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"
51 * A bound gcwq is either associated or disassociated with its CPU.
52 * While associated (!DISASSOCIATED), all workers are bound to the
53 * CPU and none has %WORKER_UNBOUND set and concurrency management
56 * While DISASSOCIATED, the cpu may be offline and all workers have
57 * %WORKER_UNBOUND set and concurrency management disabled, and may
58 * be executing on any CPU. The gcwq behaves as an unbound one.
60 * Note that DISASSOCIATED can be flipped only while holding
61 * managership of all pools on the gcwq to avoid changing binding
62 * state while create_worker() is in progress.
64 GCWQ_DISASSOCIATED
= 1 << 0, /* cpu can't serve workers */
65 GCWQ_FREEZING
= 1 << 1, /* freeze in progress */
68 POOL_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
71 WORKER_STARTED
= 1 << 0, /* started */
72 WORKER_DIE
= 1 << 1, /* die die die */
73 WORKER_IDLE
= 1 << 2, /* is idle */
74 WORKER_PREP
= 1 << 3, /* preparing to run works */
75 WORKER_REBIND
= 1 << 5, /* mom is home, come back */
76 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
77 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
79 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_REBIND
| WORKER_UNBOUND
|
82 NR_WORKER_POOLS
= 2, /* # worker pools per gcwq */
84 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
85 BUSY_WORKER_HASH_SIZE
= 1 << BUSY_WORKER_HASH_ORDER
,
86 BUSY_WORKER_HASH_MASK
= BUSY_WORKER_HASH_SIZE
- 1,
88 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
89 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
91 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
92 /* call for help after 10ms
94 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
95 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
98 * Rescue workers are used only on emergencies and shared by
101 RESCUER_NICE_LEVEL
= -20,
102 HIGHPRI_NICE_LEVEL
= -20,
106 * Structure fields follow one of the following exclusion rules.
108 * I: Modifiable by initialization/destruction paths and read-only for
111 * P: Preemption protected. Disabling preemption is enough and should
112 * only be modified and accessed from the local cpu.
114 * L: gcwq->lock protected. Access with gcwq->lock held.
116 * X: During normal operation, modification requires gcwq->lock and
117 * should be done only from local cpu. Either disabling preemption
118 * on local cpu or grabbing gcwq->lock is enough for read access.
119 * If GCWQ_DISASSOCIATED is set, it's identical to L.
121 * F: wq->flush_mutex protected.
123 * W: workqueue_lock protected.
131 * The poor guys doing the actual heavy lifting. All on-duty workers
132 * are either serving the manager role, on idle list or on busy hash.
135 /* on idle list while idle, on busy hash table while busy */
137 struct list_head entry
; /* L: while idle */
138 struct hlist_node hentry
; /* L: while busy */
141 struct work_struct
*current_work
; /* L: work being processed */
142 struct cpu_workqueue_struct
*current_cwq
; /* L: current_work's cwq */
143 struct list_head scheduled
; /* L: scheduled works */
144 struct task_struct
*task
; /* I: worker task */
145 struct worker_pool
*pool
; /* I: the associated pool */
146 /* 64 bytes boundary on 64bit, 32 on 32bit */
147 unsigned long last_active
; /* L: last active timestamp */
148 unsigned int flags
; /* X: flags */
149 int id
; /* I: worker id */
151 /* for rebinding worker to CPU */
152 struct idle_rebind
*idle_rebind
; /* L: for idle worker */
153 struct work_struct rebind_work
; /* L: for busy worker */
157 struct global_cwq
*gcwq
; /* I: the owning gcwq */
158 unsigned int flags
; /* X: flags */
160 struct list_head worklist
; /* L: list of pending works */
161 int nr_workers
; /* L: total number of workers */
162 int nr_idle
; /* L: currently idle ones */
164 struct list_head idle_list
; /* X: list of idle workers */
165 struct timer_list idle_timer
; /* L: worker idle timeout */
166 struct timer_list mayday_timer
; /* L: SOS timer for workers */
168 struct mutex manager_mutex
; /* mutex manager should hold */
169 struct ida worker_ida
; /* L: for worker IDs */
173 * Global per-cpu workqueue. There's one and only one for each cpu
174 * and all works are queued and processed here regardless of their
178 spinlock_t lock
; /* the gcwq lock */
179 unsigned int cpu
; /* I: the associated cpu */
180 unsigned int flags
; /* L: GCWQ_* flags */
182 /* workers are chained either in busy_hash or pool idle_list */
183 struct hlist_head busy_hash
[BUSY_WORKER_HASH_SIZE
];
184 /* L: hash of busy workers */
186 struct worker_pool pools
[2]; /* normal and highpri pools */
188 wait_queue_head_t rebind_hold
; /* rebind hold wait */
189 } ____cacheline_aligned_in_smp
;
192 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
193 * work_struct->data are used for flags and thus cwqs need to be
194 * aligned at two's power of the number of flag bits.
196 struct cpu_workqueue_struct
{
197 struct worker_pool
*pool
; /* I: the associated pool */
198 struct workqueue_struct
*wq
; /* I: the owning workqueue */
199 int work_color
; /* L: current color */
200 int flush_color
; /* L: flushing color */
201 int nr_in_flight
[WORK_NR_COLORS
];
202 /* L: nr of in_flight works */
203 int nr_active
; /* L: nr of active works */
204 int max_active
; /* L: max active works */
205 struct list_head delayed_works
; /* L: delayed works */
209 * Structure used to wait for workqueue flush.
212 struct list_head list
; /* F: list of flushers */
213 int flush_color
; /* F: flush color waiting for */
214 struct completion done
; /* flush completion */
218 * All cpumasks are assumed to be always set on UP and thus can't be
219 * used to determine whether there's something to be done.
222 typedef cpumask_var_t mayday_mask_t
;
223 #define mayday_test_and_set_cpu(cpu, mask) \
224 cpumask_test_and_set_cpu((cpu), (mask))
225 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
226 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
227 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
228 #define free_mayday_mask(mask) free_cpumask_var((mask))
230 typedef unsigned long mayday_mask_t
;
231 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
232 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
233 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
234 #define alloc_mayday_mask(maskp, gfp) true
235 #define free_mayday_mask(mask) do { } while (0)
239 * The externally visible workqueue abstraction is an array of
240 * per-CPU workqueues:
242 struct workqueue_struct
{
243 unsigned int flags
; /* W: WQ_* flags */
245 struct cpu_workqueue_struct __percpu
*pcpu
;
246 struct cpu_workqueue_struct
*single
;
248 } cpu_wq
; /* I: cwq's */
249 struct list_head list
; /* W: list of all workqueues */
251 struct mutex flush_mutex
; /* protects wq flushing */
252 int work_color
; /* F: current work color */
253 int flush_color
; /* F: current flush color */
254 atomic_t nr_cwqs_to_flush
; /* flush in progress */
255 struct wq_flusher
*first_flusher
; /* F: first flusher */
256 struct list_head flusher_queue
; /* F: flush waiters */
257 struct list_head flusher_overflow
; /* F: flush overflow list */
259 mayday_mask_t mayday_mask
; /* cpus requesting rescue */
260 struct worker
*rescuer
; /* I: rescue worker */
262 int nr_drainers
; /* W: drain in progress */
263 int saved_max_active
; /* W: saved cwq max_active */
264 #ifdef CONFIG_LOCKDEP
265 struct lockdep_map lockdep_map
;
267 char name
[]; /* I: workqueue name */
270 struct workqueue_struct
*system_wq __read_mostly
;
271 struct workqueue_struct
*system_long_wq __read_mostly
;
272 struct workqueue_struct
*system_nrt_wq __read_mostly
;
273 struct workqueue_struct
*system_unbound_wq __read_mostly
;
274 struct workqueue_struct
*system_freezable_wq __read_mostly
;
275 struct workqueue_struct
*system_nrt_freezable_wq __read_mostly
;
276 EXPORT_SYMBOL_GPL(system_wq
);
277 EXPORT_SYMBOL_GPL(system_long_wq
);
278 EXPORT_SYMBOL_GPL(system_nrt_wq
);
279 EXPORT_SYMBOL_GPL(system_unbound_wq
);
280 EXPORT_SYMBOL_GPL(system_freezable_wq
);
281 EXPORT_SYMBOL_GPL(system_nrt_freezable_wq
);
283 #define CREATE_TRACE_POINTS
284 #include <trace/events/workqueue.h>
286 #define for_each_worker_pool(pool, gcwq) \
287 for ((pool) = &(gcwq)->pools[0]; \
288 (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
290 #define for_each_busy_worker(worker, i, pos, gcwq) \
291 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
292 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
294 static inline int __next_gcwq_cpu(int cpu
, const struct cpumask
*mask
,
297 if (cpu
< nr_cpu_ids
) {
299 cpu
= cpumask_next(cpu
, mask
);
300 if (cpu
< nr_cpu_ids
)
304 return WORK_CPU_UNBOUND
;
306 return WORK_CPU_NONE
;
309 static inline int __next_wq_cpu(int cpu
, const struct cpumask
*mask
,
310 struct workqueue_struct
*wq
)
312 return __next_gcwq_cpu(cpu
, mask
, !(wq
->flags
& WQ_UNBOUND
) ? 1 : 2);
318 * An extra gcwq is defined for an invalid cpu number
319 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
320 * specific CPU. The following iterators are similar to
321 * for_each_*_cpu() iterators but also considers the unbound gcwq.
323 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
324 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
325 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
326 * WORK_CPU_UNBOUND for unbound workqueues
328 #define for_each_gcwq_cpu(cpu) \
329 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
330 (cpu) < WORK_CPU_NONE; \
331 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
333 #define for_each_online_gcwq_cpu(cpu) \
334 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
335 (cpu) < WORK_CPU_NONE; \
336 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
338 #define for_each_cwq_cpu(cpu, wq) \
339 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
340 (cpu) < WORK_CPU_NONE; \
341 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
343 #ifdef CONFIG_DEBUG_OBJECTS_WORK
345 static struct debug_obj_descr work_debug_descr
;
347 static void *work_debug_hint(void *addr
)
349 return ((struct work_struct
*) addr
)->func
;
353 * fixup_init is called when:
354 * - an active object is initialized
356 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
358 struct work_struct
*work
= addr
;
361 case ODEBUG_STATE_ACTIVE
:
362 cancel_work_sync(work
);
363 debug_object_init(work
, &work_debug_descr
);
371 * fixup_activate is called when:
372 * - an active object is activated
373 * - an unknown object is activated (might be a statically initialized object)
375 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
377 struct work_struct
*work
= addr
;
381 case ODEBUG_STATE_NOTAVAILABLE
:
383 * This is not really a fixup. The work struct was
384 * statically initialized. We just make sure that it
385 * is tracked in the object tracker.
387 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
388 debug_object_init(work
, &work_debug_descr
);
389 debug_object_activate(work
, &work_debug_descr
);
395 case ODEBUG_STATE_ACTIVE
:
404 * fixup_free is called when:
405 * - an active object is freed
407 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
409 struct work_struct
*work
= addr
;
412 case ODEBUG_STATE_ACTIVE
:
413 cancel_work_sync(work
);
414 debug_object_free(work
, &work_debug_descr
);
421 static struct debug_obj_descr work_debug_descr
= {
422 .name
= "work_struct",
423 .debug_hint
= work_debug_hint
,
424 .fixup_init
= work_fixup_init
,
425 .fixup_activate
= work_fixup_activate
,
426 .fixup_free
= work_fixup_free
,
429 static inline void debug_work_activate(struct work_struct
*work
)
431 debug_object_activate(work
, &work_debug_descr
);
434 static inline void debug_work_deactivate(struct work_struct
*work
)
436 debug_object_deactivate(work
, &work_debug_descr
);
439 void __init_work(struct work_struct
*work
, int onstack
)
442 debug_object_init_on_stack(work
, &work_debug_descr
);
444 debug_object_init(work
, &work_debug_descr
);
446 EXPORT_SYMBOL_GPL(__init_work
);
448 void destroy_work_on_stack(struct work_struct
*work
)
450 debug_object_free(work
, &work_debug_descr
);
452 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
455 static inline void debug_work_activate(struct work_struct
*work
) { }
456 static inline void debug_work_deactivate(struct work_struct
*work
) { }
459 /* Serializes the accesses to the list of workqueues. */
460 static DEFINE_SPINLOCK(workqueue_lock
);
461 static LIST_HEAD(workqueues
);
462 static bool workqueue_freezing
; /* W: have wqs started freezing? */
465 * The almighty global cpu workqueues. nr_running is the only field
466 * which is expected to be used frequently by other cpus via
467 * try_to_wake_up(). Put it in a separate cacheline.
469 static DEFINE_PER_CPU(struct global_cwq
, global_cwq
);
470 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t
, pool_nr_running
[NR_WORKER_POOLS
]);
473 * Global cpu workqueue and nr_running counter for unbound gcwq. The
474 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
475 * workers have WORKER_UNBOUND set.
477 static struct global_cwq unbound_global_cwq
;
478 static atomic_t unbound_pool_nr_running
[NR_WORKER_POOLS
] = {
479 [0 ... NR_WORKER_POOLS
- 1] = ATOMIC_INIT(0), /* always 0 */
482 static int worker_thread(void *__worker
);
484 static int worker_pool_pri(struct worker_pool
*pool
)
486 return pool
- pool
->gcwq
->pools
;
489 static struct global_cwq
*get_gcwq(unsigned int cpu
)
491 if (cpu
!= WORK_CPU_UNBOUND
)
492 return &per_cpu(global_cwq
, cpu
);
494 return &unbound_global_cwq
;
497 static atomic_t
*get_pool_nr_running(struct worker_pool
*pool
)
499 int cpu
= pool
->gcwq
->cpu
;
500 int idx
= worker_pool_pri(pool
);
502 if (cpu
!= WORK_CPU_UNBOUND
)
503 return &per_cpu(pool_nr_running
, cpu
)[idx
];
505 return &unbound_pool_nr_running
[idx
];
508 static struct cpu_workqueue_struct
*get_cwq(unsigned int cpu
,
509 struct workqueue_struct
*wq
)
511 if (!(wq
->flags
& WQ_UNBOUND
)) {
512 if (likely(cpu
< nr_cpu_ids
))
513 return per_cpu_ptr(wq
->cpu_wq
.pcpu
, cpu
);
514 } else if (likely(cpu
== WORK_CPU_UNBOUND
))
515 return wq
->cpu_wq
.single
;
519 static unsigned int work_color_to_flags(int color
)
521 return color
<< WORK_STRUCT_COLOR_SHIFT
;
524 static int get_work_color(struct work_struct
*work
)
526 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
527 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
530 static int work_next_color(int color
)
532 return (color
+ 1) % WORK_NR_COLORS
;
536 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
537 * contain the pointer to the queued cwq. Once execution starts, the flag
538 * is cleared and the high bits contain OFFQ flags and CPU number.
540 * set_work_cwq(), set_work_cpu_and_clear_pending() and clear_work_data()
541 * can be used to set the cwq, cpu or clear work->data. These functions
542 * should only be called while the work is owned - ie. while the PENDING
545 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
546 * corresponding to a work. gcwq is available once the work has been
547 * queued anywhere after initialization. cwq is available only from
548 * queueing until execution starts.
550 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
553 BUG_ON(!work_pending(work
));
554 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
557 static void set_work_cwq(struct work_struct
*work
,
558 struct cpu_workqueue_struct
*cwq
,
559 unsigned long extra_flags
)
561 set_work_data(work
, (unsigned long)cwq
,
562 WORK_STRUCT_PENDING
| WORK_STRUCT_CWQ
| extra_flags
);
565 static void set_work_cpu_and_clear_pending(struct work_struct
*work
,
568 set_work_data(work
, (unsigned long)cpu
<< WORK_OFFQ_CPU_SHIFT
, 0);
571 static void clear_work_data(struct work_struct
*work
)
573 set_work_data(work
, WORK_STRUCT_NO_CPU
, 0);
576 static struct cpu_workqueue_struct
*get_work_cwq(struct work_struct
*work
)
578 unsigned long data
= atomic_long_read(&work
->data
);
580 if (data
& WORK_STRUCT_CWQ
)
581 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
586 static struct global_cwq
*get_work_gcwq(struct work_struct
*work
)
588 unsigned long data
= atomic_long_read(&work
->data
);
591 if (data
& WORK_STRUCT_CWQ
)
592 return ((struct cpu_workqueue_struct
*)
593 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->gcwq
;
595 cpu
= data
>> WORK_OFFQ_CPU_SHIFT
;
596 if (cpu
== WORK_CPU_NONE
)
599 BUG_ON(cpu
>= nr_cpu_ids
&& cpu
!= WORK_CPU_UNBOUND
);
600 return get_gcwq(cpu
);
604 * Policy functions. These define the policies on how the global worker
605 * pools are managed. Unless noted otherwise, these functions assume that
606 * they're being called with gcwq->lock held.
609 static bool __need_more_worker(struct worker_pool
*pool
)
611 return !atomic_read(get_pool_nr_running(pool
));
615 * Need to wake up a worker? Called from anything but currently
618 * Note that, because unbound workers never contribute to nr_running, this
619 * function will always return %true for unbound gcwq as long as the
620 * worklist isn't empty.
622 static bool need_more_worker(struct worker_pool
*pool
)
624 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
627 /* Can I start working? Called from busy but !running workers. */
628 static bool may_start_working(struct worker_pool
*pool
)
630 return pool
->nr_idle
;
633 /* Do I need to keep working? Called from currently running workers. */
634 static bool keep_working(struct worker_pool
*pool
)
636 atomic_t
*nr_running
= get_pool_nr_running(pool
);
638 return !list_empty(&pool
->worklist
) && atomic_read(nr_running
) <= 1;
641 /* Do we need a new worker? Called from manager. */
642 static bool need_to_create_worker(struct worker_pool
*pool
)
644 return need_more_worker(pool
) && !may_start_working(pool
);
647 /* Do I need to be the manager? */
648 static bool need_to_manage_workers(struct worker_pool
*pool
)
650 return need_to_create_worker(pool
) ||
651 (pool
->flags
& POOL_MANAGE_WORKERS
);
654 /* Do we have too many workers and should some go away? */
655 static bool too_many_workers(struct worker_pool
*pool
)
657 bool managing
= mutex_is_locked(&pool
->manager_mutex
);
658 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
659 int nr_busy
= pool
->nr_workers
- nr_idle
;
661 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
668 /* Return the first worker. Safe with preemption disabled */
669 static struct worker
*first_worker(struct worker_pool
*pool
)
671 if (unlikely(list_empty(&pool
->idle_list
)))
674 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
678 * wake_up_worker - wake up an idle worker
679 * @pool: worker pool to wake worker from
681 * Wake up the first idle worker of @pool.
684 * spin_lock_irq(gcwq->lock).
686 static void wake_up_worker(struct worker_pool
*pool
)
688 struct worker
*worker
= first_worker(pool
);
691 wake_up_process(worker
->task
);
695 * wq_worker_waking_up - a worker is waking up
696 * @task: task waking up
697 * @cpu: CPU @task is waking up to
699 * This function is called during try_to_wake_up() when a worker is
703 * spin_lock_irq(rq->lock)
705 void wq_worker_waking_up(struct task_struct
*task
, unsigned int cpu
)
707 struct worker
*worker
= kthread_data(task
);
709 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
710 atomic_inc(get_pool_nr_running(worker
->pool
));
714 * wq_worker_sleeping - a worker is going to sleep
715 * @task: task going to sleep
716 * @cpu: CPU in question, must be the current CPU number
718 * This function is called during schedule() when a busy worker is
719 * going to sleep. Worker on the same cpu can be woken up by
720 * returning pointer to its task.
723 * spin_lock_irq(rq->lock)
726 * Worker task on @cpu to wake up, %NULL if none.
728 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
,
731 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
732 struct worker_pool
*pool
= worker
->pool
;
733 atomic_t
*nr_running
= get_pool_nr_running(pool
);
735 if (worker
->flags
& WORKER_NOT_RUNNING
)
738 /* this can only happen on the local cpu */
739 BUG_ON(cpu
!= raw_smp_processor_id());
742 * The counterpart of the following dec_and_test, implied mb,
743 * worklist not empty test sequence is in insert_work().
744 * Please read comment there.
746 * NOT_RUNNING is clear. This means that we're bound to and
747 * running on the local cpu w/ rq lock held and preemption
748 * disabled, which in turn means that none else could be
749 * manipulating idle_list, so dereferencing idle_list without gcwq
752 if (atomic_dec_and_test(nr_running
) && !list_empty(&pool
->worklist
))
753 to_wakeup
= first_worker(pool
);
754 return to_wakeup
? to_wakeup
->task
: NULL
;
758 * worker_set_flags - set worker flags and adjust nr_running accordingly
760 * @flags: flags to set
761 * @wakeup: wakeup an idle worker if necessary
763 * Set @flags in @worker->flags and adjust nr_running accordingly. If
764 * nr_running becomes zero and @wakeup is %true, an idle worker is
768 * spin_lock_irq(gcwq->lock)
770 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
773 struct worker_pool
*pool
= worker
->pool
;
775 WARN_ON_ONCE(worker
->task
!= current
);
778 * If transitioning into NOT_RUNNING, adjust nr_running and
779 * wake up an idle worker as necessary if requested by
782 if ((flags
& WORKER_NOT_RUNNING
) &&
783 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
784 atomic_t
*nr_running
= get_pool_nr_running(pool
);
787 if (atomic_dec_and_test(nr_running
) &&
788 !list_empty(&pool
->worklist
))
789 wake_up_worker(pool
);
791 atomic_dec(nr_running
);
794 worker
->flags
|= flags
;
798 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
800 * @flags: flags to clear
802 * Clear @flags in @worker->flags and adjust nr_running accordingly.
805 * spin_lock_irq(gcwq->lock)
807 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
809 struct worker_pool
*pool
= worker
->pool
;
810 unsigned int oflags
= worker
->flags
;
812 WARN_ON_ONCE(worker
->task
!= current
);
814 worker
->flags
&= ~flags
;
817 * If transitioning out of NOT_RUNNING, increment nr_running. Note
818 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
819 * of multiple flags, not a single flag.
821 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
822 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
823 atomic_inc(get_pool_nr_running(pool
));
827 * busy_worker_head - return the busy hash head for a work
828 * @gcwq: gcwq of interest
829 * @work: work to be hashed
831 * Return hash head of @gcwq for @work.
834 * spin_lock_irq(gcwq->lock).
837 * Pointer to the hash head.
839 static struct hlist_head
*busy_worker_head(struct global_cwq
*gcwq
,
840 struct work_struct
*work
)
842 const int base_shift
= ilog2(sizeof(struct work_struct
));
843 unsigned long v
= (unsigned long)work
;
845 /* simple shift and fold hash, do we need something better? */
847 v
+= v
>> BUSY_WORKER_HASH_ORDER
;
848 v
&= BUSY_WORKER_HASH_MASK
;
850 return &gcwq
->busy_hash
[v
];
854 * __find_worker_executing_work - find worker which is executing a work
855 * @gcwq: gcwq of interest
856 * @bwh: hash head as returned by busy_worker_head()
857 * @work: work to find worker for
859 * Find a worker which is executing @work on @gcwq. @bwh should be
860 * the hash head obtained by calling busy_worker_head() with the same
864 * spin_lock_irq(gcwq->lock).
867 * Pointer to worker which is executing @work if found, NULL
870 static struct worker
*__find_worker_executing_work(struct global_cwq
*gcwq
,
871 struct hlist_head
*bwh
,
872 struct work_struct
*work
)
874 struct worker
*worker
;
875 struct hlist_node
*tmp
;
877 hlist_for_each_entry(worker
, tmp
, bwh
, hentry
)
878 if (worker
->current_work
== work
)
884 * find_worker_executing_work - find worker which is executing a work
885 * @gcwq: gcwq of interest
886 * @work: work to find worker for
888 * Find a worker which is executing @work on @gcwq. This function is
889 * identical to __find_worker_executing_work() except that this
890 * function calculates @bwh itself.
893 * spin_lock_irq(gcwq->lock).
896 * Pointer to worker which is executing @work if found, NULL
899 static struct worker
*find_worker_executing_work(struct global_cwq
*gcwq
,
900 struct work_struct
*work
)
902 return __find_worker_executing_work(gcwq
, busy_worker_head(gcwq
, work
),
907 * move_linked_works - move linked works to a list
908 * @work: start of series of works to be scheduled
909 * @head: target list to append @work to
910 * @nextp: out paramter for nested worklist walking
912 * Schedule linked works starting from @work to @head. Work series to
913 * be scheduled starts at @work and includes any consecutive work with
914 * WORK_STRUCT_LINKED set in its predecessor.
916 * If @nextp is not NULL, it's updated to point to the next work of
917 * the last scheduled work. This allows move_linked_works() to be
918 * nested inside outer list_for_each_entry_safe().
921 * spin_lock_irq(gcwq->lock).
923 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
924 struct work_struct
**nextp
)
926 struct work_struct
*n
;
929 * Linked worklist will always end before the end of the list,
930 * use NULL for list head.
932 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
933 list_move_tail(&work
->entry
, head
);
934 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
939 * If we're already inside safe list traversal and have moved
940 * multiple works to the scheduled queue, the next position
941 * needs to be updated.
947 static void cwq_activate_first_delayed(struct cpu_workqueue_struct
*cwq
)
949 struct work_struct
*work
= list_first_entry(&cwq
->delayed_works
,
950 struct work_struct
, entry
);
952 trace_workqueue_activate_work(work
);
953 move_linked_works(work
, &cwq
->pool
->worklist
, NULL
);
954 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
959 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
960 * @cwq: cwq of interest
961 * @color: color of work which left the queue
962 * @delayed: for a delayed work
964 * A work either has completed or is removed from pending queue,
965 * decrement nr_in_flight of its cwq and handle workqueue flushing.
968 * spin_lock_irq(gcwq->lock).
970 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct
*cwq
, int color
,
973 /* ignore uncolored works */
974 if (color
== WORK_NO_COLOR
)
977 cwq
->nr_in_flight
[color
]--;
981 if (!list_empty(&cwq
->delayed_works
)) {
982 /* one down, submit a delayed one */
983 if (cwq
->nr_active
< cwq
->max_active
)
984 cwq_activate_first_delayed(cwq
);
988 /* is flush in progress and are we at the flushing tip? */
989 if (likely(cwq
->flush_color
!= color
))
992 /* are there still in-flight works? */
993 if (cwq
->nr_in_flight
[color
])
996 /* this cwq is done, clear flush_color */
997 cwq
->flush_color
= -1;
1000 * If this was the last cwq, wake up the first flusher. It
1001 * will handle the rest.
1003 if (atomic_dec_and_test(&cwq
->wq
->nr_cwqs_to_flush
))
1004 complete(&cwq
->wq
->first_flusher
->done
);
1008 * try_to_grab_pending - steal work item from worklist
1009 * @work: work item to steal
1010 * @is_dwork: @work is a delayed_work
1012 * Try to grab PENDING bit of @work. This function can handle @work in any
1013 * stable state - idle, on timer or on worklist. Return values are
1015 * 1 if @work was pending and we successfully stole PENDING
1016 * 0 if @work was idle and we claimed PENDING
1017 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1019 * On >= 0 return, the caller owns @work's PENDING bit.
1021 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
)
1023 struct global_cwq
*gcwq
;
1025 /* try to steal the timer if it exists */
1027 struct delayed_work
*dwork
= to_delayed_work(work
);
1029 if (likely(del_timer(&dwork
->timer
)))
1033 /* try to claim PENDING the normal way */
1034 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1038 * The queueing is in progress, or it is already queued. Try to
1039 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1041 gcwq
= get_work_gcwq(work
);
1045 spin_lock_irq(&gcwq
->lock
);
1046 if (!list_empty(&work
->entry
)) {
1048 * This work is queued, but perhaps we locked the wrong gcwq.
1049 * In that case we must see the new value after rmb(), see
1050 * insert_work()->wmb().
1053 if (gcwq
== get_work_gcwq(work
)) {
1054 debug_work_deactivate(work
);
1055 list_del_init(&work
->entry
);
1056 cwq_dec_nr_in_flight(get_work_cwq(work
),
1057 get_work_color(work
),
1058 *work_data_bits(work
) & WORK_STRUCT_DELAYED
);
1060 spin_unlock_irq(&gcwq
->lock
);
1064 spin_unlock_irq(&gcwq
->lock
);
1070 * insert_work - insert a work into gcwq
1071 * @cwq: cwq @work belongs to
1072 * @work: work to insert
1073 * @head: insertion point
1074 * @extra_flags: extra WORK_STRUCT_* flags to set
1076 * Insert @work which belongs to @cwq into @gcwq after @head.
1077 * @extra_flags is or'd to work_struct flags.
1080 * spin_lock_irq(gcwq->lock).
1082 static void insert_work(struct cpu_workqueue_struct
*cwq
,
1083 struct work_struct
*work
, struct list_head
*head
,
1084 unsigned int extra_flags
)
1086 struct worker_pool
*pool
= cwq
->pool
;
1088 /* we own @work, set data and link */
1089 set_work_cwq(work
, cwq
, extra_flags
);
1092 * Ensure that we get the right work->data if we see the
1093 * result of list_add() below, see try_to_grab_pending().
1097 list_add_tail(&work
->entry
, head
);
1100 * Ensure either worker_sched_deactivated() sees the above
1101 * list_add_tail() or we see zero nr_running to avoid workers
1102 * lying around lazily while there are works to be processed.
1106 if (__need_more_worker(pool
))
1107 wake_up_worker(pool
);
1111 * Test whether @work is being queued from another work executing on the
1112 * same workqueue. This is rather expensive and should only be used from
1115 static bool is_chained_work(struct workqueue_struct
*wq
)
1117 unsigned long flags
;
1120 for_each_gcwq_cpu(cpu
) {
1121 struct global_cwq
*gcwq
= get_gcwq(cpu
);
1122 struct worker
*worker
;
1123 struct hlist_node
*pos
;
1126 spin_lock_irqsave(&gcwq
->lock
, flags
);
1127 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
1128 if (worker
->task
!= current
)
1130 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1132 * I'm @worker, no locking necessary. See if @work
1133 * is headed to the same workqueue.
1135 return worker
->current_cwq
->wq
== wq
;
1137 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1142 static void __queue_work(unsigned int cpu
, struct workqueue_struct
*wq
,
1143 struct work_struct
*work
)
1145 struct global_cwq
*gcwq
;
1146 struct cpu_workqueue_struct
*cwq
;
1147 struct list_head
*worklist
;
1148 unsigned int work_flags
;
1151 * While a work item is PENDING && off queue, a task trying to
1152 * steal the PENDING will busy-loop waiting for it to either get
1153 * queued or lose PENDING. Grabbing PENDING and queueing should
1154 * happen with IRQ disabled.
1156 WARN_ON_ONCE(!irqs_disabled());
1158 debug_work_activate(work
);
1160 /* if dying, only works from the same workqueue are allowed */
1161 if (unlikely(wq
->flags
& WQ_DRAINING
) &&
1162 WARN_ON_ONCE(!is_chained_work(wq
)))
1165 /* determine gcwq to use */
1166 if (!(wq
->flags
& WQ_UNBOUND
)) {
1167 struct global_cwq
*last_gcwq
;
1169 if (cpu
== WORK_CPU_UNBOUND
)
1170 cpu
= raw_smp_processor_id();
1173 * It's multi cpu. If @wq is non-reentrant and @work
1174 * was previously on a different cpu, it might still
1175 * be running there, in which case the work needs to
1176 * be queued on that cpu to guarantee non-reentrance.
1178 gcwq
= get_gcwq(cpu
);
1179 if (wq
->flags
& WQ_NON_REENTRANT
&&
1180 (last_gcwq
= get_work_gcwq(work
)) && last_gcwq
!= gcwq
) {
1181 struct worker
*worker
;
1183 spin_lock(&last_gcwq
->lock
);
1185 worker
= find_worker_executing_work(last_gcwq
, work
);
1187 if (worker
&& worker
->current_cwq
->wq
== wq
)
1190 /* meh... not running there, queue here */
1191 spin_unlock(&last_gcwq
->lock
);
1192 spin_lock(&gcwq
->lock
);
1195 spin_lock(&gcwq
->lock
);
1198 gcwq
= get_gcwq(WORK_CPU_UNBOUND
);
1199 spin_lock(&gcwq
->lock
);
1202 /* gcwq determined, get cwq and queue */
1203 cwq
= get_cwq(gcwq
->cpu
, wq
);
1204 trace_workqueue_queue_work(cpu
, cwq
, work
);
1206 if (WARN_ON(!list_empty(&work
->entry
))) {
1207 spin_unlock(&gcwq
->lock
);
1211 cwq
->nr_in_flight
[cwq
->work_color
]++;
1212 work_flags
= work_color_to_flags(cwq
->work_color
);
1214 if (likely(cwq
->nr_active
< cwq
->max_active
)) {
1215 trace_workqueue_activate_work(work
);
1217 worklist
= &cwq
->pool
->worklist
;
1219 work_flags
|= WORK_STRUCT_DELAYED
;
1220 worklist
= &cwq
->delayed_works
;
1223 insert_work(cwq
, work
, worklist
, work_flags
);
1225 spin_unlock(&gcwq
->lock
);
1229 * queue_work_on - queue work on specific cpu
1230 * @cpu: CPU number to execute work on
1231 * @wq: workqueue to use
1232 * @work: work to queue
1234 * Returns %false if @work was already on a queue, %true otherwise.
1236 * We queue the work to a specific CPU, the caller must ensure it
1239 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1240 struct work_struct
*work
)
1243 unsigned long flags
;
1245 local_irq_save(flags
);
1247 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1248 __queue_work(cpu
, wq
, work
);
1252 local_irq_restore(flags
);
1255 EXPORT_SYMBOL_GPL(queue_work_on
);
1258 * queue_work - queue work on a workqueue
1259 * @wq: workqueue to use
1260 * @work: work to queue
1262 * Returns %false if @work was already on a queue, %true otherwise.
1264 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1265 * it can be processed by another CPU.
1267 bool queue_work(struct workqueue_struct
*wq
, struct work_struct
*work
)
1269 return queue_work_on(WORK_CPU_UNBOUND
, wq
, work
);
1271 EXPORT_SYMBOL_GPL(queue_work
);
1273 void delayed_work_timer_fn(unsigned long __data
)
1275 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1276 struct cpu_workqueue_struct
*cwq
= get_work_cwq(&dwork
->work
);
1278 local_irq_disable();
1279 __queue_work(WORK_CPU_UNBOUND
, cwq
->wq
, &dwork
->work
);
1282 EXPORT_SYMBOL_GPL(delayed_work_timer_fn
);
1284 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1285 struct delayed_work
*dwork
, unsigned long delay
)
1287 struct timer_list
*timer
= &dwork
->timer
;
1288 struct work_struct
*work
= &dwork
->work
;
1291 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1292 timer
->data
!= (unsigned long)dwork
);
1293 BUG_ON(timer_pending(timer
));
1294 BUG_ON(!list_empty(&work
->entry
));
1296 timer_stats_timer_set_start_info(&dwork
->timer
);
1299 * This stores cwq for the moment, for the timer_fn. Note that the
1300 * work's gcwq is preserved to allow reentrance detection for
1303 if (!(wq
->flags
& WQ_UNBOUND
)) {
1304 struct global_cwq
*gcwq
= get_work_gcwq(work
);
1306 if (gcwq
&& gcwq
->cpu
!= WORK_CPU_UNBOUND
)
1309 lcpu
= raw_smp_processor_id();
1311 lcpu
= WORK_CPU_UNBOUND
;
1314 set_work_cwq(work
, get_cwq(lcpu
, wq
), 0);
1316 timer
->expires
= jiffies
+ delay
;
1318 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1319 add_timer_on(timer
, cpu
);
1325 * queue_delayed_work_on - queue work on specific CPU after delay
1326 * @cpu: CPU number to execute work on
1327 * @wq: workqueue to use
1328 * @dwork: work to queue
1329 * @delay: number of jiffies to wait before queueing
1331 * Returns %false if @work was already on a queue, %true otherwise. If
1332 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1335 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1336 struct delayed_work
*dwork
, unsigned long delay
)
1338 struct work_struct
*work
= &dwork
->work
;
1340 unsigned long flags
;
1343 return queue_work_on(cpu
, wq
, &dwork
->work
);
1345 /* read the comment in __queue_work() */
1346 local_irq_save(flags
);
1348 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1349 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1353 local_irq_restore(flags
);
1356 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1359 * queue_delayed_work - queue work on a workqueue after delay
1360 * @wq: workqueue to use
1361 * @dwork: delayable work to queue
1362 * @delay: number of jiffies to wait before queueing
1364 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1366 bool queue_delayed_work(struct workqueue_struct
*wq
,
1367 struct delayed_work
*dwork
, unsigned long delay
)
1369 return queue_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1371 EXPORT_SYMBOL_GPL(queue_delayed_work
);
1374 * worker_enter_idle - enter idle state
1375 * @worker: worker which is entering idle state
1377 * @worker is entering idle state. Update stats and idle timer if
1381 * spin_lock_irq(gcwq->lock).
1383 static void worker_enter_idle(struct worker
*worker
)
1385 struct worker_pool
*pool
= worker
->pool
;
1386 struct global_cwq
*gcwq
= pool
->gcwq
;
1388 BUG_ON(worker
->flags
& WORKER_IDLE
);
1389 BUG_ON(!list_empty(&worker
->entry
) &&
1390 (worker
->hentry
.next
|| worker
->hentry
.pprev
));
1392 /* can't use worker_set_flags(), also called from start_worker() */
1393 worker
->flags
|= WORKER_IDLE
;
1395 worker
->last_active
= jiffies
;
1397 /* idle_list is LIFO */
1398 list_add(&worker
->entry
, &pool
->idle_list
);
1400 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1401 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1404 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1405 * gcwq->lock between setting %WORKER_UNBOUND and zapping
1406 * nr_running, the warning may trigger spuriously. Check iff
1407 * unbind is not in progress.
1409 WARN_ON_ONCE(!(gcwq
->flags
& GCWQ_DISASSOCIATED
) &&
1410 pool
->nr_workers
== pool
->nr_idle
&&
1411 atomic_read(get_pool_nr_running(pool
)));
1415 * worker_leave_idle - leave idle state
1416 * @worker: worker which is leaving idle state
1418 * @worker is leaving idle state. Update stats.
1421 * spin_lock_irq(gcwq->lock).
1423 static void worker_leave_idle(struct worker
*worker
)
1425 struct worker_pool
*pool
= worker
->pool
;
1427 BUG_ON(!(worker
->flags
& WORKER_IDLE
));
1428 worker_clr_flags(worker
, WORKER_IDLE
);
1430 list_del_init(&worker
->entry
);
1434 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1437 * Works which are scheduled while the cpu is online must at least be
1438 * scheduled to a worker which is bound to the cpu so that if they are
1439 * flushed from cpu callbacks while cpu is going down, they are
1440 * guaranteed to execute on the cpu.
1442 * This function is to be used by rogue workers and rescuers to bind
1443 * themselves to the target cpu and may race with cpu going down or
1444 * coming online. kthread_bind() can't be used because it may put the
1445 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1446 * verbatim as it's best effort and blocking and gcwq may be
1447 * [dis]associated in the meantime.
1449 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1450 * binding against %GCWQ_DISASSOCIATED which is set during
1451 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1452 * enters idle state or fetches works without dropping lock, it can
1453 * guarantee the scheduling requirement described in the first paragraph.
1456 * Might sleep. Called without any lock but returns with gcwq->lock
1460 * %true if the associated gcwq is online (@worker is successfully
1461 * bound), %false if offline.
1463 static bool worker_maybe_bind_and_lock(struct worker
*worker
)
1464 __acquires(&gcwq
->lock
)
1466 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1467 struct task_struct
*task
= worker
->task
;
1471 * The following call may fail, succeed or succeed
1472 * without actually migrating the task to the cpu if
1473 * it races with cpu hotunplug operation. Verify
1474 * against GCWQ_DISASSOCIATED.
1476 if (!(gcwq
->flags
& GCWQ_DISASSOCIATED
))
1477 set_cpus_allowed_ptr(task
, get_cpu_mask(gcwq
->cpu
));
1479 spin_lock_irq(&gcwq
->lock
);
1480 if (gcwq
->flags
& GCWQ_DISASSOCIATED
)
1482 if (task_cpu(task
) == gcwq
->cpu
&&
1483 cpumask_equal(¤t
->cpus_allowed
,
1484 get_cpu_mask(gcwq
->cpu
)))
1486 spin_unlock_irq(&gcwq
->lock
);
1489 * We've raced with CPU hot[un]plug. Give it a breather
1490 * and retry migration. cond_resched() is required here;
1491 * otherwise, we might deadlock against cpu_stop trying to
1492 * bring down the CPU on non-preemptive kernel.
1499 struct idle_rebind
{
1500 int cnt
; /* # workers to be rebound */
1501 struct completion done
; /* all workers rebound */
1505 * Rebind an idle @worker to its CPU. During CPU onlining, this has to
1506 * happen synchronously for idle workers. worker_thread() will test
1507 * %WORKER_REBIND before leaving idle and call this function.
1509 static void idle_worker_rebind(struct worker
*worker
)
1511 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1513 /* CPU must be online at this point */
1514 WARN_ON(!worker_maybe_bind_and_lock(worker
));
1515 if (!--worker
->idle_rebind
->cnt
)
1516 complete(&worker
->idle_rebind
->done
);
1517 spin_unlock_irq(&worker
->pool
->gcwq
->lock
);
1519 /* we did our part, wait for rebind_workers() to finish up */
1520 wait_event(gcwq
->rebind_hold
, !(worker
->flags
& WORKER_REBIND
));
1524 * Function for @worker->rebind.work used to rebind unbound busy workers to
1525 * the associated cpu which is coming back online. This is scheduled by
1526 * cpu up but can race with other cpu hotplug operations and may be
1527 * executed twice without intervening cpu down.
1529 static void busy_worker_rebind_fn(struct work_struct
*work
)
1531 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1532 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1534 if (worker_maybe_bind_and_lock(worker
))
1535 worker_clr_flags(worker
, WORKER_REBIND
);
1537 spin_unlock_irq(&gcwq
->lock
);
1541 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1542 * @gcwq: gcwq of interest
1544 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1545 * is different for idle and busy ones.
1547 * The idle ones should be rebound synchronously and idle rebinding should
1548 * be complete before any worker starts executing work items with
1549 * concurrency management enabled; otherwise, scheduler may oops trying to
1550 * wake up non-local idle worker from wq_worker_sleeping().
1552 * This is achieved by repeatedly requesting rebinding until all idle
1553 * workers are known to have been rebound under @gcwq->lock and holding all
1554 * idle workers from becoming busy until idle rebinding is complete.
1556 * Once idle workers are rebound, busy workers can be rebound as they
1557 * finish executing their current work items. Queueing the rebind work at
1558 * the head of their scheduled lists is enough. Note that nr_running will
1559 * be properbly bumped as busy workers rebind.
1561 * On return, all workers are guaranteed to either be bound or have rebind
1562 * work item scheduled.
1564 static void rebind_workers(struct global_cwq
*gcwq
)
1565 __releases(&gcwq
->lock
) __acquires(&gcwq
->lock
)
1567 struct idle_rebind idle_rebind
;
1568 struct worker_pool
*pool
;
1569 struct worker
*worker
;
1570 struct hlist_node
*pos
;
1573 lockdep_assert_held(&gcwq
->lock
);
1575 for_each_worker_pool(pool
, gcwq
)
1576 lockdep_assert_held(&pool
->manager_mutex
);
1579 * Rebind idle workers. Interlocked both ways. We wait for
1580 * workers to rebind via @idle_rebind.done. Workers will wait for
1581 * us to finish up by watching %WORKER_REBIND.
1583 init_completion(&idle_rebind
.done
);
1585 idle_rebind
.cnt
= 1;
1586 INIT_COMPLETION(idle_rebind
.done
);
1588 /* set REBIND and kick idle ones, we'll wait for these later */
1589 for_each_worker_pool(pool
, gcwq
) {
1590 list_for_each_entry(worker
, &pool
->idle_list
, entry
) {
1591 if (worker
->flags
& WORKER_REBIND
)
1594 /* morph UNBOUND to REBIND */
1595 worker
->flags
&= ~WORKER_UNBOUND
;
1596 worker
->flags
|= WORKER_REBIND
;
1599 worker
->idle_rebind
= &idle_rebind
;
1601 /* worker_thread() will call idle_worker_rebind() */
1602 wake_up_process(worker
->task
);
1606 if (--idle_rebind
.cnt
) {
1607 spin_unlock_irq(&gcwq
->lock
);
1608 wait_for_completion(&idle_rebind
.done
);
1609 spin_lock_irq(&gcwq
->lock
);
1610 /* busy ones might have become idle while waiting, retry */
1615 * All idle workers are rebound and waiting for %WORKER_REBIND to
1616 * be cleared inside idle_worker_rebind(). Clear and release.
1617 * Clearing %WORKER_REBIND from this foreign context is safe
1618 * because these workers are still guaranteed to be idle.
1620 for_each_worker_pool(pool
, gcwq
)
1621 list_for_each_entry(worker
, &pool
->idle_list
, entry
)
1622 worker
->flags
&= ~WORKER_REBIND
;
1624 wake_up_all(&gcwq
->rebind_hold
);
1626 /* rebind busy workers */
1627 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
1628 struct work_struct
*rebind_work
= &worker
->rebind_work
;
1630 /* morph UNBOUND to REBIND */
1631 worker
->flags
&= ~WORKER_UNBOUND
;
1632 worker
->flags
|= WORKER_REBIND
;
1634 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
1635 work_data_bits(rebind_work
)))
1638 /* wq doesn't matter, use the default one */
1639 debug_work_activate(rebind_work
);
1640 insert_work(get_cwq(gcwq
->cpu
, system_wq
), rebind_work
,
1641 worker
->scheduled
.next
,
1642 work_color_to_flags(WORK_NO_COLOR
));
1646 static struct worker
*alloc_worker(void)
1648 struct worker
*worker
;
1650 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1652 INIT_LIST_HEAD(&worker
->entry
);
1653 INIT_LIST_HEAD(&worker
->scheduled
);
1654 INIT_WORK(&worker
->rebind_work
, busy_worker_rebind_fn
);
1655 /* on creation a worker is in !idle && prep state */
1656 worker
->flags
= WORKER_PREP
;
1662 * create_worker - create a new workqueue worker
1663 * @pool: pool the new worker will belong to
1665 * Create a new worker which is bound to @pool. The returned worker
1666 * can be started by calling start_worker() or destroyed using
1670 * Might sleep. Does GFP_KERNEL allocations.
1673 * Pointer to the newly created worker.
1675 static struct worker
*create_worker(struct worker_pool
*pool
)
1677 struct global_cwq
*gcwq
= pool
->gcwq
;
1678 const char *pri
= worker_pool_pri(pool
) ? "H" : "";
1679 struct worker
*worker
= NULL
;
1682 spin_lock_irq(&gcwq
->lock
);
1683 while (ida_get_new(&pool
->worker_ida
, &id
)) {
1684 spin_unlock_irq(&gcwq
->lock
);
1685 if (!ida_pre_get(&pool
->worker_ida
, GFP_KERNEL
))
1687 spin_lock_irq(&gcwq
->lock
);
1689 spin_unlock_irq(&gcwq
->lock
);
1691 worker
= alloc_worker();
1695 worker
->pool
= pool
;
1698 if (gcwq
->cpu
!= WORK_CPU_UNBOUND
)
1699 worker
->task
= kthread_create_on_node(worker_thread
,
1700 worker
, cpu_to_node(gcwq
->cpu
),
1701 "kworker/%u:%d%s", gcwq
->cpu
, id
, pri
);
1703 worker
->task
= kthread_create(worker_thread
, worker
,
1704 "kworker/u:%d%s", id
, pri
);
1705 if (IS_ERR(worker
->task
))
1708 if (worker_pool_pri(pool
))
1709 set_user_nice(worker
->task
, HIGHPRI_NICE_LEVEL
);
1712 * Determine CPU binding of the new worker depending on
1713 * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the
1714 * flag remains stable across this function. See the comments
1715 * above the flag definition for details.
1717 * As an unbound worker may later become a regular one if CPU comes
1718 * online, make sure every worker has %PF_THREAD_BOUND set.
1720 if (!(gcwq
->flags
& GCWQ_DISASSOCIATED
)) {
1721 kthread_bind(worker
->task
, gcwq
->cpu
);
1723 worker
->task
->flags
|= PF_THREAD_BOUND
;
1724 worker
->flags
|= WORKER_UNBOUND
;
1730 spin_lock_irq(&gcwq
->lock
);
1731 ida_remove(&pool
->worker_ida
, id
);
1732 spin_unlock_irq(&gcwq
->lock
);
1739 * start_worker - start a newly created worker
1740 * @worker: worker to start
1742 * Make the gcwq aware of @worker and start it.
1745 * spin_lock_irq(gcwq->lock).
1747 static void start_worker(struct worker
*worker
)
1749 worker
->flags
|= WORKER_STARTED
;
1750 worker
->pool
->nr_workers
++;
1751 worker_enter_idle(worker
);
1752 wake_up_process(worker
->task
);
1756 * destroy_worker - destroy a workqueue worker
1757 * @worker: worker to be destroyed
1759 * Destroy @worker and adjust @gcwq stats accordingly.
1762 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1764 static void destroy_worker(struct worker
*worker
)
1766 struct worker_pool
*pool
= worker
->pool
;
1767 struct global_cwq
*gcwq
= pool
->gcwq
;
1768 int id
= worker
->id
;
1770 /* sanity check frenzy */
1771 BUG_ON(worker
->current_work
);
1772 BUG_ON(!list_empty(&worker
->scheduled
));
1774 if (worker
->flags
& WORKER_STARTED
)
1776 if (worker
->flags
& WORKER_IDLE
)
1779 list_del_init(&worker
->entry
);
1780 worker
->flags
|= WORKER_DIE
;
1782 spin_unlock_irq(&gcwq
->lock
);
1784 kthread_stop(worker
->task
);
1787 spin_lock_irq(&gcwq
->lock
);
1788 ida_remove(&pool
->worker_ida
, id
);
1791 static void idle_worker_timeout(unsigned long __pool
)
1793 struct worker_pool
*pool
= (void *)__pool
;
1794 struct global_cwq
*gcwq
= pool
->gcwq
;
1796 spin_lock_irq(&gcwq
->lock
);
1798 if (too_many_workers(pool
)) {
1799 struct worker
*worker
;
1800 unsigned long expires
;
1802 /* idle_list is kept in LIFO order, check the last one */
1803 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1804 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1806 if (time_before(jiffies
, expires
))
1807 mod_timer(&pool
->idle_timer
, expires
);
1809 /* it's been idle for too long, wake up manager */
1810 pool
->flags
|= POOL_MANAGE_WORKERS
;
1811 wake_up_worker(pool
);
1815 spin_unlock_irq(&gcwq
->lock
);
1818 static bool send_mayday(struct work_struct
*work
)
1820 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1821 struct workqueue_struct
*wq
= cwq
->wq
;
1824 if (!(wq
->flags
& WQ_RESCUER
))
1827 /* mayday mayday mayday */
1828 cpu
= cwq
->pool
->gcwq
->cpu
;
1829 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1830 if (cpu
== WORK_CPU_UNBOUND
)
1832 if (!mayday_test_and_set_cpu(cpu
, wq
->mayday_mask
))
1833 wake_up_process(wq
->rescuer
->task
);
1837 static void gcwq_mayday_timeout(unsigned long __pool
)
1839 struct worker_pool
*pool
= (void *)__pool
;
1840 struct global_cwq
*gcwq
= pool
->gcwq
;
1841 struct work_struct
*work
;
1843 spin_lock_irq(&gcwq
->lock
);
1845 if (need_to_create_worker(pool
)) {
1847 * We've been trying to create a new worker but
1848 * haven't been successful. We might be hitting an
1849 * allocation deadlock. Send distress signals to
1852 list_for_each_entry(work
, &pool
->worklist
, entry
)
1856 spin_unlock_irq(&gcwq
->lock
);
1858 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1862 * maybe_create_worker - create a new worker if necessary
1863 * @pool: pool to create a new worker for
1865 * Create a new worker for @pool if necessary. @pool is guaranteed to
1866 * have at least one idle worker on return from this function. If
1867 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1868 * sent to all rescuers with works scheduled on @pool to resolve
1869 * possible allocation deadlock.
1871 * On return, need_to_create_worker() is guaranteed to be false and
1872 * may_start_working() true.
1875 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1876 * multiple times. Does GFP_KERNEL allocations. Called only from
1880 * false if no action was taken and gcwq->lock stayed locked, true
1883 static bool maybe_create_worker(struct worker_pool
*pool
)
1884 __releases(&gcwq
->lock
)
1885 __acquires(&gcwq
->lock
)
1887 struct global_cwq
*gcwq
= pool
->gcwq
;
1889 if (!need_to_create_worker(pool
))
1892 spin_unlock_irq(&gcwq
->lock
);
1894 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1895 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1898 struct worker
*worker
;
1900 worker
= create_worker(pool
);
1902 del_timer_sync(&pool
->mayday_timer
);
1903 spin_lock_irq(&gcwq
->lock
);
1904 start_worker(worker
);
1905 BUG_ON(need_to_create_worker(pool
));
1909 if (!need_to_create_worker(pool
))
1912 __set_current_state(TASK_INTERRUPTIBLE
);
1913 schedule_timeout(CREATE_COOLDOWN
);
1915 if (!need_to_create_worker(pool
))
1919 del_timer_sync(&pool
->mayday_timer
);
1920 spin_lock_irq(&gcwq
->lock
);
1921 if (need_to_create_worker(pool
))
1927 * maybe_destroy_worker - destroy workers which have been idle for a while
1928 * @pool: pool to destroy workers for
1930 * Destroy @pool workers which have been idle for longer than
1931 * IDLE_WORKER_TIMEOUT.
1934 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1935 * multiple times. Called only from manager.
1938 * false if no action was taken and gcwq->lock stayed locked, true
1941 static bool maybe_destroy_workers(struct worker_pool
*pool
)
1945 while (too_many_workers(pool
)) {
1946 struct worker
*worker
;
1947 unsigned long expires
;
1949 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1950 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1952 if (time_before(jiffies
, expires
)) {
1953 mod_timer(&pool
->idle_timer
, expires
);
1957 destroy_worker(worker
);
1965 * manage_workers - manage worker pool
1968 * Assume the manager role and manage gcwq worker pool @worker belongs
1969 * to. At any given time, there can be only zero or one manager per
1970 * gcwq. The exclusion is handled automatically by this function.
1972 * The caller can safely start processing works on false return. On
1973 * true return, it's guaranteed that need_to_create_worker() is false
1974 * and may_start_working() is true.
1977 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1978 * multiple times. Does GFP_KERNEL allocations.
1981 * false if no action was taken and gcwq->lock stayed locked, true if
1982 * some action was taken.
1984 static bool manage_workers(struct worker
*worker
)
1986 struct worker_pool
*pool
= worker
->pool
;
1989 if (!mutex_trylock(&pool
->manager_mutex
))
1992 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
1995 * Destroy and then create so that may_start_working() is true
1998 ret
|= maybe_destroy_workers(pool
);
1999 ret
|= maybe_create_worker(pool
);
2001 mutex_unlock(&pool
->manager_mutex
);
2006 * process_one_work - process single work
2008 * @work: work to process
2010 * Process @work. This function contains all the logics necessary to
2011 * process a single work including synchronization against and
2012 * interaction with other workers on the same cpu, queueing and
2013 * flushing. As long as context requirement is met, any worker can
2014 * call this function to process a work.
2017 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2019 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2020 __releases(&gcwq
->lock
)
2021 __acquires(&gcwq
->lock
)
2023 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
2024 struct worker_pool
*pool
= worker
->pool
;
2025 struct global_cwq
*gcwq
= pool
->gcwq
;
2026 struct hlist_head
*bwh
= busy_worker_head(gcwq
, work
);
2027 bool cpu_intensive
= cwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2028 work_func_t f
= work
->func
;
2030 struct worker
*collision
;
2031 #ifdef CONFIG_LOCKDEP
2033 * It is permissible to free the struct work_struct from
2034 * inside the function that is called from it, this we need to
2035 * take into account for lockdep too. To avoid bogus "held
2036 * lock freed" warnings as well as problems when looking into
2037 * work->lockdep_map, make a copy and use that here.
2039 struct lockdep_map lockdep_map
;
2041 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2044 * Ensure we're on the correct CPU. DISASSOCIATED test is
2045 * necessary to avoid spurious warnings from rescuers servicing the
2046 * unbound or a disassociated gcwq.
2048 WARN_ON_ONCE(!(worker
->flags
& (WORKER_UNBOUND
| WORKER_REBIND
)) &&
2049 !(gcwq
->flags
& GCWQ_DISASSOCIATED
) &&
2050 raw_smp_processor_id() != gcwq
->cpu
);
2053 * A single work shouldn't be executed concurrently by
2054 * multiple workers on a single cpu. Check whether anyone is
2055 * already processing the work. If so, defer the work to the
2056 * currently executing one.
2058 collision
= __find_worker_executing_work(gcwq
, bwh
, work
);
2059 if (unlikely(collision
)) {
2060 move_linked_works(work
, &collision
->scheduled
, NULL
);
2064 /* claim and dequeue */
2065 debug_work_deactivate(work
);
2066 hlist_add_head(&worker
->hentry
, bwh
);
2067 worker
->current_work
= work
;
2068 worker
->current_cwq
= cwq
;
2069 work_color
= get_work_color(work
);
2071 list_del_init(&work
->entry
);
2074 * CPU intensive works don't participate in concurrency
2075 * management. They're the scheduler's responsibility.
2077 if (unlikely(cpu_intensive
))
2078 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2081 * Unbound gcwq isn't concurrency managed and work items should be
2082 * executed ASAP. Wake up another worker if necessary.
2084 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2085 wake_up_worker(pool
);
2088 * Record the last CPU and clear PENDING. The following wmb is
2089 * paired with the implied mb in test_and_set_bit(PENDING) and
2090 * ensures all updates to @work made here are visible to and
2091 * precede any updates by the next PENDING owner. Also, clear
2092 * PENDING inside @gcwq->lock so that PENDING and queued state
2093 * changes happen together while IRQ is disabled.
2096 set_work_cpu_and_clear_pending(work
, gcwq
->cpu
);
2098 spin_unlock_irq(&gcwq
->lock
);
2100 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2101 lock_map_acquire(&lockdep_map
);
2102 trace_workqueue_execute_start(work
);
2105 * While we must be careful to not use "work" after this, the trace
2106 * point will only record its address.
2108 trace_workqueue_execute_end(work
);
2109 lock_map_release(&lockdep_map
);
2110 lock_map_release(&cwq
->wq
->lockdep_map
);
2112 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2113 printk(KERN_ERR
"BUG: workqueue leaked lock or atomic: "
2115 current
->comm
, preempt_count(), task_pid_nr(current
));
2116 printk(KERN_ERR
" last function: ");
2117 print_symbol("%s\n", (unsigned long)f
);
2118 debug_show_held_locks(current
);
2122 spin_lock_irq(&gcwq
->lock
);
2124 /* clear cpu intensive status */
2125 if (unlikely(cpu_intensive
))
2126 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2128 /* we're done with it, release */
2129 hlist_del_init(&worker
->hentry
);
2130 worker
->current_work
= NULL
;
2131 worker
->current_cwq
= NULL
;
2132 cwq_dec_nr_in_flight(cwq
, work_color
, false);
2136 * process_scheduled_works - process scheduled works
2139 * Process all scheduled works. Please note that the scheduled list
2140 * may change while processing a work, so this function repeatedly
2141 * fetches a work from the top and executes it.
2144 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2147 static void process_scheduled_works(struct worker
*worker
)
2149 while (!list_empty(&worker
->scheduled
)) {
2150 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2151 struct work_struct
, entry
);
2152 process_one_work(worker
, work
);
2157 * worker_thread - the worker thread function
2160 * The gcwq worker thread function. There's a single dynamic pool of
2161 * these per each cpu. These workers process all works regardless of
2162 * their specific target workqueue. The only exception is works which
2163 * belong to workqueues with a rescuer which will be explained in
2166 static int worker_thread(void *__worker
)
2168 struct worker
*worker
= __worker
;
2169 struct worker_pool
*pool
= worker
->pool
;
2170 struct global_cwq
*gcwq
= pool
->gcwq
;
2172 /* tell the scheduler that this is a workqueue worker */
2173 worker
->task
->flags
|= PF_WQ_WORKER
;
2175 spin_lock_irq(&gcwq
->lock
);
2178 * DIE can be set only while idle and REBIND set while busy has
2179 * @worker->rebind_work scheduled. Checking here is enough.
2181 if (unlikely(worker
->flags
& (WORKER_REBIND
| WORKER_DIE
))) {
2182 spin_unlock_irq(&gcwq
->lock
);
2184 if (worker
->flags
& WORKER_DIE
) {
2185 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2189 idle_worker_rebind(worker
);
2193 worker_leave_idle(worker
);
2195 /* no more worker necessary? */
2196 if (!need_more_worker(pool
))
2199 /* do we need to manage? */
2200 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2204 * ->scheduled list can only be filled while a worker is
2205 * preparing to process a work or actually processing it.
2206 * Make sure nobody diddled with it while I was sleeping.
2208 BUG_ON(!list_empty(&worker
->scheduled
));
2211 * When control reaches this point, we're guaranteed to have
2212 * at least one idle worker or that someone else has already
2213 * assumed the manager role.
2215 worker_clr_flags(worker
, WORKER_PREP
);
2218 struct work_struct
*work
=
2219 list_first_entry(&pool
->worklist
,
2220 struct work_struct
, entry
);
2222 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2223 /* optimization path, not strictly necessary */
2224 process_one_work(worker
, work
);
2225 if (unlikely(!list_empty(&worker
->scheduled
)))
2226 process_scheduled_works(worker
);
2228 move_linked_works(work
, &worker
->scheduled
, NULL
);
2229 process_scheduled_works(worker
);
2231 } while (keep_working(pool
));
2233 worker_set_flags(worker
, WORKER_PREP
, false);
2235 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2239 * gcwq->lock is held and there's no work to process and no
2240 * need to manage, sleep. Workers are woken up only while
2241 * holding gcwq->lock or from local cpu, so setting the
2242 * current state before releasing gcwq->lock is enough to
2243 * prevent losing any event.
2245 worker_enter_idle(worker
);
2246 __set_current_state(TASK_INTERRUPTIBLE
);
2247 spin_unlock_irq(&gcwq
->lock
);
2253 * rescuer_thread - the rescuer thread function
2254 * @__wq: the associated workqueue
2256 * Workqueue rescuer thread function. There's one rescuer for each
2257 * workqueue which has WQ_RESCUER set.
2259 * Regular work processing on a gcwq may block trying to create a new
2260 * worker which uses GFP_KERNEL allocation which has slight chance of
2261 * developing into deadlock if some works currently on the same queue
2262 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2263 * the problem rescuer solves.
2265 * When such condition is possible, the gcwq summons rescuers of all
2266 * workqueues which have works queued on the gcwq and let them process
2267 * those works so that forward progress can be guaranteed.
2269 * This should happen rarely.
2271 static int rescuer_thread(void *__wq
)
2273 struct workqueue_struct
*wq
= __wq
;
2274 struct worker
*rescuer
= wq
->rescuer
;
2275 struct list_head
*scheduled
= &rescuer
->scheduled
;
2276 bool is_unbound
= wq
->flags
& WQ_UNBOUND
;
2279 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2281 set_current_state(TASK_INTERRUPTIBLE
);
2283 if (kthread_should_stop())
2287 * See whether any cpu is asking for help. Unbounded
2288 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2290 for_each_mayday_cpu(cpu
, wq
->mayday_mask
) {
2291 unsigned int tcpu
= is_unbound
? WORK_CPU_UNBOUND
: cpu
;
2292 struct cpu_workqueue_struct
*cwq
= get_cwq(tcpu
, wq
);
2293 struct worker_pool
*pool
= cwq
->pool
;
2294 struct global_cwq
*gcwq
= pool
->gcwq
;
2295 struct work_struct
*work
, *n
;
2297 __set_current_state(TASK_RUNNING
);
2298 mayday_clear_cpu(cpu
, wq
->mayday_mask
);
2300 /* migrate to the target cpu if possible */
2301 rescuer
->pool
= pool
;
2302 worker_maybe_bind_and_lock(rescuer
);
2305 * Slurp in all works issued via this workqueue and
2308 BUG_ON(!list_empty(&rescuer
->scheduled
));
2309 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2310 if (get_work_cwq(work
) == cwq
)
2311 move_linked_works(work
, scheduled
, &n
);
2313 process_scheduled_works(rescuer
);
2316 * Leave this gcwq. If keep_working() is %true, notify a
2317 * regular worker; otherwise, we end up with 0 concurrency
2318 * and stalling the execution.
2320 if (keep_working(pool
))
2321 wake_up_worker(pool
);
2323 spin_unlock_irq(&gcwq
->lock
);
2331 struct work_struct work
;
2332 struct completion done
;
2335 static void wq_barrier_func(struct work_struct
*work
)
2337 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2338 complete(&barr
->done
);
2342 * insert_wq_barrier - insert a barrier work
2343 * @cwq: cwq to insert barrier into
2344 * @barr: wq_barrier to insert
2345 * @target: target work to attach @barr to
2346 * @worker: worker currently executing @target, NULL if @target is not executing
2348 * @barr is linked to @target such that @barr is completed only after
2349 * @target finishes execution. Please note that the ordering
2350 * guarantee is observed only with respect to @target and on the local
2353 * Currently, a queued barrier can't be canceled. This is because
2354 * try_to_grab_pending() can't determine whether the work to be
2355 * grabbed is at the head of the queue and thus can't clear LINKED
2356 * flag of the previous work while there must be a valid next work
2357 * after a work with LINKED flag set.
2359 * Note that when @worker is non-NULL, @target may be modified
2360 * underneath us, so we can't reliably determine cwq from @target.
2363 * spin_lock_irq(gcwq->lock).
2365 static void insert_wq_barrier(struct cpu_workqueue_struct
*cwq
,
2366 struct wq_barrier
*barr
,
2367 struct work_struct
*target
, struct worker
*worker
)
2369 struct list_head
*head
;
2370 unsigned int linked
= 0;
2373 * debugobject calls are safe here even with gcwq->lock locked
2374 * as we know for sure that this will not trigger any of the
2375 * checks and call back into the fixup functions where we
2378 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2379 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2380 init_completion(&barr
->done
);
2383 * If @target is currently being executed, schedule the
2384 * barrier to the worker; otherwise, put it after @target.
2387 head
= worker
->scheduled
.next
;
2389 unsigned long *bits
= work_data_bits(target
);
2391 head
= target
->entry
.next
;
2392 /* there can already be other linked works, inherit and set */
2393 linked
= *bits
& WORK_STRUCT_LINKED
;
2394 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2397 debug_work_activate(&barr
->work
);
2398 insert_work(cwq
, &barr
->work
, head
,
2399 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2403 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2404 * @wq: workqueue being flushed
2405 * @flush_color: new flush color, < 0 for no-op
2406 * @work_color: new work color, < 0 for no-op
2408 * Prepare cwqs for workqueue flushing.
2410 * If @flush_color is non-negative, flush_color on all cwqs should be
2411 * -1. If no cwq has in-flight commands at the specified color, all
2412 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2413 * has in flight commands, its cwq->flush_color is set to
2414 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2415 * wakeup logic is armed and %true is returned.
2417 * The caller should have initialized @wq->first_flusher prior to
2418 * calling this function with non-negative @flush_color. If
2419 * @flush_color is negative, no flush color update is done and %false
2422 * If @work_color is non-negative, all cwqs should have the same
2423 * work_color which is previous to @work_color and all will be
2424 * advanced to @work_color.
2427 * mutex_lock(wq->flush_mutex).
2430 * %true if @flush_color >= 0 and there's something to flush. %false
2433 static bool flush_workqueue_prep_cwqs(struct workqueue_struct
*wq
,
2434 int flush_color
, int work_color
)
2439 if (flush_color
>= 0) {
2440 BUG_ON(atomic_read(&wq
->nr_cwqs_to_flush
));
2441 atomic_set(&wq
->nr_cwqs_to_flush
, 1);
2444 for_each_cwq_cpu(cpu
, wq
) {
2445 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2446 struct global_cwq
*gcwq
= cwq
->pool
->gcwq
;
2448 spin_lock_irq(&gcwq
->lock
);
2450 if (flush_color
>= 0) {
2451 BUG_ON(cwq
->flush_color
!= -1);
2453 if (cwq
->nr_in_flight
[flush_color
]) {
2454 cwq
->flush_color
= flush_color
;
2455 atomic_inc(&wq
->nr_cwqs_to_flush
);
2460 if (work_color
>= 0) {
2461 BUG_ON(work_color
!= work_next_color(cwq
->work_color
));
2462 cwq
->work_color
= work_color
;
2465 spin_unlock_irq(&gcwq
->lock
);
2468 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_cwqs_to_flush
))
2469 complete(&wq
->first_flusher
->done
);
2475 * flush_workqueue - ensure that any scheduled work has run to completion.
2476 * @wq: workqueue to flush
2478 * Forces execution of the workqueue and blocks until its completion.
2479 * This is typically used in driver shutdown handlers.
2481 * We sleep until all works which were queued on entry have been handled,
2482 * but we are not livelocked by new incoming ones.
2484 void flush_workqueue(struct workqueue_struct
*wq
)
2486 struct wq_flusher this_flusher
= {
2487 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2489 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2493 lock_map_acquire(&wq
->lockdep_map
);
2494 lock_map_release(&wq
->lockdep_map
);
2496 mutex_lock(&wq
->flush_mutex
);
2499 * Start-to-wait phase
2501 next_color
= work_next_color(wq
->work_color
);
2503 if (next_color
!= wq
->flush_color
) {
2505 * Color space is not full. The current work_color
2506 * becomes our flush_color and work_color is advanced
2509 BUG_ON(!list_empty(&wq
->flusher_overflow
));
2510 this_flusher
.flush_color
= wq
->work_color
;
2511 wq
->work_color
= next_color
;
2513 if (!wq
->first_flusher
) {
2514 /* no flush in progress, become the first flusher */
2515 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2517 wq
->first_flusher
= &this_flusher
;
2519 if (!flush_workqueue_prep_cwqs(wq
, wq
->flush_color
,
2521 /* nothing to flush, done */
2522 wq
->flush_color
= next_color
;
2523 wq
->first_flusher
= NULL
;
2528 BUG_ON(wq
->flush_color
== this_flusher
.flush_color
);
2529 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2530 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2534 * Oops, color space is full, wait on overflow queue.
2535 * The next flush completion will assign us
2536 * flush_color and transfer to flusher_queue.
2538 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2541 mutex_unlock(&wq
->flush_mutex
);
2543 wait_for_completion(&this_flusher
.done
);
2546 * Wake-up-and-cascade phase
2548 * First flushers are responsible for cascading flushes and
2549 * handling overflow. Non-first flushers can simply return.
2551 if (wq
->first_flusher
!= &this_flusher
)
2554 mutex_lock(&wq
->flush_mutex
);
2556 /* we might have raced, check again with mutex held */
2557 if (wq
->first_flusher
!= &this_flusher
)
2560 wq
->first_flusher
= NULL
;
2562 BUG_ON(!list_empty(&this_flusher
.list
));
2563 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2566 struct wq_flusher
*next
, *tmp
;
2568 /* complete all the flushers sharing the current flush color */
2569 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2570 if (next
->flush_color
!= wq
->flush_color
)
2572 list_del_init(&next
->list
);
2573 complete(&next
->done
);
2576 BUG_ON(!list_empty(&wq
->flusher_overflow
) &&
2577 wq
->flush_color
!= work_next_color(wq
->work_color
));
2579 /* this flush_color is finished, advance by one */
2580 wq
->flush_color
= work_next_color(wq
->flush_color
);
2582 /* one color has been freed, handle overflow queue */
2583 if (!list_empty(&wq
->flusher_overflow
)) {
2585 * Assign the same color to all overflowed
2586 * flushers, advance work_color and append to
2587 * flusher_queue. This is the start-to-wait
2588 * phase for these overflowed flushers.
2590 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2591 tmp
->flush_color
= wq
->work_color
;
2593 wq
->work_color
= work_next_color(wq
->work_color
);
2595 list_splice_tail_init(&wq
->flusher_overflow
,
2596 &wq
->flusher_queue
);
2597 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2600 if (list_empty(&wq
->flusher_queue
)) {
2601 BUG_ON(wq
->flush_color
!= wq
->work_color
);
2606 * Need to flush more colors. Make the next flusher
2607 * the new first flusher and arm cwqs.
2609 BUG_ON(wq
->flush_color
== wq
->work_color
);
2610 BUG_ON(wq
->flush_color
!= next
->flush_color
);
2612 list_del_init(&next
->list
);
2613 wq
->first_flusher
= next
;
2615 if (flush_workqueue_prep_cwqs(wq
, wq
->flush_color
, -1))
2619 * Meh... this color is already done, clear first
2620 * flusher and repeat cascading.
2622 wq
->first_flusher
= NULL
;
2626 mutex_unlock(&wq
->flush_mutex
);
2628 EXPORT_SYMBOL_GPL(flush_workqueue
);
2631 * drain_workqueue - drain a workqueue
2632 * @wq: workqueue to drain
2634 * Wait until the workqueue becomes empty. While draining is in progress,
2635 * only chain queueing is allowed. IOW, only currently pending or running
2636 * work items on @wq can queue further work items on it. @wq is flushed
2637 * repeatedly until it becomes empty. The number of flushing is detemined
2638 * by the depth of chaining and should be relatively short. Whine if it
2641 void drain_workqueue(struct workqueue_struct
*wq
)
2643 unsigned int flush_cnt
= 0;
2647 * __queue_work() needs to test whether there are drainers, is much
2648 * hotter than drain_workqueue() and already looks at @wq->flags.
2649 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2651 spin_lock(&workqueue_lock
);
2652 if (!wq
->nr_drainers
++)
2653 wq
->flags
|= WQ_DRAINING
;
2654 spin_unlock(&workqueue_lock
);
2656 flush_workqueue(wq
);
2658 for_each_cwq_cpu(cpu
, wq
) {
2659 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2662 spin_lock_irq(&cwq
->pool
->gcwq
->lock
);
2663 drained
= !cwq
->nr_active
&& list_empty(&cwq
->delayed_works
);
2664 spin_unlock_irq(&cwq
->pool
->gcwq
->lock
);
2669 if (++flush_cnt
== 10 ||
2670 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2671 pr_warning("workqueue %s: flush on destruction isn't complete after %u tries\n",
2672 wq
->name
, flush_cnt
);
2676 spin_lock(&workqueue_lock
);
2677 if (!--wq
->nr_drainers
)
2678 wq
->flags
&= ~WQ_DRAINING
;
2679 spin_unlock(&workqueue_lock
);
2681 EXPORT_SYMBOL_GPL(drain_workqueue
);
2683 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
,
2684 bool wait_executing
)
2686 struct worker
*worker
= NULL
;
2687 struct global_cwq
*gcwq
;
2688 struct cpu_workqueue_struct
*cwq
;
2691 gcwq
= get_work_gcwq(work
);
2695 spin_lock_irq(&gcwq
->lock
);
2696 if (!list_empty(&work
->entry
)) {
2698 * See the comment near try_to_grab_pending()->smp_rmb().
2699 * If it was re-queued to a different gcwq under us, we
2700 * are not going to wait.
2703 cwq
= get_work_cwq(work
);
2704 if (unlikely(!cwq
|| gcwq
!= cwq
->pool
->gcwq
))
2706 } else if (wait_executing
) {
2707 worker
= find_worker_executing_work(gcwq
, work
);
2710 cwq
= worker
->current_cwq
;
2714 insert_wq_barrier(cwq
, barr
, work
, worker
);
2715 spin_unlock_irq(&gcwq
->lock
);
2718 * If @max_active is 1 or rescuer is in use, flushing another work
2719 * item on the same workqueue may lead to deadlock. Make sure the
2720 * flusher is not running on the same workqueue by verifying write
2723 if (cwq
->wq
->saved_max_active
== 1 || cwq
->wq
->flags
& WQ_RESCUER
)
2724 lock_map_acquire(&cwq
->wq
->lockdep_map
);
2726 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2727 lock_map_release(&cwq
->wq
->lockdep_map
);
2731 spin_unlock_irq(&gcwq
->lock
);
2736 * flush_work - wait for a work to finish executing the last queueing instance
2737 * @work: the work to flush
2739 * Wait until @work has finished execution. This function considers
2740 * only the last queueing instance of @work. If @work has been
2741 * enqueued across different CPUs on a non-reentrant workqueue or on
2742 * multiple workqueues, @work might still be executing on return on
2743 * some of the CPUs from earlier queueing.
2745 * If @work was queued only on a non-reentrant, ordered or unbound
2746 * workqueue, @work is guaranteed to be idle on return if it hasn't
2747 * been requeued since flush started.
2750 * %true if flush_work() waited for the work to finish execution,
2751 * %false if it was already idle.
2753 bool flush_work(struct work_struct
*work
)
2755 struct wq_barrier barr
;
2757 lock_map_acquire(&work
->lockdep_map
);
2758 lock_map_release(&work
->lockdep_map
);
2760 if (start_flush_work(work
, &barr
, true)) {
2761 wait_for_completion(&barr
.done
);
2762 destroy_work_on_stack(&barr
.work
);
2767 EXPORT_SYMBOL_GPL(flush_work
);
2769 static bool wait_on_cpu_work(struct global_cwq
*gcwq
, struct work_struct
*work
)
2771 struct wq_barrier barr
;
2772 struct worker
*worker
;
2774 spin_lock_irq(&gcwq
->lock
);
2776 worker
= find_worker_executing_work(gcwq
, work
);
2777 if (unlikely(worker
))
2778 insert_wq_barrier(worker
->current_cwq
, &barr
, work
, worker
);
2780 spin_unlock_irq(&gcwq
->lock
);
2782 if (unlikely(worker
)) {
2783 wait_for_completion(&barr
.done
);
2784 destroy_work_on_stack(&barr
.work
);
2790 static bool wait_on_work(struct work_struct
*work
)
2797 lock_map_acquire(&work
->lockdep_map
);
2798 lock_map_release(&work
->lockdep_map
);
2800 for_each_gcwq_cpu(cpu
)
2801 ret
|= wait_on_cpu_work(get_gcwq(cpu
), work
);
2806 * flush_work_sync - wait until a work has finished execution
2807 * @work: the work to flush
2809 * Wait until @work has finished execution. On return, it's
2810 * guaranteed that all queueing instances of @work which happened
2811 * before this function is called are finished. In other words, if
2812 * @work hasn't been requeued since this function was called, @work is
2813 * guaranteed to be idle on return.
2816 * %true if flush_work_sync() waited for the work to finish execution,
2817 * %false if it was already idle.
2819 bool flush_work_sync(struct work_struct
*work
)
2821 struct wq_barrier barr
;
2822 bool pending
, waited
;
2824 /* we'll wait for executions separately, queue barr only if pending */
2825 pending
= start_flush_work(work
, &barr
, false);
2827 /* wait for executions to finish */
2828 waited
= wait_on_work(work
);
2830 /* wait for the pending one */
2832 wait_for_completion(&barr
.done
);
2833 destroy_work_on_stack(&barr
.work
);
2836 return pending
|| waited
;
2838 EXPORT_SYMBOL_GPL(flush_work_sync
);
2840 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2845 ret
= try_to_grab_pending(work
, is_dwork
);
2847 } while (unlikely(ret
< 0));
2849 clear_work_data(work
);
2854 * cancel_work_sync - cancel a work and wait for it to finish
2855 * @work: the work to cancel
2857 * Cancel @work and wait for its execution to finish. This function
2858 * can be used even if the work re-queues itself or migrates to
2859 * another workqueue. On return from this function, @work is
2860 * guaranteed to be not pending or executing on any CPU.
2862 * cancel_work_sync(&delayed_work->work) must not be used for
2863 * delayed_work's. Use cancel_delayed_work_sync() instead.
2865 * The caller must ensure that the workqueue on which @work was last
2866 * queued can't be destroyed before this function returns.
2869 * %true if @work was pending, %false otherwise.
2871 bool cancel_work_sync(struct work_struct
*work
)
2873 return __cancel_work_timer(work
, false);
2875 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2878 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2879 * @dwork: the delayed work to flush
2881 * Delayed timer is cancelled and the pending work is queued for
2882 * immediate execution. Like flush_work(), this function only
2883 * considers the last queueing instance of @dwork.
2886 * %true if flush_work() waited for the work to finish execution,
2887 * %false if it was already idle.
2889 bool flush_delayed_work(struct delayed_work
*dwork
)
2891 local_irq_disable();
2892 if (del_timer_sync(&dwork
->timer
))
2893 __queue_work(WORK_CPU_UNBOUND
,
2894 get_work_cwq(&dwork
->work
)->wq
, &dwork
->work
);
2896 return flush_work(&dwork
->work
);
2898 EXPORT_SYMBOL(flush_delayed_work
);
2901 * flush_delayed_work_sync - wait for a dwork to finish
2902 * @dwork: the delayed work to flush
2904 * Delayed timer is cancelled and the pending work is queued for
2905 * execution immediately. Other than timer handling, its behavior
2906 * is identical to flush_work_sync().
2909 * %true if flush_work_sync() waited for the work to finish execution,
2910 * %false if it was already idle.
2912 bool flush_delayed_work_sync(struct delayed_work
*dwork
)
2914 local_irq_disable();
2915 if (del_timer_sync(&dwork
->timer
))
2916 __queue_work(WORK_CPU_UNBOUND
,
2917 get_work_cwq(&dwork
->work
)->wq
, &dwork
->work
);
2919 return flush_work_sync(&dwork
->work
);
2921 EXPORT_SYMBOL(flush_delayed_work_sync
);
2924 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2925 * @dwork: the delayed work cancel
2927 * This is cancel_work_sync() for delayed works.
2930 * %true if @dwork was pending, %false otherwise.
2932 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2934 return __cancel_work_timer(&dwork
->work
, true);
2936 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2939 * schedule_work_on - put work task on a specific cpu
2940 * @cpu: cpu to put the work task on
2941 * @work: job to be done
2943 * This puts a job on a specific cpu
2945 bool schedule_work_on(int cpu
, struct work_struct
*work
)
2947 return queue_work_on(cpu
, system_wq
, work
);
2949 EXPORT_SYMBOL(schedule_work_on
);
2952 * schedule_work - put work task in global workqueue
2953 * @work: job to be done
2955 * Returns %false if @work was already on the kernel-global workqueue and
2958 * This puts a job in the kernel-global workqueue if it was not already
2959 * queued and leaves it in the same position on the kernel-global
2960 * workqueue otherwise.
2962 bool schedule_work(struct work_struct
*work
)
2964 return queue_work(system_wq
, work
);
2966 EXPORT_SYMBOL(schedule_work
);
2969 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2971 * @dwork: job to be done
2972 * @delay: number of jiffies to wait
2974 * After waiting for a given time this puts a job in the kernel-global
2975 * workqueue on the specified CPU.
2977 bool schedule_delayed_work_on(int cpu
, struct delayed_work
*dwork
,
2978 unsigned long delay
)
2980 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
2982 EXPORT_SYMBOL(schedule_delayed_work_on
);
2985 * schedule_delayed_work - put work task in global workqueue after delay
2986 * @dwork: job to be done
2987 * @delay: number of jiffies to wait or 0 for immediate execution
2989 * After waiting for a given time this puts a job in the kernel-global
2992 bool schedule_delayed_work(struct delayed_work
*dwork
, unsigned long delay
)
2994 return queue_delayed_work(system_wq
, dwork
, delay
);
2996 EXPORT_SYMBOL(schedule_delayed_work
);
2999 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3000 * @func: the function to call
3002 * schedule_on_each_cpu() executes @func on each online CPU using the
3003 * system workqueue and blocks until all CPUs have completed.
3004 * schedule_on_each_cpu() is very slow.
3007 * 0 on success, -errno on failure.
3009 int schedule_on_each_cpu(work_func_t func
)
3012 struct work_struct __percpu
*works
;
3014 works
= alloc_percpu(struct work_struct
);
3020 for_each_online_cpu(cpu
) {
3021 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3023 INIT_WORK(work
, func
);
3024 schedule_work_on(cpu
, work
);
3027 for_each_online_cpu(cpu
)
3028 flush_work(per_cpu_ptr(works
, cpu
));
3036 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3038 * Forces execution of the kernel-global workqueue and blocks until its
3041 * Think twice before calling this function! It's very easy to get into
3042 * trouble if you don't take great care. Either of the following situations
3043 * will lead to deadlock:
3045 * One of the work items currently on the workqueue needs to acquire
3046 * a lock held by your code or its caller.
3048 * Your code is running in the context of a work routine.
3050 * They will be detected by lockdep when they occur, but the first might not
3051 * occur very often. It depends on what work items are on the workqueue and
3052 * what locks they need, which you have no control over.
3054 * In most situations flushing the entire workqueue is overkill; you merely
3055 * need to know that a particular work item isn't queued and isn't running.
3056 * In such cases you should use cancel_delayed_work_sync() or
3057 * cancel_work_sync() instead.
3059 void flush_scheduled_work(void)
3061 flush_workqueue(system_wq
);
3063 EXPORT_SYMBOL(flush_scheduled_work
);
3066 * execute_in_process_context - reliably execute the routine with user context
3067 * @fn: the function to execute
3068 * @ew: guaranteed storage for the execute work structure (must
3069 * be available when the work executes)
3071 * Executes the function immediately if process context is available,
3072 * otherwise schedules the function for delayed execution.
3074 * Returns: 0 - function was executed
3075 * 1 - function was scheduled for execution
3077 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3079 if (!in_interrupt()) {
3084 INIT_WORK(&ew
->work
, fn
);
3085 schedule_work(&ew
->work
);
3089 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3091 int keventd_up(void)
3093 return system_wq
!= NULL
;
3096 static int alloc_cwqs(struct workqueue_struct
*wq
)
3099 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3100 * Make sure that the alignment isn't lower than that of
3101 * unsigned long long.
3103 const size_t size
= sizeof(struct cpu_workqueue_struct
);
3104 const size_t align
= max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS
,
3105 __alignof__(unsigned long long));
3107 if (!(wq
->flags
& WQ_UNBOUND
))
3108 wq
->cpu_wq
.pcpu
= __alloc_percpu(size
, align
);
3113 * Allocate enough room to align cwq and put an extra
3114 * pointer at the end pointing back to the originally
3115 * allocated pointer which will be used for free.
3117 ptr
= kzalloc(size
+ align
+ sizeof(void *), GFP_KERNEL
);
3119 wq
->cpu_wq
.single
= PTR_ALIGN(ptr
, align
);
3120 *(void **)(wq
->cpu_wq
.single
+ 1) = ptr
;
3124 /* just in case, make sure it's actually aligned */
3125 BUG_ON(!IS_ALIGNED(wq
->cpu_wq
.v
, align
));
3126 return wq
->cpu_wq
.v
? 0 : -ENOMEM
;
3129 static void free_cwqs(struct workqueue_struct
*wq
)
3131 if (!(wq
->flags
& WQ_UNBOUND
))
3132 free_percpu(wq
->cpu_wq
.pcpu
);
3133 else if (wq
->cpu_wq
.single
) {
3134 /* the pointer to free is stored right after the cwq */
3135 kfree(*(void **)(wq
->cpu_wq
.single
+ 1));
3139 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3142 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3144 if (max_active
< 1 || max_active
> lim
)
3145 printk(KERN_WARNING
"workqueue: max_active %d requested for %s "
3146 "is out of range, clamping between %d and %d\n",
3147 max_active
, name
, 1, lim
);
3149 return clamp_val(max_active
, 1, lim
);
3152 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3155 struct lock_class_key
*key
,
3156 const char *lock_name
, ...)
3158 va_list args
, args1
;
3159 struct workqueue_struct
*wq
;
3163 /* determine namelen, allocate wq and format name */
3164 va_start(args
, lock_name
);
3165 va_copy(args1
, args
);
3166 namelen
= vsnprintf(NULL
, 0, fmt
, args
) + 1;
3168 wq
= kzalloc(sizeof(*wq
) + namelen
, GFP_KERNEL
);
3172 vsnprintf(wq
->name
, namelen
, fmt
, args1
);
3177 * Workqueues which may be used during memory reclaim should
3178 * have a rescuer to guarantee forward progress.
3180 if (flags
& WQ_MEM_RECLAIM
)
3181 flags
|= WQ_RESCUER
;
3183 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3184 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3188 wq
->saved_max_active
= max_active
;
3189 mutex_init(&wq
->flush_mutex
);
3190 atomic_set(&wq
->nr_cwqs_to_flush
, 0);
3191 INIT_LIST_HEAD(&wq
->flusher_queue
);
3192 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3194 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3195 INIT_LIST_HEAD(&wq
->list
);
3197 if (alloc_cwqs(wq
) < 0)
3200 for_each_cwq_cpu(cpu
, wq
) {
3201 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3202 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3203 int pool_idx
= (bool)(flags
& WQ_HIGHPRI
);
3205 BUG_ON((unsigned long)cwq
& WORK_STRUCT_FLAG_MASK
);
3206 cwq
->pool
= &gcwq
->pools
[pool_idx
];
3208 cwq
->flush_color
= -1;
3209 cwq
->max_active
= max_active
;
3210 INIT_LIST_HEAD(&cwq
->delayed_works
);
3213 if (flags
& WQ_RESCUER
) {
3214 struct worker
*rescuer
;
3216 if (!alloc_mayday_mask(&wq
->mayday_mask
, GFP_KERNEL
))
3219 wq
->rescuer
= rescuer
= alloc_worker();
3223 rescuer
->task
= kthread_create(rescuer_thread
, wq
, "%s",
3225 if (IS_ERR(rescuer
->task
))
3228 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
3229 wake_up_process(rescuer
->task
);
3233 * workqueue_lock protects global freeze state and workqueues
3234 * list. Grab it, set max_active accordingly and add the new
3235 * workqueue to workqueues list.
3237 spin_lock(&workqueue_lock
);
3239 if (workqueue_freezing
&& wq
->flags
& WQ_FREEZABLE
)
3240 for_each_cwq_cpu(cpu
, wq
)
3241 get_cwq(cpu
, wq
)->max_active
= 0;
3243 list_add(&wq
->list
, &workqueues
);
3245 spin_unlock(&workqueue_lock
);
3251 free_mayday_mask(wq
->mayday_mask
);
3257 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3260 * destroy_workqueue - safely terminate a workqueue
3261 * @wq: target workqueue
3263 * Safely destroy a workqueue. All work currently pending will be done first.
3265 void destroy_workqueue(struct workqueue_struct
*wq
)
3269 /* drain it before proceeding with destruction */
3270 drain_workqueue(wq
);
3273 * wq list is used to freeze wq, remove from list after
3274 * flushing is complete in case freeze races us.
3276 spin_lock(&workqueue_lock
);
3277 list_del(&wq
->list
);
3278 spin_unlock(&workqueue_lock
);
3281 for_each_cwq_cpu(cpu
, wq
) {
3282 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3285 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
3286 BUG_ON(cwq
->nr_in_flight
[i
]);
3287 BUG_ON(cwq
->nr_active
);
3288 BUG_ON(!list_empty(&cwq
->delayed_works
));
3291 if (wq
->flags
& WQ_RESCUER
) {
3292 kthread_stop(wq
->rescuer
->task
);
3293 free_mayday_mask(wq
->mayday_mask
);
3300 EXPORT_SYMBOL_GPL(destroy_workqueue
);
3303 * workqueue_set_max_active - adjust max_active of a workqueue
3304 * @wq: target workqueue
3305 * @max_active: new max_active value.
3307 * Set max_active of @wq to @max_active.
3310 * Don't call from IRQ context.
3312 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
3316 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
3318 spin_lock(&workqueue_lock
);
3320 wq
->saved_max_active
= max_active
;
3322 for_each_cwq_cpu(cpu
, wq
) {
3323 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3325 spin_lock_irq(&gcwq
->lock
);
3327 if (!(wq
->flags
& WQ_FREEZABLE
) ||
3328 !(gcwq
->flags
& GCWQ_FREEZING
))
3329 get_cwq(gcwq
->cpu
, wq
)->max_active
= max_active
;
3331 spin_unlock_irq(&gcwq
->lock
);
3334 spin_unlock(&workqueue_lock
);
3336 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
3339 * workqueue_congested - test whether a workqueue is congested
3340 * @cpu: CPU in question
3341 * @wq: target workqueue
3343 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3344 * no synchronization around this function and the test result is
3345 * unreliable and only useful as advisory hints or for debugging.
3348 * %true if congested, %false otherwise.
3350 bool workqueue_congested(unsigned int cpu
, struct workqueue_struct
*wq
)
3352 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3354 return !list_empty(&cwq
->delayed_works
);
3356 EXPORT_SYMBOL_GPL(workqueue_congested
);
3359 * work_cpu - return the last known associated cpu for @work
3360 * @work: the work of interest
3363 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3365 unsigned int work_cpu(struct work_struct
*work
)
3367 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3369 return gcwq
? gcwq
->cpu
: WORK_CPU_NONE
;
3371 EXPORT_SYMBOL_GPL(work_cpu
);
3374 * work_busy - test whether a work is currently pending or running
3375 * @work: the work to be tested
3377 * Test whether @work is currently pending or running. There is no
3378 * synchronization around this function and the test result is
3379 * unreliable and only useful as advisory hints or for debugging.
3380 * Especially for reentrant wqs, the pending state might hide the
3384 * OR'd bitmask of WORK_BUSY_* bits.
3386 unsigned int work_busy(struct work_struct
*work
)
3388 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3389 unsigned long flags
;
3390 unsigned int ret
= 0;
3395 spin_lock_irqsave(&gcwq
->lock
, flags
);
3397 if (work_pending(work
))
3398 ret
|= WORK_BUSY_PENDING
;
3399 if (find_worker_executing_work(gcwq
, work
))
3400 ret
|= WORK_BUSY_RUNNING
;
3402 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
3406 EXPORT_SYMBOL_GPL(work_busy
);
3411 * There are two challenges in supporting CPU hotplug. Firstly, there
3412 * are a lot of assumptions on strong associations among work, cwq and
3413 * gcwq which make migrating pending and scheduled works very
3414 * difficult to implement without impacting hot paths. Secondly,
3415 * gcwqs serve mix of short, long and very long running works making
3416 * blocked draining impractical.
3418 * This is solved by allowing a gcwq to be disassociated from the CPU
3419 * running as an unbound one and allowing it to be reattached later if the
3420 * cpu comes back online.
3423 /* claim manager positions of all pools */
3424 static void gcwq_claim_management_and_lock(struct global_cwq
*gcwq
)
3426 struct worker_pool
*pool
;
3428 for_each_worker_pool(pool
, gcwq
)
3429 mutex_lock_nested(&pool
->manager_mutex
, pool
- gcwq
->pools
);
3430 spin_lock_irq(&gcwq
->lock
);
3433 /* release manager positions */
3434 static void gcwq_release_management_and_unlock(struct global_cwq
*gcwq
)
3436 struct worker_pool
*pool
;
3438 spin_unlock_irq(&gcwq
->lock
);
3439 for_each_worker_pool(pool
, gcwq
)
3440 mutex_unlock(&pool
->manager_mutex
);
3443 static void gcwq_unbind_fn(struct work_struct
*work
)
3445 struct global_cwq
*gcwq
= get_gcwq(smp_processor_id());
3446 struct worker_pool
*pool
;
3447 struct worker
*worker
;
3448 struct hlist_node
*pos
;
3451 BUG_ON(gcwq
->cpu
!= smp_processor_id());
3453 gcwq_claim_management_and_lock(gcwq
);
3456 * We've claimed all manager positions. Make all workers unbound
3457 * and set DISASSOCIATED. Before this, all workers except for the
3458 * ones which are still executing works from before the last CPU
3459 * down must be on the cpu. After this, they may become diasporas.
3461 for_each_worker_pool(pool
, gcwq
)
3462 list_for_each_entry(worker
, &pool
->idle_list
, entry
)
3463 worker
->flags
|= WORKER_UNBOUND
;
3465 for_each_busy_worker(worker
, i
, pos
, gcwq
)
3466 worker
->flags
|= WORKER_UNBOUND
;
3468 gcwq
->flags
|= GCWQ_DISASSOCIATED
;
3470 gcwq_release_management_and_unlock(gcwq
);
3473 * Call schedule() so that we cross rq->lock and thus can guarantee
3474 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3475 * as scheduler callbacks may be invoked from other cpus.
3480 * Sched callbacks are disabled now. Zap nr_running. After this,
3481 * nr_running stays zero and need_more_worker() and keep_working()
3482 * are always true as long as the worklist is not empty. @gcwq now
3483 * behaves as unbound (in terms of concurrency management) gcwq
3484 * which is served by workers tied to the CPU.
3486 * On return from this function, the current worker would trigger
3487 * unbound chain execution of pending work items if other workers
3490 for_each_worker_pool(pool
, gcwq
)
3491 atomic_set(get_pool_nr_running(pool
), 0);
3495 * Workqueues should be brought up before normal priority CPU notifiers.
3496 * This will be registered high priority CPU notifier.
3498 static int __devinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
3499 unsigned long action
,
3502 unsigned int cpu
= (unsigned long)hcpu
;
3503 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3504 struct worker_pool
*pool
;
3506 switch (action
& ~CPU_TASKS_FROZEN
) {
3507 case CPU_UP_PREPARE
:
3508 for_each_worker_pool(pool
, gcwq
) {
3509 struct worker
*worker
;
3511 if (pool
->nr_workers
)
3514 worker
= create_worker(pool
);
3518 spin_lock_irq(&gcwq
->lock
);
3519 start_worker(worker
);
3520 spin_unlock_irq(&gcwq
->lock
);
3524 case CPU_DOWN_FAILED
:
3526 gcwq_claim_management_and_lock(gcwq
);
3527 gcwq
->flags
&= ~GCWQ_DISASSOCIATED
;
3528 rebind_workers(gcwq
);
3529 gcwq_release_management_and_unlock(gcwq
);
3536 * Workqueues should be brought down after normal priority CPU notifiers.
3537 * This will be registered as low priority CPU notifier.
3539 static int __devinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
3540 unsigned long action
,
3543 unsigned int cpu
= (unsigned long)hcpu
;
3544 struct work_struct unbind_work
;
3546 switch (action
& ~CPU_TASKS_FROZEN
) {
3547 case CPU_DOWN_PREPARE
:
3548 /* unbinding should happen on the local CPU */
3549 INIT_WORK_ONSTACK(&unbind_work
, gcwq_unbind_fn
);
3550 schedule_work_on(cpu
, &unbind_work
);
3551 flush_work(&unbind_work
);
3559 struct work_for_cpu
{
3560 struct completion completion
;
3566 static int do_work_for_cpu(void *_wfc
)
3568 struct work_for_cpu
*wfc
= _wfc
;
3569 wfc
->ret
= wfc
->fn(wfc
->arg
);
3570 complete(&wfc
->completion
);
3575 * work_on_cpu - run a function in user context on a particular cpu
3576 * @cpu: the cpu to run on
3577 * @fn: the function to run
3578 * @arg: the function arg
3580 * This will return the value @fn returns.
3581 * It is up to the caller to ensure that the cpu doesn't go offline.
3582 * The caller must not hold any locks which would prevent @fn from completing.
3584 long work_on_cpu(unsigned int cpu
, long (*fn
)(void *), void *arg
)
3586 struct task_struct
*sub_thread
;
3587 struct work_for_cpu wfc
= {
3588 .completion
= COMPLETION_INITIALIZER_ONSTACK(wfc
.completion
),
3593 sub_thread
= kthread_create(do_work_for_cpu
, &wfc
, "work_for_cpu");
3594 if (IS_ERR(sub_thread
))
3595 return PTR_ERR(sub_thread
);
3596 kthread_bind(sub_thread
, cpu
);
3597 wake_up_process(sub_thread
);
3598 wait_for_completion(&wfc
.completion
);
3601 EXPORT_SYMBOL_GPL(work_on_cpu
);
3602 #endif /* CONFIG_SMP */
3604 #ifdef CONFIG_FREEZER
3607 * freeze_workqueues_begin - begin freezing workqueues
3609 * Start freezing workqueues. After this function returns, all freezable
3610 * workqueues will queue new works to their frozen_works list instead of
3614 * Grabs and releases workqueue_lock and gcwq->lock's.
3616 void freeze_workqueues_begin(void)
3620 spin_lock(&workqueue_lock
);
3622 BUG_ON(workqueue_freezing
);
3623 workqueue_freezing
= true;
3625 for_each_gcwq_cpu(cpu
) {
3626 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3627 struct workqueue_struct
*wq
;
3629 spin_lock_irq(&gcwq
->lock
);
3631 BUG_ON(gcwq
->flags
& GCWQ_FREEZING
);
3632 gcwq
->flags
|= GCWQ_FREEZING
;
3634 list_for_each_entry(wq
, &workqueues
, list
) {
3635 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3637 if (cwq
&& wq
->flags
& WQ_FREEZABLE
)
3638 cwq
->max_active
= 0;
3641 spin_unlock_irq(&gcwq
->lock
);
3644 spin_unlock(&workqueue_lock
);
3648 * freeze_workqueues_busy - are freezable workqueues still busy?
3650 * Check whether freezing is complete. This function must be called
3651 * between freeze_workqueues_begin() and thaw_workqueues().
3654 * Grabs and releases workqueue_lock.
3657 * %true if some freezable workqueues are still busy. %false if freezing
3660 bool freeze_workqueues_busy(void)
3665 spin_lock(&workqueue_lock
);
3667 BUG_ON(!workqueue_freezing
);
3669 for_each_gcwq_cpu(cpu
) {
3670 struct workqueue_struct
*wq
;
3672 * nr_active is monotonically decreasing. It's safe
3673 * to peek without lock.
3675 list_for_each_entry(wq
, &workqueues
, list
) {
3676 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3678 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3681 BUG_ON(cwq
->nr_active
< 0);
3682 if (cwq
->nr_active
) {
3689 spin_unlock(&workqueue_lock
);
3694 * thaw_workqueues - thaw workqueues
3696 * Thaw workqueues. Normal queueing is restored and all collected
3697 * frozen works are transferred to their respective gcwq worklists.
3700 * Grabs and releases workqueue_lock and gcwq->lock's.
3702 void thaw_workqueues(void)
3706 spin_lock(&workqueue_lock
);
3708 if (!workqueue_freezing
)
3711 for_each_gcwq_cpu(cpu
) {
3712 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3713 struct worker_pool
*pool
;
3714 struct workqueue_struct
*wq
;
3716 spin_lock_irq(&gcwq
->lock
);
3718 BUG_ON(!(gcwq
->flags
& GCWQ_FREEZING
));
3719 gcwq
->flags
&= ~GCWQ_FREEZING
;
3721 list_for_each_entry(wq
, &workqueues
, list
) {
3722 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3724 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3727 /* restore max_active and repopulate worklist */
3728 cwq
->max_active
= wq
->saved_max_active
;
3730 while (!list_empty(&cwq
->delayed_works
) &&
3731 cwq
->nr_active
< cwq
->max_active
)
3732 cwq_activate_first_delayed(cwq
);
3735 for_each_worker_pool(pool
, gcwq
)
3736 wake_up_worker(pool
);
3738 spin_unlock_irq(&gcwq
->lock
);
3741 workqueue_freezing
= false;
3743 spin_unlock(&workqueue_lock
);
3745 #endif /* CONFIG_FREEZER */
3747 static int __init
init_workqueues(void)
3752 /* make sure we have enough bits for OFFQ CPU number */
3753 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_CPU_SHIFT
)) <
3756 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
3757 cpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
3759 /* initialize gcwqs */
3760 for_each_gcwq_cpu(cpu
) {
3761 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3762 struct worker_pool
*pool
;
3764 spin_lock_init(&gcwq
->lock
);
3766 gcwq
->flags
|= GCWQ_DISASSOCIATED
;
3768 for (i
= 0; i
< BUSY_WORKER_HASH_SIZE
; i
++)
3769 INIT_HLIST_HEAD(&gcwq
->busy_hash
[i
]);
3771 for_each_worker_pool(pool
, gcwq
) {
3773 INIT_LIST_HEAD(&pool
->worklist
);
3774 INIT_LIST_HEAD(&pool
->idle_list
);
3776 init_timer_deferrable(&pool
->idle_timer
);
3777 pool
->idle_timer
.function
= idle_worker_timeout
;
3778 pool
->idle_timer
.data
= (unsigned long)pool
;
3780 setup_timer(&pool
->mayday_timer
, gcwq_mayday_timeout
,
3781 (unsigned long)pool
);
3783 mutex_init(&pool
->manager_mutex
);
3784 ida_init(&pool
->worker_ida
);
3787 init_waitqueue_head(&gcwq
->rebind_hold
);
3790 /* create the initial worker */
3791 for_each_online_gcwq_cpu(cpu
) {
3792 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3793 struct worker_pool
*pool
;
3795 if (cpu
!= WORK_CPU_UNBOUND
)
3796 gcwq
->flags
&= ~GCWQ_DISASSOCIATED
;
3798 for_each_worker_pool(pool
, gcwq
) {
3799 struct worker
*worker
;
3801 worker
= create_worker(pool
);
3803 spin_lock_irq(&gcwq
->lock
);
3804 start_worker(worker
);
3805 spin_unlock_irq(&gcwq
->lock
);
3809 system_wq
= alloc_workqueue("events", 0, 0);
3810 system_long_wq
= alloc_workqueue("events_long", 0, 0);
3811 system_nrt_wq
= alloc_workqueue("events_nrt", WQ_NON_REENTRANT
, 0);
3812 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
3813 WQ_UNBOUND_MAX_ACTIVE
);
3814 system_freezable_wq
= alloc_workqueue("events_freezable",
3816 system_nrt_freezable_wq
= alloc_workqueue("events_nrt_freezable",
3817 WQ_NON_REENTRANT
| WQ_FREEZABLE
, 0);
3818 BUG_ON(!system_wq
|| !system_long_wq
|| !system_nrt_wq
||
3819 !system_unbound_wq
|| !system_freezable_wq
||
3820 !system_nrt_freezable_wq
);
3823 early_initcall(init_workqueues
);