2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/hashtable.h>
46 #include "workqueue_internal.h"
52 * A bound pool is either associated or disassociated with its CPU.
53 * While associated (!DISASSOCIATED), all workers are bound to the
54 * CPU and none has %WORKER_UNBOUND set and concurrency management
57 * While DISASSOCIATED, the cpu may be offline and all workers have
58 * %WORKER_UNBOUND set and concurrency management disabled, and may
59 * be executing on any CPU. The pool behaves as an unbound one.
61 * Note that DISASSOCIATED can be flipped only while holding
62 * assoc_mutex to avoid changing binding state while
63 * create_worker() is in progress.
65 POOL_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
66 POOL_MANAGING_WORKERS
= 1 << 1, /* managing workers */
67 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
68 POOL_FREEZING
= 1 << 3, /* freeze in progress */
71 WORKER_STARTED
= 1 << 0, /* started */
72 WORKER_DIE
= 1 << 1, /* die die die */
73 WORKER_IDLE
= 1 << 2, /* is idle */
74 WORKER_PREP
= 1 << 3, /* preparing to run works */
75 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
76 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
78 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_UNBOUND
|
81 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
83 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
85 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
86 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
88 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
89 /* call for help after 10ms
91 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
92 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
95 * Rescue workers are used only on emergencies and shared by
98 RESCUER_NICE_LEVEL
= -20,
99 HIGHPRI_NICE_LEVEL
= -20,
103 * Structure fields follow one of the following exclusion rules.
105 * I: Modifiable by initialization/destruction paths and read-only for
108 * P: Preemption protected. Disabling preemption is enough and should
109 * only be modified and accessed from the local cpu.
111 * L: pool->lock protected. Access with pool->lock held.
113 * X: During normal operation, modification requires pool->lock and should
114 * be done only from local cpu. Either disabling preemption on local
115 * cpu or grabbing pool->lock is enough for read access. If
116 * POOL_DISASSOCIATED is set, it's identical to L.
118 * F: wq->flush_mutex protected.
120 * W: workqueue_lock protected.
123 /* struct worker is defined in workqueue_internal.h */
126 spinlock_t lock
; /* the pool lock */
127 unsigned int cpu
; /* I: the associated cpu */
128 int id
; /* I: pool ID */
129 unsigned int flags
; /* X: flags */
131 struct list_head worklist
; /* L: list of pending works */
132 int nr_workers
; /* L: total number of workers */
134 /* nr_idle includes the ones off idle_list for rebinding */
135 int nr_idle
; /* L: currently idle ones */
137 struct list_head idle_list
; /* X: list of idle workers */
138 struct timer_list idle_timer
; /* L: worker idle timeout */
139 struct timer_list mayday_timer
; /* L: SOS timer for workers */
141 /* workers are chained either in busy_hash or idle_list */
142 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
143 /* L: hash of busy workers */
145 struct mutex assoc_mutex
; /* protect POOL_DISASSOCIATED */
146 struct ida worker_ida
; /* L: for worker IDs */
150 * Global per-cpu workqueue. There's one and only one for each cpu
151 * and all works are queued and processed here regardless of their
155 struct worker_pool pools
[NR_STD_WORKER_POOLS
];
156 /* normal and highpri pools */
157 } ____cacheline_aligned_in_smp
;
160 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
161 * work_struct->data are used for flags and thus cwqs need to be
162 * aligned at two's power of the number of flag bits.
164 struct cpu_workqueue_struct
{
165 struct worker_pool
*pool
; /* I: the associated pool */
166 struct workqueue_struct
*wq
; /* I: the owning workqueue */
167 int work_color
; /* L: current color */
168 int flush_color
; /* L: flushing color */
169 int nr_in_flight
[WORK_NR_COLORS
];
170 /* L: nr of in_flight works */
171 int nr_active
; /* L: nr of active works */
172 int max_active
; /* L: max active works */
173 struct list_head delayed_works
; /* L: delayed works */
177 * Structure used to wait for workqueue flush.
180 struct list_head list
; /* F: list of flushers */
181 int flush_color
; /* F: flush color waiting for */
182 struct completion done
; /* flush completion */
186 * All cpumasks are assumed to be always set on UP and thus can't be
187 * used to determine whether there's something to be done.
190 typedef cpumask_var_t mayday_mask_t
;
191 #define mayday_test_and_set_cpu(cpu, mask) \
192 cpumask_test_and_set_cpu((cpu), (mask))
193 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
194 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
195 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
196 #define free_mayday_mask(mask) free_cpumask_var((mask))
198 typedef unsigned long mayday_mask_t
;
199 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
200 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
201 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
202 #define alloc_mayday_mask(maskp, gfp) true
203 #define free_mayday_mask(mask) do { } while (0)
207 * The externally visible workqueue abstraction is an array of
208 * per-CPU workqueues:
210 struct workqueue_struct
{
211 unsigned int flags
; /* W: WQ_* flags */
213 struct cpu_workqueue_struct __percpu
*pcpu
;
214 struct cpu_workqueue_struct
*single
;
216 } cpu_wq
; /* I: cwq's */
217 struct list_head list
; /* W: list of all workqueues */
219 struct mutex flush_mutex
; /* protects wq flushing */
220 int work_color
; /* F: current work color */
221 int flush_color
; /* F: current flush color */
222 atomic_t nr_cwqs_to_flush
; /* flush in progress */
223 struct wq_flusher
*first_flusher
; /* F: first flusher */
224 struct list_head flusher_queue
; /* F: flush waiters */
225 struct list_head flusher_overflow
; /* F: flush overflow list */
227 mayday_mask_t mayday_mask
; /* cpus requesting rescue */
228 struct worker
*rescuer
; /* I: rescue worker */
230 int nr_drainers
; /* W: drain in progress */
231 int saved_max_active
; /* W: saved cwq max_active */
232 #ifdef CONFIG_LOCKDEP
233 struct lockdep_map lockdep_map
;
235 char name
[]; /* I: workqueue name */
238 struct workqueue_struct
*system_wq __read_mostly
;
239 EXPORT_SYMBOL_GPL(system_wq
);
240 struct workqueue_struct
*system_highpri_wq __read_mostly
;
241 EXPORT_SYMBOL_GPL(system_highpri_wq
);
242 struct workqueue_struct
*system_long_wq __read_mostly
;
243 EXPORT_SYMBOL_GPL(system_long_wq
);
244 struct workqueue_struct
*system_unbound_wq __read_mostly
;
245 EXPORT_SYMBOL_GPL(system_unbound_wq
);
246 struct workqueue_struct
*system_freezable_wq __read_mostly
;
247 EXPORT_SYMBOL_GPL(system_freezable_wq
);
249 #define CREATE_TRACE_POINTS
250 #include <trace/events/workqueue.h>
252 #define for_each_std_worker_pool(pool, cpu) \
253 for ((pool) = &get_gcwq((cpu))->pools[0]; \
254 (pool) < &get_gcwq((cpu))->pools[NR_STD_WORKER_POOLS]; (pool)++)
256 #define for_each_busy_worker(worker, i, pos, pool) \
257 hash_for_each(pool->busy_hash, i, pos, worker, hentry)
259 static inline int __next_gcwq_cpu(int cpu
, const struct cpumask
*mask
,
262 if (cpu
< nr_cpu_ids
) {
264 cpu
= cpumask_next(cpu
, mask
);
265 if (cpu
< nr_cpu_ids
)
269 return WORK_CPU_UNBOUND
;
271 return WORK_CPU_NONE
;
274 static inline int __next_wq_cpu(int cpu
, const struct cpumask
*mask
,
275 struct workqueue_struct
*wq
)
277 return __next_gcwq_cpu(cpu
, mask
, !(wq
->flags
& WQ_UNBOUND
) ? 1 : 2);
283 * An extra gcwq is defined for an invalid cpu number
284 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
285 * specific CPU. The following iterators are similar to
286 * for_each_*_cpu() iterators but also considers the unbound gcwq.
288 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
289 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
290 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
291 * WORK_CPU_UNBOUND for unbound workqueues
293 #define for_each_gcwq_cpu(cpu) \
294 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
295 (cpu) < WORK_CPU_NONE; \
296 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
298 #define for_each_online_gcwq_cpu(cpu) \
299 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
300 (cpu) < WORK_CPU_NONE; \
301 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
303 #define for_each_cwq_cpu(cpu, wq) \
304 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
305 (cpu) < WORK_CPU_NONE; \
306 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
308 #ifdef CONFIG_DEBUG_OBJECTS_WORK
310 static struct debug_obj_descr work_debug_descr
;
312 static void *work_debug_hint(void *addr
)
314 return ((struct work_struct
*) addr
)->func
;
318 * fixup_init is called when:
319 * - an active object is initialized
321 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
323 struct work_struct
*work
= addr
;
326 case ODEBUG_STATE_ACTIVE
:
327 cancel_work_sync(work
);
328 debug_object_init(work
, &work_debug_descr
);
336 * fixup_activate is called when:
337 * - an active object is activated
338 * - an unknown object is activated (might be a statically initialized object)
340 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
342 struct work_struct
*work
= addr
;
346 case ODEBUG_STATE_NOTAVAILABLE
:
348 * This is not really a fixup. The work struct was
349 * statically initialized. We just make sure that it
350 * is tracked in the object tracker.
352 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
353 debug_object_init(work
, &work_debug_descr
);
354 debug_object_activate(work
, &work_debug_descr
);
360 case ODEBUG_STATE_ACTIVE
:
369 * fixup_free is called when:
370 * - an active object is freed
372 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
374 struct work_struct
*work
= addr
;
377 case ODEBUG_STATE_ACTIVE
:
378 cancel_work_sync(work
);
379 debug_object_free(work
, &work_debug_descr
);
386 static struct debug_obj_descr work_debug_descr
= {
387 .name
= "work_struct",
388 .debug_hint
= work_debug_hint
,
389 .fixup_init
= work_fixup_init
,
390 .fixup_activate
= work_fixup_activate
,
391 .fixup_free
= work_fixup_free
,
394 static inline void debug_work_activate(struct work_struct
*work
)
396 debug_object_activate(work
, &work_debug_descr
);
399 static inline void debug_work_deactivate(struct work_struct
*work
)
401 debug_object_deactivate(work
, &work_debug_descr
);
404 void __init_work(struct work_struct
*work
, int onstack
)
407 debug_object_init_on_stack(work
, &work_debug_descr
);
409 debug_object_init(work
, &work_debug_descr
);
411 EXPORT_SYMBOL_GPL(__init_work
);
413 void destroy_work_on_stack(struct work_struct
*work
)
415 debug_object_free(work
, &work_debug_descr
);
417 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
420 static inline void debug_work_activate(struct work_struct
*work
) { }
421 static inline void debug_work_deactivate(struct work_struct
*work
) { }
424 /* Serializes the accesses to the list of workqueues. */
425 static DEFINE_SPINLOCK(workqueue_lock
);
426 static LIST_HEAD(workqueues
);
427 static bool workqueue_freezing
; /* W: have wqs started freezing? */
430 * The almighty global cpu workqueues. nr_running is the only field
431 * which is expected to be used frequently by other cpus via
432 * try_to_wake_up(). Put it in a separate cacheline.
434 static DEFINE_PER_CPU(struct global_cwq
, global_cwq
);
435 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t
, pool_nr_running
[NR_STD_WORKER_POOLS
]);
438 * Global cpu workqueue and nr_running counter for unbound gcwq. The pools
439 * for online CPUs have POOL_DISASSOCIATED set, and all their workers have
440 * WORKER_UNBOUND set.
442 static struct global_cwq unbound_global_cwq
;
443 static atomic_t unbound_pool_nr_running
[NR_STD_WORKER_POOLS
] = {
444 [0 ... NR_STD_WORKER_POOLS
- 1] = ATOMIC_INIT(0), /* always 0 */
447 /* idr of all pools */
448 static DEFINE_MUTEX(worker_pool_idr_mutex
);
449 static DEFINE_IDR(worker_pool_idr
);
451 static int worker_thread(void *__worker
);
453 static struct global_cwq
*get_gcwq(unsigned int cpu
)
455 if (cpu
!= WORK_CPU_UNBOUND
)
456 return &per_cpu(global_cwq
, cpu
);
458 return &unbound_global_cwq
;
461 static int std_worker_pool_pri(struct worker_pool
*pool
)
463 return pool
- get_gcwq(pool
->cpu
)->pools
;
466 /* allocate ID and assign it to @pool */
467 static int worker_pool_assign_id(struct worker_pool
*pool
)
471 mutex_lock(&worker_pool_idr_mutex
);
472 idr_pre_get(&worker_pool_idr
, GFP_KERNEL
);
473 ret
= idr_get_new(&worker_pool_idr
, pool
, &pool
->id
);
474 mutex_unlock(&worker_pool_idr_mutex
);
480 * Lookup worker_pool by id. The idr currently is built during boot and
481 * never modified. Don't worry about locking for now.
483 static struct worker_pool
*worker_pool_by_id(int pool_id
)
485 return idr_find(&worker_pool_idr
, pool_id
);
488 static struct worker_pool
*get_std_worker_pool(int cpu
, bool highpri
)
490 struct global_cwq
*gcwq
= get_gcwq(cpu
);
492 return &gcwq
->pools
[highpri
];
495 static atomic_t
*get_pool_nr_running(struct worker_pool
*pool
)
498 int idx
= std_worker_pool_pri(pool
);
500 if (cpu
!= WORK_CPU_UNBOUND
)
501 return &per_cpu(pool_nr_running
, cpu
)[idx
];
503 return &unbound_pool_nr_running
[idx
];
506 static struct cpu_workqueue_struct
*get_cwq(unsigned int cpu
,
507 struct workqueue_struct
*wq
)
509 if (!(wq
->flags
& WQ_UNBOUND
)) {
510 if (likely(cpu
< nr_cpu_ids
))
511 return per_cpu_ptr(wq
->cpu_wq
.pcpu
, cpu
);
512 } else if (likely(cpu
== WORK_CPU_UNBOUND
))
513 return wq
->cpu_wq
.single
;
517 static unsigned int work_color_to_flags(int color
)
519 return color
<< WORK_STRUCT_COLOR_SHIFT
;
522 static int get_work_color(struct work_struct
*work
)
524 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
525 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
528 static int work_next_color(int color
)
530 return (color
+ 1) % WORK_NR_COLORS
;
534 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
535 * contain the pointer to the queued cwq. Once execution starts, the flag
536 * is cleared and the high bits contain OFFQ flags and pool ID.
538 * set_work_cwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
539 * and clear_work_data() can be used to set the cwq, pool or clear
540 * work->data. These functions should only be called while the work is
541 * owned - ie. while the PENDING bit is set.
543 * get_work_pool() and get_work_cwq() can be used to obtain the pool or cwq
544 * corresponding to a work. Pool is available once the work has been
545 * queued anywhere after initialization until it is sync canceled. cwq is
546 * available only while the work item is queued.
548 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
549 * canceled. While being canceled, a work item may have its PENDING set
550 * but stay off timer and worklist for arbitrarily long and nobody should
551 * try to steal the PENDING bit.
553 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
556 BUG_ON(!work_pending(work
));
557 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
560 static void set_work_cwq(struct work_struct
*work
,
561 struct cpu_workqueue_struct
*cwq
,
562 unsigned long extra_flags
)
564 set_work_data(work
, (unsigned long)cwq
,
565 WORK_STRUCT_PENDING
| WORK_STRUCT_CWQ
| extra_flags
);
568 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
572 * The following wmb is paired with the implied mb in
573 * test_and_set_bit(PENDING) and ensures all updates to @work made
574 * here are visible to and precede any updates by the next PENDING
578 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
581 static void clear_work_data(struct work_struct
*work
)
583 smp_wmb(); /* see set_work_pool_and_clear_pending() */
584 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
587 static struct cpu_workqueue_struct
*get_work_cwq(struct work_struct
*work
)
589 unsigned long data
= atomic_long_read(&work
->data
);
591 if (data
& WORK_STRUCT_CWQ
)
592 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
598 * get_work_pool - return the worker_pool a given work was associated with
599 * @work: the work item of interest
601 * Return the worker_pool @work was last associated with. %NULL if none.
603 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
605 unsigned long data
= atomic_long_read(&work
->data
);
606 struct worker_pool
*pool
;
609 if (data
& WORK_STRUCT_CWQ
)
610 return ((struct cpu_workqueue_struct
*)
611 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
613 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
614 if (pool_id
== WORK_OFFQ_POOL_NONE
)
617 pool
= worker_pool_by_id(pool_id
);
623 * get_work_pool_id - return the worker pool ID a given work is associated with
624 * @work: the work item of interest
626 * Return the worker_pool ID @work was last associated with.
627 * %WORK_OFFQ_POOL_NONE if none.
629 static int get_work_pool_id(struct work_struct
*work
)
631 struct worker_pool
*pool
= get_work_pool(work
);
633 return pool
? pool
->id
: WORK_OFFQ_POOL_NONE
;
636 static void mark_work_canceling(struct work_struct
*work
)
638 unsigned long pool_id
= get_work_pool_id(work
);
640 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
641 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
644 static bool work_is_canceling(struct work_struct
*work
)
646 unsigned long data
= atomic_long_read(&work
->data
);
648 return !(data
& WORK_STRUCT_CWQ
) && (data
& WORK_OFFQ_CANCELING
);
652 * Policy functions. These define the policies on how the global worker
653 * pools are managed. Unless noted otherwise, these functions assume that
654 * they're being called with pool->lock held.
657 static bool __need_more_worker(struct worker_pool
*pool
)
659 return !atomic_read(get_pool_nr_running(pool
));
663 * Need to wake up a worker? Called from anything but currently
666 * Note that, because unbound workers never contribute to nr_running, this
667 * function will always return %true for unbound gcwq as long as the
668 * worklist isn't empty.
670 static bool need_more_worker(struct worker_pool
*pool
)
672 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
675 /* Can I start working? Called from busy but !running workers. */
676 static bool may_start_working(struct worker_pool
*pool
)
678 return pool
->nr_idle
;
681 /* Do I need to keep working? Called from currently running workers. */
682 static bool keep_working(struct worker_pool
*pool
)
684 atomic_t
*nr_running
= get_pool_nr_running(pool
);
686 return !list_empty(&pool
->worklist
) && atomic_read(nr_running
) <= 1;
689 /* Do we need a new worker? Called from manager. */
690 static bool need_to_create_worker(struct worker_pool
*pool
)
692 return need_more_worker(pool
) && !may_start_working(pool
);
695 /* Do I need to be the manager? */
696 static bool need_to_manage_workers(struct worker_pool
*pool
)
698 return need_to_create_worker(pool
) ||
699 (pool
->flags
& POOL_MANAGE_WORKERS
);
702 /* Do we have too many workers and should some go away? */
703 static bool too_many_workers(struct worker_pool
*pool
)
705 bool managing
= pool
->flags
& POOL_MANAGING_WORKERS
;
706 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
707 int nr_busy
= pool
->nr_workers
- nr_idle
;
710 * nr_idle and idle_list may disagree if idle rebinding is in
711 * progress. Never return %true if idle_list is empty.
713 if (list_empty(&pool
->idle_list
))
716 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
723 /* Return the first worker. Safe with preemption disabled */
724 static struct worker
*first_worker(struct worker_pool
*pool
)
726 if (unlikely(list_empty(&pool
->idle_list
)))
729 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
733 * wake_up_worker - wake up an idle worker
734 * @pool: worker pool to wake worker from
736 * Wake up the first idle worker of @pool.
739 * spin_lock_irq(pool->lock).
741 static void wake_up_worker(struct worker_pool
*pool
)
743 struct worker
*worker
= first_worker(pool
);
746 wake_up_process(worker
->task
);
750 * wq_worker_waking_up - a worker is waking up
751 * @task: task waking up
752 * @cpu: CPU @task is waking up to
754 * This function is called during try_to_wake_up() when a worker is
758 * spin_lock_irq(rq->lock)
760 void wq_worker_waking_up(struct task_struct
*task
, unsigned int cpu
)
762 struct worker
*worker
= kthread_data(task
);
764 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
765 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
766 atomic_inc(get_pool_nr_running(worker
->pool
));
771 * wq_worker_sleeping - a worker is going to sleep
772 * @task: task going to sleep
773 * @cpu: CPU in question, must be the current CPU number
775 * This function is called during schedule() when a busy worker is
776 * going to sleep. Worker on the same cpu can be woken up by
777 * returning pointer to its task.
780 * spin_lock_irq(rq->lock)
783 * Worker task on @cpu to wake up, %NULL if none.
785 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
,
788 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
789 struct worker_pool
*pool
;
790 atomic_t
*nr_running
;
793 * Rescuers, which may not have all the fields set up like normal
794 * workers, also reach here, let's not access anything before
795 * checking NOT_RUNNING.
797 if (worker
->flags
& WORKER_NOT_RUNNING
)
801 nr_running
= get_pool_nr_running(pool
);
803 /* this can only happen on the local cpu */
804 BUG_ON(cpu
!= raw_smp_processor_id());
807 * The counterpart of the following dec_and_test, implied mb,
808 * worklist not empty test sequence is in insert_work().
809 * Please read comment there.
811 * NOT_RUNNING is clear. This means that we're bound to and
812 * running on the local cpu w/ rq lock held and preemption
813 * disabled, which in turn means that none else could be
814 * manipulating idle_list, so dereferencing idle_list without pool
817 if (atomic_dec_and_test(nr_running
) && !list_empty(&pool
->worklist
))
818 to_wakeup
= first_worker(pool
);
819 return to_wakeup
? to_wakeup
->task
: NULL
;
823 * worker_set_flags - set worker flags and adjust nr_running accordingly
825 * @flags: flags to set
826 * @wakeup: wakeup an idle worker if necessary
828 * Set @flags in @worker->flags and adjust nr_running accordingly. If
829 * nr_running becomes zero and @wakeup is %true, an idle worker is
833 * spin_lock_irq(pool->lock)
835 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
838 struct worker_pool
*pool
= worker
->pool
;
840 WARN_ON_ONCE(worker
->task
!= current
);
843 * If transitioning into NOT_RUNNING, adjust nr_running and
844 * wake up an idle worker as necessary if requested by
847 if ((flags
& WORKER_NOT_RUNNING
) &&
848 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
849 atomic_t
*nr_running
= get_pool_nr_running(pool
);
852 if (atomic_dec_and_test(nr_running
) &&
853 !list_empty(&pool
->worklist
))
854 wake_up_worker(pool
);
856 atomic_dec(nr_running
);
859 worker
->flags
|= flags
;
863 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
865 * @flags: flags to clear
867 * Clear @flags in @worker->flags and adjust nr_running accordingly.
870 * spin_lock_irq(pool->lock)
872 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
874 struct worker_pool
*pool
= worker
->pool
;
875 unsigned int oflags
= worker
->flags
;
877 WARN_ON_ONCE(worker
->task
!= current
);
879 worker
->flags
&= ~flags
;
882 * If transitioning out of NOT_RUNNING, increment nr_running. Note
883 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
884 * of multiple flags, not a single flag.
886 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
887 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
888 atomic_inc(get_pool_nr_running(pool
));
892 * find_worker_executing_work - find worker which is executing a work
893 * @pool: pool of interest
894 * @work: work to find worker for
896 * Find a worker which is executing @work on @pool by searching
897 * @pool->busy_hash which is keyed by the address of @work. For a worker
898 * to match, its current execution should match the address of @work and
899 * its work function. This is to avoid unwanted dependency between
900 * unrelated work executions through a work item being recycled while still
903 * This is a bit tricky. A work item may be freed once its execution
904 * starts and nothing prevents the freed area from being recycled for
905 * another work item. If the same work item address ends up being reused
906 * before the original execution finishes, workqueue will identify the
907 * recycled work item as currently executing and make it wait until the
908 * current execution finishes, introducing an unwanted dependency.
910 * This function checks the work item address, work function and workqueue
911 * to avoid false positives. Note that this isn't complete as one may
912 * construct a work function which can introduce dependency onto itself
913 * through a recycled work item. Well, if somebody wants to shoot oneself
914 * in the foot that badly, there's only so much we can do, and if such
915 * deadlock actually occurs, it should be easy to locate the culprit work
919 * spin_lock_irq(pool->lock).
922 * Pointer to worker which is executing @work if found, NULL
925 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
926 struct work_struct
*work
)
928 struct worker
*worker
;
929 struct hlist_node
*tmp
;
931 hash_for_each_possible(pool
->busy_hash
, worker
, tmp
, hentry
,
933 if (worker
->current_work
== work
&&
934 worker
->current_func
== work
->func
)
941 * move_linked_works - move linked works to a list
942 * @work: start of series of works to be scheduled
943 * @head: target list to append @work to
944 * @nextp: out paramter for nested worklist walking
946 * Schedule linked works starting from @work to @head. Work series to
947 * be scheduled starts at @work and includes any consecutive work with
948 * WORK_STRUCT_LINKED set in its predecessor.
950 * If @nextp is not NULL, it's updated to point to the next work of
951 * the last scheduled work. This allows move_linked_works() to be
952 * nested inside outer list_for_each_entry_safe().
955 * spin_lock_irq(pool->lock).
957 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
958 struct work_struct
**nextp
)
960 struct work_struct
*n
;
963 * Linked worklist will always end before the end of the list,
964 * use NULL for list head.
966 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
967 list_move_tail(&work
->entry
, head
);
968 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
973 * If we're already inside safe list traversal and have moved
974 * multiple works to the scheduled queue, the next position
975 * needs to be updated.
981 static void cwq_activate_delayed_work(struct work_struct
*work
)
983 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
985 trace_workqueue_activate_work(work
);
986 move_linked_works(work
, &cwq
->pool
->worklist
, NULL
);
987 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
991 static void cwq_activate_first_delayed(struct cpu_workqueue_struct
*cwq
)
993 struct work_struct
*work
= list_first_entry(&cwq
->delayed_works
,
994 struct work_struct
, entry
);
996 cwq_activate_delayed_work(work
);
1000 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1001 * @cwq: cwq of interest
1002 * @color: color of work which left the queue
1004 * A work either has completed or is removed from pending queue,
1005 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1008 * spin_lock_irq(pool->lock).
1010 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct
*cwq
, int color
)
1012 /* ignore uncolored works */
1013 if (color
== WORK_NO_COLOR
)
1016 cwq
->nr_in_flight
[color
]--;
1019 if (!list_empty(&cwq
->delayed_works
)) {
1020 /* one down, submit a delayed one */
1021 if (cwq
->nr_active
< cwq
->max_active
)
1022 cwq_activate_first_delayed(cwq
);
1025 /* is flush in progress and are we at the flushing tip? */
1026 if (likely(cwq
->flush_color
!= color
))
1029 /* are there still in-flight works? */
1030 if (cwq
->nr_in_flight
[color
])
1033 /* this cwq is done, clear flush_color */
1034 cwq
->flush_color
= -1;
1037 * If this was the last cwq, wake up the first flusher. It
1038 * will handle the rest.
1040 if (atomic_dec_and_test(&cwq
->wq
->nr_cwqs_to_flush
))
1041 complete(&cwq
->wq
->first_flusher
->done
);
1045 * try_to_grab_pending - steal work item from worklist and disable irq
1046 * @work: work item to steal
1047 * @is_dwork: @work is a delayed_work
1048 * @flags: place to store irq state
1050 * Try to grab PENDING bit of @work. This function can handle @work in any
1051 * stable state - idle, on timer or on worklist. Return values are
1053 * 1 if @work was pending and we successfully stole PENDING
1054 * 0 if @work was idle and we claimed PENDING
1055 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1056 * -ENOENT if someone else is canceling @work, this state may persist
1057 * for arbitrarily long
1059 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1060 * interrupted while holding PENDING and @work off queue, irq must be
1061 * disabled on entry. This, combined with delayed_work->timer being
1062 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1064 * On successful return, >= 0, irq is disabled and the caller is
1065 * responsible for releasing it using local_irq_restore(*@flags).
1067 * This function is safe to call from any context including IRQ handler.
1069 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1070 unsigned long *flags
)
1072 struct worker_pool
*pool
;
1074 local_irq_save(*flags
);
1076 /* try to steal the timer if it exists */
1078 struct delayed_work
*dwork
= to_delayed_work(work
);
1081 * dwork->timer is irqsafe. If del_timer() fails, it's
1082 * guaranteed that the timer is not queued anywhere and not
1083 * running on the local CPU.
1085 if (likely(del_timer(&dwork
->timer
)))
1089 /* try to claim PENDING the normal way */
1090 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1094 * The queueing is in progress, or it is already queued. Try to
1095 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1097 pool
= get_work_pool(work
);
1101 spin_lock(&pool
->lock
);
1102 if (!list_empty(&work
->entry
)) {
1104 * This work is queued, but perhaps we locked the wrong
1105 * pool. In that case we must see the new value after
1106 * rmb(), see insert_work()->wmb().
1109 if (pool
== get_work_pool(work
)) {
1110 debug_work_deactivate(work
);
1113 * A delayed work item cannot be grabbed directly
1114 * because it might have linked NO_COLOR work items
1115 * which, if left on the delayed_list, will confuse
1116 * cwq->nr_active management later on and cause
1117 * stall. Make sure the work item is activated
1120 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1121 cwq_activate_delayed_work(work
);
1123 list_del_init(&work
->entry
);
1124 cwq_dec_nr_in_flight(get_work_cwq(work
),
1125 get_work_color(work
));
1127 spin_unlock(&pool
->lock
);
1131 spin_unlock(&pool
->lock
);
1133 local_irq_restore(*flags
);
1134 if (work_is_canceling(work
))
1141 * insert_work - insert a work into gcwq
1142 * @cwq: cwq @work belongs to
1143 * @work: work to insert
1144 * @head: insertion point
1145 * @extra_flags: extra WORK_STRUCT_* flags to set
1147 * Insert @work which belongs to @cwq into @gcwq after @head.
1148 * @extra_flags is or'd to work_struct flags.
1151 * spin_lock_irq(pool->lock).
1153 static void insert_work(struct cpu_workqueue_struct
*cwq
,
1154 struct work_struct
*work
, struct list_head
*head
,
1155 unsigned int extra_flags
)
1157 struct worker_pool
*pool
= cwq
->pool
;
1159 /* we own @work, set data and link */
1160 set_work_cwq(work
, cwq
, extra_flags
);
1163 * Ensure that we get the right work->data if we see the
1164 * result of list_add() below, see try_to_grab_pending().
1168 list_add_tail(&work
->entry
, head
);
1171 * Ensure either worker_sched_deactivated() sees the above
1172 * list_add_tail() or we see zero nr_running to avoid workers
1173 * lying around lazily while there are works to be processed.
1177 if (__need_more_worker(pool
))
1178 wake_up_worker(pool
);
1182 * Test whether @work is being queued from another work executing on the
1183 * same workqueue. This is rather expensive and should only be used from
1186 static bool is_chained_work(struct workqueue_struct
*wq
)
1188 unsigned long flags
;
1191 for_each_gcwq_cpu(cpu
) {
1192 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
1193 struct worker_pool
*pool
= cwq
->pool
;
1194 struct worker
*worker
;
1195 struct hlist_node
*pos
;
1198 spin_lock_irqsave(&pool
->lock
, flags
);
1199 for_each_busy_worker(worker
, i
, pos
, pool
) {
1200 if (worker
->task
!= current
)
1202 spin_unlock_irqrestore(&pool
->lock
, flags
);
1204 * I'm @worker, no locking necessary. See if @work
1205 * is headed to the same workqueue.
1207 return worker
->current_cwq
->wq
== wq
;
1209 spin_unlock_irqrestore(&pool
->lock
, flags
);
1214 static void __queue_work(unsigned int cpu
, struct workqueue_struct
*wq
,
1215 struct work_struct
*work
)
1217 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
1218 struct worker_pool
*pool
;
1219 struct cpu_workqueue_struct
*cwq
;
1220 struct list_head
*worklist
;
1221 unsigned int work_flags
;
1222 unsigned int req_cpu
= cpu
;
1225 * While a work item is PENDING && off queue, a task trying to
1226 * steal the PENDING will busy-loop waiting for it to either get
1227 * queued or lose PENDING. Grabbing PENDING and queueing should
1228 * happen with IRQ disabled.
1230 WARN_ON_ONCE(!irqs_disabled());
1232 debug_work_activate(work
);
1234 /* if dying, only works from the same workqueue are allowed */
1235 if (unlikely(wq
->flags
& WQ_DRAINING
) &&
1236 WARN_ON_ONCE(!is_chained_work(wq
)))
1239 /* determine pool to use */
1240 if (!(wq
->flags
& WQ_UNBOUND
)) {
1241 struct worker_pool
*last_pool
;
1243 if (cpu
== WORK_CPU_UNBOUND
)
1244 cpu
= raw_smp_processor_id();
1247 * It's multi cpu. If @work was previously on a different
1248 * cpu, it might still be running there, in which case the
1249 * work needs to be queued on that cpu to guarantee
1252 pool
= get_std_worker_pool(cpu
, highpri
);
1253 last_pool
= get_work_pool(work
);
1255 if (last_pool
&& last_pool
!= pool
) {
1256 struct worker
*worker
;
1258 spin_lock(&last_pool
->lock
);
1260 worker
= find_worker_executing_work(last_pool
, work
);
1262 if (worker
&& worker
->current_cwq
->wq
== wq
)
1265 /* meh... not running there, queue here */
1266 spin_unlock(&last_pool
->lock
);
1267 spin_lock(&pool
->lock
);
1270 spin_lock(&pool
->lock
);
1273 pool
= get_std_worker_pool(WORK_CPU_UNBOUND
, highpri
);
1274 spin_lock(&pool
->lock
);
1277 /* pool determined, get cwq and queue */
1278 cwq
= get_cwq(pool
->cpu
, wq
);
1279 trace_workqueue_queue_work(req_cpu
, cwq
, work
);
1281 if (WARN_ON(!list_empty(&work
->entry
))) {
1282 spin_unlock(&pool
->lock
);
1286 cwq
->nr_in_flight
[cwq
->work_color
]++;
1287 work_flags
= work_color_to_flags(cwq
->work_color
);
1289 if (likely(cwq
->nr_active
< cwq
->max_active
)) {
1290 trace_workqueue_activate_work(work
);
1292 worklist
= &cwq
->pool
->worklist
;
1294 work_flags
|= WORK_STRUCT_DELAYED
;
1295 worklist
= &cwq
->delayed_works
;
1298 insert_work(cwq
, work
, worklist
, work_flags
);
1300 spin_unlock(&pool
->lock
);
1304 * queue_work_on - queue work on specific cpu
1305 * @cpu: CPU number to execute work on
1306 * @wq: workqueue to use
1307 * @work: work to queue
1309 * Returns %false if @work was already on a queue, %true otherwise.
1311 * We queue the work to a specific CPU, the caller must ensure it
1314 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1315 struct work_struct
*work
)
1318 unsigned long flags
;
1320 local_irq_save(flags
);
1322 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1323 __queue_work(cpu
, wq
, work
);
1327 local_irq_restore(flags
);
1330 EXPORT_SYMBOL_GPL(queue_work_on
);
1333 * queue_work - queue work on a workqueue
1334 * @wq: workqueue to use
1335 * @work: work to queue
1337 * Returns %false if @work was already on a queue, %true otherwise.
1339 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1340 * it can be processed by another CPU.
1342 bool queue_work(struct workqueue_struct
*wq
, struct work_struct
*work
)
1344 return queue_work_on(WORK_CPU_UNBOUND
, wq
, work
);
1346 EXPORT_SYMBOL_GPL(queue_work
);
1348 void delayed_work_timer_fn(unsigned long __data
)
1350 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1351 struct cpu_workqueue_struct
*cwq
= get_work_cwq(&dwork
->work
);
1353 /* should have been called from irqsafe timer with irq already off */
1354 __queue_work(dwork
->cpu
, cwq
->wq
, &dwork
->work
);
1356 EXPORT_SYMBOL_GPL(delayed_work_timer_fn
);
1358 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1359 struct delayed_work
*dwork
, unsigned long delay
)
1361 struct timer_list
*timer
= &dwork
->timer
;
1362 struct work_struct
*work
= &dwork
->work
;
1365 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1366 timer
->data
!= (unsigned long)dwork
);
1367 WARN_ON_ONCE(timer_pending(timer
));
1368 WARN_ON_ONCE(!list_empty(&work
->entry
));
1371 * If @delay is 0, queue @dwork->work immediately. This is for
1372 * both optimization and correctness. The earliest @timer can
1373 * expire is on the closest next tick and delayed_work users depend
1374 * on that there's no such delay when @delay is 0.
1377 __queue_work(cpu
, wq
, &dwork
->work
);
1381 timer_stats_timer_set_start_info(&dwork
->timer
);
1384 * This stores cwq for the moment, for the timer_fn. Note that the
1385 * work's pool is preserved to allow reentrance detection for
1388 if (!(wq
->flags
& WQ_UNBOUND
)) {
1389 struct worker_pool
*pool
= get_work_pool(work
);
1392 * If we cannot get the last pool from @work directly,
1393 * select the last CPU such that it avoids unnecessarily
1394 * triggering non-reentrancy check in __queue_work().
1399 if (lcpu
== WORK_CPU_UNBOUND
)
1400 lcpu
= raw_smp_processor_id();
1402 lcpu
= WORK_CPU_UNBOUND
;
1405 set_work_cwq(work
, get_cwq(lcpu
, wq
), 0);
1408 timer
->expires
= jiffies
+ delay
;
1410 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1411 add_timer_on(timer
, cpu
);
1417 * queue_delayed_work_on - queue work on specific CPU after delay
1418 * @cpu: CPU number to execute work on
1419 * @wq: workqueue to use
1420 * @dwork: work to queue
1421 * @delay: number of jiffies to wait before queueing
1423 * Returns %false if @work was already on a queue, %true otherwise. If
1424 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1427 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1428 struct delayed_work
*dwork
, unsigned long delay
)
1430 struct work_struct
*work
= &dwork
->work
;
1432 unsigned long flags
;
1434 /* read the comment in __queue_work() */
1435 local_irq_save(flags
);
1437 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1438 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1442 local_irq_restore(flags
);
1445 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1448 * queue_delayed_work - queue work on a workqueue after delay
1449 * @wq: workqueue to use
1450 * @dwork: delayable work to queue
1451 * @delay: number of jiffies to wait before queueing
1453 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1455 bool queue_delayed_work(struct workqueue_struct
*wq
,
1456 struct delayed_work
*dwork
, unsigned long delay
)
1458 return queue_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1460 EXPORT_SYMBOL_GPL(queue_delayed_work
);
1463 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1464 * @cpu: CPU number to execute work on
1465 * @wq: workqueue to use
1466 * @dwork: work to queue
1467 * @delay: number of jiffies to wait before queueing
1469 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1470 * modify @dwork's timer so that it expires after @delay. If @delay is
1471 * zero, @work is guaranteed to be scheduled immediately regardless of its
1474 * Returns %false if @dwork was idle and queued, %true if @dwork was
1475 * pending and its timer was modified.
1477 * This function is safe to call from any context including IRQ handler.
1478 * See try_to_grab_pending() for details.
1480 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1481 struct delayed_work
*dwork
, unsigned long delay
)
1483 unsigned long flags
;
1487 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1488 } while (unlikely(ret
== -EAGAIN
));
1490 if (likely(ret
>= 0)) {
1491 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1492 local_irq_restore(flags
);
1495 /* -ENOENT from try_to_grab_pending() becomes %true */
1498 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1501 * mod_delayed_work - modify delay of or queue a delayed work
1502 * @wq: workqueue to use
1503 * @dwork: work to queue
1504 * @delay: number of jiffies to wait before queueing
1506 * mod_delayed_work_on() on local CPU.
1508 bool mod_delayed_work(struct workqueue_struct
*wq
, struct delayed_work
*dwork
,
1509 unsigned long delay
)
1511 return mod_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1513 EXPORT_SYMBOL_GPL(mod_delayed_work
);
1516 * worker_enter_idle - enter idle state
1517 * @worker: worker which is entering idle state
1519 * @worker is entering idle state. Update stats and idle timer if
1523 * spin_lock_irq(pool->lock).
1525 static void worker_enter_idle(struct worker
*worker
)
1527 struct worker_pool
*pool
= worker
->pool
;
1529 BUG_ON(worker
->flags
& WORKER_IDLE
);
1530 BUG_ON(!list_empty(&worker
->entry
) &&
1531 (worker
->hentry
.next
|| worker
->hentry
.pprev
));
1533 /* can't use worker_set_flags(), also called from start_worker() */
1534 worker
->flags
|= WORKER_IDLE
;
1536 worker
->last_active
= jiffies
;
1538 /* idle_list is LIFO */
1539 list_add(&worker
->entry
, &pool
->idle_list
);
1541 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1542 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1545 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1546 * pool->lock between setting %WORKER_UNBOUND and zapping
1547 * nr_running, the warning may trigger spuriously. Check iff
1548 * unbind is not in progress.
1550 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1551 pool
->nr_workers
== pool
->nr_idle
&&
1552 atomic_read(get_pool_nr_running(pool
)));
1556 * worker_leave_idle - leave idle state
1557 * @worker: worker which is leaving idle state
1559 * @worker is leaving idle state. Update stats.
1562 * spin_lock_irq(pool->lock).
1564 static void worker_leave_idle(struct worker
*worker
)
1566 struct worker_pool
*pool
= worker
->pool
;
1568 BUG_ON(!(worker
->flags
& WORKER_IDLE
));
1569 worker_clr_flags(worker
, WORKER_IDLE
);
1571 list_del_init(&worker
->entry
);
1575 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1578 * Works which are scheduled while the cpu is online must at least be
1579 * scheduled to a worker which is bound to the cpu so that if they are
1580 * flushed from cpu callbacks while cpu is going down, they are
1581 * guaranteed to execute on the cpu.
1583 * This function is to be used by rogue workers and rescuers to bind
1584 * themselves to the target cpu and may race with cpu going down or
1585 * coming online. kthread_bind() can't be used because it may put the
1586 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1587 * verbatim as it's best effort and blocking and gcwq may be
1588 * [dis]associated in the meantime.
1590 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1591 * binding against %POOL_DISASSOCIATED which is set during
1592 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1593 * enters idle state or fetches works without dropping lock, it can
1594 * guarantee the scheduling requirement described in the first paragraph.
1597 * Might sleep. Called without any lock but returns with pool->lock
1601 * %true if the associated gcwq is online (@worker is successfully
1602 * bound), %false if offline.
1604 static bool worker_maybe_bind_and_lock(struct worker
*worker
)
1605 __acquires(&pool
->lock
)
1607 struct worker_pool
*pool
= worker
->pool
;
1608 struct task_struct
*task
= worker
->task
;
1612 * The following call may fail, succeed or succeed
1613 * without actually migrating the task to the cpu if
1614 * it races with cpu hotunplug operation. Verify
1615 * against POOL_DISASSOCIATED.
1617 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1618 set_cpus_allowed_ptr(task
, get_cpu_mask(pool
->cpu
));
1620 spin_lock_irq(&pool
->lock
);
1621 if (pool
->flags
& POOL_DISASSOCIATED
)
1623 if (task_cpu(task
) == pool
->cpu
&&
1624 cpumask_equal(¤t
->cpus_allowed
,
1625 get_cpu_mask(pool
->cpu
)))
1627 spin_unlock_irq(&pool
->lock
);
1630 * We've raced with CPU hot[un]plug. Give it a breather
1631 * and retry migration. cond_resched() is required here;
1632 * otherwise, we might deadlock against cpu_stop trying to
1633 * bring down the CPU on non-preemptive kernel.
1641 * Rebind an idle @worker to its CPU. worker_thread() will test
1642 * list_empty(@worker->entry) before leaving idle and call this function.
1644 static void idle_worker_rebind(struct worker
*worker
)
1646 /* CPU may go down again inbetween, clear UNBOUND only on success */
1647 if (worker_maybe_bind_and_lock(worker
))
1648 worker_clr_flags(worker
, WORKER_UNBOUND
);
1650 /* rebind complete, become available again */
1651 list_add(&worker
->entry
, &worker
->pool
->idle_list
);
1652 spin_unlock_irq(&worker
->pool
->lock
);
1656 * Function for @worker->rebind.work used to rebind unbound busy workers to
1657 * the associated cpu which is coming back online. This is scheduled by
1658 * cpu up but can race with other cpu hotplug operations and may be
1659 * executed twice without intervening cpu down.
1661 static void busy_worker_rebind_fn(struct work_struct
*work
)
1663 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1665 if (worker_maybe_bind_and_lock(worker
))
1666 worker_clr_flags(worker
, WORKER_UNBOUND
);
1668 spin_unlock_irq(&worker
->pool
->lock
);
1672 * rebind_workers - rebind all workers of a pool to the associated CPU
1673 * @pool: pool of interest
1675 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1676 * is different for idle and busy ones.
1678 * Idle ones will be removed from the idle_list and woken up. They will
1679 * add themselves back after completing rebind. This ensures that the
1680 * idle_list doesn't contain any unbound workers when re-bound busy workers
1681 * try to perform local wake-ups for concurrency management.
1683 * Busy workers can rebind after they finish their current work items.
1684 * Queueing the rebind work item at the head of the scheduled list is
1685 * enough. Note that nr_running will be properly bumped as busy workers
1688 * On return, all non-manager workers are scheduled for rebind - see
1689 * manage_workers() for the manager special case. Any idle worker
1690 * including the manager will not appear on @idle_list until rebind is
1691 * complete, making local wake-ups safe.
1693 static void rebind_workers(struct worker_pool
*pool
)
1695 struct worker
*worker
, *n
;
1696 struct hlist_node
*pos
;
1699 lockdep_assert_held(&pool
->assoc_mutex
);
1700 lockdep_assert_held(&pool
->lock
);
1702 /* dequeue and kick idle ones */
1703 list_for_each_entry_safe(worker
, n
, &pool
->idle_list
, entry
) {
1705 * idle workers should be off @pool->idle_list until rebind
1706 * is complete to avoid receiving premature local wake-ups.
1708 list_del_init(&worker
->entry
);
1711 * worker_thread() will see the above dequeuing and call
1712 * idle_worker_rebind().
1714 wake_up_process(worker
->task
);
1717 /* rebind busy workers */
1718 for_each_busy_worker(worker
, i
, pos
, pool
) {
1719 struct work_struct
*rebind_work
= &worker
->rebind_work
;
1720 struct workqueue_struct
*wq
;
1722 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
1723 work_data_bits(rebind_work
)))
1726 debug_work_activate(rebind_work
);
1729 * wq doesn't really matter but let's keep @worker->pool
1730 * and @cwq->pool consistent for sanity.
1732 if (std_worker_pool_pri(worker
->pool
))
1733 wq
= system_highpri_wq
;
1737 insert_work(get_cwq(pool
->cpu
, wq
), rebind_work
,
1738 worker
->scheduled
.next
,
1739 work_color_to_flags(WORK_NO_COLOR
));
1743 static struct worker
*alloc_worker(void)
1745 struct worker
*worker
;
1747 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1749 INIT_LIST_HEAD(&worker
->entry
);
1750 INIT_LIST_HEAD(&worker
->scheduled
);
1751 INIT_WORK(&worker
->rebind_work
, busy_worker_rebind_fn
);
1752 /* on creation a worker is in !idle && prep state */
1753 worker
->flags
= WORKER_PREP
;
1759 * create_worker - create a new workqueue worker
1760 * @pool: pool the new worker will belong to
1762 * Create a new worker which is bound to @pool. The returned worker
1763 * can be started by calling start_worker() or destroyed using
1767 * Might sleep. Does GFP_KERNEL allocations.
1770 * Pointer to the newly created worker.
1772 static struct worker
*create_worker(struct worker_pool
*pool
)
1774 const char *pri
= std_worker_pool_pri(pool
) ? "H" : "";
1775 struct worker
*worker
= NULL
;
1778 spin_lock_irq(&pool
->lock
);
1779 while (ida_get_new(&pool
->worker_ida
, &id
)) {
1780 spin_unlock_irq(&pool
->lock
);
1781 if (!ida_pre_get(&pool
->worker_ida
, GFP_KERNEL
))
1783 spin_lock_irq(&pool
->lock
);
1785 spin_unlock_irq(&pool
->lock
);
1787 worker
= alloc_worker();
1791 worker
->pool
= pool
;
1794 if (pool
->cpu
!= WORK_CPU_UNBOUND
)
1795 worker
->task
= kthread_create_on_node(worker_thread
,
1796 worker
, cpu_to_node(pool
->cpu
),
1797 "kworker/%u:%d%s", pool
->cpu
, id
, pri
);
1799 worker
->task
= kthread_create(worker_thread
, worker
,
1800 "kworker/u:%d%s", id
, pri
);
1801 if (IS_ERR(worker
->task
))
1804 if (std_worker_pool_pri(pool
))
1805 set_user_nice(worker
->task
, HIGHPRI_NICE_LEVEL
);
1808 * Determine CPU binding of the new worker depending on
1809 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1810 * flag remains stable across this function. See the comments
1811 * above the flag definition for details.
1813 * As an unbound worker may later become a regular one if CPU comes
1814 * online, make sure every worker has %PF_THREAD_BOUND set.
1816 if (!(pool
->flags
& POOL_DISASSOCIATED
)) {
1817 kthread_bind(worker
->task
, pool
->cpu
);
1819 worker
->task
->flags
|= PF_THREAD_BOUND
;
1820 worker
->flags
|= WORKER_UNBOUND
;
1826 spin_lock_irq(&pool
->lock
);
1827 ida_remove(&pool
->worker_ida
, id
);
1828 spin_unlock_irq(&pool
->lock
);
1835 * start_worker - start a newly created worker
1836 * @worker: worker to start
1838 * Make the gcwq aware of @worker and start it.
1841 * spin_lock_irq(pool->lock).
1843 static void start_worker(struct worker
*worker
)
1845 worker
->flags
|= WORKER_STARTED
;
1846 worker
->pool
->nr_workers
++;
1847 worker_enter_idle(worker
);
1848 wake_up_process(worker
->task
);
1852 * destroy_worker - destroy a workqueue worker
1853 * @worker: worker to be destroyed
1855 * Destroy @worker and adjust @gcwq stats accordingly.
1858 * spin_lock_irq(pool->lock) which is released and regrabbed.
1860 static void destroy_worker(struct worker
*worker
)
1862 struct worker_pool
*pool
= worker
->pool
;
1863 int id
= worker
->id
;
1865 /* sanity check frenzy */
1866 BUG_ON(worker
->current_work
);
1867 BUG_ON(!list_empty(&worker
->scheduled
));
1869 if (worker
->flags
& WORKER_STARTED
)
1871 if (worker
->flags
& WORKER_IDLE
)
1874 list_del_init(&worker
->entry
);
1875 worker
->flags
|= WORKER_DIE
;
1877 spin_unlock_irq(&pool
->lock
);
1879 kthread_stop(worker
->task
);
1882 spin_lock_irq(&pool
->lock
);
1883 ida_remove(&pool
->worker_ida
, id
);
1886 static void idle_worker_timeout(unsigned long __pool
)
1888 struct worker_pool
*pool
= (void *)__pool
;
1890 spin_lock_irq(&pool
->lock
);
1892 if (too_many_workers(pool
)) {
1893 struct worker
*worker
;
1894 unsigned long expires
;
1896 /* idle_list is kept in LIFO order, check the last one */
1897 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1898 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1900 if (time_before(jiffies
, expires
))
1901 mod_timer(&pool
->idle_timer
, expires
);
1903 /* it's been idle for too long, wake up manager */
1904 pool
->flags
|= POOL_MANAGE_WORKERS
;
1905 wake_up_worker(pool
);
1909 spin_unlock_irq(&pool
->lock
);
1912 static bool send_mayday(struct work_struct
*work
)
1914 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1915 struct workqueue_struct
*wq
= cwq
->wq
;
1918 if (!(wq
->flags
& WQ_RESCUER
))
1921 /* mayday mayday mayday */
1922 cpu
= cwq
->pool
->cpu
;
1923 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1924 if (cpu
== WORK_CPU_UNBOUND
)
1926 if (!mayday_test_and_set_cpu(cpu
, wq
->mayday_mask
))
1927 wake_up_process(wq
->rescuer
->task
);
1931 static void gcwq_mayday_timeout(unsigned long __pool
)
1933 struct worker_pool
*pool
= (void *)__pool
;
1934 struct work_struct
*work
;
1936 spin_lock_irq(&pool
->lock
);
1938 if (need_to_create_worker(pool
)) {
1940 * We've been trying to create a new worker but
1941 * haven't been successful. We might be hitting an
1942 * allocation deadlock. Send distress signals to
1945 list_for_each_entry(work
, &pool
->worklist
, entry
)
1949 spin_unlock_irq(&pool
->lock
);
1951 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1955 * maybe_create_worker - create a new worker if necessary
1956 * @pool: pool to create a new worker for
1958 * Create a new worker for @pool if necessary. @pool is guaranteed to
1959 * have at least one idle worker on return from this function. If
1960 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1961 * sent to all rescuers with works scheduled on @pool to resolve
1962 * possible allocation deadlock.
1964 * On return, need_to_create_worker() is guaranteed to be false and
1965 * may_start_working() true.
1968 * spin_lock_irq(pool->lock) which may be released and regrabbed
1969 * multiple times. Does GFP_KERNEL allocations. Called only from
1973 * false if no action was taken and pool->lock stayed locked, true
1976 static bool maybe_create_worker(struct worker_pool
*pool
)
1977 __releases(&pool
->lock
)
1978 __acquires(&pool
->lock
)
1980 if (!need_to_create_worker(pool
))
1983 spin_unlock_irq(&pool
->lock
);
1985 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1986 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1989 struct worker
*worker
;
1991 worker
= create_worker(pool
);
1993 del_timer_sync(&pool
->mayday_timer
);
1994 spin_lock_irq(&pool
->lock
);
1995 start_worker(worker
);
1996 BUG_ON(need_to_create_worker(pool
));
2000 if (!need_to_create_worker(pool
))
2003 __set_current_state(TASK_INTERRUPTIBLE
);
2004 schedule_timeout(CREATE_COOLDOWN
);
2006 if (!need_to_create_worker(pool
))
2010 del_timer_sync(&pool
->mayday_timer
);
2011 spin_lock_irq(&pool
->lock
);
2012 if (need_to_create_worker(pool
))
2018 * maybe_destroy_worker - destroy workers which have been idle for a while
2019 * @pool: pool to destroy workers for
2021 * Destroy @pool workers which have been idle for longer than
2022 * IDLE_WORKER_TIMEOUT.
2025 * spin_lock_irq(pool->lock) which may be released and regrabbed
2026 * multiple times. Called only from manager.
2029 * false if no action was taken and pool->lock stayed locked, true
2032 static bool maybe_destroy_workers(struct worker_pool
*pool
)
2036 while (too_many_workers(pool
)) {
2037 struct worker
*worker
;
2038 unsigned long expires
;
2040 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2041 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2043 if (time_before(jiffies
, expires
)) {
2044 mod_timer(&pool
->idle_timer
, expires
);
2048 destroy_worker(worker
);
2056 * manage_workers - manage worker pool
2059 * Assume the manager role and manage gcwq worker pool @worker belongs
2060 * to. At any given time, there can be only zero or one manager per
2061 * gcwq. The exclusion is handled automatically by this function.
2063 * The caller can safely start processing works on false return. On
2064 * true return, it's guaranteed that need_to_create_worker() is false
2065 * and may_start_working() is true.
2068 * spin_lock_irq(pool->lock) which may be released and regrabbed
2069 * multiple times. Does GFP_KERNEL allocations.
2072 * spin_lock_irq(pool->lock) which may be released and regrabbed
2073 * multiple times. Does GFP_KERNEL allocations.
2075 static bool manage_workers(struct worker
*worker
)
2077 struct worker_pool
*pool
= worker
->pool
;
2080 if (pool
->flags
& POOL_MANAGING_WORKERS
)
2083 pool
->flags
|= POOL_MANAGING_WORKERS
;
2086 * To simplify both worker management and CPU hotplug, hold off
2087 * management while hotplug is in progress. CPU hotplug path can't
2088 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2089 * lead to idle worker depletion (all become busy thinking someone
2090 * else is managing) which in turn can result in deadlock under
2091 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2092 * manager against CPU hotplug.
2094 * assoc_mutex would always be free unless CPU hotplug is in
2095 * progress. trylock first without dropping @pool->lock.
2097 if (unlikely(!mutex_trylock(&pool
->assoc_mutex
))) {
2098 spin_unlock_irq(&pool
->lock
);
2099 mutex_lock(&pool
->assoc_mutex
);
2101 * CPU hotplug could have happened while we were waiting
2102 * for assoc_mutex. Hotplug itself can't handle us
2103 * because manager isn't either on idle or busy list, and
2104 * @gcwq's state and ours could have deviated.
2106 * As hotplug is now excluded via assoc_mutex, we can
2107 * simply try to bind. It will succeed or fail depending
2108 * on @gcwq's current state. Try it and adjust
2109 * %WORKER_UNBOUND accordingly.
2111 if (worker_maybe_bind_and_lock(worker
))
2112 worker
->flags
&= ~WORKER_UNBOUND
;
2114 worker
->flags
|= WORKER_UNBOUND
;
2119 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2122 * Destroy and then create so that may_start_working() is true
2125 ret
|= maybe_destroy_workers(pool
);
2126 ret
|= maybe_create_worker(pool
);
2128 pool
->flags
&= ~POOL_MANAGING_WORKERS
;
2129 mutex_unlock(&pool
->assoc_mutex
);
2134 * process_one_work - process single work
2136 * @work: work to process
2138 * Process @work. This function contains all the logics necessary to
2139 * process a single work including synchronization against and
2140 * interaction with other workers on the same cpu, queueing and
2141 * flushing. As long as context requirement is met, any worker can
2142 * call this function to process a work.
2145 * spin_lock_irq(pool->lock) which is released and regrabbed.
2147 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2148 __releases(&pool
->lock
)
2149 __acquires(&pool
->lock
)
2151 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
2152 struct worker_pool
*pool
= worker
->pool
;
2153 bool cpu_intensive
= cwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2155 struct worker
*collision
;
2156 #ifdef CONFIG_LOCKDEP
2158 * It is permissible to free the struct work_struct from
2159 * inside the function that is called from it, this we need to
2160 * take into account for lockdep too. To avoid bogus "held
2161 * lock freed" warnings as well as problems when looking into
2162 * work->lockdep_map, make a copy and use that here.
2164 struct lockdep_map lockdep_map
;
2166 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2169 * Ensure we're on the correct CPU. DISASSOCIATED test is
2170 * necessary to avoid spurious warnings from rescuers servicing the
2171 * unbound or a disassociated pool.
2173 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2174 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2175 raw_smp_processor_id() != pool
->cpu
);
2178 * A single work shouldn't be executed concurrently by
2179 * multiple workers on a single cpu. Check whether anyone is
2180 * already processing the work. If so, defer the work to the
2181 * currently executing one.
2183 collision
= find_worker_executing_work(pool
, work
);
2184 if (unlikely(collision
)) {
2185 move_linked_works(work
, &collision
->scheduled
, NULL
);
2189 /* claim and dequeue */
2190 debug_work_deactivate(work
);
2191 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2192 worker
->current_work
= work
;
2193 worker
->current_func
= work
->func
;
2194 worker
->current_cwq
= cwq
;
2195 work_color
= get_work_color(work
);
2197 list_del_init(&work
->entry
);
2200 * CPU intensive works don't participate in concurrency
2201 * management. They're the scheduler's responsibility.
2203 if (unlikely(cpu_intensive
))
2204 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2207 * Unbound pool isn't concurrency managed and work items should be
2208 * executed ASAP. Wake up another worker if necessary.
2210 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2211 wake_up_worker(pool
);
2214 * Record the last pool and clear PENDING which should be the last
2215 * update to @work. Also, do this inside @pool->lock so that
2216 * PENDING and queued state changes happen together while IRQ is
2219 set_work_pool_and_clear_pending(work
, pool
->id
);
2221 spin_unlock_irq(&pool
->lock
);
2223 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2224 lock_map_acquire(&lockdep_map
);
2225 trace_workqueue_execute_start(work
);
2226 worker
->current_func(work
);
2228 * While we must be careful to not use "work" after this, the trace
2229 * point will only record its address.
2231 trace_workqueue_execute_end(work
);
2232 lock_map_release(&lockdep_map
);
2233 lock_map_release(&cwq
->wq
->lockdep_map
);
2235 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2236 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2237 " last function: %pf\n",
2238 current
->comm
, preempt_count(), task_pid_nr(current
),
2239 worker
->current_func
);
2240 debug_show_held_locks(current
);
2244 spin_lock_irq(&pool
->lock
);
2246 /* clear cpu intensive status */
2247 if (unlikely(cpu_intensive
))
2248 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2250 /* we're done with it, release */
2251 hash_del(&worker
->hentry
);
2252 worker
->current_work
= NULL
;
2253 worker
->current_func
= NULL
;
2254 worker
->current_cwq
= NULL
;
2255 cwq_dec_nr_in_flight(cwq
, work_color
);
2259 * process_scheduled_works - process scheduled works
2262 * Process all scheduled works. Please note that the scheduled list
2263 * may change while processing a work, so this function repeatedly
2264 * fetches a work from the top and executes it.
2267 * spin_lock_irq(pool->lock) which may be released and regrabbed
2270 static void process_scheduled_works(struct worker
*worker
)
2272 while (!list_empty(&worker
->scheduled
)) {
2273 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2274 struct work_struct
, entry
);
2275 process_one_work(worker
, work
);
2280 * worker_thread - the worker thread function
2283 * The gcwq worker thread function. There's a single dynamic pool of
2284 * these per each cpu. These workers process all works regardless of
2285 * their specific target workqueue. The only exception is works which
2286 * belong to workqueues with a rescuer which will be explained in
2289 static int worker_thread(void *__worker
)
2291 struct worker
*worker
= __worker
;
2292 struct worker_pool
*pool
= worker
->pool
;
2294 /* tell the scheduler that this is a workqueue worker */
2295 worker
->task
->flags
|= PF_WQ_WORKER
;
2297 spin_lock_irq(&pool
->lock
);
2299 /* we are off idle list if destruction or rebind is requested */
2300 if (unlikely(list_empty(&worker
->entry
))) {
2301 spin_unlock_irq(&pool
->lock
);
2303 /* if DIE is set, destruction is requested */
2304 if (worker
->flags
& WORKER_DIE
) {
2305 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2309 /* otherwise, rebind */
2310 idle_worker_rebind(worker
);
2314 worker_leave_idle(worker
);
2316 /* no more worker necessary? */
2317 if (!need_more_worker(pool
))
2320 /* do we need to manage? */
2321 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2325 * ->scheduled list can only be filled while a worker is
2326 * preparing to process a work or actually processing it.
2327 * Make sure nobody diddled with it while I was sleeping.
2329 BUG_ON(!list_empty(&worker
->scheduled
));
2332 * When control reaches this point, we're guaranteed to have
2333 * at least one idle worker or that someone else has already
2334 * assumed the manager role.
2336 worker_clr_flags(worker
, WORKER_PREP
);
2339 struct work_struct
*work
=
2340 list_first_entry(&pool
->worklist
,
2341 struct work_struct
, entry
);
2343 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2344 /* optimization path, not strictly necessary */
2345 process_one_work(worker
, work
);
2346 if (unlikely(!list_empty(&worker
->scheduled
)))
2347 process_scheduled_works(worker
);
2349 move_linked_works(work
, &worker
->scheduled
, NULL
);
2350 process_scheduled_works(worker
);
2352 } while (keep_working(pool
));
2354 worker_set_flags(worker
, WORKER_PREP
, false);
2356 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2360 * pool->lock is held and there's no work to process and no need to
2361 * manage, sleep. Workers are woken up only while holding
2362 * pool->lock or from local cpu, so setting the current state
2363 * before releasing pool->lock is enough to prevent losing any
2366 worker_enter_idle(worker
);
2367 __set_current_state(TASK_INTERRUPTIBLE
);
2368 spin_unlock_irq(&pool
->lock
);
2374 * rescuer_thread - the rescuer thread function
2377 * Workqueue rescuer thread function. There's one rescuer for each
2378 * workqueue which has WQ_RESCUER set.
2380 * Regular work processing on a gcwq may block trying to create a new
2381 * worker which uses GFP_KERNEL allocation which has slight chance of
2382 * developing into deadlock if some works currently on the same queue
2383 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2384 * the problem rescuer solves.
2386 * When such condition is possible, the gcwq summons rescuers of all
2387 * workqueues which have works queued on the gcwq and let them process
2388 * those works so that forward progress can be guaranteed.
2390 * This should happen rarely.
2392 static int rescuer_thread(void *__rescuer
)
2394 struct worker
*rescuer
= __rescuer
;
2395 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2396 struct list_head
*scheduled
= &rescuer
->scheduled
;
2397 bool is_unbound
= wq
->flags
& WQ_UNBOUND
;
2400 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2403 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2404 * doesn't participate in concurrency management.
2406 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2408 set_current_state(TASK_INTERRUPTIBLE
);
2410 if (kthread_should_stop()) {
2411 __set_current_state(TASK_RUNNING
);
2412 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2417 * See whether any cpu is asking for help. Unbounded
2418 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2420 for_each_mayday_cpu(cpu
, wq
->mayday_mask
) {
2421 unsigned int tcpu
= is_unbound
? WORK_CPU_UNBOUND
: cpu
;
2422 struct cpu_workqueue_struct
*cwq
= get_cwq(tcpu
, wq
);
2423 struct worker_pool
*pool
= cwq
->pool
;
2424 struct work_struct
*work
, *n
;
2426 __set_current_state(TASK_RUNNING
);
2427 mayday_clear_cpu(cpu
, wq
->mayday_mask
);
2429 /* migrate to the target cpu if possible */
2430 rescuer
->pool
= pool
;
2431 worker_maybe_bind_and_lock(rescuer
);
2434 * Slurp in all works issued via this workqueue and
2437 BUG_ON(!list_empty(&rescuer
->scheduled
));
2438 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2439 if (get_work_cwq(work
) == cwq
)
2440 move_linked_works(work
, scheduled
, &n
);
2442 process_scheduled_works(rescuer
);
2445 * Leave this pool. If keep_working() is %true, notify a
2446 * regular worker; otherwise, we end up with 0 concurrency
2447 * and stalling the execution.
2449 if (keep_working(pool
))
2450 wake_up_worker(pool
);
2452 spin_unlock_irq(&pool
->lock
);
2455 /* rescuers should never participate in concurrency management */
2456 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2462 struct work_struct work
;
2463 struct completion done
;
2466 static void wq_barrier_func(struct work_struct
*work
)
2468 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2469 complete(&barr
->done
);
2473 * insert_wq_barrier - insert a barrier work
2474 * @cwq: cwq to insert barrier into
2475 * @barr: wq_barrier to insert
2476 * @target: target work to attach @barr to
2477 * @worker: worker currently executing @target, NULL if @target is not executing
2479 * @barr is linked to @target such that @barr is completed only after
2480 * @target finishes execution. Please note that the ordering
2481 * guarantee is observed only with respect to @target and on the local
2484 * Currently, a queued barrier can't be canceled. This is because
2485 * try_to_grab_pending() can't determine whether the work to be
2486 * grabbed is at the head of the queue and thus can't clear LINKED
2487 * flag of the previous work while there must be a valid next work
2488 * after a work with LINKED flag set.
2490 * Note that when @worker is non-NULL, @target may be modified
2491 * underneath us, so we can't reliably determine cwq from @target.
2494 * spin_lock_irq(pool->lock).
2496 static void insert_wq_barrier(struct cpu_workqueue_struct
*cwq
,
2497 struct wq_barrier
*barr
,
2498 struct work_struct
*target
, struct worker
*worker
)
2500 struct list_head
*head
;
2501 unsigned int linked
= 0;
2504 * debugobject calls are safe here even with pool->lock locked
2505 * as we know for sure that this will not trigger any of the
2506 * checks and call back into the fixup functions where we
2509 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2510 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2511 init_completion(&barr
->done
);
2514 * If @target is currently being executed, schedule the
2515 * barrier to the worker; otherwise, put it after @target.
2518 head
= worker
->scheduled
.next
;
2520 unsigned long *bits
= work_data_bits(target
);
2522 head
= target
->entry
.next
;
2523 /* there can already be other linked works, inherit and set */
2524 linked
= *bits
& WORK_STRUCT_LINKED
;
2525 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2528 debug_work_activate(&barr
->work
);
2529 insert_work(cwq
, &barr
->work
, head
,
2530 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2534 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2535 * @wq: workqueue being flushed
2536 * @flush_color: new flush color, < 0 for no-op
2537 * @work_color: new work color, < 0 for no-op
2539 * Prepare cwqs for workqueue flushing.
2541 * If @flush_color is non-negative, flush_color on all cwqs should be
2542 * -1. If no cwq has in-flight commands at the specified color, all
2543 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2544 * has in flight commands, its cwq->flush_color is set to
2545 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2546 * wakeup logic is armed and %true is returned.
2548 * The caller should have initialized @wq->first_flusher prior to
2549 * calling this function with non-negative @flush_color. If
2550 * @flush_color is negative, no flush color update is done and %false
2553 * If @work_color is non-negative, all cwqs should have the same
2554 * work_color which is previous to @work_color and all will be
2555 * advanced to @work_color.
2558 * mutex_lock(wq->flush_mutex).
2561 * %true if @flush_color >= 0 and there's something to flush. %false
2564 static bool flush_workqueue_prep_cwqs(struct workqueue_struct
*wq
,
2565 int flush_color
, int work_color
)
2570 if (flush_color
>= 0) {
2571 BUG_ON(atomic_read(&wq
->nr_cwqs_to_flush
));
2572 atomic_set(&wq
->nr_cwqs_to_flush
, 1);
2575 for_each_cwq_cpu(cpu
, wq
) {
2576 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2577 struct worker_pool
*pool
= cwq
->pool
;
2579 spin_lock_irq(&pool
->lock
);
2581 if (flush_color
>= 0) {
2582 BUG_ON(cwq
->flush_color
!= -1);
2584 if (cwq
->nr_in_flight
[flush_color
]) {
2585 cwq
->flush_color
= flush_color
;
2586 atomic_inc(&wq
->nr_cwqs_to_flush
);
2591 if (work_color
>= 0) {
2592 BUG_ON(work_color
!= work_next_color(cwq
->work_color
));
2593 cwq
->work_color
= work_color
;
2596 spin_unlock_irq(&pool
->lock
);
2599 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_cwqs_to_flush
))
2600 complete(&wq
->first_flusher
->done
);
2606 * flush_workqueue - ensure that any scheduled work has run to completion.
2607 * @wq: workqueue to flush
2609 * Forces execution of the workqueue and blocks until its completion.
2610 * This is typically used in driver shutdown handlers.
2612 * We sleep until all works which were queued on entry have been handled,
2613 * but we are not livelocked by new incoming ones.
2615 void flush_workqueue(struct workqueue_struct
*wq
)
2617 struct wq_flusher this_flusher
= {
2618 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2620 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2624 lock_map_acquire(&wq
->lockdep_map
);
2625 lock_map_release(&wq
->lockdep_map
);
2627 mutex_lock(&wq
->flush_mutex
);
2630 * Start-to-wait phase
2632 next_color
= work_next_color(wq
->work_color
);
2634 if (next_color
!= wq
->flush_color
) {
2636 * Color space is not full. The current work_color
2637 * becomes our flush_color and work_color is advanced
2640 BUG_ON(!list_empty(&wq
->flusher_overflow
));
2641 this_flusher
.flush_color
= wq
->work_color
;
2642 wq
->work_color
= next_color
;
2644 if (!wq
->first_flusher
) {
2645 /* no flush in progress, become the first flusher */
2646 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2648 wq
->first_flusher
= &this_flusher
;
2650 if (!flush_workqueue_prep_cwqs(wq
, wq
->flush_color
,
2652 /* nothing to flush, done */
2653 wq
->flush_color
= next_color
;
2654 wq
->first_flusher
= NULL
;
2659 BUG_ON(wq
->flush_color
== this_flusher
.flush_color
);
2660 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2661 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2665 * Oops, color space is full, wait on overflow queue.
2666 * The next flush completion will assign us
2667 * flush_color and transfer to flusher_queue.
2669 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2672 mutex_unlock(&wq
->flush_mutex
);
2674 wait_for_completion(&this_flusher
.done
);
2677 * Wake-up-and-cascade phase
2679 * First flushers are responsible for cascading flushes and
2680 * handling overflow. Non-first flushers can simply return.
2682 if (wq
->first_flusher
!= &this_flusher
)
2685 mutex_lock(&wq
->flush_mutex
);
2687 /* we might have raced, check again with mutex held */
2688 if (wq
->first_flusher
!= &this_flusher
)
2691 wq
->first_flusher
= NULL
;
2693 BUG_ON(!list_empty(&this_flusher
.list
));
2694 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2697 struct wq_flusher
*next
, *tmp
;
2699 /* complete all the flushers sharing the current flush color */
2700 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2701 if (next
->flush_color
!= wq
->flush_color
)
2703 list_del_init(&next
->list
);
2704 complete(&next
->done
);
2707 BUG_ON(!list_empty(&wq
->flusher_overflow
) &&
2708 wq
->flush_color
!= work_next_color(wq
->work_color
));
2710 /* this flush_color is finished, advance by one */
2711 wq
->flush_color
= work_next_color(wq
->flush_color
);
2713 /* one color has been freed, handle overflow queue */
2714 if (!list_empty(&wq
->flusher_overflow
)) {
2716 * Assign the same color to all overflowed
2717 * flushers, advance work_color and append to
2718 * flusher_queue. This is the start-to-wait
2719 * phase for these overflowed flushers.
2721 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2722 tmp
->flush_color
= wq
->work_color
;
2724 wq
->work_color
= work_next_color(wq
->work_color
);
2726 list_splice_tail_init(&wq
->flusher_overflow
,
2727 &wq
->flusher_queue
);
2728 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2731 if (list_empty(&wq
->flusher_queue
)) {
2732 BUG_ON(wq
->flush_color
!= wq
->work_color
);
2737 * Need to flush more colors. Make the next flusher
2738 * the new first flusher and arm cwqs.
2740 BUG_ON(wq
->flush_color
== wq
->work_color
);
2741 BUG_ON(wq
->flush_color
!= next
->flush_color
);
2743 list_del_init(&next
->list
);
2744 wq
->first_flusher
= next
;
2746 if (flush_workqueue_prep_cwqs(wq
, wq
->flush_color
, -1))
2750 * Meh... this color is already done, clear first
2751 * flusher and repeat cascading.
2753 wq
->first_flusher
= NULL
;
2757 mutex_unlock(&wq
->flush_mutex
);
2759 EXPORT_SYMBOL_GPL(flush_workqueue
);
2762 * drain_workqueue - drain a workqueue
2763 * @wq: workqueue to drain
2765 * Wait until the workqueue becomes empty. While draining is in progress,
2766 * only chain queueing is allowed. IOW, only currently pending or running
2767 * work items on @wq can queue further work items on it. @wq is flushed
2768 * repeatedly until it becomes empty. The number of flushing is detemined
2769 * by the depth of chaining and should be relatively short. Whine if it
2772 void drain_workqueue(struct workqueue_struct
*wq
)
2774 unsigned int flush_cnt
= 0;
2778 * __queue_work() needs to test whether there are drainers, is much
2779 * hotter than drain_workqueue() and already looks at @wq->flags.
2780 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2782 spin_lock(&workqueue_lock
);
2783 if (!wq
->nr_drainers
++)
2784 wq
->flags
|= WQ_DRAINING
;
2785 spin_unlock(&workqueue_lock
);
2787 flush_workqueue(wq
);
2789 for_each_cwq_cpu(cpu
, wq
) {
2790 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2793 spin_lock_irq(&cwq
->pool
->lock
);
2794 drained
= !cwq
->nr_active
&& list_empty(&cwq
->delayed_works
);
2795 spin_unlock_irq(&cwq
->pool
->lock
);
2800 if (++flush_cnt
== 10 ||
2801 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2802 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2803 wq
->name
, flush_cnt
);
2807 spin_lock(&workqueue_lock
);
2808 if (!--wq
->nr_drainers
)
2809 wq
->flags
&= ~WQ_DRAINING
;
2810 spin_unlock(&workqueue_lock
);
2812 EXPORT_SYMBOL_GPL(drain_workqueue
);
2814 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2816 struct worker
*worker
= NULL
;
2817 struct worker_pool
*pool
;
2818 struct cpu_workqueue_struct
*cwq
;
2821 pool
= get_work_pool(work
);
2825 spin_lock_irq(&pool
->lock
);
2826 if (!list_empty(&work
->entry
)) {
2828 * See the comment near try_to_grab_pending()->smp_rmb().
2829 * If it was re-queued to a different pool under us, we
2830 * are not going to wait.
2833 cwq
= get_work_cwq(work
);
2834 if (unlikely(!cwq
|| pool
!= cwq
->pool
))
2837 worker
= find_worker_executing_work(pool
, work
);
2840 cwq
= worker
->current_cwq
;
2843 insert_wq_barrier(cwq
, barr
, work
, worker
);
2844 spin_unlock_irq(&pool
->lock
);
2847 * If @max_active is 1 or rescuer is in use, flushing another work
2848 * item on the same workqueue may lead to deadlock. Make sure the
2849 * flusher is not running on the same workqueue by verifying write
2852 if (cwq
->wq
->saved_max_active
== 1 || cwq
->wq
->flags
& WQ_RESCUER
)
2853 lock_map_acquire(&cwq
->wq
->lockdep_map
);
2855 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2856 lock_map_release(&cwq
->wq
->lockdep_map
);
2860 spin_unlock_irq(&pool
->lock
);
2865 * flush_work - wait for a work to finish executing the last queueing instance
2866 * @work: the work to flush
2868 * Wait until @work has finished execution. @work is guaranteed to be idle
2869 * on return if it hasn't been requeued since flush started.
2872 * %true if flush_work() waited for the work to finish execution,
2873 * %false if it was already idle.
2875 bool flush_work(struct work_struct
*work
)
2877 struct wq_barrier barr
;
2879 lock_map_acquire(&work
->lockdep_map
);
2880 lock_map_release(&work
->lockdep_map
);
2882 if (start_flush_work(work
, &barr
)) {
2883 wait_for_completion(&barr
.done
);
2884 destroy_work_on_stack(&barr
.work
);
2890 EXPORT_SYMBOL_GPL(flush_work
);
2892 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2894 unsigned long flags
;
2898 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2900 * If someone else is canceling, wait for the same event it
2901 * would be waiting for before retrying.
2903 if (unlikely(ret
== -ENOENT
))
2905 } while (unlikely(ret
< 0));
2907 /* tell other tasks trying to grab @work to back off */
2908 mark_work_canceling(work
);
2909 local_irq_restore(flags
);
2912 clear_work_data(work
);
2917 * cancel_work_sync - cancel a work and wait for it to finish
2918 * @work: the work to cancel
2920 * Cancel @work and wait for its execution to finish. This function
2921 * can be used even if the work re-queues itself or migrates to
2922 * another workqueue. On return from this function, @work is
2923 * guaranteed to be not pending or executing on any CPU.
2925 * cancel_work_sync(&delayed_work->work) must not be used for
2926 * delayed_work's. Use cancel_delayed_work_sync() instead.
2928 * The caller must ensure that the workqueue on which @work was last
2929 * queued can't be destroyed before this function returns.
2932 * %true if @work was pending, %false otherwise.
2934 bool cancel_work_sync(struct work_struct
*work
)
2936 return __cancel_work_timer(work
, false);
2938 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2941 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2942 * @dwork: the delayed work to flush
2944 * Delayed timer is cancelled and the pending work is queued for
2945 * immediate execution. Like flush_work(), this function only
2946 * considers the last queueing instance of @dwork.
2949 * %true if flush_work() waited for the work to finish execution,
2950 * %false if it was already idle.
2952 bool flush_delayed_work(struct delayed_work
*dwork
)
2954 local_irq_disable();
2955 if (del_timer_sync(&dwork
->timer
))
2956 __queue_work(dwork
->cpu
,
2957 get_work_cwq(&dwork
->work
)->wq
, &dwork
->work
);
2959 return flush_work(&dwork
->work
);
2961 EXPORT_SYMBOL(flush_delayed_work
);
2964 * cancel_delayed_work - cancel a delayed work
2965 * @dwork: delayed_work to cancel
2967 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2968 * and canceled; %false if wasn't pending. Note that the work callback
2969 * function may still be running on return, unless it returns %true and the
2970 * work doesn't re-arm itself. Explicitly flush or use
2971 * cancel_delayed_work_sync() to wait on it.
2973 * This function is safe to call from any context including IRQ handler.
2975 bool cancel_delayed_work(struct delayed_work
*dwork
)
2977 unsigned long flags
;
2981 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2982 } while (unlikely(ret
== -EAGAIN
));
2984 if (unlikely(ret
< 0))
2987 set_work_pool_and_clear_pending(&dwork
->work
,
2988 get_work_pool_id(&dwork
->work
));
2989 local_irq_restore(flags
);
2992 EXPORT_SYMBOL(cancel_delayed_work
);
2995 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2996 * @dwork: the delayed work cancel
2998 * This is cancel_work_sync() for delayed works.
3001 * %true if @dwork was pending, %false otherwise.
3003 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
3005 return __cancel_work_timer(&dwork
->work
, true);
3007 EXPORT_SYMBOL(cancel_delayed_work_sync
);
3010 * schedule_work_on - put work task on a specific cpu
3011 * @cpu: cpu to put the work task on
3012 * @work: job to be done
3014 * This puts a job on a specific cpu
3016 bool schedule_work_on(int cpu
, struct work_struct
*work
)
3018 return queue_work_on(cpu
, system_wq
, work
);
3020 EXPORT_SYMBOL(schedule_work_on
);
3023 * schedule_work - put work task in global workqueue
3024 * @work: job to be done
3026 * Returns %false if @work was already on the kernel-global workqueue and
3029 * This puts a job in the kernel-global workqueue if it was not already
3030 * queued and leaves it in the same position on the kernel-global
3031 * workqueue otherwise.
3033 bool schedule_work(struct work_struct
*work
)
3035 return queue_work(system_wq
, work
);
3037 EXPORT_SYMBOL(schedule_work
);
3040 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3042 * @dwork: job to be done
3043 * @delay: number of jiffies to wait
3045 * After waiting for a given time this puts a job in the kernel-global
3046 * workqueue on the specified CPU.
3048 bool schedule_delayed_work_on(int cpu
, struct delayed_work
*dwork
,
3049 unsigned long delay
)
3051 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
3053 EXPORT_SYMBOL(schedule_delayed_work_on
);
3056 * schedule_delayed_work - put work task in global workqueue after delay
3057 * @dwork: job to be done
3058 * @delay: number of jiffies to wait or 0 for immediate execution
3060 * After waiting for a given time this puts a job in the kernel-global
3063 bool schedule_delayed_work(struct delayed_work
*dwork
, unsigned long delay
)
3065 return queue_delayed_work(system_wq
, dwork
, delay
);
3067 EXPORT_SYMBOL(schedule_delayed_work
);
3070 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3071 * @func: the function to call
3073 * schedule_on_each_cpu() executes @func on each online CPU using the
3074 * system workqueue and blocks until all CPUs have completed.
3075 * schedule_on_each_cpu() is very slow.
3078 * 0 on success, -errno on failure.
3080 int schedule_on_each_cpu(work_func_t func
)
3083 struct work_struct __percpu
*works
;
3085 works
= alloc_percpu(struct work_struct
);
3091 for_each_online_cpu(cpu
) {
3092 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3094 INIT_WORK(work
, func
);
3095 schedule_work_on(cpu
, work
);
3098 for_each_online_cpu(cpu
)
3099 flush_work(per_cpu_ptr(works
, cpu
));
3107 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3109 * Forces execution of the kernel-global workqueue and blocks until its
3112 * Think twice before calling this function! It's very easy to get into
3113 * trouble if you don't take great care. Either of the following situations
3114 * will lead to deadlock:
3116 * One of the work items currently on the workqueue needs to acquire
3117 * a lock held by your code or its caller.
3119 * Your code is running in the context of a work routine.
3121 * They will be detected by lockdep when they occur, but the first might not
3122 * occur very often. It depends on what work items are on the workqueue and
3123 * what locks they need, which you have no control over.
3125 * In most situations flushing the entire workqueue is overkill; you merely
3126 * need to know that a particular work item isn't queued and isn't running.
3127 * In such cases you should use cancel_delayed_work_sync() or
3128 * cancel_work_sync() instead.
3130 void flush_scheduled_work(void)
3132 flush_workqueue(system_wq
);
3134 EXPORT_SYMBOL(flush_scheduled_work
);
3137 * execute_in_process_context - reliably execute the routine with user context
3138 * @fn: the function to execute
3139 * @ew: guaranteed storage for the execute work structure (must
3140 * be available when the work executes)
3142 * Executes the function immediately if process context is available,
3143 * otherwise schedules the function for delayed execution.
3145 * Returns: 0 - function was executed
3146 * 1 - function was scheduled for execution
3148 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3150 if (!in_interrupt()) {
3155 INIT_WORK(&ew
->work
, fn
);
3156 schedule_work(&ew
->work
);
3160 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3162 int keventd_up(void)
3164 return system_wq
!= NULL
;
3167 static int alloc_cwqs(struct workqueue_struct
*wq
)
3170 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3171 * Make sure that the alignment isn't lower than that of
3172 * unsigned long long.
3174 const size_t size
= sizeof(struct cpu_workqueue_struct
);
3175 const size_t align
= max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS
,
3176 __alignof__(unsigned long long));
3178 if (!(wq
->flags
& WQ_UNBOUND
))
3179 wq
->cpu_wq
.pcpu
= __alloc_percpu(size
, align
);
3184 * Allocate enough room to align cwq and put an extra
3185 * pointer at the end pointing back to the originally
3186 * allocated pointer which will be used for free.
3188 ptr
= kzalloc(size
+ align
+ sizeof(void *), GFP_KERNEL
);
3190 wq
->cpu_wq
.single
= PTR_ALIGN(ptr
, align
);
3191 *(void **)(wq
->cpu_wq
.single
+ 1) = ptr
;
3195 /* just in case, make sure it's actually aligned */
3196 BUG_ON(!IS_ALIGNED(wq
->cpu_wq
.v
, align
));
3197 return wq
->cpu_wq
.v
? 0 : -ENOMEM
;
3200 static void free_cwqs(struct workqueue_struct
*wq
)
3202 if (!(wq
->flags
& WQ_UNBOUND
))
3203 free_percpu(wq
->cpu_wq
.pcpu
);
3204 else if (wq
->cpu_wq
.single
) {
3205 /* the pointer to free is stored right after the cwq */
3206 kfree(*(void **)(wq
->cpu_wq
.single
+ 1));
3210 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3213 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3215 if (max_active
< 1 || max_active
> lim
)
3216 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3217 max_active
, name
, 1, lim
);
3219 return clamp_val(max_active
, 1, lim
);
3222 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3225 struct lock_class_key
*key
,
3226 const char *lock_name
, ...)
3228 va_list args
, args1
;
3229 struct workqueue_struct
*wq
;
3233 /* determine namelen, allocate wq and format name */
3234 va_start(args
, lock_name
);
3235 va_copy(args1
, args
);
3236 namelen
= vsnprintf(NULL
, 0, fmt
, args
) + 1;
3238 wq
= kzalloc(sizeof(*wq
) + namelen
, GFP_KERNEL
);
3242 vsnprintf(wq
->name
, namelen
, fmt
, args1
);
3247 * Workqueues which may be used during memory reclaim should
3248 * have a rescuer to guarantee forward progress.
3250 if (flags
& WQ_MEM_RECLAIM
)
3251 flags
|= WQ_RESCUER
;
3253 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3254 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3258 wq
->saved_max_active
= max_active
;
3259 mutex_init(&wq
->flush_mutex
);
3260 atomic_set(&wq
->nr_cwqs_to_flush
, 0);
3261 INIT_LIST_HEAD(&wq
->flusher_queue
);
3262 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3264 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3265 INIT_LIST_HEAD(&wq
->list
);
3267 if (alloc_cwqs(wq
) < 0)
3270 for_each_cwq_cpu(cpu
, wq
) {
3271 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3272 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3273 int pool_idx
= (bool)(flags
& WQ_HIGHPRI
);
3275 BUG_ON((unsigned long)cwq
& WORK_STRUCT_FLAG_MASK
);
3276 cwq
->pool
= &gcwq
->pools
[pool_idx
];
3278 cwq
->flush_color
= -1;
3279 cwq
->max_active
= max_active
;
3280 INIT_LIST_HEAD(&cwq
->delayed_works
);
3283 if (flags
& WQ_RESCUER
) {
3284 struct worker
*rescuer
;
3286 if (!alloc_mayday_mask(&wq
->mayday_mask
, GFP_KERNEL
))
3289 wq
->rescuer
= rescuer
= alloc_worker();
3293 rescuer
->rescue_wq
= wq
;
3294 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
3296 if (IS_ERR(rescuer
->task
))
3299 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
3300 wake_up_process(rescuer
->task
);
3304 * workqueue_lock protects global freeze state and workqueues
3305 * list. Grab it, set max_active accordingly and add the new
3306 * workqueue to workqueues list.
3308 spin_lock(&workqueue_lock
);
3310 if (workqueue_freezing
&& wq
->flags
& WQ_FREEZABLE
)
3311 for_each_cwq_cpu(cpu
, wq
)
3312 get_cwq(cpu
, wq
)->max_active
= 0;
3314 list_add(&wq
->list
, &workqueues
);
3316 spin_unlock(&workqueue_lock
);
3322 free_mayday_mask(wq
->mayday_mask
);
3328 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3331 * destroy_workqueue - safely terminate a workqueue
3332 * @wq: target workqueue
3334 * Safely destroy a workqueue. All work currently pending will be done first.
3336 void destroy_workqueue(struct workqueue_struct
*wq
)
3340 /* drain it before proceeding with destruction */
3341 drain_workqueue(wq
);
3344 * wq list is used to freeze wq, remove from list after
3345 * flushing is complete in case freeze races us.
3347 spin_lock(&workqueue_lock
);
3348 list_del(&wq
->list
);
3349 spin_unlock(&workqueue_lock
);
3352 for_each_cwq_cpu(cpu
, wq
) {
3353 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3356 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
3357 BUG_ON(cwq
->nr_in_flight
[i
]);
3358 BUG_ON(cwq
->nr_active
);
3359 BUG_ON(!list_empty(&cwq
->delayed_works
));
3362 if (wq
->flags
& WQ_RESCUER
) {
3363 kthread_stop(wq
->rescuer
->task
);
3364 free_mayday_mask(wq
->mayday_mask
);
3371 EXPORT_SYMBOL_GPL(destroy_workqueue
);
3374 * cwq_set_max_active - adjust max_active of a cwq
3375 * @cwq: target cpu_workqueue_struct
3376 * @max_active: new max_active value.
3378 * Set @cwq->max_active to @max_active and activate delayed works if
3382 * spin_lock_irq(pool->lock).
3384 static void cwq_set_max_active(struct cpu_workqueue_struct
*cwq
, int max_active
)
3386 cwq
->max_active
= max_active
;
3388 while (!list_empty(&cwq
->delayed_works
) &&
3389 cwq
->nr_active
< cwq
->max_active
)
3390 cwq_activate_first_delayed(cwq
);
3394 * workqueue_set_max_active - adjust max_active of a workqueue
3395 * @wq: target workqueue
3396 * @max_active: new max_active value.
3398 * Set max_active of @wq to @max_active.
3401 * Don't call from IRQ context.
3403 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
3407 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
3409 spin_lock(&workqueue_lock
);
3411 wq
->saved_max_active
= max_active
;
3413 for_each_cwq_cpu(cpu
, wq
) {
3414 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3415 struct worker_pool
*pool
= cwq
->pool
;
3417 spin_lock_irq(&pool
->lock
);
3419 if (!(wq
->flags
& WQ_FREEZABLE
) ||
3420 !(pool
->flags
& POOL_FREEZING
))
3421 cwq_set_max_active(cwq
, max_active
);
3423 spin_unlock_irq(&pool
->lock
);
3426 spin_unlock(&workqueue_lock
);
3428 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
3431 * workqueue_congested - test whether a workqueue is congested
3432 * @cpu: CPU in question
3433 * @wq: target workqueue
3435 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3436 * no synchronization around this function and the test result is
3437 * unreliable and only useful as advisory hints or for debugging.
3440 * %true if congested, %false otherwise.
3442 bool workqueue_congested(unsigned int cpu
, struct workqueue_struct
*wq
)
3444 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3446 return !list_empty(&cwq
->delayed_works
);
3448 EXPORT_SYMBOL_GPL(workqueue_congested
);
3451 * work_busy - test whether a work is currently pending or running
3452 * @work: the work to be tested
3454 * Test whether @work is currently pending or running. There is no
3455 * synchronization around this function and the test result is
3456 * unreliable and only useful as advisory hints or for debugging.
3457 * Especially for reentrant wqs, the pending state might hide the
3461 * OR'd bitmask of WORK_BUSY_* bits.
3463 unsigned int work_busy(struct work_struct
*work
)
3465 struct worker_pool
*pool
= get_work_pool(work
);
3466 unsigned long flags
;
3467 unsigned int ret
= 0;
3472 spin_lock_irqsave(&pool
->lock
, flags
);
3474 if (work_pending(work
))
3475 ret
|= WORK_BUSY_PENDING
;
3476 if (find_worker_executing_work(pool
, work
))
3477 ret
|= WORK_BUSY_RUNNING
;
3479 spin_unlock_irqrestore(&pool
->lock
, flags
);
3483 EXPORT_SYMBOL_GPL(work_busy
);
3488 * There are two challenges in supporting CPU hotplug. Firstly, there
3489 * are a lot of assumptions on strong associations among work, cwq and
3490 * gcwq which make migrating pending and scheduled works very
3491 * difficult to implement without impacting hot paths. Secondly,
3492 * worker pools serve mix of short, long and very long running works making
3493 * blocked draining impractical.
3495 * This is solved by allowing the pools to be disassociated from the CPU
3496 * running as an unbound one and allowing it to be reattached later if the
3497 * cpu comes back online.
3500 static void gcwq_unbind_fn(struct work_struct
*work
)
3502 int cpu
= smp_processor_id();
3503 struct worker_pool
*pool
;
3504 struct worker
*worker
;
3505 struct hlist_node
*pos
;
3508 for_each_std_worker_pool(pool
, cpu
) {
3509 BUG_ON(cpu
!= smp_processor_id());
3511 mutex_lock(&pool
->assoc_mutex
);
3512 spin_lock_irq(&pool
->lock
);
3515 * We've claimed all manager positions. Make all workers
3516 * unbound and set DISASSOCIATED. Before this, all workers
3517 * except for the ones which are still executing works from
3518 * before the last CPU down must be on the cpu. After
3519 * this, they may become diasporas.
3521 list_for_each_entry(worker
, &pool
->idle_list
, entry
)
3522 worker
->flags
|= WORKER_UNBOUND
;
3524 for_each_busy_worker(worker
, i
, pos
, pool
)
3525 worker
->flags
|= WORKER_UNBOUND
;
3527 pool
->flags
|= POOL_DISASSOCIATED
;
3529 spin_unlock_irq(&pool
->lock
);
3530 mutex_unlock(&pool
->assoc_mutex
);
3534 * Call schedule() so that we cross rq->lock and thus can guarantee
3535 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3536 * as scheduler callbacks may be invoked from other cpus.
3541 * Sched callbacks are disabled now. Zap nr_running. After this,
3542 * nr_running stays zero and need_more_worker() and keep_working()
3543 * are always true as long as the worklist is not empty. Pools on
3544 * @cpu now behave as unbound (in terms of concurrency management)
3545 * pools which are served by workers tied to the CPU.
3547 * On return from this function, the current worker would trigger
3548 * unbound chain execution of pending work items if other workers
3551 for_each_std_worker_pool(pool
, cpu
)
3552 atomic_set(get_pool_nr_running(pool
), 0);
3556 * Workqueues should be brought up before normal priority CPU notifiers.
3557 * This will be registered high priority CPU notifier.
3559 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
3560 unsigned long action
,
3563 unsigned int cpu
= (unsigned long)hcpu
;
3564 struct worker_pool
*pool
;
3566 switch (action
& ~CPU_TASKS_FROZEN
) {
3567 case CPU_UP_PREPARE
:
3568 for_each_std_worker_pool(pool
, cpu
) {
3569 struct worker
*worker
;
3571 if (pool
->nr_workers
)
3574 worker
= create_worker(pool
);
3578 spin_lock_irq(&pool
->lock
);
3579 start_worker(worker
);
3580 spin_unlock_irq(&pool
->lock
);
3584 case CPU_DOWN_FAILED
:
3586 for_each_std_worker_pool(pool
, cpu
) {
3587 mutex_lock(&pool
->assoc_mutex
);
3588 spin_lock_irq(&pool
->lock
);
3590 pool
->flags
&= ~POOL_DISASSOCIATED
;
3591 rebind_workers(pool
);
3593 spin_unlock_irq(&pool
->lock
);
3594 mutex_unlock(&pool
->assoc_mutex
);
3602 * Workqueues should be brought down after normal priority CPU notifiers.
3603 * This will be registered as low priority CPU notifier.
3605 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
3606 unsigned long action
,
3609 unsigned int cpu
= (unsigned long)hcpu
;
3610 struct work_struct unbind_work
;
3612 switch (action
& ~CPU_TASKS_FROZEN
) {
3613 case CPU_DOWN_PREPARE
:
3614 /* unbinding should happen on the local CPU */
3615 INIT_WORK_ONSTACK(&unbind_work
, gcwq_unbind_fn
);
3616 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
3617 flush_work(&unbind_work
);
3625 struct work_for_cpu
{
3626 struct work_struct work
;
3632 static void work_for_cpu_fn(struct work_struct
*work
)
3634 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
3636 wfc
->ret
= wfc
->fn(wfc
->arg
);
3640 * work_on_cpu - run a function in user context on a particular cpu
3641 * @cpu: the cpu to run on
3642 * @fn: the function to run
3643 * @arg: the function arg
3645 * This will return the value @fn returns.
3646 * It is up to the caller to ensure that the cpu doesn't go offline.
3647 * The caller must not hold any locks which would prevent @fn from completing.
3649 long work_on_cpu(unsigned int cpu
, long (*fn
)(void *), void *arg
)
3651 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
3653 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
3654 schedule_work_on(cpu
, &wfc
.work
);
3655 flush_work(&wfc
.work
);
3658 EXPORT_SYMBOL_GPL(work_on_cpu
);
3659 #endif /* CONFIG_SMP */
3661 #ifdef CONFIG_FREEZER
3664 * freeze_workqueues_begin - begin freezing workqueues
3666 * Start freezing workqueues. After this function returns, all freezable
3667 * workqueues will queue new works to their frozen_works list instead of
3671 * Grabs and releases workqueue_lock and pool->lock's.
3673 void freeze_workqueues_begin(void)
3677 spin_lock(&workqueue_lock
);
3679 BUG_ON(workqueue_freezing
);
3680 workqueue_freezing
= true;
3682 for_each_gcwq_cpu(cpu
) {
3683 struct worker_pool
*pool
;
3684 struct workqueue_struct
*wq
;
3686 for_each_std_worker_pool(pool
, cpu
) {
3687 spin_lock_irq(&pool
->lock
);
3689 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
3690 pool
->flags
|= POOL_FREEZING
;
3692 list_for_each_entry(wq
, &workqueues
, list
) {
3693 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3695 if (cwq
&& cwq
->pool
== pool
&&
3696 (wq
->flags
& WQ_FREEZABLE
))
3697 cwq
->max_active
= 0;
3700 spin_unlock_irq(&pool
->lock
);
3704 spin_unlock(&workqueue_lock
);
3708 * freeze_workqueues_busy - are freezable workqueues still busy?
3710 * Check whether freezing is complete. This function must be called
3711 * between freeze_workqueues_begin() and thaw_workqueues().
3714 * Grabs and releases workqueue_lock.
3717 * %true if some freezable workqueues are still busy. %false if freezing
3720 bool freeze_workqueues_busy(void)
3725 spin_lock(&workqueue_lock
);
3727 BUG_ON(!workqueue_freezing
);
3729 for_each_gcwq_cpu(cpu
) {
3730 struct workqueue_struct
*wq
;
3732 * nr_active is monotonically decreasing. It's safe
3733 * to peek without lock.
3735 list_for_each_entry(wq
, &workqueues
, list
) {
3736 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3738 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3741 BUG_ON(cwq
->nr_active
< 0);
3742 if (cwq
->nr_active
) {
3749 spin_unlock(&workqueue_lock
);
3754 * thaw_workqueues - thaw workqueues
3756 * Thaw workqueues. Normal queueing is restored and all collected
3757 * frozen works are transferred to their respective gcwq worklists.
3760 * Grabs and releases workqueue_lock and pool->lock's.
3762 void thaw_workqueues(void)
3766 spin_lock(&workqueue_lock
);
3768 if (!workqueue_freezing
)
3771 for_each_gcwq_cpu(cpu
) {
3772 struct worker_pool
*pool
;
3773 struct workqueue_struct
*wq
;
3775 for_each_std_worker_pool(pool
, cpu
) {
3776 spin_lock_irq(&pool
->lock
);
3778 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
3779 pool
->flags
&= ~POOL_FREEZING
;
3781 list_for_each_entry(wq
, &workqueues
, list
) {
3782 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3784 if (!cwq
|| cwq
->pool
!= pool
||
3785 !(wq
->flags
& WQ_FREEZABLE
))
3788 /* restore max_active and repopulate worklist */
3789 cwq_set_max_active(cwq
, wq
->saved_max_active
);
3792 wake_up_worker(pool
);
3794 spin_unlock_irq(&pool
->lock
);
3798 workqueue_freezing
= false;
3800 spin_unlock(&workqueue_lock
);
3802 #endif /* CONFIG_FREEZER */
3804 static int __init
init_workqueues(void)
3808 /* make sure we have enough bits for OFFQ pool ID */
3809 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_POOL_SHIFT
)) <
3810 WORK_CPU_LAST
* NR_STD_WORKER_POOLS
);
3812 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
3813 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
3815 /* initialize gcwqs */
3816 for_each_gcwq_cpu(cpu
) {
3817 struct worker_pool
*pool
;
3819 for_each_std_worker_pool(pool
, cpu
) {
3820 spin_lock_init(&pool
->lock
);
3822 pool
->flags
|= POOL_DISASSOCIATED
;
3823 INIT_LIST_HEAD(&pool
->worklist
);
3824 INIT_LIST_HEAD(&pool
->idle_list
);
3825 hash_init(pool
->busy_hash
);
3827 init_timer_deferrable(&pool
->idle_timer
);
3828 pool
->idle_timer
.function
= idle_worker_timeout
;
3829 pool
->idle_timer
.data
= (unsigned long)pool
;
3831 setup_timer(&pool
->mayday_timer
, gcwq_mayday_timeout
,
3832 (unsigned long)pool
);
3834 mutex_init(&pool
->assoc_mutex
);
3835 ida_init(&pool
->worker_ida
);
3838 BUG_ON(worker_pool_assign_id(pool
));
3842 /* create the initial worker */
3843 for_each_online_gcwq_cpu(cpu
) {
3844 struct worker_pool
*pool
;
3846 for_each_std_worker_pool(pool
, cpu
) {
3847 struct worker
*worker
;
3849 if (cpu
!= WORK_CPU_UNBOUND
)
3850 pool
->flags
&= ~POOL_DISASSOCIATED
;
3852 worker
= create_worker(pool
);
3854 spin_lock_irq(&pool
->lock
);
3855 start_worker(worker
);
3856 spin_unlock_irq(&pool
->lock
);
3860 system_wq
= alloc_workqueue("events", 0, 0);
3861 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
3862 system_long_wq
= alloc_workqueue("events_long", 0, 0);
3863 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
3864 WQ_UNBOUND_MAX_ACTIVE
);
3865 system_freezable_wq
= alloc_workqueue("events_freezable",
3867 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
3868 !system_unbound_wq
|| !system_freezable_wq
);
3871 early_initcall(init_workqueues
);