 atomic_t
*nr_running
= get_pool_nr_running(pool
);
725 if (worker
->flags
& WORKER_NOT_RUNNING
)
728 /* this can only happen on the local cpu */
729 BUG_ON(cpu
!= raw_smp_processor_id());
732 * The counterpart of the following dec_and_test, implied mb,
733 * worklist not empty test sequence is in insert_work().
734 * Please read comment there.
736 * NOT_RUNNING is clear. This means that trustee is not in
737 * charge and we're running on the local cpu w/ rq lock held
738 * and preemption disabled, which in turn means that none else
739 * could be manipulating idle_list, so dereferencing idle_list
740 * without gcwq lock is safe.
742 if (atomic_dec_and_test(nr_running
) && !list_empty(&pool
->worklist
))
743 to_wakeup
= first_worker(pool
);
744 return to_wakeup
? to_wakeup
->task
: NULL
;
748 * worker_set_flags - set worker flags and adjust nr_running accordingly
750 * @flags: flags to set
751 * @wakeup: wakeup an idle worker if necessary
753 * Set @flags in @worker->flags and adjust nr_running accordingly. If
754 * nr_running becomes zero and @wakeup is %true, an idle worker is
758 * spin_lock_irq(gcwq->lock)
760 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
763 struct worker_pool
*pool
= worker
->pool
;
765 WARN_ON_ONCE(worker
->task
!= current
);
768 * If transitioning into NOT_RUNNING, adjust nr_running and
769 * wake up an idle worker as necessary if requested by
772 if ((flags
& WORKER_NOT_RUNNING
) &&
773 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
774 atomic_t
*nr_running
= get_pool_nr_running(pool
);
777 if (atomic_dec_and_test(nr_running
) &&
778 !list_empty(&pool
->worklist
))
779 wake_up_worker(pool
);
781 atomic_dec(nr_running
);
784 worker
->flags
|= flags
;
788 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
790 * @flags: flags to clear
792 * Clear @flags in @worker->flags and adjust nr_running accordingly.
795 * spin_lock_irq(gcwq->lock)
797 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
799 struct worker_pool
*pool
= worker
->pool
;
800 unsigned int oflags
= worker
->flags
;
802 WARN_ON_ONCE(worker
->task
!= current
);
804 worker
->flags
&= ~flags
;
807 * If transitioning out of NOT_RUNNING, increment nr_running. Note
808 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
809 * of multiple flags, not a single flag.
811 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
812 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
813 atomic_inc(get_pool_nr_running(pool
));
817 * busy_worker_head - return the busy hash head for a work
818 * @gcwq: gcwq of interest
819 * @work: work to be hashed
821 * Return hash head of @gcwq for @work.
824 * spin_lock_irq(gcwq->lock).
827 * Pointer to the hash head.
829 static struct hlist_head
*busy_worker_head(struct global_cwq
*gcwq
,
830 struct work_struct
*work
)
832 const int base_shift
= ilog2(sizeof(struct work_struct
));
833 unsigned long v
= (unsigned long)work
;
835 /* simple shift and fold hash, do we need something better? */
837 v
+= v
>> BUSY_WORKER_HASH_ORDER
;
838 v
&= BUSY_WORKER_HASH_MASK
;
840 return &gcwq
->busy_hash
[v
];
844 * __find_worker_executing_work - find worker which is executing a work
845 * @gcwq: gcwq of interest
846 * @bwh: hash head as returned by busy_worker_head()
847 * @work: work to find worker for
849 * Find a worker which is executing @work on @gcwq. @bwh should be
850 * the hash head obtained by calling busy_worker_head() with the same
854 * spin_lock_irq(gcwq->lock).
857 * Pointer to worker which is executing @work if found, NULL
860 static struct worker
*__find_worker_executing_work(struct global_cwq
*gcwq
,
861 struct hlist_head
*bwh
,
862 struct work_struct
*work
)
864 struct worker
*worker
;
865 struct hlist_node
*tmp
;
867 hlist_for_each_entry(worker
, tmp
, bwh
, hentry
)
868 if (worker
->current_work
== work
)
874 * find_worker_executing_work - find worker which is executing a work
875 * @gcwq: gcwq of interest
876 * @work: work to find worker for
878 * Find a worker which is executing @work on @gcwq. This function is
879 * identical to __find_worker_executing_work() except that this
880 * function calculates @bwh itself.
883 * spin_lock_irq(gcwq->lock).
886 * Pointer to worker which is executing @work if found, NULL
889 static struct worker
*find_worker_executing_work(struct global_cwq
*gcwq
,
890 struct work_struct
*work
)
892 return __find_worker_executing_work(gcwq
, busy_worker_head(gcwq
, work
),
897 * insert_work - insert a work into gcwq
898 * @cwq: cwq @work belongs to
899 * @work: work to insert
900 * @head: insertion point
901 * @extra_flags: extra WORK_STRUCT_* flags to set
903 * Insert @work which belongs to @cwq into @gcwq after @head.
904 * @extra_flags is or'd to work_struct flags.
907 * spin_lock_irq(gcwq->lock).
909 static void insert_work(struct cpu_workqueue_struct
*cwq
,
910 struct work_struct
*work
, struct list_head
*head
,
911 unsigned int extra_flags
)
913 struct worker_pool
*pool
= cwq
->pool
;
915 /* we own @work, set data and link */
916 set_work_cwq(work
, cwq
, extra_flags
);
919 * Ensure that we get the right work->data if we see the
920 * result of list_add() below, see try_to_grab_pending().
924 list_add_tail(&work
->entry
, head
);
927 * Ensure either worker_sched_deactivated() sees the above
928 * list_add_tail() or we see zero nr_running to avoid workers
929 * lying around lazily while there are works to be processed.
933 if (__need_more_worker(pool
))
934 wake_up_worker(pool
);
938 * Test whether @work is being queued from another work executing on the
939 * same workqueue. This is rather expensive and should only be used from
942 static bool is_chained_work(struct workqueue_struct
*wq
)
947 for_each_gcwq_cpu(cpu
) {
948 struct global_cwq
*gcwq
= get_gcwq(cpu
);
949 struct worker
*worker
;
950 struct hlist_node
*pos
;
953 spin_lock_irqsave(&gcwq
->lock
, flags
);
954 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
955 if (worker
->task
!= current
)
957 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
959 * I'm @worker, no locking necessary. See if @work
960 * is headed to the same workqueue.
962 return worker
->current_cwq
->wq
== wq
;
964 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
969 static void __queue_work(unsigned int cpu
, struct workqueue_struct
*wq
,
970 struct work_struct
*work
)
972 struct global_cwq
*gcwq
;
973 struct cpu_workqueue_struct
*cwq
;
974 struct list_head
*worklist
;
975 unsigned int work_flags
;
978 debug_work_activate(work
);
980 /* if dying, only works from the same workqueue are allowed */
981 if (unlikely(wq
->flags
& WQ_DRAINING
) &&
982 WARN_ON_ONCE(!is_chained_work(wq
)))
985 /* determine gcwq to use */
986 if (!(wq
->flags
& WQ_UNBOUND
)) {
987 struct global_cwq
*last_gcwq
;
989 if (unlikely(cpu
== WORK_CPU_UNBOUND
))
990 cpu
= raw_smp_processor_id();
993 * It's multi cpu. If @wq is non-reentrant and @work
994 * was previously on a different cpu, it might still
995 * be running there, in which case the work needs to
996 * be queued on that cpu to guarantee non-reentrance.
998 gcwq
= get_gcwq(cpu
);
999 if (wq
->flags
& WQ_NON_REENTRANT
&&
1000 (last_gcwq
= get_work_gcwq(work
)) && last_gcwq
!= gcwq
) {
1001 struct worker
*worker
;
1003 spin_lock_irqsave(&last_gcwq
->lock
, flags
);
1005 worker
= find_worker_executing_work(last_gcwq
, work
);
1007 if (worker
&& worker
->current_cwq
->wq
== wq
)
1010 /* meh... not running there, queue here */
1011 spin_unlock_irqrestore(&last_gcwq
->lock
, flags
);
1012 spin_lock_irqsave(&gcwq
->lock
, flags
);
1015 spin_lock_irqsave(&gcwq
->lock
, flags
);
1017 gcwq
= get_gcwq(WORK_CPU_UNBOUND
);
1018 spin_lock_irqsave(&gcwq
->lock
, flags
);
1021 /* gcwq determined, get cwq and queue */
1022 cwq
= get_cwq(gcwq
->cpu
, wq
);
1023 trace_workqueue_queue_work(cpu
, cwq
, work
);
1025 if (WARN_ON(!list_empty(&work
->entry
))) {
1026 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1030 cwq
->nr_in_flight
[cwq
->work_color
]++;
1031 work_flags
= work_color_to_flags(cwq
->work_color
);
1033 if (likely(cwq
->nr_active
< cwq
->max_active
)) {
1034 trace_workqueue_activate_work(work
);
1036 worklist
= &cwq
->pool
->worklist
;
1038 work_flags
|= WORK_STRUCT_DELAYED
;
1039 worklist
= &cwq
->delayed_works
;
1042 insert_work(cwq
, work
, worklist
, work_flags
);
1044 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1048 * queue_work - queue work on a workqueue
1049 * @wq: workqueue to use
1050 * @work: work to queue
1052 * Returns 0 if @work was already on a queue, non-zero otherwise.
1054 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1055 * it can be processed by another CPU.
1057 int queue_work(struct workqueue_struct
*wq
, struct work_struct
*work
)
1061 ret
= queue_work_on(get_cpu(), wq
, work
);
1066 EXPORT_SYMBOL_GPL(queue_work
);
1069 * queue_work_on - queue work on specific cpu
1070 * @cpu: CPU number to execute work on
1071 * @wq: workqueue to use
1072 * @work: work to queue
1074 * Returns 0 if @work was already on a queue, non-zero otherwise.
1076 * We queue the work to a specific CPU, the caller must ensure it
1080 queue_work_on(int cpu
, struct workqueue_struct
*wq
, struct work_struct
*work
)
1084 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1085 __queue_work(cpu
, wq
, work
);
1090 EXPORT_SYMBOL_GPL(queue_work_on
);
1092 static void delayed_work_timer_fn(unsigned long __data
)
1094 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1095 struct cpu_workqueue_struct
*cwq
= get_work_cwq(&dwork
->work
);
1097 __queue_work(smp_processor_id(), cwq
->wq
, &dwork
->work
);
1101 * queue_delayed_work - queue work on a workqueue after delay
1102 * @wq: workqueue to use
1103 * @dwork: delayable work to queue
1104 * @delay: number of jiffies to wait before queueing
1106 * Returns 0 if @work was already on a queue, non-zero otherwise.
1108 int queue_delayed_work(struct workqueue_struct
*wq
,
1109 struct delayed_work
*dwork
, unsigned long delay
)
1112 return queue_work(wq
, &dwork
->work
);
1114 return queue_delayed_work_on(-1, wq
, dwork
, delay
);
1116 EXPORT_SYMBOL_GPL(queue_delayed_work
);
1119 * queue_delayed_work_on - queue work on specific CPU after delay
1120 * @cpu: CPU number to execute work on
1121 * @wq: workqueue to use
1122 * @dwork: work to queue
1123 * @delay: number of jiffies to wait before queueing
1125 * Returns 0 if @work was already on a queue, non-zero otherwise.
1127 int queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1128 struct delayed_work
*dwork
, unsigned long delay
)
1131 struct timer_list
*timer
= &dwork
->timer
;
1132 struct work_struct
*work
= &dwork
->work
;
1134 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1137 BUG_ON(timer_pending(timer
));
1138 BUG_ON(!list_empty(&work
->entry
));
1140 timer_stats_timer_set_start_info(&dwork
->timer
);
1143 * This stores cwq for the moment, for the timer_fn.
1144 * Note that the work's gcwq is preserved to allow
1145 * reentrance detection for delayed works.
1147 if (!(wq
->flags
& WQ_UNBOUND
)) {
1148 struct global_cwq
*gcwq
= get_work_gcwq(work
);
1150 if (gcwq
&& gcwq
->cpu
!= WORK_CPU_UNBOUND
)
1153 lcpu
= raw_smp_processor_id();
1155 lcpu
= WORK_CPU_UNBOUND
;
1157 set_work_cwq(work
, get_cwq(lcpu
, wq
), 0);
1159 timer
->expires
= jiffies
+ delay
;
1160 timer
->data
= (unsigned long)dwork
;
1161 timer
->function
= delayed_work_timer_fn
;
1163 if (unlikely(cpu
>= 0))
1164 add_timer_on(timer
, cpu
);
1171 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1174 * worker_enter_idle - enter idle state
1175 * @worker: worker which is entering idle state
1177 * @worker is entering idle state. Update stats and idle timer if
1181 * spin_lock_irq(gcwq->lock).
1183 static void worker_enter_idle(struct worker
*worker
)
1185 struct worker_pool
*pool
= worker
->pool
;
1186 struct global_cwq
*gcwq
= pool
->gcwq
;
1188 BUG_ON(worker
->flags
& WORKER_IDLE
);
1189 BUG_ON(!list_empty(&worker
->entry
) &&
1190 (worker
->hentry
.next
|| worker
->hentry
.pprev
));
1192 /* can't use worker_set_flags(), also called from start_worker() */
1193 worker
->flags
|= WORKER_IDLE
;
1195 worker
->last_active
= jiffies
;
1197 /* idle_list is LIFO */
1198 list_add(&worker
->entry
, &pool
->idle_list
);
1200 if (likely(gcwq
->trustee_state
!= TRUSTEE_DONE
)) {
1201 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1202 mod_timer(&pool
->idle_timer
,
1203 jiffies
+ IDLE_WORKER_TIMEOUT
);
1205 wake_up_all(&gcwq
->trustee_wait
);
1208 * Sanity check nr_running. Because trustee releases gcwq->lock
1209 * between setting %WORKER_UNBOUND and zapping nr_running, the
1210 * warning may trigger spuriously. Check iff trustee is idle.
1212 WARN_ON_ONCE(gcwq
->trustee_state
== TRUSTEE_DONE
&&
1213 pool
->nr_workers
== pool
->nr_idle
&&
1214 atomic_read(get_pool_nr_running(pool
)));
1218 * worker_leave_idle - leave idle state
1219 * @worker: worker which is leaving idle state
1221 * @worker is leaving idle state. Update stats.
1224 * spin_lock_irq(gcwq->lock).
1226 static void worker_leave_idle(struct worker
*worker
)
1228 struct worker_pool
*pool
= worker
->pool
;
1230 BUG_ON(!(worker
->flags
& WORKER_IDLE
));
1231 worker_clr_flags(worker
, WORKER_IDLE
);
1233 list_del_init(&worker
->entry
);
1237 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1240 * Works which are scheduled while the cpu is online must at least be
1241 * scheduled to a worker which is bound to the cpu so that if they are
1242 * flushed from cpu callbacks while cpu is going down, they are
1243 * guaranteed to execute on the cpu.
1245 * This function is to be used by rogue workers and rescuers to bind
1246 * themselves to the target cpu and may race with cpu going down or
1247 * coming online. kthread_bind() can't be used because it may put the
1248 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1249 * verbatim as it's best effort and blocking and gcwq may be
1250 * [dis]associated in the meantime.
1252 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1253 * binding against %GCWQ_DISASSOCIATED which is set during
1254 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1255 * enters idle state or fetches works without dropping lock, it can
1256 * guarantee the scheduling requirement described in the first paragraph.
1259 * Might sleep. Called without any lock but returns with gcwq->lock
1263 * %true if the associated gcwq is online (@worker is successfully
1264 * bound), %false if offline.
1266 static bool worker_maybe_bind_and_lock(struct worker
*worker
)
1267 __acquires(&gcwq
->lock
)
1269 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1270 struct task_struct
*task
= worker
->task
;
1274 * The following call may fail, succeed or succeed
1275 * without actually migrating the task to the cpu if
1276 * it races with cpu hotunplug operation. Verify
1277 * against GCWQ_DISASSOCIATED.
1279 if (!(gcwq
->flags
& GCWQ_DISASSOCIATED
))
1280 set_cpus_allowed_ptr(task
, get_cpu_mask(gcwq
->cpu
));
1282 spin_lock_irq(&gcwq
->lock
);
1283 if (gcwq
->flags
& GCWQ_DISASSOCIATED
)
1285 if (task_cpu(task
) == gcwq
->cpu
&&
1286 cpumask_equal(¤t
->cpus_allowed
,
1287 get_cpu_mask(gcwq
->cpu
)))
1289 spin_unlock_irq(&gcwq
->lock
);
1292 * We've raced with CPU hot[un]plug. Give it a breather
1293 * and retry migration. cond_resched() is required here;
1294 * otherwise, we might deadlock against cpu_stop trying to
1295 * bring down the CPU on non-preemptive kernel.
1303 * Function for worker->rebind_work used to rebind unbound busy workers to
1304 * the associated cpu which is coming back online. This is scheduled by
1305 * cpu up but can race with other cpu hotplug operations and may be
1306 * executed twice without intervening cpu down.
1308 static void worker_rebind_fn(struct work_struct
*work
)
1310 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1311 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1313 if (worker_maybe_bind_and_lock(worker
))
1314 worker_clr_flags(worker
, WORKER_REBIND
);
1316 spin_unlock_irq(&gcwq
->lock
);
1319 static struct worker
*alloc_worker(void)
1321 struct worker
*worker
;
1323 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1325 INIT_LIST_HEAD(&worker
->entry
);
1326 INIT_LIST_HEAD(&worker
->scheduled
);
1327 INIT_WORK(&worker
->rebind_work
, worker_rebind_fn
);
1328 /* on creation a worker is in !idle && prep state */
1329 worker
->flags
= WORKER_PREP
;
1335 * create_worker - create a new workqueue worker
1336 * @pool: pool the new worker will belong to
1337 * @bind: whether to set affinity to @cpu or not
1339 * Create a new worker which is bound to @pool. The returned worker
1340 * can be started by calling start_worker() or destroyed using
1344 * Might sleep. Does GFP_KERNEL allocations.
1347 * Pointer to the newly created worker.
1349 static struct worker
*create_worker(struct worker_pool
*pool
, bool bind
)
1351 struct global_cwq
*gcwq
= pool
->gcwq
;
1352 bool on_unbound_cpu
= gcwq
->cpu
== WORK_CPU_UNBOUND
;
1353 const char *pri
= worker_pool_pri(pool
) ? "H" : "";
1354 struct worker
*worker
= NULL
;
1357 spin_lock_irq(&gcwq
->lock
);
1358 while (ida_get_new(&pool
->worker_ida
, &id
)) {
1359 spin_unlock_irq(&gcwq
->lock
);
1360 if (!ida_pre_get(&pool
->worker_ida
, GFP_KERNEL
))
1362 spin_lock_irq(&gcwq
->lock
);
1364 spin_unlock_irq(&gcwq
->lock
);
1366 worker
= alloc_worker();
1370 worker
->pool
= pool
;
1373 if (!on_unbound_cpu
)
1374 worker
->task
= kthread_create_on_node(worker_thread
,
1375 worker
, cpu_to_node(gcwq
->cpu
),
1376 "kworker/%u:%d%s", gcwq
->cpu
, id
, pri
);
1378 worker
->task
= kthread_create(worker_thread
, worker
,
1379 "kworker/u:%d%s", id
, pri
);
1380 if (IS_ERR(worker
->task
))
1383 if (worker_pool_pri(pool
))
1384 set_user_nice(worker
->task
, HIGHPRI_NICE_LEVEL
);
1387 * An unbound worker will become a regular one if CPU comes online
1388 * later on. Make sure every worker has PF_THREAD_BOUND set.
1390 if (bind
&& !on_unbound_cpu
)
1391 kthread_bind(worker
->task
, gcwq
->cpu
);
1393 worker
->task
->flags
|= PF_THREAD_BOUND
;
1395 worker
->flags
|= WORKER_UNBOUND
;
1401 spin_lock_irq(&gcwq
->lock
);
1402 ida_remove(&pool
->worker_ida
, id
);
1403 spin_unlock_irq(&gcwq
->lock
);
1410 * start_worker - start a newly created worker
1411 * @worker: worker to start
1413 * Make the gcwq aware of @worker and start it.
1416 * spin_lock_irq(gcwq->lock).
1418 static void start_worker(struct worker
*worker
)
1420 worker
->flags
|= WORKER_STARTED
;
1421 worker
->pool
->nr_workers
++;
1422 worker_enter_idle(worker
);
1423 wake_up_process(worker
->task
);
1427 * destroy_worker - destroy a workqueue worker
1428 * @worker: worker to be destroyed
1430 * Destroy @worker and adjust @gcwq stats accordingly.
1433 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1435 static void destroy_worker(struct worker
*worker
)
1437 struct worker_pool
*pool
= worker
->pool
;
1438 struct global_cwq
*gcwq
= pool
->gcwq
;
1439 int id
= worker
->id
;
1441 /* sanity check frenzy */
1442 BUG_ON(worker
->current_work
);
1443 BUG_ON(!list_empty(&worker
->scheduled
));
1445 if (worker
->flags
& WORKER_STARTED
)
1447 if (worker
->flags
& WORKER_IDLE
)
1450 list_del_init(&worker
->entry
);
1451 worker
->flags
|= WORKER_DIE
;
1453 spin_unlock_irq(&gcwq
->lock
);
1455 kthread_stop(worker
->task
);
1458 spin_lock_irq(&gcwq
->lock
);
1459 ida_remove(&pool
->worker_ida
, id
);
1462 static void idle_worker_timeout(unsigned long __pool
)
1464 struct worker_pool
*pool
= (void *)__pool
;
1465 struct global_cwq
*gcwq
= pool
->gcwq
;
1467 spin_lock_irq(&gcwq
->lock
);
1469 if (too_many_workers(pool
)) {
1470 struct worker
*worker
;
1471 unsigned long expires
;
1473 /* idle_list is kept in LIFO order, check the last one */
1474 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1475 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1477 if (time_before(jiffies
, expires
))
1478 mod_timer(&pool
->idle_timer
, expires
);
1480 /* it's been idle for too long, wake up manager */
1481 pool
->flags
|= POOL_MANAGE_WORKERS
;
1482 wake_up_worker(pool
);
1486 spin_unlock_irq(&gcwq
->lock
);
1489 static bool send_mayday(struct work_struct
*work
)
1491 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1492 struct workqueue_struct
*wq
= cwq
->wq
;
1495 if (!(wq
->flags
& WQ_RESCUER
))
1498 /* mayday mayday mayday */
1499 cpu
= cwq
->pool
->gcwq
->cpu
;
1500 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1501 if (cpu
== WORK_CPU_UNBOUND
)
1503 if (!mayday_test_and_set_cpu(cpu
, wq
->mayday_mask
))
1504 wake_up_process(wq
->rescuer
->task
);
1508 static void gcwq_mayday_timeout(unsigned long __pool
)
1510 struct worker_pool
*pool
= (void *)__pool
;
1511 struct global_cwq
*gcwq
= pool
->gcwq
;
1512 struct work_struct
*work
;
1514 spin_lock_irq(&gcwq
->lock
);
1516 if (need_to_create_worker(pool
)) {
1518 * We've been trying to create a new worker but
1519 * haven't been successful. We might be hitting an
1520 * allocation deadlock. Send distress signals to
1523 list_for_each_entry(work
, &pool
->worklist
, entry
)
1527 spin_unlock_irq(&gcwq
->lock
);
1529 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1533 * maybe_create_worker - create a new worker if necessary
1534 * @pool: pool to create a new worker for
1536 * Create a new worker for @pool if necessary. @pool is guaranteed to
1537 * have at least one idle worker on return from this function. If
1538 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1539 * sent to all rescuers with works scheduled on @pool to resolve
1540 * possible allocation deadlock.
1542 * On return, need_to_create_worker() is guaranteed to be false and
1543 * may_start_working() true.
1546 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1547 * multiple times. Does GFP_KERNEL allocations. Called only from
1551 * false if no action was taken and gcwq->lock stayed locked, true
1554 static bool maybe_create_worker(struct worker_pool
*pool
)
1555 __releases(&gcwq
->lock
)
1556 __acquires(&gcwq
->lock
)
1558 struct global_cwq
*gcwq
= pool
->gcwq
;
1560 if (!need_to_create_worker(pool
))
1563 spin_unlock_irq(&gcwq
->lock
);
1565 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1566 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1569 struct worker
*worker
;
1571 worker
= create_worker(pool
, true);
1573 del_timer_sync(&pool
->mayday_timer
);
1574 spin_lock_irq(&gcwq
->lock
);
1575 start_worker(worker
);
1576 BUG_ON(need_to_create_worker(pool
));
1580 if (!need_to_create_worker(pool
))
1583 __set_current_state(TASK_INTERRUPTIBLE
);
1584 schedule_timeout(CREATE_COOLDOWN
);
1586 if (!need_to_create_worker(pool
))
1590 del_timer_sync(&pool
->mayday_timer
);
1591 spin_lock_irq(&gcwq
->lock
);
1592 if (need_to_create_worker(pool
))
1598 * maybe_destroy_worker - destroy workers which have been idle for a while
1599 * @pool: pool to destroy workers for
1601 * Destroy @pool workers which have been idle for longer than
1602 * IDLE_WORKER_TIMEOUT.
1605 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1606 * multiple times. Called only from manager.
1609 * false if no action was taken and gcwq->lock stayed locked, true
1612 static bool maybe_destroy_workers(struct worker_pool
*pool
)
1616 while (too_many_workers(pool
)) {
1617 struct worker
*worker
;
1618 unsigned long expires
;
1620 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1621 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1623 if (time_before(jiffies
, expires
)) {
1624 mod_timer(&pool
->idle_timer
, expires
);
1628 destroy_worker(worker
);
1636 * manage_workers - manage worker pool
1639 * Assume the manager role and manage gcwq worker pool @worker belongs
1640 * to. At any given time, there can be only zero or one manager per
1641 * gcwq. The exclusion is handled automatically by this function.
1643 * The caller can safely start processing works on false return. On
1644 * true return, it's guaranteed that need_to_create_worker() is false
1645 * and may_start_working() is true.
1648 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1649 * multiple times. Does GFP_KERNEL allocations.
1652 * false if no action was taken and gcwq->lock stayed locked, true if
1653 * some action was taken.
1655 static bool manage_workers(struct worker
*worker
)
1657 struct worker_pool
*pool
= worker
->pool
;
1660 if (!mutex_trylock(&pool
->manager_mutex
))
1663 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
1666 * Destroy and then create so that may_start_working() is true
1669 ret
|= maybe_destroy_workers(pool
);
1670 ret
|= maybe_create_worker(pool
);
1672 mutex_unlock(&pool
->manager_mutex
);
1677 * move_linked_works - move linked works to a list
1678 * @work: start of series of works to be scheduled
1679 * @head: target list to append @work to
1680 * @nextp: out paramter for nested worklist walking
1682 * Schedule linked works starting from @work to @head. Work series to
1683 * be scheduled starts at @work and includes any consecutive work with
1684 * WORK_STRUCT_LINKED set in its predecessor.
1686 * If @nextp is not NULL, it's updated to point to the next work of
1687 * the last scheduled work. This allows move_linked_works() to be
1688 * nested inside outer list_for_each_entry_safe().
1691 * spin_lock_irq(gcwq->lock).
1693 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1694 struct work_struct
**nextp
)
1696 struct work_struct
*n
;
1699 * Linked worklist will always end before the end of the list,
1700 * use NULL for list head.
1702 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1703 list_move_tail(&work
->entry
, head
);
1704 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1709 * If we're already inside safe list traversal and have moved
1710 * multiple works to the scheduled queue, the next position
1711 * needs to be updated.
1717 static void cwq_activate_first_delayed(struct cpu_workqueue_struct
*cwq
)
1719 struct work_struct
*work
= list_first_entry(&cwq
->delayed_works
,
1720 struct work_struct
, entry
);
1722 trace_workqueue_activate_work(work
);
1723 move_linked_works(work
, &cwq
->pool
->worklist
, NULL
);
1724 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1729 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1730 * @cwq: cwq of interest
1731 * @color: color of work which left the queue
1732 * @delayed: for a delayed work
1734 * A work either has completed or is removed from pending queue,
1735 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1738 * spin_lock_irq(gcwq->lock).
1740 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct
*cwq
, int color
,
1743 /* ignore uncolored works */
1744 if (color
== WORK_NO_COLOR
)
1747 cwq
->nr_in_flight
[color
]--;
1751 if (!list_empty(&cwq
->delayed_works
)) {
1752 /* one down, submit a delayed one */
1753 if (cwq
->nr_active
< cwq
->max_active
)
1754 cwq_activate_first_delayed(cwq
);
1758 /* is flush in progress and are we at the flushing tip? */
1759 if (likely(cwq
->flush_color
!= color
))
1762 /* are there still in-flight works? */
1763 if (cwq
->nr_in_flight
[color
])
1766 /* this cwq is done, clear flush_color */
1767 cwq
->flush_color
= -1;
1770 * If this was the last cwq, wake up the first flusher. It
1771 * will handle the rest.
1773 if (atomic_dec_and_test(&cwq
->wq
->nr_cwqs_to_flush
))
1774 complete(&cwq
->wq
->first_flusher
->done
);
1778 * process_one_work - process single work
1780 * @work: work to process
1782 * Process @work. This function contains all the logics necessary to
1783 * process a single work including synchronization against and
1784 * interaction with other workers on the same cpu, queueing and
1785 * flushing. As long as context requirement is met, any worker can
1786 * call this function to process a work.
1789 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1791 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
1792 __releases(&gcwq
->lock
)
1793 __acquires(&gcwq
->lock
)
1795 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1796 struct worker_pool
*pool
= worker
->pool
;
1797 struct global_cwq
*gcwq
= pool
->gcwq
;
1798 struct hlist_head
*bwh
= busy_worker_head(gcwq
, work
);
1799 bool cpu_intensive
= cwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
1800 work_func_t f
= work
->func
;
1802 struct worker
*collision
;
1803 #ifdef CONFIG_LOCKDEP
1805 * It is permissible to free the struct work_struct from
1806 * inside the function that is called from it, this we need to
1807 * take into account for lockdep too. To avoid bogus "held
1808 * lock freed" warnings as well as problems when looking into
1809 * work->lockdep_map, make a copy and use that here.
1811 struct lockdep_map lockdep_map
;
1813 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
1816 * A single work shouldn't be executed concurrently by
1817 * multiple workers on a single cpu. Check whether anyone is
1818 * already processing the work. If so, defer the work to the
1819 * currently executing one.
1821 collision
= __find_worker_executing_work(gcwq
, bwh
, work
);
1822 if (unlikely(collision
)) {
1823 move_linked_works(work
, &collision
->scheduled
, NULL
);
1827 /* claim and process */
1828 debug_work_deactivate(work
);
1829 hlist_add_head(&worker
->hentry
, bwh
);
1830 worker
->current_work
= work
;
1831 worker
->current_cwq
= cwq
;
1832 work_color
= get_work_color(work
);
1834 /* record the current cpu number in the work data and dequeue */
1835 set_work_cpu(work
, gcwq
->cpu
);
1836 list_del_init(&work
->entry
);
1839 * CPU intensive works don't participate in concurrency
1840 * management. They're the scheduler's responsibility.
1842 if (unlikely(cpu_intensive
))
1843 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
1846 * Unbound gcwq isn't concurrency managed and work items should be
1847 * executed ASAP. Wake up another worker if necessary.
1849 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
1850 wake_up_worker(pool
);
1852 spin_unlock_irq(&gcwq
->lock
);
1854 work_clear_pending(work
);
1855 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
1856 lock_map_acquire(&lockdep_map
);
1857 trace_workqueue_execute_start(work
);
1860 * While we must be careful to not use "work" after this, the trace
1861 * point will only record its address.
1863 trace_workqueue_execute_end(work
);
1864 lock_map_release(&lockdep_map
);
1865 lock_map_release(&cwq
->wq
->lockdep_map
);
1867 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
1868 printk(KERN_ERR
"BUG: workqueue leaked lock or atomic: "
1870 current
->comm
, preempt_count(), task_pid_nr(current
));
1871 printk(KERN_ERR
" last function: ");
1872 print_symbol("%s\n", (unsigned long)f
);
1873 debug_show_held_locks(current
);
1877 spin_lock_irq(&gcwq
->lock
);
1879 /* clear cpu intensive status */
1880 if (unlikely(cpu_intensive
))
1881 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
1883 /* we're done with it, release */
1884 hlist_del_init(&worker
->hentry
);
1885 worker
->current_work
= NULL
;
1886 worker
->current_cwq
= NULL
;
1887 cwq_dec_nr_in_flight(cwq
, work_color
, false);
1891 * process_scheduled_works - process scheduled works
1894 * Process all scheduled works. Please note that the scheduled list
1895 * may change while processing a work, so this function repeatedly
1896 * fetches a work from the top and executes it.
1899 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1902 static void process_scheduled_works(struct worker
*worker
)
1904 while (!list_empty(&worker
->scheduled
)) {
1905 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
1906 struct work_struct
, entry
);
1907 process_one_work(worker
, work
);
1912 * worker_thread - the worker thread function
1915 * The gcwq worker thread function. There's a single dynamic pool of
1916 * these per each cpu. These workers process all works regardless of
1917 * their specific target workqueue. The only exception is works which
1918 * belong to workqueues with a rescuer which will be explained in
1921 static int worker_thread(void *__worker
)
1923 struct worker
*worker
= __worker
;
1924 struct worker_pool
*pool
= worker
->pool
;
1925 struct global_cwq
*gcwq
= pool
->gcwq
;
1927 /* tell the scheduler that this is a workqueue worker */
1928 worker
->task
->flags
|= PF_WQ_WORKER
;
1930 spin_lock_irq(&gcwq
->lock
);
1932 /* DIE can be set only while we're idle, checking here is enough */
1933 if (worker
->flags
& WORKER_DIE
) {
1934 spin_unlock_irq(&gcwq
->lock
);
1935 worker
->task
->flags
&= ~PF_WQ_WORKER
;
1939 worker_leave_idle(worker
);
1941 /* no more worker necessary? */
1942 if (!need_more_worker(pool
))
1945 /* do we need to manage? */
1946 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
1950 * ->scheduled list can only be filled while a worker is
1951 * preparing to process a work or actually processing it.
1952 * Make sure nobody diddled with it while I was sleeping.
1954 BUG_ON(!list_empty(&worker
->scheduled
));
1957 * When control reaches this point, we're guaranteed to have
1958 * at least one idle worker or that someone else has already
1959 * assumed the manager role.
1961 worker_clr_flags(worker
, WORKER_PREP
);
1964 struct work_struct
*work
=
1965 list_first_entry(&pool
->worklist
,
1966 struct work_struct
, entry
);
1968 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
1969 /* optimization path, not strictly necessary */
1970 process_one_work(worker
, work
);
1971 if (unlikely(!list_empty(&worker
->scheduled
)))
1972 process_scheduled_works(worker
);
1974 move_linked_works(work
, &worker
->scheduled
, NULL
);
1975 process_scheduled_works(worker
);
1977 } while (keep_working(pool
));
1979 worker_set_flags(worker
, WORKER_PREP
, false);
1981 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
1985 * gcwq->lock is held and there's no work to process and no
1986 * need to manage, sleep. Workers are woken up only while
1987 * holding gcwq->lock or from local cpu, so setting the
1988 * current state before releasing gcwq->lock is enough to
1989 * prevent losing any event.
1991 worker_enter_idle(worker
);
1992 __set_current_state(TASK_INTERRUPTIBLE
);
1993 spin_unlock_irq(&gcwq
->lock
);
1999 * rescuer_thread - the rescuer thread function
2000 * @__wq: the associated workqueue
2002 * Workqueue rescuer thread function. There's one rescuer for each
2003 * workqueue which has WQ_RESCUER set.
2005 * Regular work processing on a gcwq may block trying to create a new
2006 * worker which uses GFP_KERNEL allocation which has slight chance of
2007 * developing into deadlock if some works currently on the same queue
2008 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2009 * the problem rescuer solves.
2011 * When such condition is possible, the gcwq summons rescuers of all
2012 * workqueues which have works queued on the gcwq and let them process
2013 * those works so that forward progress can be guaranteed.
2015 * This should happen rarely.
2017 static int rescuer_thread(void *__wq
)
2019 struct workqueue_struct
*wq
= __wq
;
2020 struct worker
*rescuer
= wq
->rescuer
;
2021 struct list_head
*scheduled
= &rescuer
->scheduled
;
2022 bool is_unbound
= wq
->flags
& WQ_UNBOUND
;
2025 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2027 set_current_state(TASK_INTERRUPTIBLE
);
2029 if (kthread_should_stop())
2033 * See whether any cpu is asking for help. Unbounded
2034 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2036 for_each_mayday_cpu(cpu
, wq
->mayday_mask
) {
2037 unsigned int tcpu
= is_unbound
? WORK_CPU_UNBOUND
: cpu
;
2038 struct cpu_workqueue_struct
*cwq
= get_cwq(tcpu
, wq
);
2039 struct worker_pool
*pool
= cwq
->pool
;
2040 struct global_cwq
*gcwq
= pool
->gcwq
;
2041 struct work_struct
*work
, *n
;
2043 __set_current_state(TASK_RUNNING
);
2044 mayday_clear_cpu(cpu
, wq
->mayday_mask
);
2046 /* migrate to the target cpu if possible */
2047 rescuer
->pool
= pool
;
2048 worker_maybe_bind_and_lock(rescuer
);
2051 * Slurp in all works issued via this workqueue and
2054 BUG_ON(!list_empty(&rescuer
->scheduled
));
2055 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2056 if (get_work_cwq(work
) == cwq
)
2057 move_linked_works(work
, scheduled
, &n
);
2059 process_scheduled_works(rescuer
);
2062 * Leave this gcwq. If keep_working() is %true, notify a
2063 * regular worker; otherwise, we end up with 0 concurrency
2064 * and stalling the execution.
2066 if (keep_working(pool
))
2067 wake_up_worker(pool
);
2069 spin_unlock_irq(&gcwq
->lock
);
2077 struct work_struct work
;
2078 struct completion done
;
2081 static void wq_barrier_func(struct work_struct
*work
)
2083 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2084 complete(&barr
->done
);
2088 * insert_wq_barrier - insert a barrier work
2089 * @cwq: cwq to insert barrier into
2090 * @barr: wq_barrier to insert
2091 * @target: target work to attach @barr to
2092 * @worker: worker currently executing @target, NULL if @target is not executing
2094 * @barr is linked to @target such that @barr is completed only after
2095 * @target finishes execution. Please note that the ordering
2096 * guarantee is observed only with respect to @target and on the local
2099 * Currently, a queued barrier can't be canceled. This is because
2100 * try_to_grab_pending() can't determine whether the work to be
2101 * grabbed is at the head of the queue and thus can't clear LINKED
2102 * flag of the previous work while there must be a valid next work
2103 * after a work with LINKED flag set.
2105 * Note that when @worker is non-NULL, @target may be modified
2106 * underneath us, so we can't reliably determine cwq from @target.
2109 * spin_lock_irq(gcwq->lock).
2111 static void insert_wq_barrier(struct cpu_workqueue_struct
*cwq
,
2112 struct wq_barrier
*barr
,
2113 struct work_struct
*target
, struct worker
*worker
)
2115 struct list_head
*head
;
2116 unsigned int linked
= 0;
2119 * debugobject calls are safe here even with gcwq->lock locked
2120 * as we know for sure that this will not trigger any of the
2121 * checks and call back into the fixup functions where we
2124 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2125 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2126 init_completion(&barr
->done
);
2129 * If @target is currently being executed, schedule the
2130 * barrier to the worker; otherwise, put it after @target.
2133 head
= worker
->scheduled
.next
;
2135 unsigned long *bits
= work_data_bits(target
);
2137 head
= target
->entry
.next
;
2138 /* there can already be other linked works, inherit and set */
2139 linked
= *bits
& WORK_STRUCT_LINKED
;
2140 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2143 debug_work_activate(&barr
->work
);
2144 insert_work(cwq
, &barr
->work
, head
,
2145 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2149 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2150 * @wq: workqueue being flushed
2151 * @flush_color: new flush color, < 0 for no-op
2152 * @work_color: new work color, < 0 for no-op
2154 * Prepare cwqs for workqueue flushing.
2156 * If @flush_color is non-negative, flush_color on all cwqs should be
2157 * -1. If no cwq has in-flight commands at the specified color, all
2158 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2159 * has in flight commands, its cwq->flush_color is set to
2160 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2161 * wakeup logic is armed and %true is returned.
2163 * The caller should have initialized @wq->first_flusher prior to
2164 * calling this function with non-negative @flush_color. If
2165 * @flush_color is negative, no flush color update is done and %false
2168 * If @work_color is non-negative, all cwqs should have the same
2169 * work_color which is previous to @work_color and all will be
2170 * advanced to @work_color.
2173 * mutex_lock(wq->flush_mutex).
2176 * %true if @flush_color >= 0 and there's something to flush. %false
2179 static bool flush_workqueue_prep_cwqs(struct workqueue_struct
*wq
,
2180 int flush_color
, int work_color
)
2185 if (flush_color
>= 0) {
2186 BUG_ON(atomic_read(&wq
->nr_cwqs_to_flush
));
2187 atomic_set(&wq
->nr_cwqs_to_flush
, 1);
2190 for_each_cwq_cpu(cpu
, wq
) {
2191 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2192 struct global_cwq
*gcwq
= cwq
->pool
->gcwq
;
2194 spin_lock_irq(&gcwq
->lock
);
2196 if (flush_color
>= 0) {
2197 BUG_ON(cwq
->flush_color
!= -1);
2199 if (cwq
->nr_in_flight
[flush_color
]) {
2200 cwq
->flush_color
= flush_color
;
2201 atomic_inc(&wq
->nr_cwqs_to_flush
);
2206 if (work_color
>= 0) {
2207 BUG_ON(work_color
!= work_next_color(cwq
->work_color
));
2208 cwq
->work_color
= work_color
;
2211 spin_unlock_irq(&gcwq
->lock
);
2214 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_cwqs_to_flush
))
2215 complete(&wq
->first_flusher
->done
);
2221 * flush_workqueue - ensure that any scheduled work has run to completion.
2222 * @wq: workqueue to flush
2224 * Forces execution of the workqueue and blocks until its completion.
2225 * This is typically used in driver shutdown handlers.
2227 * We sleep until all works which were queued on entry have been handled,
2228 * but we are not livelocked by new incoming ones.
2230 void flush_workqueue(struct workqueue_struct
*wq
)
2232 struct wq_flusher this_flusher
= {
2233 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2235 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2239 lock_map_acquire(&wq
->lockdep_map
);
2240 lock_map_release(&wq
->lockdep_map
);
2242 mutex_lock(&wq
->flush_mutex
);
2245 * Start-to-wait phase
2247 next_color
= work_next_color(wq
->work_color
);
2249 if (next_color
!= wq
->flush_color
) {
2251 * Color space is not full. The current work_color
2252 * becomes our flush_color and work_color is advanced
2255 BUG_ON(!list_empty(&wq
->flusher_overflow
));
2256 this_flusher
.flush_color
= wq
->work_color
;
2257 wq
->work_color
= next_color
;
2259 if (!wq
->first_flusher
) {
2260 /* no flush in progress, become the first flusher */
2261 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2263 wq
->first_flusher
= &this_flusher
;
2265 if (!flush_workqueue_prep_cwqs(wq
, wq
->flush_color
,
2267 /* nothing to flush, done */
2268 wq
->flush_color
= next_color
;
2269 wq
->first_flusher
= NULL
;
2274 BUG_ON(wq
->flush_color
== this_flusher
.flush_color
);
2275 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2276 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2280 * Oops, color space is full, wait on overflow queue.
2281 * The next flush completion will assign us
2282 * flush_color and transfer to flusher_queue.
2284 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2287 mutex_unlock(&wq
->flush_mutex
);
2289 wait_for_completion(&this_flusher
.done
);
2292 * Wake-up-and-cascade phase
2294 * First flushers are responsible for cascading flushes and
2295 * handling overflow. Non-first flushers can simply return.
2297 if (wq
->first_flusher
!= &this_flusher
)
2300 mutex_lock(&wq
->flush_mutex
);
2302 /* we might have raced, check again with mutex held */
2303 if (wq
->first_flusher
!= &this_flusher
)
2306 wq
->first_flusher
= NULL
;
2308 BUG_ON(!list_empty(&this_flusher
.list
));
2309 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2312 struct wq_flusher
*next
, *tmp
;
2314 /* complete all the flushers sharing the current flush color */
2315 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2316 if (next
->flush_color
!= wq
->flush_color
)
2318 list_del_init(&next
->list
);
2319 complete(&next
->done
);
2322 BUG_ON(!list_empty(&wq
->flusher_overflow
) &&
2323 wq
->flush_color
!= work_next_color(wq
->work_color
));
2325 /* this flush_color is finished, advance by one */
2326 wq
->flush_color
= work_next_color(wq
->flush_color
);
2328 /* one color has been freed, handle overflow queue */
2329 if (!list_empty(&wq
->flusher_overflow
)) {
2331 * Assign the same color to all overflowed
2332 * flushers, advance work_color and append to
2333 * flusher_queue. This is the start-to-wait
2334 * phase for these overflowed flushers.
2336 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2337 tmp
->flush_color
= wq
->work_color
;
2339 wq
->work_color
= work_next_color(wq
->work_color
);
2341 list_splice_tail_init(&wq
->flusher_overflow
,
2342 &wq
->flusher_queue
);
2343 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2346 if (list_empty(&wq
->flusher_queue
)) {
2347 BUG_ON(wq
->flush_color
!= wq
->work_color
);
2352 * Need to flush more colors. Make the next flusher
2353 * the new first flusher and arm cwqs.
2355 BUG_ON(wq
->flush_color
== wq
->work_color
);
2356 BUG_ON(wq
->flush_color
!= next
->flush_color
);
2358 list_del_init(&next
->list
);
2359 wq
->first_flusher
= next
;
2361 if (flush_workqueue_prep_cwqs(wq
, wq
->flush_color
, -1))
2365 * Meh... this color is already done, clear first
2366 * flusher and repeat cascading.
2368 wq
->first_flusher
= NULL
;
2372 mutex_unlock(&wq
->flush_mutex
);
2374 EXPORT_SYMBOL_GPL(flush_workqueue
);
2377 * drain_workqueue - drain a workqueue
2378 * @wq: workqueue to drain
2380 * Wait until the workqueue becomes empty. While draining is in progress,
2381 * only chain queueing is allowed. IOW, only currently pending or running
2382 * work items on @wq can queue further work items on it. @wq is flushed
2383 * repeatedly until it becomes empty. The number of flushing is detemined
2384 * by the depth of chaining and should be relatively short. Whine if it
2387 void drain_workqueue(struct workqueue_struct
*wq
)
2389 unsigned int flush_cnt
= 0;
2393 * __queue_work() needs to test whether there are drainers, is much
2394 * hotter than drain_workqueue() and already looks at @wq->flags.
2395 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2397 spin_lock(&workqueue_lock
);
2398 if (!wq
->nr_drainers
++)
2399 wq
->flags
|= WQ_DRAINING
;
2400 spin_unlock(&workqueue_lock
);
2402 flush_workqueue(wq
);
2404 for_each_cwq_cpu(cpu
, wq
) {
2405 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2408 spin_lock_irq(&cwq
->pool
->gcwq
->lock
);
2409 drained
= !cwq
->nr_active
&& list_empty(&cwq
->delayed_works
);
2410 spin_unlock_irq(&cwq
->pool
->gcwq
->lock
);
2415 if (++flush_cnt
== 10 ||
2416 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2417 pr_warning("workqueue %s: flush on destruction isn't complete after %u tries\n",
2418 wq
->name
, flush_cnt
);
2422 spin_lock(&workqueue_lock
);
2423 if (!--wq
->nr_drainers
)
2424 wq
->flags
&= ~WQ_DRAINING
;
2425 spin_unlock(&workqueue_lock
);
2427 EXPORT_SYMBOL_GPL(drain_workqueue
);
2429 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
,
2430 bool wait_executing
)
2432 struct worker
*worker
= NULL
;
2433 struct global_cwq
*gcwq
;
2434 struct cpu_workqueue_struct
*cwq
;
2437 gcwq
= get_work_gcwq(work
);
2441 spin_lock_irq(&gcwq
->lock
);
2442 if (!list_empty(&work
->entry
)) {
2444 * See the comment near try_to_grab_pending()->smp_rmb().
2445 * If it was re-queued to a different gcwq under us, we
2446 * are not going to wait.
2449 cwq
= get_work_cwq(work
);
2450 if (unlikely(!cwq
|| gcwq
!= cwq
->pool
->gcwq
))
2452 } else if (wait_executing
) {
2453 worker
= find_worker_executing_work(gcwq
, work
);
2456 cwq
= worker
->current_cwq
;
2460 insert_wq_barrier(cwq
, barr
, work
, worker
);
2461 spin_unlock_irq(&gcwq
->lock
);
2464 * If @max_active is 1 or rescuer is in use, flushing another work
2465 * item on the same workqueue may lead to deadlock. Make sure the
2466 * flusher is not running on the same workqueue by verifying write
2469 if (cwq
->wq
->saved_max_active
== 1 || cwq
->wq
->flags
& WQ_RESCUER
)
2470 lock_map_acquire(&cwq
->wq
->lockdep_map
);
2472 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2473 lock_map_release(&cwq
->wq
->lockdep_map
);
2477 spin_unlock_irq(&gcwq
->lock
);
2482 * flush_work - wait for a work to finish executing the last queueing instance
2483 * @work: the work to flush
2485 * Wait until @work has finished execution. This function considers
2486 * only the last queueing instance of @work. If @work has been
2487 * enqueued across different CPUs on a non-reentrant workqueue or on
2488 * multiple workqueues, @work might still be executing on return on
2489 * some of the CPUs from earlier queueing.
2491 * If @work was queued only on a non-reentrant, ordered or unbound
2492 * workqueue, @work is guaranteed to be idle on return if it hasn't
2493 * been requeued since flush started.
2496 * %true if flush_work() waited for the work to finish execution,
2497 * %false if it was already idle.
2499 bool flush_work(struct work_struct
*work
)
2501 struct wq_barrier barr
;
2503 lock_map_acquire(&work
->lockdep_map
);
2504 lock_map_release(&work
->lockdep_map
);
2506 if (start_flush_work(work
, &barr
, true)) {
2507 wait_for_completion(&barr
.done
);
2508 destroy_work_on_stack(&barr
.work
);
2513 EXPORT_SYMBOL_GPL(flush_work
);
2515 static bool wait_on_cpu_work(struct global_cwq
*gcwq
, struct work_struct
*work
)
2517 struct wq_barrier barr
;
2518 struct worker
*worker
;
2520 spin_lock_irq(&gcwq
->lock
);
2522 worker
= find_worker_executing_work(gcwq
, work
);
2523 if (unlikely(worker
))
2524 insert_wq_barrier(worker
->current_cwq
, &barr
, work
, worker
);
2526 spin_unlock_irq(&gcwq
->lock
);
2528 if (unlikely(worker
)) {
2529 wait_for_completion(&barr
.done
);
2530 destroy_work_on_stack(&barr
.work
);
2536 static bool wait_on_work(struct work_struct
*work
)
2543 lock_map_acquire(&work
->lockdep_map
);
2544 lock_map_release(&work
->lockdep_map
);
2546 for_each_gcwq_cpu(cpu
)
2547 ret
|= wait_on_cpu_work(get_gcwq(cpu
), work
);
2552 * flush_work_sync - wait until a work has finished execution
2553 * @work: the work to flush
2555 * Wait until @work has finished execution. On return, it's
2556 * guaranteed that all queueing instances of @work which happened
2557 * before this function is called are finished. In other words, if
2558 * @work hasn't been requeued since this function was called, @work is
2559 * guaranteed to be idle on return.
2562 * %true if flush_work_sync() waited for the work to finish execution,
2563 * %false if it was already idle.
2565 bool flush_work_sync(struct work_struct
*work
)
2567 struct wq_barrier barr
;
2568 bool pending
, waited
;
2570 /* we'll wait for executions separately, queue barr only if pending */
2571 pending
= start_flush_work(work
, &barr
, false);
2573 /* wait for executions to finish */
2574 waited
= wait_on_work(work
);
2576 /* wait for the pending one */
2578 wait_for_completion(&barr
.done
);
2579 destroy_work_on_stack(&barr
.work
);
2582 return pending
|| waited
;
2584 EXPORT_SYMBOL_GPL(flush_work_sync
);
2587 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2588 * so this work can't be re-armed in any way.
2590 static int try_to_grab_pending(struct work_struct
*work
)
2592 struct global_cwq
*gcwq
;
2595 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
2599 * The queueing is in progress, or it is already queued. Try to
2600 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2602 gcwq
= get_work_gcwq(work
);
2606 spin_lock_irq(&gcwq
->lock
);
2607 if (!list_empty(&work
->entry
)) {
2609 * This work is queued, but perhaps we locked the wrong gcwq.
2610 * In that case we must see the new value after rmb(), see
2611 * insert_work()->wmb().
2614 if (gcwq
== get_work_gcwq(work
)) {
2615 debug_work_deactivate(work
);
2616 list_del_init(&work
->entry
);
2617 cwq_dec_nr_in_flight(get_work_cwq(work
),
2618 get_work_color(work
),
2619 *work_data_bits(work
) & WORK_STRUCT_DELAYED
);
2623 spin_unlock_irq(&gcwq
->lock
);
2628 static bool __cancel_work_timer(struct work_struct
*work
,
2629 struct timer_list
* timer
)
2634 ret
= (timer
&& likely(del_timer(timer
)));
2636 ret
= try_to_grab_pending(work
);
2638 } while (unlikely(ret
< 0));
2640 clear_work_data(work
);
2645 * cancel_work_sync - cancel a work and wait for it to finish
2646 * @work: the work to cancel
2648 * Cancel @work and wait for its execution to finish. This function
2649 * can be used even if the work re-queues itself or migrates to
2650 * another workqueue. On return from this function, @work is
2651 * guaranteed to be not pending or executing on any CPU.
2653 * cancel_work_sync(&delayed_work->work) must not be used for
2654 * delayed_work's. Use cancel_delayed_work_sync() instead.
2656 * The caller must ensure that the workqueue on which @work was last
2657 * queued can't be destroyed before this function returns.
2660 * %true if @work was pending, %false otherwise.
2662 bool cancel_work_sync(struct work_struct
*work
)
2664 return __cancel_work_timer(work
, NULL
);
2666 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2669 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2670 * @dwork: the delayed work to flush
2672 * Delayed timer is cancelled and the pending work is queued for
2673 * immediate execution. Like flush_work(), this function only
2674 * considers the last queueing instance of @dwork.
2677 * %true if flush_work() waited for the work to finish execution,
2678 * %false if it was already idle.
2680 bool flush_delayed_work(struct delayed_work
*dwork
)
2682 if (del_timer_sync(&dwork
->timer
))
2683 __queue_work(raw_smp_processor_id(),
2684 get_work_cwq(&dwork
->work
)->wq
, &dwork
->work
);
2685 return flush_work(&dwork
->work
);
2687 EXPORT_SYMBOL(flush_delayed_work
);
2690 * flush_delayed_work_sync - wait for a dwork to finish
2691 * @dwork: the delayed work to flush
2693 * Delayed timer is cancelled and the pending work is queued for
2694 * execution immediately. Other than timer handling, its behavior
2695 * is identical to flush_work_sync().
2698 * %true if flush_work_sync() waited for the work to finish execution,
2699 * %false if it was already idle.
2701 bool flush_delayed_work_sync(struct delayed_work
*dwork
)
2703 if (del_timer_sync(&dwork
->timer
))
2704 __queue_work(raw_smp_processor_id(),
2705 get_work_cwq(&dwork
->work
)->wq
, &dwork
->work
);
2706 return flush_work_sync(&dwork
->work
);
2708 EXPORT_SYMBOL(flush_delayed_work_sync
);
2711 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2712 * @dwork: the delayed work cancel
2714 * This is cancel_work_sync() for delayed works.
2717 * %true if @dwork was pending, %false otherwise.
2719 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2721 return __cancel_work_timer(&dwork
->work
, &dwork
->timer
);
2723 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2726 * schedule_work - put work task in global workqueue
2727 * @work: job to be done
2729 * Returns zero if @work was already on the kernel-global workqueue and
2730 * non-zero otherwise.
2732 * This puts a job in the kernel-global workqueue if it was not already
2733 * queued and leaves it in the same position on the kernel-global
2734 * workqueue otherwise.
2736 int schedule_work(struct work_struct
*work
)
2738 return queue_work(system_wq
, work
);
2740 EXPORT_SYMBOL(schedule_work
);
2743 * schedule_work_on - put work task on a specific cpu
2744 * @cpu: cpu to put the work task on
2745 * @work: job to be done
2747 * This puts a job on a specific cpu
2749 int schedule_work_on(int cpu
, struct work_struct
*work
)
2751 return queue_work_on(cpu
, system_wq
, work
);
2753 EXPORT_SYMBOL(schedule_work_on
);
2756 * schedule_delayed_work - put work task in global workqueue after delay
2757 * @dwork: job to be done
2758 * @delay: number of jiffies to wait or 0 for immediate execution
2760 * After waiting for a given time this puts a job in the kernel-global
2763 int schedule_delayed_work(struct delayed_work
*dwork
,
2764 unsigned long delay
)
2766 return queue_delayed_work(system_wq
, dwork
, delay
);
2768 EXPORT_SYMBOL(schedule_delayed_work
);
2771 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2773 * @dwork: job to be done
2774 * @delay: number of jiffies to wait
2776 * After waiting for a given time this puts a job in the kernel-global
2777 * workqueue on the specified CPU.
2779 int schedule_delayed_work_on(int cpu
,
2780 struct delayed_work
*dwork
, unsigned long delay
)
2782 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
2784 EXPORT_SYMBOL(schedule_delayed_work_on
);
2787 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2788 * @func: the function to call
2790 * schedule_on_each_cpu() executes @func on each online CPU using the
2791 * system workqueue and blocks until all CPUs have completed.
2792 * schedule_on_each_cpu() is very slow.
2795 * 0 on success, -errno on failure.
2797 int schedule_on_each_cpu(work_func_t func
)
2800 struct work_struct __percpu
*works
;
2802 works
= alloc_percpu(struct work_struct
);
2808 for_each_online_cpu(cpu
) {
2809 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
2811 INIT_WORK(work
, func
);
2812 schedule_work_on(cpu
, work
);
2815 for_each_online_cpu(cpu
)
2816 flush_work(per_cpu_ptr(works
, cpu
));
2824 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2826 * Forces execution of the kernel-global workqueue and blocks until its
2829 * Think twice before calling this function! It's very easy to get into
2830 * trouble if you don't take great care. Either of the following situations
2831 * will lead to deadlock:
2833 * One of the work items currently on the workqueue needs to acquire
2834 * a lock held by your code or its caller.
2836 * Your code is running in the context of a work routine.
2838 * They will be detected by lockdep when they occur, but the first might not
2839 * occur very often. It depends on what work items are on the workqueue and
2840 * what locks they need, which you have no control over.
2842 * In most situations flushing the entire workqueue is overkill; you merely
2843 * need to know that a particular work item isn't queued and isn't running.
2844 * In such cases you should use cancel_delayed_work_sync() or
2845 * cancel_work_sync() instead.
2847 void flush_scheduled_work(void)
2849 flush_workqueue(system_wq
);
2851 EXPORT_SYMBOL(flush_scheduled_work
);
2854 * execute_in_process_context - reliably execute the routine with user context
2855 * @fn: the function to execute
2856 * @ew: guaranteed storage for the execute work structure (must
2857 * be available when the work executes)
2859 * Executes the function immediately if process context is available,
2860 * otherwise schedules the function for delayed execution.
2862 * Returns: 0 - function was executed
2863 * 1 - function was scheduled for execution
2865 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
2867 if (!in_interrupt()) {
2872 INIT_WORK(&ew
->work
, fn
);
2873 schedule_work(&ew
->work
);
2877 EXPORT_SYMBOL_GPL(execute_in_process_context
);
2879 int keventd_up(void)
2881 return system_wq
!= NULL
;
2884 static int alloc_cwqs(struct workqueue_struct
*wq
)
2887 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2888 * Make sure that the alignment isn't lower than that of
2889 * unsigned long long.
2891 const size_t size
= sizeof(struct cpu_workqueue_struct
);
2892 const size_t align
= max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS
,
2893 __alignof__(unsigned long long));
2895 if (!(wq
->flags
& WQ_UNBOUND
))
2896 wq
->cpu_wq
.pcpu
= __alloc_percpu(size
, align
);
2901 * Allocate enough room to align cwq and put an extra
2902 * pointer at the end pointing back to the originally
2903 * allocated pointer which will be used for free.
2905 ptr
= kzalloc(size
+ align
+ sizeof(void *), GFP_KERNEL
);
2907 wq
->cpu_wq
.single
= PTR_ALIGN(ptr
, align
);
2908 *(void **)(wq
->cpu_wq
.single
+ 1) = ptr
;
2912 /* just in case, make sure it's actually aligned */
2913 BUG_ON(!IS_ALIGNED(wq
->cpu_wq
.v
, align
));
2914 return wq
->cpu_wq
.v
? 0 : -ENOMEM
;
2917 static void free_cwqs(struct workqueue_struct
*wq
)
2919 if (!(wq
->flags
& WQ_UNBOUND
))
2920 free_percpu(wq
->cpu_wq
.pcpu
);
2921 else if (wq
->cpu_wq
.single
) {
2922 /* the pointer to free is stored right after the cwq */
2923 kfree(*(void **)(wq
->cpu_wq
.single
+ 1));
2927 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
2930 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
2932 if (max_active
< 1 || max_active
> lim
)
2933 printk(KERN_WARNING
"workqueue: max_active %d requested for %s "
2934 "is out of range, clamping between %d and %d\n",
2935 max_active
, name
, 1, lim
);
2937 return clamp_val(max_active
, 1, lim
);
2940 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
2943 struct lock_class_key
*key
,
2944 const char *lock_name
, ...)
2946 va_list args
, args1
;
2947 struct workqueue_struct
*wq
;
2951 /* determine namelen, allocate wq and format name */
2952 va_start(args
, lock_name
);
2953 va_copy(args1
, args
);
2954 namelen
= vsnprintf(NULL
, 0, fmt
, args
) + 1;
2956 wq
= kzalloc(sizeof(*wq
) + namelen
, GFP_KERNEL
);
2960 vsnprintf(wq
->name
, namelen
, fmt
, args1
);
2965 * Workqueues which may be used during memory reclaim should
2966 * have a rescuer to guarantee forward progress.
2968 if (flags
& WQ_MEM_RECLAIM
)
2969 flags
|= WQ_RESCUER
;
2971 max_active
= max_active
?: WQ_DFL_ACTIVE
;
2972 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
2976 wq
->saved_max_active
= max_active
;
2977 mutex_init(&wq
->flush_mutex
);
2978 atomic_set(&wq
->nr_cwqs_to_flush
, 0);
2979 INIT_LIST_HEAD(&wq
->flusher_queue
);
2980 INIT_LIST_HEAD(&wq
->flusher_overflow
);
2982 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
2983 INIT_LIST_HEAD(&wq
->list
);
2985 if (alloc_cwqs(wq
) < 0)
2988 for_each_cwq_cpu(cpu
, wq
) {
2989 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2990 struct global_cwq
*gcwq
= get_gcwq(cpu
);
2991 int pool_idx
= (bool)(flags
& WQ_HIGHPRI
);
2993 BUG_ON((unsigned long)cwq
& WORK_STRUCT_FLAG_MASK
);
2994 cwq
->pool
= &gcwq
->pools
[pool_idx
];
2996 cwq
->flush_color
= -1;
2997 cwq
->max_active
= max_active
;
2998 INIT_LIST_HEAD(&cwq
->delayed_works
);
3001 if (flags
& WQ_RESCUER
) {
3002 struct worker
*rescuer
;
3004 if (!alloc_mayday_mask(&wq
->mayday_mask
, GFP_KERNEL
))
3007 wq
->rescuer
= rescuer
= alloc_worker();
3011 rescuer
->task
= kthread_create(rescuer_thread
, wq
, "%s",
3013 if (IS_ERR(rescuer
->task
))
3016 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
3017 wake_up_process(rescuer
->task
);
3021 * workqueue_lock protects global freeze state and workqueues
3022 * list. Grab it, set max_active accordingly and add the new
3023 * workqueue to workqueues list.
3025 spin_lock(&workqueue_lock
);
3027 if (workqueue_freezing
&& wq
->flags
& WQ_FREEZABLE
)
3028 for_each_cwq_cpu(cpu
, wq
)
3029 get_cwq(cpu
, wq
)->max_active
= 0;
3031 list_add(&wq
->list
, &workqueues
);
3033 spin_unlock(&workqueue_lock
);
3039 free_mayday_mask(wq
->mayday_mask
);
3045 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3048 * destroy_workqueue - safely terminate a workqueue
3049 * @wq: target workqueue
3051 * Safely destroy a workqueue. All work currently pending will be done first.
3053 void destroy_workqueue(struct workqueue_struct
*wq
)
3057 /* drain it before proceeding with destruction */
3058 drain_workqueue(wq
);
3061 * wq list is used to freeze wq, remove from list after
3062 * flushing is complete in case freeze races us.
3064 spin_lock(&workqueue_lock
);
3065 list_del(&wq
->list
);
3066 spin_unlock(&workqueue_lock
);
3069 for_each_cwq_cpu(cpu
, wq
) {
3070 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3073 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
3074 BUG_ON(cwq
->nr_in_flight
[i
]);
3075 BUG_ON(cwq
->nr_active
);
3076 BUG_ON(!list_empty(&cwq
->delayed_works
));
3079 if (wq
->flags
& WQ_RESCUER
) {
3080 kthread_stop(wq
->rescuer
->task
);
3081 free_mayday_mask(wq
->mayday_mask
);
3088 EXPORT_SYMBOL_GPL(destroy_workqueue
);
3091 * workqueue_set_max_active - adjust max_active of a workqueue
3092 * @wq: target workqueue
3093 * @max_active: new max_active value.
3095 * Set max_active of @wq to @max_active.
3098 * Don't call from IRQ context.
3100 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
3104 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
3106 spin_lock(&workqueue_lock
);
3108 wq
->saved_max_active
= max_active
;
3110 for_each_cwq_cpu(cpu
, wq
) {
3111 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3113 spin_lock_irq(&gcwq
->lock
);
3115 if (!(wq
->flags
& WQ_FREEZABLE
) ||
3116 !(gcwq
->flags
& GCWQ_FREEZING
))
3117 get_cwq(gcwq
->cpu
, wq
)->max_active
= max_active
;
3119 spin_unlock_irq(&gcwq
->lock
);
3122 spin_unlock(&workqueue_lock
);
3124 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
3127 * workqueue_congested - test whether a workqueue is congested
3128 * @cpu: CPU in question
3129 * @wq: target workqueue
3131 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3132 * no synchronization around this function and the test result is
3133 * unreliable and only useful as advisory hints or for debugging.
3136 * %true if congested, %false otherwise.
3138 bool workqueue_congested(unsigned int cpu
, struct workqueue_struct
*wq
)
3140 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3142 return !list_empty(&cwq
->delayed_works
);
3144 EXPORT_SYMBOL_GPL(workqueue_congested
);
3147 * work_cpu - return the last known associated cpu for @work
3148 * @work: the work of interest
3151 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3153 unsigned int work_cpu(struct work_struct
*work
)
3155 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3157 return gcwq
? gcwq
->cpu
: WORK_CPU_NONE
;
3159 EXPORT_SYMBOL_GPL(work_cpu
);
3162 * work_busy - test whether a work is currently pending or running
3163 * @work: the work to be tested
3165 * Test whether @work is currently pending or running. There is no
3166 * synchronization around this function and the test result is
3167 * unreliable and only useful as advisory hints or for debugging.
3168 * Especially for reentrant wqs, the pending state might hide the
3172 * OR'd bitmask of WORK_BUSY_* bits.
3174 unsigned int work_busy(struct work_struct
*work
)
3176 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3177 unsigned long flags
;
3178 unsigned int ret
= 0;
3183 spin_lock_irqsave(&gcwq
->lock
, flags
);
3185 if (work_pending(work
))
3186 ret
|= WORK_BUSY_PENDING
;
3187 if (find_worker_executing_work(gcwq
, work
))
3188 ret
|= WORK_BUSY_RUNNING
;
3190 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
3194 EXPORT_SYMBOL_GPL(work_busy
);
3199 * There are two challenges in supporting CPU hotplug. Firstly, there
3200 * are a lot of assumptions on strong associations among work, cwq and
3201 * gcwq which make migrating pending and scheduled works very
3202 * difficult to implement without impacting hot paths. Secondly,
3203 * gcwqs serve mix of short, long and very long running works making
3204 * blocked draining impractical.
3206 * This is solved by allowing a gcwq to be detached from CPU, running it
3207 * with unbound workers and allowing it to be reattached later if the cpu
3208 * comes back online. A separate thread is created to govern a gcwq in
3209 * such state and is called the trustee of the gcwq.
3211 * Trustee states and their descriptions.
3213 * START Command state used on startup. On CPU_DOWN_PREPARE, a
3214 * new trustee is started with this state.
3216 * IN_CHARGE Once started, trustee will enter this state after
3217 * assuming the manager role and making all existing
3218 * workers rogue. DOWN_PREPARE waits for trustee to
3219 * enter this state. After reaching IN_CHARGE, trustee
3220 * tries to execute the pending worklist until it's empty
3221 * and the state is set to BUTCHER, or the state is set
3224 * BUTCHER Command state which is set by the cpu callback after
3225 * the cpu has went down. Once this state is set trustee
3226 * knows that there will be no new works on the worklist
3227 * and once the worklist is empty it can proceed to
3228 * killing idle workers.
3230 * RELEASE Command state which is set by the cpu callback if the
3231 * cpu down has been canceled or it has come online
3232 * again. After recognizing this state, trustee stops
3233 * trying to drain or butcher and clears ROGUE, rebinds
3234 * all remaining workers back to the cpu and releases
3237 * DONE Trustee will enter this state after BUTCHER or RELEASE
3240 * trustee CPU draining
3241 * took over down complete
3242 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3244 * | CPU is back online v return workers |
3245 * ----------------> RELEASE --------------
3248 /* claim manager positions of all pools */
3249 static void gcwq_claim_management(struct global_cwq
*gcwq
)
3251 struct worker_pool
*pool
;
3253 for_each_worker_pool(pool
, gcwq
)
3254 mutex_lock_nested(&pool
->manager_mutex
, pool
- gcwq
->pools
);
3257 /* release manager positions */
3258 static void gcwq_release_management(struct global_cwq
*gcwq
)
3260 struct worker_pool
*pool
;
3262 for_each_worker_pool(pool
, gcwq
)
3263 mutex_unlock(&pool
->manager_mutex
);
3267 * trustee_wait_event_timeout - timed event wait for trustee
3268 * @cond: condition to wait for
3269 * @timeout: timeout in jiffies
3271 * wait_event_timeout() for trustee to use. Handles locking and
3272 * checks for RELEASE request.
3275 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3276 * multiple times. To be used by trustee.
3279 * Positive indicating left time if @cond is satisfied, 0 if timed
3280 * out, -1 if canceled.
3282 #define trustee_wait_event_timeout(cond, timeout) ({ \
3283 long __ret = (timeout); \
3284 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3286 spin_unlock_irq(&gcwq->lock); \
3287 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3288 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3290 spin_lock_irq(&gcwq->lock); \
3292 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3296 * trustee_wait_event - event wait for trustee
3297 * @cond: condition to wait for
3299 * wait_event() for trustee to use. Automatically handles locking and
3300 * checks for CANCEL request.
3303 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3304 * multiple times. To be used by trustee.
3307 * 0 if @cond is satisfied, -1 if canceled.
3309 #define trustee_wait_event(cond) ({ \
3311 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3312 __ret1 < 0 ? -1 : 0; \
3315 static bool gcwq_has_idle_workers(struct global_cwq
*gcwq
)
3317 struct worker_pool
*pool
;
3319 for_each_worker_pool(pool
, gcwq
)
3320 if (!list_empty(&pool
->idle_list
))
3325 static int __cpuinit
trustee_thread(void *__gcwq
)
3327 struct global_cwq
*gcwq
= __gcwq
;
3328 struct worker_pool
*pool
;
3329 struct worker
*worker
;
3330 struct work_struct
*work
;
3331 struct hlist_node
*pos
;
3335 BUG_ON(gcwq
->cpu
!= smp_processor_id());
3337 gcwq_claim_management(gcwq
);
3338 spin_lock_irq(&gcwq
->lock
);
3341 * We've claimed all manager positions. Make all workers unbound
3342 * and set DISASSOCIATED. Before this, all workers except for the
3343 * ones which are still executing works from before the last CPU
3344 * down must be on the cpu. After this, they may become diasporas.
3346 for_each_worker_pool(pool
, gcwq
)
3347 list_for_each_entry(worker
, &pool
->idle_list
, entry
)
3348 worker
->flags
|= WORKER_UNBOUND
;
3350 for_each_busy_worker(worker
, i
, pos
, gcwq
)
3351 worker
->flags
|= WORKER_UNBOUND
;
3353 gcwq
->flags
|= GCWQ_DISASSOCIATED
;
3356 * Call schedule() so that we cross rq->lock and thus can guarantee
3357 * sched callbacks see the unbound flag. This is necessary as
3358 * scheduler callbacks may be invoked from other cpus.
3360 spin_unlock_irq(&gcwq
->lock
);
3362 spin_lock_irq(&gcwq
->lock
);
3365 * Sched callbacks are disabled now. Zap nr_running. After
3366 * this, nr_running stays zero and need_more_worker() and
3367 * keep_working() are always true as long as the worklist is
3370 for_each_worker_pool(pool
, gcwq
)
3371 atomic_set(get_pool_nr_running(pool
), 0);
3373 spin_unlock_irq(&gcwq
->lock
);
3374 for_each_worker_pool(pool
, gcwq
)
3375 del_timer_sync(&pool
->idle_timer
);
3376 spin_lock_irq(&gcwq
->lock
);
3379 * We're now in charge. Notify and proceed to drain. We need
3380 * to keep the gcwq running during the whole CPU down
3381 * procedure as other cpu hotunplug callbacks may need to
3382 * flush currently running tasks.
3384 gcwq
->trustee_state
= TRUSTEE_IN_CHARGE
;
3385 wake_up_all(&gcwq
->trustee_wait
);
3388 * The original cpu is in the process of dying and may go away
3389 * anytime now. When that happens, we and all workers would
3390 * be migrated to other cpus. Try draining any left work. We
3391 * want to get it over with ASAP - spam rescuers, wake up as
3392 * many idlers as necessary and create new ones till the
3393 * worklist is empty. Note that if the gcwq is frozen, there
3394 * may be frozen works in freezable cwqs. Don't declare
3395 * completion while frozen.
3400 for_each_worker_pool(pool
, gcwq
)
3401 busy
|= pool
->nr_workers
!= pool
->nr_idle
;
3403 if (!busy
&& !(gcwq
->flags
& GCWQ_FREEZING
) &&
3404 gcwq
->trustee_state
!= TRUSTEE_IN_CHARGE
)
3407 for_each_worker_pool(pool
, gcwq
) {
3410 list_for_each_entry(work
, &pool
->worklist
, entry
) {
3415 list_for_each_entry(worker
, &pool
->idle_list
, entry
) {
3418 wake_up_process(worker
->task
);
3421 if (need_to_create_worker(pool
)) {
3422 spin_unlock_irq(&gcwq
->lock
);
3423 worker
= create_worker(pool
, false);
3424 spin_lock_irq(&gcwq
->lock
);
3426 worker
->flags
|= WORKER_UNBOUND
;
3427 start_worker(worker
);
3432 /* give a breather */
3433 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN
) < 0)
3438 * Either all works have been scheduled and cpu is down, or
3439 * cpu down has already been canceled. Wait for and butcher
3440 * all workers till we're canceled.
3443 rc
= trustee_wait_event(gcwq_has_idle_workers(gcwq
));
3446 for_each_worker_pool(pool
, gcwq
) {
3447 while (!list_empty(&pool
->idle_list
)) {
3448 worker
= list_first_entry(&pool
->idle_list
,
3449 struct worker
, entry
);
3450 destroy_worker(worker
);
3452 i
|= pool
->nr_workers
;
3454 } while (i
&& rc
>= 0);
3457 * At this point, either draining has completed and no worker
3458 * is left, or cpu down has been canceled or the cpu is being
3459 * brought back up. There shouldn't be any idle one left.
3460 * Tell the remaining busy ones to rebind once it finishes the
3461 * currently scheduled works by scheduling the rebind_work.
3463 for_each_worker_pool(pool
, gcwq
)
3464 WARN_ON(!list_empty(&pool
->idle_list
));
3466 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
3467 struct work_struct
*rebind_work
= &worker
->rebind_work
;
3470 * Rebind_work may race with future cpu hotplug
3471 * operations. Use a separate flag to mark that
3472 * rebinding is scheduled.
3474 worker
->flags
|= WORKER_REBIND
;
3475 worker
->flags
&= ~WORKER_UNBOUND
;
3477 /* queue rebind_work, wq doesn't matter, use the default one */
3478 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
3479 work_data_bits(rebind_work
)))
3482 debug_work_activate(rebind_work
);
3483 insert_work(get_cwq(gcwq
->cpu
, system_wq
), rebind_work
,
3484 worker
->scheduled
.next
,
3485 work_color_to_flags(WORK_NO_COLOR
));
3488 gcwq_release_management(gcwq
);
3490 /* notify completion */
3491 gcwq
->trustee
= NULL
;
3492 gcwq
->trustee_state
= TRUSTEE_DONE
;
3493 wake_up_all(&gcwq
->trustee_wait
);
3494 spin_unlock_irq(&gcwq
->lock
);
3499 * wait_trustee_state - wait for trustee to enter the specified state
3500 * @gcwq: gcwq the trustee of interest belongs to
3501 * @state: target state to wait for
3503 * Wait for the trustee to reach @state. DONE is already matched.
3506 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3507 * multiple times. To be used by cpu_callback.
3509 static void __cpuinit
wait_trustee_state(struct global_cwq
*gcwq
, int state
)
3510 __releases(&gcwq
->lock
)
3511 __acquires(&gcwq
->lock
)
3513 if (!(gcwq
->trustee_state
== state
||
3514 gcwq
->trustee_state
== TRUSTEE_DONE
)) {
3515 spin_unlock_irq(&gcwq
->lock
);
3516 __wait_event(gcwq
->trustee_wait
,
3517 gcwq
->trustee_state
== state
||
3518 gcwq
->trustee_state
== TRUSTEE_DONE
);
3519 spin_lock_irq(&gcwq
->lock
);
3523 static int __devinit
workqueue_cpu_callback(struct notifier_block
*nfb
,
3524 unsigned long action
,
3527 unsigned int cpu
= (unsigned long)hcpu
;
3528 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3529 struct task_struct
*new_trustee
= NULL
;
3530 struct worker
*new_workers
[NR_WORKER_POOLS
] = { };
3531 struct worker_pool
*pool
;
3532 unsigned long flags
;
3535 action
&= ~CPU_TASKS_FROZEN
;
3538 case CPU_DOWN_PREPARE
:
3539 new_trustee
= kthread_create(trustee_thread
, gcwq
,
3540 "workqueue_trustee/%d\n", cpu
);
3541 if (IS_ERR(new_trustee
))
3542 return notifier_from_errno(PTR_ERR(new_trustee
));
3543 kthread_bind(new_trustee
, cpu
);
3545 case CPU_UP_PREPARE
:
3547 for_each_worker_pool(pool
, gcwq
) {
3548 BUG_ON(pool
->first_idle
);
3549 new_workers
[i
] = create_worker(pool
, false);
3550 if (!new_workers
[i
++])
3555 /* some are called w/ irq disabled, don't disturb irq status */
3556 spin_lock_irqsave(&gcwq
->lock
, flags
);
3559 case CPU_DOWN_PREPARE
:
3560 /* initialize trustee and tell it to acquire the gcwq */
3561 BUG_ON(gcwq
->trustee
|| gcwq
->trustee_state
!= TRUSTEE_DONE
);
3562 gcwq
->trustee
= new_trustee
;
3563 gcwq
->trustee_state
= TRUSTEE_START
;
3564 wake_up_process(gcwq
->trustee
);
3565 wait_trustee_state(gcwq
, TRUSTEE_IN_CHARGE
);
3567 case CPU_UP_PREPARE
:
3569 for_each_worker_pool(pool
, gcwq
) {
3570 BUG_ON(pool
->first_idle
);
3571 pool
->first_idle
= new_workers
[i
++];
3576 gcwq
->trustee_state
= TRUSTEE_BUTCHER
;
3578 case CPU_UP_CANCELED
:
3579 for_each_worker_pool(pool
, gcwq
) {
3580 destroy_worker(pool
->first_idle
);
3581 pool
->first_idle
= NULL
;
3585 case CPU_DOWN_FAILED
:
3587 gcwq
->flags
&= ~GCWQ_DISASSOCIATED
;
3588 if (gcwq
->trustee_state
!= TRUSTEE_DONE
) {
3589 gcwq
->trustee_state
= TRUSTEE_RELEASE
;
3590 wake_up_process(gcwq
->trustee
);
3591 wait_trustee_state(gcwq
, TRUSTEE_DONE
);
3595 * Trustee is done and there might be no worker left.
3596 * Put the first_idle in and request a real manager to
3599 for_each_worker_pool(pool
, gcwq
) {
3600 spin_unlock_irq(&gcwq
->lock
);
3601 kthread_bind(pool
->first_idle
->task
, cpu
);
3602 spin_lock_irq(&gcwq
->lock
);
3603 pool
->flags
|= POOL_MANAGE_WORKERS
;
3604 start_worker(pool
->first_idle
);
3605 pool
->first_idle
= NULL
;
3610 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
3612 return notifier_from_errno(0);
3616 kthread_stop(new_trustee
);
3618 spin_lock_irqsave(&gcwq
->lock
, flags
);
3619 for (i
= 0; i
< NR_WORKER_POOLS
; i
++)
3621 destroy_worker(new_workers
[i
]);
3622 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
3628 * Workqueues should be brought up before normal priority CPU notifiers.
3629 * This will be registered high priority CPU notifier.
3631 static int __devinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
3632 unsigned long action
,
3635 switch (action
& ~CPU_TASKS_FROZEN
) {
3636 case CPU_UP_PREPARE
:
3637 case CPU_UP_CANCELED
:
3638 case CPU_DOWN_FAILED
:
3640 return workqueue_cpu_callback(nfb
, action
, hcpu
);
3646 * Workqueues should be brought down after normal priority CPU notifiers.
3647 * This will be registered as low priority CPU notifier.
3649 static int __devinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
3650 unsigned long action
,
3653 switch (action
& ~CPU_TASKS_FROZEN
) {
3654 case CPU_DOWN_PREPARE
:
3656 return workqueue_cpu_callback(nfb
, action
, hcpu
);
3663 struct work_for_cpu
{
3664 struct completion completion
;
3670 static int do_work_for_cpu(void *_wfc
)
3672 struct work_for_cpu
*wfc
= _wfc
;
3673 wfc
->ret
= wfc
->fn(wfc
->arg
);
3674 complete(&wfc
->completion
);
3679 * work_on_cpu - run a function in user context on a particular cpu
3680 * @cpu: the cpu to run on
3681 * @fn: the function to run
3682 * @arg: the function arg
3684 * This will return the value @fn returns.
3685 * It is up to the caller to ensure that the cpu doesn't go offline.
3686 * The caller must not hold any locks which would prevent @fn from completing.
3688 long work_on_cpu(unsigned int cpu
, long (*fn
)(void *), void *arg
)
3690 struct task_struct
*sub_thread
;
3691 struct work_for_cpu wfc
= {
3692 .completion
= COMPLETION_INITIALIZER_ONSTACK(wfc
.completion
),
3697 sub_thread
= kthread_create(do_work_for_cpu
, &wfc
, "work_for_cpu");
3698 if (IS_ERR(sub_thread
))
3699 return PTR_ERR(sub_thread
);
3700 kthread_bind(sub_thread
, cpu
);
3701 wake_up_process(sub_thread
);
3702 wait_for_completion(&wfc
.completion
);
3705 EXPORT_SYMBOL_GPL(work_on_cpu
);
3706 #endif /* CONFIG_SMP */
3708 #ifdef CONFIG_FREEZER
3711 * freeze_workqueues_begin - begin freezing workqueues
3713 * Start freezing workqueues. After this function returns, all freezable
3714 * workqueues will queue new works to their frozen_works list instead of
3718 * Grabs and releases workqueue_lock and gcwq->lock's.
3720 void freeze_workqueues_begin(void)
3724 spin_lock(&workqueue_lock
);
3726 BUG_ON(workqueue_freezing
);
3727 workqueue_freezing
= true;
3729 for_each_gcwq_cpu(cpu
) {
3730 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3731 struct workqueue_struct
*wq
;
3733 spin_lock_irq(&gcwq
->lock
);
3735 BUG_ON(gcwq
->flags
& GCWQ_FREEZING
);
3736 gcwq
->flags
|= GCWQ_FREEZING
;
3738 list_for_each_entry(wq
, &workqueues
, list
) {
3739 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3741 if (cwq
&& wq
->flags
& WQ_FREEZABLE
)
3742 cwq
->max_active
= 0;
3745 spin_unlock_irq(&gcwq
->lock
);
3748 spin_unlock(&workqueue_lock
);
3752 * freeze_workqueues_busy - are freezable workqueues still busy?
3754 * Check whether freezing is complete. This function must be called
3755 * between freeze_workqueues_begin() and thaw_workqueues().
3758 * Grabs and releases workqueue_lock.
3761 * %true if some freezable workqueues are still busy. %false if freezing
3764 bool freeze_workqueues_busy(void)
3769 spin_lock(&workqueue_lock
);
3771 BUG_ON(!workqueue_freezing
);
3773 for_each_gcwq_cpu(cpu
) {
3774 struct workqueue_struct
*wq
;
3776 * nr_active is monotonically decreasing. It's safe
3777 * to peek without lock.
3779 list_for_each_entry(wq
, &workqueues
, list
) {
3780 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3782 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3785 BUG_ON(cwq
->nr_active
< 0);
3786 if (cwq
->nr_active
) {
3793 spin_unlock(&workqueue_lock
);
3798 * thaw_workqueues - thaw workqueues
3800 * Thaw workqueues. Normal queueing is restored and all collected
3801 * frozen works are transferred to their respective gcwq worklists.
3804 * Grabs and releases workqueue_lock and gcwq->lock's.
3806 void thaw_workqueues(void)
3810 spin_lock(&workqueue_lock
);
3812 if (!workqueue_freezing
)
3815 for_each_gcwq_cpu(cpu
) {
3816 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3817 struct worker_pool
*pool
;
3818 struct workqueue_struct
*wq
;
3820 spin_lock_irq(&gcwq
->lock
);
3822 BUG_ON(!(gcwq
->flags
& GCWQ_FREEZING
));
3823 gcwq
->flags
&= ~GCWQ_FREEZING
;
3825 list_for_each_entry(wq
, &workqueues
, list
) {
3826 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3828 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3831 /* restore max_active and repopulate worklist */
3832 cwq
->max_active
= wq
->saved_max_active
;
3834 while (!list_empty(&cwq
->delayed_works
) &&
3835 cwq
->nr_active
< cwq
->max_active
)
3836 cwq_activate_first_delayed(cwq
);
3839 for_each_worker_pool(pool
, gcwq
)
3840 wake_up_worker(pool
);
3842 spin_unlock_irq(&gcwq
->lock
);
3845 workqueue_freezing
= false;
3847 spin_unlock(&workqueue_lock
);
3849 #endif /* CONFIG_FREEZER */
3851 static int __init
init_workqueues(void)
3856 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
3857 cpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
3859 /* initialize gcwqs */
3860 for_each_gcwq_cpu(cpu
) {
3861 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3862 struct worker_pool
*pool
;
3864 spin_lock_init(&gcwq
->lock
);
3866 gcwq
->flags
|= GCWQ_DISASSOCIATED
;
3868 for (i
= 0; i
< BUSY_WORKER_HASH_SIZE
; i
++)
3869 INIT_HLIST_HEAD(&gcwq
->busy_hash
[i
]);
3871 for_each_worker_pool(pool
, gcwq
) {
3873 INIT_LIST_HEAD(&pool
->worklist
);
3874 INIT_LIST_HEAD(&pool
->idle_list
);
3876 init_timer_deferrable(&pool
->idle_timer
);
3877 pool
->idle_timer
.function
= idle_worker_timeout
;
3878 pool
->idle_timer
.data
= (unsigned long)pool
;
3880 setup_timer(&pool
->mayday_timer
, gcwq_mayday_timeout
,
3881 (unsigned long)pool
);
3883 mutex_init(&pool
->manager_mutex
);
3884 ida_init(&pool
->worker_ida
);
3887 gcwq
->trustee_state
= TRUSTEE_DONE
;
3888 init_waitqueue_head(&gcwq
->trustee_wait
);
3891 /* create the initial worker */
3892 for_each_online_gcwq_cpu(cpu
) {
3893 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3894 struct worker_pool
*pool
;
3896 if (cpu
!= WORK_CPU_UNBOUND
)
3897 gcwq
->flags
&= ~GCWQ_DISASSOCIATED
;
3899 for_each_worker_pool(pool
, gcwq
) {
3900 struct worker
*worker
;
3902 worker
= create_worker(pool
, true);
3904 spin_lock_irq(&gcwq
->lock
);
3905 start_worker(worker
);
3906 spin_unlock_irq(&gcwq
->lock
);
3910 system_wq
= alloc_workqueue("events", 0, 0);
3911 system_long_wq
= alloc_workqueue("events_long", 0, 0);
3912 system_nrt_wq
= alloc_workqueue("events_nrt", WQ_NON_REENTRANT
, 0);
3913 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
3914 WQ_UNBOUND_MAX_ACTIVE
);
3915 system_freezable_wq
= alloc_workqueue("events_freezable",
3917 system_nrt_freezable_wq
= alloc_workqueue("events_nrt_freezable",
3918 WQ_NON_REENTRANT
| WQ_FREEZABLE
, 0);
3919 BUG_ON(!system_wq
|| !system_long_wq
|| !system_nrt_wq
||
3920 !system_unbound_wq
|| !system_freezable_wq
||
3921 !system_nrt_freezable_wq
);
3924 early_initcall(init_workqueues
);