2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/hashtable.h>
46 #include "workqueue_internal.h"
52 * A bound pool is either associated or disassociated with its CPU.
53 * While associated (!DISASSOCIATED), all workers are bound to the
54 * CPU and none has %WORKER_UNBOUND set and concurrency management
57 * While DISASSOCIATED, the cpu may be offline and all workers have
58 * %WORKER_UNBOUND set and concurrency management disabled, and may
59 * be executing on any CPU. The pool behaves as an unbound one.
61 * Note that DISASSOCIATED can be flipped only while holding
62 * assoc_mutex to avoid changing binding state while
63 * create_worker() is in progress.
65 POOL_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
66 POOL_MANAGING_WORKERS
= 1 << 1, /* managing workers */
67 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
68 POOL_FREEZING
= 1 << 3, /* freeze in progress */
71 WORKER_STARTED
= 1 << 0, /* started */
72 WORKER_DIE
= 1 << 1, /* die die die */
73 WORKER_IDLE
= 1 << 2, /* is idle */
74 WORKER_PREP
= 1 << 3, /* preparing to run works */
75 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
76 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
78 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_UNBOUND
|
81 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
83 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
85 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
86 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
88 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
89 /* call for help after 10ms
91 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
92 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
95 * Rescue workers are used only on emergencies and shared by
98 RESCUER_NICE_LEVEL
= -20,
99 HIGHPRI_NICE_LEVEL
= -20,
103 * Structure fields follow one of the following exclusion rules.
105 * I: Modifiable by initialization/destruction paths and read-only for
108 * P: Preemption protected. Disabling preemption is enough and should
109 * only be modified and accessed from the local cpu.
111 * L: gcwq->lock protected. Access with gcwq->lock held.
113 * X: During normal operation, modification requires gcwq->lock and
114 * should be done only from local cpu. Either disabling preemption
115 * on local cpu or grabbing gcwq->lock is enough for read access.
116 * If POOL_DISASSOCIATED is set, it's identical to L.
118 * F: wq->flush_mutex protected.
120 * W: workqueue_lock protected.
123 /* struct worker is defined in workqueue_internal.h */
126 struct global_cwq
*gcwq
; /* I: the owning gcwq */
127 int id
; /* I: pool ID */
128 unsigned int flags
; /* X: flags */
130 struct list_head worklist
; /* L: list of pending works */
131 int nr_workers
; /* L: total number of workers */
133 /* nr_idle includes the ones off idle_list for rebinding */
134 int nr_idle
; /* L: currently idle ones */
136 struct list_head idle_list
; /* X: list of idle workers */
137 struct timer_list idle_timer
; /* L: worker idle timeout */
138 struct timer_list mayday_timer
; /* L: SOS timer for workers */
140 struct mutex assoc_mutex
; /* protect POOL_DISASSOCIATED */
141 struct ida worker_ida
; /* L: for worker IDs */
145 * Global per-cpu workqueue. There's one and only one for each cpu
146 * and all works are queued and processed here regardless of their
150 spinlock_t lock
; /* the gcwq lock */
151 unsigned int cpu
; /* I: the associated cpu */
153 /* workers are chained either in busy_hash or pool idle_list */
154 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
155 /* L: hash of busy workers */
157 struct worker_pool pools
[NR_STD_WORKER_POOLS
];
158 /* normal and highpri pools */
159 } ____cacheline_aligned_in_smp
;
162 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
163 * work_struct->data are used for flags and thus cwqs need to be
164 * aligned at two's power of the number of flag bits.
166 struct cpu_workqueue_struct
{
167 struct worker_pool
*pool
; /* I: the associated pool */
168 struct workqueue_struct
*wq
; /* I: the owning workqueue */
169 int work_color
; /* L: current color */
170 int flush_color
; /* L: flushing color */
171 int nr_in_flight
[WORK_NR_COLORS
];
172 /* L: nr of in_flight works */
173 int nr_active
; /* L: nr of active works */
174 int max_active
; /* L: max active works */
175 struct list_head delayed_works
; /* L: delayed works */
179 * Structure used to wait for workqueue flush.
182 struct list_head list
; /* F: list of flushers */
183 int flush_color
; /* F: flush color waiting for */
184 struct completion done
; /* flush completion */
188 * All cpumasks are assumed to be always set on UP and thus can't be
189 * used to determine whether there's something to be done.
192 typedef cpumask_var_t mayday_mask_t
;
193 #define mayday_test_and_set_cpu(cpu, mask) \
194 cpumask_test_and_set_cpu((cpu), (mask))
195 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
196 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
197 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
198 #define free_mayday_mask(mask) free_cpumask_var((mask))
200 typedef unsigned long mayday_mask_t
;
201 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
202 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
203 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
204 #define alloc_mayday_mask(maskp, gfp) true
205 #define free_mayday_mask(mask) do { } while (0)
209 * The externally visible workqueue abstraction is an array of
210 * per-CPU workqueues:
212 struct workqueue_struct
{
213 unsigned int flags
; /* W: WQ_* flags */
215 struct cpu_workqueue_struct __percpu
*pcpu
;
216 struct cpu_workqueue_struct
*single
;
218 } cpu_wq
; /* I: cwq's */
219 struct list_head list
; /* W: list of all workqueues */
221 struct mutex flush_mutex
; /* protects wq flushing */
222 int work_color
; /* F: current work color */
223 int flush_color
; /* F: current flush color */
224 atomic_t nr_cwqs_to_flush
; /* flush in progress */
225 struct wq_flusher
*first_flusher
; /* F: first flusher */
226 struct list_head flusher_queue
; /* F: flush waiters */
227 struct list_head flusher_overflow
; /* F: flush overflow list */
229 mayday_mask_t mayday_mask
; /* cpus requesting rescue */
230 struct worker
*rescuer
; /* I: rescue worker */
232 int nr_drainers
; /* W: drain in progress */
233 int saved_max_active
; /* W: saved cwq max_active */
234 #ifdef CONFIG_LOCKDEP
235 struct lockdep_map lockdep_map
;
237 char name
[]; /* I: workqueue name */
240 struct workqueue_struct
*system_wq __read_mostly
;
241 EXPORT_SYMBOL_GPL(system_wq
);
242 struct workqueue_struct
*system_highpri_wq __read_mostly
;
243 EXPORT_SYMBOL_GPL(system_highpri_wq
);
244 struct workqueue_struct
*system_long_wq __read_mostly
;
245 EXPORT_SYMBOL_GPL(system_long_wq
);
246 struct workqueue_struct
*system_unbound_wq __read_mostly
;
247 EXPORT_SYMBOL_GPL(system_unbound_wq
);
248 struct workqueue_struct
*system_freezable_wq __read_mostly
;
249 EXPORT_SYMBOL_GPL(system_freezable_wq
);
251 #define CREATE_TRACE_POINTS
252 #include <trace/events/workqueue.h>
254 #define for_each_worker_pool(pool, gcwq) \
255 for ((pool) = &(gcwq)->pools[0]; \
256 (pool) < &(gcwq)->pools[NR_STD_WORKER_POOLS]; (pool)++)
258 #define for_each_busy_worker(worker, i, pos, gcwq) \
259 hash_for_each(gcwq->busy_hash, i, pos, worker, hentry)
261 static inline int __next_gcwq_cpu(int cpu
, const struct cpumask
*mask
,
264 if (cpu
< nr_cpu_ids
) {
266 cpu
= cpumask_next(cpu
, mask
);
267 if (cpu
< nr_cpu_ids
)
271 return WORK_CPU_UNBOUND
;
273 return WORK_CPU_NONE
;
276 static inline int __next_wq_cpu(int cpu
, const struct cpumask
*mask
,
277 struct workqueue_struct
*wq
)
279 return __next_gcwq_cpu(cpu
, mask
, !(wq
->flags
& WQ_UNBOUND
) ? 1 : 2);
285 * An extra gcwq is defined for an invalid cpu number
286 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
287 * specific CPU. The following iterators are similar to
288 * for_each_*_cpu() iterators but also considers the unbound gcwq.
290 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
291 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
292 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
293 * WORK_CPU_UNBOUND for unbound workqueues
295 #define for_each_gcwq_cpu(cpu) \
296 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
297 (cpu) < WORK_CPU_NONE; \
298 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
300 #define for_each_online_gcwq_cpu(cpu) \
301 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
302 (cpu) < WORK_CPU_NONE; \
303 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
305 #define for_each_cwq_cpu(cpu, wq) \
306 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
307 (cpu) < WORK_CPU_NONE; \
308 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
310 #ifdef CONFIG_DEBUG_OBJECTS_WORK
312 static struct debug_obj_descr work_debug_descr
;
314 static void *work_debug_hint(void *addr
)
316 return ((struct work_struct
*) addr
)->func
;
320 * fixup_init is called when:
321 * - an active object is initialized
323 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
325 struct work_struct
*work
= addr
;
328 case ODEBUG_STATE_ACTIVE
:
329 cancel_work_sync(work
);
330 debug_object_init(work
, &work_debug_descr
);
338 * fixup_activate is called when:
339 * - an active object is activated
340 * - an unknown object is activated (might be a statically initialized object)
342 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
344 struct work_struct
*work
= addr
;
348 case ODEBUG_STATE_NOTAVAILABLE
:
350 * This is not really a fixup. The work struct was
351 * statically initialized. We just make sure that it
352 * is tracked in the object tracker.
354 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
355 debug_object_init(work
, &work_debug_descr
);
356 debug_object_activate(work
, &work_debug_descr
);
362 case ODEBUG_STATE_ACTIVE
:
371 * fixup_free is called when:
372 * - an active object is freed
374 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
376 struct work_struct
*work
= addr
;
379 case ODEBUG_STATE_ACTIVE
:
380 cancel_work_sync(work
);
381 debug_object_free(work
, &work_debug_descr
);
388 static struct debug_obj_descr work_debug_descr
= {
389 .name
= "work_struct",
390 .debug_hint
= work_debug_hint
,
391 .fixup_init
= work_fixup_init
,
392 .fixup_activate
= work_fixup_activate
,
393 .fixup_free
= work_fixup_free
,
396 static inline void debug_work_activate(struct work_struct
*work
)
398 debug_object_activate(work
, &work_debug_descr
);
401 static inline void debug_work_deactivate(struct work_struct
*work
)
403 debug_object_deactivate(work
, &work_debug_descr
);
406 void __init_work(struct work_struct
*work
, int onstack
)
409 debug_object_init_on_stack(work
, &work_debug_descr
);
411 debug_object_init(work
, &work_debug_descr
);
413 EXPORT_SYMBOL_GPL(__init_work
);
415 void destroy_work_on_stack(struct work_struct
*work
)
417 debug_object_free(work
, &work_debug_descr
);
419 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
422 static inline void debug_work_activate(struct work_struct
*work
) { }
423 static inline void debug_work_deactivate(struct work_struct
*work
) { }
426 /* Serializes the accesses to the list of workqueues. */
427 static DEFINE_SPINLOCK(workqueue_lock
);
428 static LIST_HEAD(workqueues
);
429 static bool workqueue_freezing
; /* W: have wqs started freezing? */
432 * The almighty global cpu workqueues. nr_running is the only field
433 * which is expected to be used frequently by other cpus via
434 * try_to_wake_up(). Put it in a separate cacheline.
436 static DEFINE_PER_CPU(struct global_cwq
, global_cwq
);
437 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t
, pool_nr_running
[NR_STD_WORKER_POOLS
]);
440 * Global cpu workqueue and nr_running counter for unbound gcwq. The pools
441 * for online CPUs have POOL_DISASSOCIATED set, and all their workers have
442 * WORKER_UNBOUND set.
444 static struct global_cwq unbound_global_cwq
;
445 static atomic_t unbound_pool_nr_running
[NR_STD_WORKER_POOLS
] = {
446 [0 ... NR_STD_WORKER_POOLS
- 1] = ATOMIC_INIT(0), /* always 0 */
449 /* idr of all pools */
450 static DEFINE_MUTEX(worker_pool_idr_mutex
);
451 static DEFINE_IDR(worker_pool_idr
);
453 static int worker_thread(void *__worker
);
455 static int std_worker_pool_pri(struct worker_pool
*pool
)
457 return pool
- pool
->gcwq
->pools
;
460 static struct global_cwq
*get_gcwq(unsigned int cpu
)
462 if (cpu
!= WORK_CPU_UNBOUND
)
463 return &per_cpu(global_cwq
, cpu
);
465 return &unbound_global_cwq
;
468 /* allocate ID and assign it to @pool */
469 static int worker_pool_assign_id(struct worker_pool
*pool
)
473 mutex_lock(&worker_pool_idr_mutex
);
474 idr_pre_get(&worker_pool_idr
, GFP_KERNEL
);
475 ret
= idr_get_new(&worker_pool_idr
, pool
, &pool
->id
);
476 mutex_unlock(&worker_pool_idr_mutex
);
482 * Lookup worker_pool by id. The idr currently is built during boot and
483 * never modified. Don't worry about locking for now.
485 static struct worker_pool
*worker_pool_by_id(int pool_id
)
487 return idr_find(&worker_pool_idr
, pool_id
);
490 static atomic_t
*get_pool_nr_running(struct worker_pool
*pool
)
492 int cpu
= pool
->gcwq
->cpu
;
493 int idx
= std_worker_pool_pri(pool
);
495 if (cpu
!= WORK_CPU_UNBOUND
)
496 return &per_cpu(pool_nr_running
, cpu
)[idx
];
498 return &unbound_pool_nr_running
[idx
];
501 static struct cpu_workqueue_struct
*get_cwq(unsigned int cpu
,
502 struct workqueue_struct
*wq
)
504 if (!(wq
->flags
& WQ_UNBOUND
)) {
505 if (likely(cpu
< nr_cpu_ids
))
506 return per_cpu_ptr(wq
->cpu_wq
.pcpu
, cpu
);
507 } else if (likely(cpu
== WORK_CPU_UNBOUND
))
508 return wq
->cpu_wq
.single
;
512 static unsigned int work_color_to_flags(int color
)
514 return color
<< WORK_STRUCT_COLOR_SHIFT
;
517 static int get_work_color(struct work_struct
*work
)
519 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
520 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
523 static int work_next_color(int color
)
525 return (color
+ 1) % WORK_NR_COLORS
;
529 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
530 * contain the pointer to the queued cwq. Once execution starts, the flag
531 * is cleared and the high bits contain OFFQ flags and pool ID.
533 * set_work_cwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
534 * and clear_work_data() can be used to set the cwq, pool or clear
535 * work->data. These functions should only be called while the work is
536 * owned - ie. while the PENDING bit is set.
538 * get_work_pool() and get_work_cwq() can be used to obtain the pool or cwq
539 * corresponding to a work. Pool is available once the work has been
540 * queued anywhere after initialization until it is sync canceled. cwq is
541 * available only while the work item is queued.
543 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
544 * canceled. While being canceled, a work item may have its PENDING set
545 * but stay off timer and worklist for arbitrarily long and nobody should
546 * try to steal the PENDING bit.
548 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
551 BUG_ON(!work_pending(work
));
552 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
555 static void set_work_cwq(struct work_struct
*work
,
556 struct cpu_workqueue_struct
*cwq
,
557 unsigned long extra_flags
)
559 set_work_data(work
, (unsigned long)cwq
,
560 WORK_STRUCT_PENDING
| WORK_STRUCT_CWQ
| extra_flags
);
563 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
567 * The following wmb is paired with the implied mb in
568 * test_and_set_bit(PENDING) and ensures all updates to @work made
569 * here are visible to and precede any updates by the next PENDING
573 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
576 static void clear_work_data(struct work_struct
*work
)
578 smp_wmb(); /* see set_work_pool_and_clear_pending() */
579 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
582 static struct cpu_workqueue_struct
*get_work_cwq(struct work_struct
*work
)
584 unsigned long data
= atomic_long_read(&work
->data
);
586 if (data
& WORK_STRUCT_CWQ
)
587 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
593 * get_work_pool - return the worker_pool a given work was associated with
594 * @work: the work item of interest
596 * Return the worker_pool @work was last associated with. %NULL if none.
598 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
600 unsigned long data
= atomic_long_read(&work
->data
);
601 struct worker_pool
*pool
;
604 if (data
& WORK_STRUCT_CWQ
)
605 return ((struct cpu_workqueue_struct
*)
606 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
608 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
609 if (pool_id
== WORK_OFFQ_POOL_NONE
)
612 pool
= worker_pool_by_id(pool_id
);
618 * get_work_pool_id - return the worker pool ID a given work is associated with
619 * @work: the work item of interest
621 * Return the worker_pool ID @work was last associated with.
622 * %WORK_OFFQ_POOL_NONE if none.
624 static int get_work_pool_id(struct work_struct
*work
)
626 struct worker_pool
*pool
= get_work_pool(work
);
628 return pool
? pool
->id
: WORK_OFFQ_POOL_NONE
;
631 static struct global_cwq
*get_work_gcwq(struct work_struct
*work
)
633 struct worker_pool
*pool
= get_work_pool(work
);
635 return pool
? pool
->gcwq
: NULL
;
638 static void mark_work_canceling(struct work_struct
*work
)
640 unsigned long pool_id
= get_work_pool_id(work
);
642 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
643 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
646 static bool work_is_canceling(struct work_struct
*work
)
648 unsigned long data
= atomic_long_read(&work
->data
);
650 return !(data
& WORK_STRUCT_CWQ
) && (data
& WORK_OFFQ_CANCELING
);
654 * Policy functions. These define the policies on how the global worker
655 * pools are managed. Unless noted otherwise, these functions assume that
656 * they're being called with gcwq->lock held.
659 static bool __need_more_worker(struct worker_pool
*pool
)
661 return !atomic_read(get_pool_nr_running(pool
));
665 * Need to wake up a worker? Called from anything but currently
668 * Note that, because unbound workers never contribute to nr_running, this
669 * function will always return %true for unbound gcwq as long as the
670 * worklist isn't empty.
672 static bool need_more_worker(struct worker_pool
*pool
)
674 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
677 /* Can I start working? Called from busy but !running workers. */
678 static bool may_start_working(struct worker_pool
*pool
)
680 return pool
->nr_idle
;
683 /* Do I need to keep working? Called from currently running workers. */
684 static bool keep_working(struct worker_pool
*pool
)
686 atomic_t
*nr_running
= get_pool_nr_running(pool
);
688 return !list_empty(&pool
->worklist
) && atomic_read(nr_running
) <= 1;
691 /* Do we need a new worker? Called from manager. */
692 static bool need_to_create_worker(struct worker_pool
*pool
)
694 return need_more_worker(pool
) && !may_start_working(pool
);
697 /* Do I need to be the manager? */
698 static bool need_to_manage_workers(struct worker_pool
*pool
)
700 return need_to_create_worker(pool
) ||
701 (pool
->flags
& POOL_MANAGE_WORKERS
);
704 /* Do we have too many workers and should some go away? */
705 static bool too_many_workers(struct worker_pool
*pool
)
707 bool managing
= pool
->flags
& POOL_MANAGING_WORKERS
;
708 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
709 int nr_busy
= pool
->nr_workers
- nr_idle
;
712 * nr_idle and idle_list may disagree if idle rebinding is in
713 * progress. Never return %true if idle_list is empty.
715 if (list_empty(&pool
->idle_list
))
718 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
725 /* Return the first worker. Safe with preemption disabled */
726 static struct worker
*first_worker(struct worker_pool
*pool
)
728 if (unlikely(list_empty(&pool
->idle_list
)))
731 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
735 * wake_up_worker - wake up an idle worker
736 * @pool: worker pool to wake worker from
738 * Wake up the first idle worker of @pool.
741 * spin_lock_irq(gcwq->lock).
743 static void wake_up_worker(struct worker_pool
*pool
)
745 struct worker
*worker
= first_worker(pool
);
748 wake_up_process(worker
->task
);
752 * wq_worker_waking_up - a worker is waking up
753 * @task: task waking up
754 * @cpu: CPU @task is waking up to
756 * This function is called during try_to_wake_up() when a worker is
760 * spin_lock_irq(rq->lock)
762 void wq_worker_waking_up(struct task_struct
*task
, unsigned int cpu
)
764 struct worker
*worker
= kthread_data(task
);
766 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
767 WARN_ON_ONCE(worker
->pool
->gcwq
->cpu
!= cpu
);
768 atomic_inc(get_pool_nr_running(worker
->pool
));
773 * wq_worker_sleeping - a worker is going to sleep
774 * @task: task going to sleep
775 * @cpu: CPU in question, must be the current CPU number
777 * This function is called during schedule() when a busy worker is
778 * going to sleep. Worker on the same cpu can be woken up by
779 * returning pointer to its task.
782 * spin_lock_irq(rq->lock)
785 * Worker task on @cpu to wake up, %NULL if none.
787 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
,
790 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
791 struct worker_pool
*pool
;
792 atomic_t
*nr_running
;
795 * Rescuers, which may not have all the fields set up like normal
796 * workers, also reach here, let's not access anything before
797 * checking NOT_RUNNING.
799 if (worker
->flags
& WORKER_NOT_RUNNING
)
803 nr_running
= get_pool_nr_running(pool
);
805 /* this can only happen on the local cpu */
806 BUG_ON(cpu
!= raw_smp_processor_id());
809 * The counterpart of the following dec_and_test, implied mb,
810 * worklist not empty test sequence is in insert_work().
811 * Please read comment there.
813 * NOT_RUNNING is clear. This means that we're bound to and
814 * running on the local cpu w/ rq lock held and preemption
815 * disabled, which in turn means that none else could be
816 * manipulating idle_list, so dereferencing idle_list without gcwq
819 if (atomic_dec_and_test(nr_running
) && !list_empty(&pool
->worklist
))
820 to_wakeup
= first_worker(pool
);
821 return to_wakeup
? to_wakeup
->task
: NULL
;
825 * worker_set_flags - set worker flags and adjust nr_running accordingly
827 * @flags: flags to set
828 * @wakeup: wakeup an idle worker if necessary
830 * Set @flags in @worker->flags and adjust nr_running accordingly. If
831 * nr_running becomes zero and @wakeup is %true, an idle worker is
835 * spin_lock_irq(gcwq->lock)
837 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
840 struct worker_pool
*pool
= worker
->pool
;
842 WARN_ON_ONCE(worker
->task
!= current
);
845 * If transitioning into NOT_RUNNING, adjust nr_running and
846 * wake up an idle worker as necessary if requested by
849 if ((flags
& WORKER_NOT_RUNNING
) &&
850 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
851 atomic_t
*nr_running
= get_pool_nr_running(pool
);
854 if (atomic_dec_and_test(nr_running
) &&
855 !list_empty(&pool
->worklist
))
856 wake_up_worker(pool
);
858 atomic_dec(nr_running
);
861 worker
->flags
|= flags
;
865 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
867 * @flags: flags to clear
869 * Clear @flags in @worker->flags and adjust nr_running accordingly.
872 * spin_lock_irq(gcwq->lock)
874 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
876 struct worker_pool
*pool
= worker
->pool
;
877 unsigned int oflags
= worker
->flags
;
879 WARN_ON_ONCE(worker
->task
!= current
);
881 worker
->flags
&= ~flags
;
884 * If transitioning out of NOT_RUNNING, increment nr_running. Note
885 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
886 * of multiple flags, not a single flag.
888 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
889 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
890 atomic_inc(get_pool_nr_running(pool
));
894 * find_worker_executing_work - find worker which is executing a work
895 * @gcwq: gcwq of interest
896 * @work: work to find worker for
898 * Find a worker which is executing @work on @gcwq by searching
899 * @gcwq->busy_hash which is keyed by the address of @work. For a worker
900 * to match, its current execution should match the address of @work and
901 * its work function. This is to avoid unwanted dependency between
902 * unrelated work executions through a work item being recycled while still
905 * This is a bit tricky. A work item may be freed once its execution
906 * starts and nothing prevents the freed area from being recycled for
907 * another work item. If the same work item address ends up being reused
908 * before the original execution finishes, workqueue will identify the
909 * recycled work item as currently executing and make it wait until the
910 * current execution finishes, introducing an unwanted dependency.
912 * This function checks the work item address, work function and workqueue
913 * to avoid false positives. Note that this isn't complete as one may
914 * construct a work function which can introduce dependency onto itself
915 * through a recycled work item. Well, if somebody wants to shoot oneself
916 * in the foot that badly, there's only so much we can do, and if such
917 * deadlock actually occurs, it should be easy to locate the culprit work
921 * spin_lock_irq(gcwq->lock).
924 * Pointer to worker which is executing @work if found, NULL
927 static struct worker
*find_worker_executing_work(struct global_cwq
*gcwq
,
928 struct work_struct
*work
)
930 struct worker
*worker
;
931 struct hlist_node
*tmp
;
933 hash_for_each_possible(gcwq
->busy_hash
, worker
, tmp
, hentry
,
935 if (worker
->current_work
== work
&&
936 worker
->current_func
== work
->func
)
943 * move_linked_works - move linked works to a list
944 * @work: start of series of works to be scheduled
945 * @head: target list to append @work to
946 * @nextp: out paramter for nested worklist walking
948 * Schedule linked works starting from @work to @head. Work series to
949 * be scheduled starts at @work and includes any consecutive work with
950 * WORK_STRUCT_LINKED set in its predecessor.
952 * If @nextp is not NULL, it's updated to point to the next work of
953 * the last scheduled work. This allows move_linked_works() to be
954 * nested inside outer list_for_each_entry_safe().
957 * spin_lock_irq(gcwq->lock).
959 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
960 struct work_struct
**nextp
)
962 struct work_struct
*n
;
965 * Linked worklist will always end before the end of the list,
966 * use NULL for list head.
968 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
969 list_move_tail(&work
->entry
, head
);
970 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
975 * If we're already inside safe list traversal and have moved
976 * multiple works to the scheduled queue, the next position
977 * needs to be updated.
983 static void cwq_activate_delayed_work(struct work_struct
*work
)
985 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
987 trace_workqueue_activate_work(work
);
988 move_linked_works(work
, &cwq
->pool
->worklist
, NULL
);
989 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
993 static void cwq_activate_first_delayed(struct cpu_workqueue_struct
*cwq
)
995 struct work_struct
*work
= list_first_entry(&cwq
->delayed_works
,
996 struct work_struct
, entry
);
998 cwq_activate_delayed_work(work
);
1002 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1003 * @cwq: cwq of interest
1004 * @color: color of work which left the queue
1006 * A work either has completed or is removed from pending queue,
1007 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1010 * spin_lock_irq(gcwq->lock).
1012 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct
*cwq
, int color
)
1014 /* ignore uncolored works */
1015 if (color
== WORK_NO_COLOR
)
1018 cwq
->nr_in_flight
[color
]--;
1021 if (!list_empty(&cwq
->delayed_works
)) {
1022 /* one down, submit a delayed one */
1023 if (cwq
->nr_active
< cwq
->max_active
)
1024 cwq_activate_first_delayed(cwq
);
1027 /* is flush in progress and are we at the flushing tip? */
1028 if (likely(cwq
->flush_color
!= color
))
1031 /* are there still in-flight works? */
1032 if (cwq
->nr_in_flight
[color
])
1035 /* this cwq is done, clear flush_color */
1036 cwq
->flush_color
= -1;
1039 * If this was the last cwq, wake up the first flusher. It
1040 * will handle the rest.
1042 if (atomic_dec_and_test(&cwq
->wq
->nr_cwqs_to_flush
))
1043 complete(&cwq
->wq
->first_flusher
->done
);
1047 * try_to_grab_pending - steal work item from worklist and disable irq
1048 * @work: work item to steal
1049 * @is_dwork: @work is a delayed_work
1050 * @flags: place to store irq state
1052 * Try to grab PENDING bit of @work. This function can handle @work in any
1053 * stable state - idle, on timer or on worklist. Return values are
1055 * 1 if @work was pending and we successfully stole PENDING
1056 * 0 if @work was idle and we claimed PENDING
1057 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1058 * -ENOENT if someone else is canceling @work, this state may persist
1059 * for arbitrarily long
1061 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1062 * interrupted while holding PENDING and @work off queue, irq must be
1063 * disabled on entry. This, combined with delayed_work->timer being
1064 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1066 * On successful return, >= 0, irq is disabled and the caller is
1067 * responsible for releasing it using local_irq_restore(*@flags).
1069 * This function is safe to call from any context including IRQ handler.
1071 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1072 unsigned long *flags
)
1074 struct global_cwq
*gcwq
;
1076 local_irq_save(*flags
);
1078 /* try to steal the timer if it exists */
1080 struct delayed_work
*dwork
= to_delayed_work(work
);
1083 * dwork->timer is irqsafe. If del_timer() fails, it's
1084 * guaranteed that the timer is not queued anywhere and not
1085 * running on the local CPU.
1087 if (likely(del_timer(&dwork
->timer
)))
1091 /* try to claim PENDING the normal way */
1092 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1096 * The queueing is in progress, or it is already queued. Try to
1097 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1099 gcwq
= get_work_gcwq(work
);
1103 spin_lock(&gcwq
->lock
);
1104 if (!list_empty(&work
->entry
)) {
1106 * This work is queued, but perhaps we locked the wrong gcwq.
1107 * In that case we must see the new value after rmb(), see
1108 * insert_work()->wmb().
1111 if (gcwq
== get_work_gcwq(work
)) {
1112 debug_work_deactivate(work
);
1115 * A delayed work item cannot be grabbed directly
1116 * because it might have linked NO_COLOR work items
1117 * which, if left on the delayed_list, will confuse
1118 * cwq->nr_active management later on and cause
1119 * stall. Make sure the work item is activated
1122 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1123 cwq_activate_delayed_work(work
);
1125 list_del_init(&work
->entry
);
1126 cwq_dec_nr_in_flight(get_work_cwq(work
),
1127 get_work_color(work
));
1129 spin_unlock(&gcwq
->lock
);
1133 spin_unlock(&gcwq
->lock
);
1135 local_irq_restore(*flags
);
1136 if (work_is_canceling(work
))
1143 * insert_work - insert a work into gcwq
1144 * @cwq: cwq @work belongs to
1145 * @work: work to insert
1146 * @head: insertion point
1147 * @extra_flags: extra WORK_STRUCT_* flags to set
1149 * Insert @work which belongs to @cwq into @gcwq after @head.
1150 * @extra_flags is or'd to work_struct flags.
1153 * spin_lock_irq(gcwq->lock).
1155 static void insert_work(struct cpu_workqueue_struct
*cwq
,
1156 struct work_struct
*work
, struct list_head
*head
,
1157 unsigned int extra_flags
)
1159 struct worker_pool
*pool
= cwq
->pool
;
1161 /* we own @work, set data and link */
1162 set_work_cwq(work
, cwq
, extra_flags
);
1165 * Ensure that we get the right work->data if we see the
1166 * result of list_add() below, see try_to_grab_pending().
1170 list_add_tail(&work
->entry
, head
);
1173 * Ensure either worker_sched_deactivated() sees the above
1174 * list_add_tail() or we see zero nr_running to avoid workers
1175 * lying around lazily while there are works to be processed.
1179 if (__need_more_worker(pool
))
1180 wake_up_worker(pool
);
1184 * Test whether @work is being queued from another work executing on the
1185 * same workqueue. This is rather expensive and should only be used from
1188 static bool is_chained_work(struct workqueue_struct
*wq
)
1190 unsigned long flags
;
1193 for_each_gcwq_cpu(cpu
) {
1194 struct global_cwq
*gcwq
= get_gcwq(cpu
);
1195 struct worker
*worker
;
1196 struct hlist_node
*pos
;
1199 spin_lock_irqsave(&gcwq
->lock
, flags
);
1200 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
1201 if (worker
->task
!= current
)
1203 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1205 * I'm @worker, no locking necessary. See if @work
1206 * is headed to the same workqueue.
1208 return worker
->current_cwq
->wq
== wq
;
1210 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1215 static void __queue_work(unsigned int cpu
, struct workqueue_struct
*wq
,
1216 struct work_struct
*work
)
1218 struct global_cwq
*gcwq
;
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 gcwq to use */
1240 if (!(wq
->flags
& WQ_UNBOUND
)) {
1241 struct global_cwq
*last_gcwq
;
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 gcwq
= get_gcwq(cpu
);
1253 last_gcwq
= get_work_gcwq(work
);
1255 if (last_gcwq
&& last_gcwq
!= gcwq
) {
1256 struct worker
*worker
;
1258 spin_lock(&last_gcwq
->lock
);
1260 worker
= find_worker_executing_work(last_gcwq
, work
);
1262 if (worker
&& worker
->current_cwq
->wq
== wq
)
1265 /* meh... not running there, queue here */
1266 spin_unlock(&last_gcwq
->lock
);
1267 spin_lock(&gcwq
->lock
);
1270 spin_lock(&gcwq
->lock
);
1273 gcwq
= get_gcwq(WORK_CPU_UNBOUND
);
1274 spin_lock(&gcwq
->lock
);
1277 /* gcwq determined, get cwq and queue */
1278 cwq
= get_cwq(gcwq
->cpu
, wq
);
1279 trace_workqueue_queue_work(req_cpu
, cwq
, work
);
1281 if (WARN_ON(!list_empty(&work
->entry
))) {
1282 spin_unlock(&gcwq
->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(&gcwq
->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 gcwq is preserved to allow reentrance detection for
1388 if (!(wq
->flags
& WQ_UNBOUND
)) {
1389 struct global_cwq
*gcwq
= get_work_gcwq(work
);
1392 * If we cannot get the last gcwq 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(gcwq->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 * gcwq->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(gcwq->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 gcwq->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(&gcwq
->lock
)
1607 struct worker_pool
*pool
= worker
->pool
;
1608 struct global_cwq
*gcwq
= pool
->gcwq
;
1609 struct task_struct
*task
= worker
->task
;
1613 * The following call may fail, succeed or succeed
1614 * without actually migrating the task to the cpu if
1615 * it races with cpu hotunplug operation. Verify
1616 * against POOL_DISASSOCIATED.
1618 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1619 set_cpus_allowed_ptr(task
, get_cpu_mask(gcwq
->cpu
));
1621 spin_lock_irq(&gcwq
->lock
);
1622 if (pool
->flags
& POOL_DISASSOCIATED
)
1624 if (task_cpu(task
) == gcwq
->cpu
&&
1625 cpumask_equal(¤t
->cpus_allowed
,
1626 get_cpu_mask(gcwq
->cpu
)))
1628 spin_unlock_irq(&gcwq
->lock
);
1631 * We've raced with CPU hot[un]plug. Give it a breather
1632 * and retry migration. cond_resched() is required here;
1633 * otherwise, we might deadlock against cpu_stop trying to
1634 * bring down the CPU on non-preemptive kernel.
1642 * Rebind an idle @worker to its CPU. worker_thread() will test
1643 * list_empty(@worker->entry) before leaving idle and call this function.
1645 static void idle_worker_rebind(struct worker
*worker
)
1647 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1649 /* CPU may go down again inbetween, clear UNBOUND only on success */
1650 if (worker_maybe_bind_and_lock(worker
))
1651 worker_clr_flags(worker
, WORKER_UNBOUND
);
1653 /* rebind complete, become available again */
1654 list_add(&worker
->entry
, &worker
->pool
->idle_list
);
1655 spin_unlock_irq(&gcwq
->lock
);
1659 * Function for @worker->rebind.work used to rebind unbound busy workers to
1660 * the associated cpu which is coming back online. This is scheduled by
1661 * cpu up but can race with other cpu hotplug operations and may be
1662 * executed twice without intervening cpu down.
1664 static void busy_worker_rebind_fn(struct work_struct
*work
)
1666 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1667 struct global_cwq
*gcwq
= worker
->pool
->gcwq
;
1669 if (worker_maybe_bind_and_lock(worker
))
1670 worker_clr_flags(worker
, WORKER_UNBOUND
);
1672 spin_unlock_irq(&gcwq
->lock
);
1676 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1677 * @gcwq: gcwq of interest
1679 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1680 * is different for idle and busy ones.
1682 * Idle ones will be removed from the idle_list and woken up. They will
1683 * add themselves back after completing rebind. This ensures that the
1684 * idle_list doesn't contain any unbound workers when re-bound busy workers
1685 * try to perform local wake-ups for concurrency management.
1687 * Busy workers can rebind after they finish their current work items.
1688 * Queueing the rebind work item at the head of the scheduled list is
1689 * enough. Note that nr_running will be properly bumped as busy workers
1692 * On return, all non-manager workers are scheduled for rebind - see
1693 * manage_workers() for the manager special case. Any idle worker
1694 * including the manager will not appear on @idle_list until rebind is
1695 * complete, making local wake-ups safe.
1697 static void rebind_workers(struct global_cwq
*gcwq
)
1699 struct worker_pool
*pool
;
1700 struct worker
*worker
, *n
;
1701 struct hlist_node
*pos
;
1704 lockdep_assert_held(&gcwq
->lock
);
1706 for_each_worker_pool(pool
, gcwq
)
1707 lockdep_assert_held(&pool
->assoc_mutex
);
1709 /* dequeue and kick idle ones */
1710 for_each_worker_pool(pool
, gcwq
) {
1711 list_for_each_entry_safe(worker
, n
, &pool
->idle_list
, entry
) {
1713 * idle workers should be off @pool->idle_list
1714 * until rebind is complete to avoid receiving
1715 * premature local wake-ups.
1717 list_del_init(&worker
->entry
);
1720 * worker_thread() will see the above dequeuing
1721 * and call idle_worker_rebind().
1723 wake_up_process(worker
->task
);
1727 /* rebind busy workers */
1728 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
1729 struct work_struct
*rebind_work
= &worker
->rebind_work
;
1730 struct workqueue_struct
*wq
;
1732 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
1733 work_data_bits(rebind_work
)))
1736 debug_work_activate(rebind_work
);
1739 * wq doesn't really matter but let's keep @worker->pool
1740 * and @cwq->pool consistent for sanity.
1742 if (std_worker_pool_pri(worker
->pool
))
1743 wq
= system_highpri_wq
;
1747 insert_work(get_cwq(gcwq
->cpu
, wq
), rebind_work
,
1748 worker
->scheduled
.next
,
1749 work_color_to_flags(WORK_NO_COLOR
));
1753 static struct worker
*alloc_worker(void)
1755 struct worker
*worker
;
1757 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1759 INIT_LIST_HEAD(&worker
->entry
);
1760 INIT_LIST_HEAD(&worker
->scheduled
);
1761 INIT_WORK(&worker
->rebind_work
, busy_worker_rebind_fn
);
1762 /* on creation a worker is in !idle && prep state */
1763 worker
->flags
= WORKER_PREP
;
1769 * create_worker - create a new workqueue worker
1770 * @pool: pool the new worker will belong to
1772 * Create a new worker which is bound to @pool. The returned worker
1773 * can be started by calling start_worker() or destroyed using
1777 * Might sleep. Does GFP_KERNEL allocations.
1780 * Pointer to the newly created worker.
1782 static struct worker
*create_worker(struct worker_pool
*pool
)
1784 struct global_cwq
*gcwq
= pool
->gcwq
;
1785 const char *pri
= std_worker_pool_pri(pool
) ? "H" : "";
1786 struct worker
*worker
= NULL
;
1789 spin_lock_irq(&gcwq
->lock
);
1790 while (ida_get_new(&pool
->worker_ida
, &id
)) {
1791 spin_unlock_irq(&gcwq
->lock
);
1792 if (!ida_pre_get(&pool
->worker_ida
, GFP_KERNEL
))
1794 spin_lock_irq(&gcwq
->lock
);
1796 spin_unlock_irq(&gcwq
->lock
);
1798 worker
= alloc_worker();
1802 worker
->pool
= pool
;
1805 if (gcwq
->cpu
!= WORK_CPU_UNBOUND
)
1806 worker
->task
= kthread_create_on_node(worker_thread
,
1807 worker
, cpu_to_node(gcwq
->cpu
),
1808 "kworker/%u:%d%s", gcwq
->cpu
, id
, pri
);
1810 worker
->task
= kthread_create(worker_thread
, worker
,
1811 "kworker/u:%d%s", id
, pri
);
1812 if (IS_ERR(worker
->task
))
1815 if (std_worker_pool_pri(pool
))
1816 set_user_nice(worker
->task
, HIGHPRI_NICE_LEVEL
);
1819 * Determine CPU binding of the new worker depending on
1820 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1821 * flag remains stable across this function. See the comments
1822 * above the flag definition for details.
1824 * As an unbound worker may later become a regular one if CPU comes
1825 * online, make sure every worker has %PF_THREAD_BOUND set.
1827 if (!(pool
->flags
& POOL_DISASSOCIATED
)) {
1828 kthread_bind(worker
->task
, gcwq
->cpu
);
1830 worker
->task
->flags
|= PF_THREAD_BOUND
;
1831 worker
->flags
|= WORKER_UNBOUND
;
1837 spin_lock_irq(&gcwq
->lock
);
1838 ida_remove(&pool
->worker_ida
, id
);
1839 spin_unlock_irq(&gcwq
->lock
);
1846 * start_worker - start a newly created worker
1847 * @worker: worker to start
1849 * Make the gcwq aware of @worker and start it.
1852 * spin_lock_irq(gcwq->lock).
1854 static void start_worker(struct worker
*worker
)
1856 worker
->flags
|= WORKER_STARTED
;
1857 worker
->pool
->nr_workers
++;
1858 worker_enter_idle(worker
);
1859 wake_up_process(worker
->task
);
1863 * destroy_worker - destroy a workqueue worker
1864 * @worker: worker to be destroyed
1866 * Destroy @worker and adjust @gcwq stats accordingly.
1869 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1871 static void destroy_worker(struct worker
*worker
)
1873 struct worker_pool
*pool
= worker
->pool
;
1874 struct global_cwq
*gcwq
= pool
->gcwq
;
1875 int id
= worker
->id
;
1877 /* sanity check frenzy */
1878 BUG_ON(worker
->current_work
);
1879 BUG_ON(!list_empty(&worker
->scheduled
));
1881 if (worker
->flags
& WORKER_STARTED
)
1883 if (worker
->flags
& WORKER_IDLE
)
1886 list_del_init(&worker
->entry
);
1887 worker
->flags
|= WORKER_DIE
;
1889 spin_unlock_irq(&gcwq
->lock
);
1891 kthread_stop(worker
->task
);
1894 spin_lock_irq(&gcwq
->lock
);
1895 ida_remove(&pool
->worker_ida
, id
);
1898 static void idle_worker_timeout(unsigned long __pool
)
1900 struct worker_pool
*pool
= (void *)__pool
;
1901 struct global_cwq
*gcwq
= pool
->gcwq
;
1903 spin_lock_irq(&gcwq
->lock
);
1905 if (too_many_workers(pool
)) {
1906 struct worker
*worker
;
1907 unsigned long expires
;
1909 /* idle_list is kept in LIFO order, check the last one */
1910 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1911 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1913 if (time_before(jiffies
, expires
))
1914 mod_timer(&pool
->idle_timer
, expires
);
1916 /* it's been idle for too long, wake up manager */
1917 pool
->flags
|= POOL_MANAGE_WORKERS
;
1918 wake_up_worker(pool
);
1922 spin_unlock_irq(&gcwq
->lock
);
1925 static bool send_mayday(struct work_struct
*work
)
1927 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1928 struct workqueue_struct
*wq
= cwq
->wq
;
1931 if (!(wq
->flags
& WQ_RESCUER
))
1934 /* mayday mayday mayday */
1935 cpu
= cwq
->pool
->gcwq
->cpu
;
1936 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1937 if (cpu
== WORK_CPU_UNBOUND
)
1939 if (!mayday_test_and_set_cpu(cpu
, wq
->mayday_mask
))
1940 wake_up_process(wq
->rescuer
->task
);
1944 static void gcwq_mayday_timeout(unsigned long __pool
)
1946 struct worker_pool
*pool
= (void *)__pool
;
1947 struct global_cwq
*gcwq
= pool
->gcwq
;
1948 struct work_struct
*work
;
1950 spin_lock_irq(&gcwq
->lock
);
1952 if (need_to_create_worker(pool
)) {
1954 * We've been trying to create a new worker but
1955 * haven't been successful. We might be hitting an
1956 * allocation deadlock. Send distress signals to
1959 list_for_each_entry(work
, &pool
->worklist
, entry
)
1963 spin_unlock_irq(&gcwq
->lock
);
1965 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1969 * maybe_create_worker - create a new worker if necessary
1970 * @pool: pool to create a new worker for
1972 * Create a new worker for @pool if necessary. @pool is guaranteed to
1973 * have at least one idle worker on return from this function. If
1974 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1975 * sent to all rescuers with works scheduled on @pool to resolve
1976 * possible allocation deadlock.
1978 * On return, need_to_create_worker() is guaranteed to be false and
1979 * may_start_working() true.
1982 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1983 * multiple times. Does GFP_KERNEL allocations. Called only from
1987 * false if no action was taken and gcwq->lock stayed locked, true
1990 static bool maybe_create_worker(struct worker_pool
*pool
)
1991 __releases(&gcwq
->lock
)
1992 __acquires(&gcwq
->lock
)
1994 struct global_cwq
*gcwq
= pool
->gcwq
;
1996 if (!need_to_create_worker(pool
))
1999 spin_unlock_irq(&gcwq
->lock
);
2001 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2002 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
2005 struct worker
*worker
;
2007 worker
= create_worker(pool
);
2009 del_timer_sync(&pool
->mayday_timer
);
2010 spin_lock_irq(&gcwq
->lock
);
2011 start_worker(worker
);
2012 BUG_ON(need_to_create_worker(pool
));
2016 if (!need_to_create_worker(pool
))
2019 __set_current_state(TASK_INTERRUPTIBLE
);
2020 schedule_timeout(CREATE_COOLDOWN
);
2022 if (!need_to_create_worker(pool
))
2026 del_timer_sync(&pool
->mayday_timer
);
2027 spin_lock_irq(&gcwq
->lock
);
2028 if (need_to_create_worker(pool
))
2034 * maybe_destroy_worker - destroy workers which have been idle for a while
2035 * @pool: pool to destroy workers for
2037 * Destroy @pool workers which have been idle for longer than
2038 * IDLE_WORKER_TIMEOUT.
2041 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2042 * multiple times. Called only from manager.
2045 * false if no action was taken and gcwq->lock stayed locked, true
2048 static bool maybe_destroy_workers(struct worker_pool
*pool
)
2052 while (too_many_workers(pool
)) {
2053 struct worker
*worker
;
2054 unsigned long expires
;
2056 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2057 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2059 if (time_before(jiffies
, expires
)) {
2060 mod_timer(&pool
->idle_timer
, expires
);
2064 destroy_worker(worker
);
2072 * manage_workers - manage worker pool
2075 * Assume the manager role and manage gcwq worker pool @worker belongs
2076 * to. At any given time, there can be only zero or one manager per
2077 * gcwq. The exclusion is handled automatically by this function.
2079 * The caller can safely start processing works on false return. On
2080 * true return, it's guaranteed that need_to_create_worker() is false
2081 * and may_start_working() is true.
2084 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2085 * multiple times. Does GFP_KERNEL allocations.
2088 * false if no action was taken and gcwq->lock stayed locked, true if
2089 * some action was taken.
2091 static bool manage_workers(struct worker
*worker
)
2093 struct worker_pool
*pool
= worker
->pool
;
2096 if (pool
->flags
& POOL_MANAGING_WORKERS
)
2099 pool
->flags
|= POOL_MANAGING_WORKERS
;
2102 * To simplify both worker management and CPU hotplug, hold off
2103 * management while hotplug is in progress. CPU hotplug path can't
2104 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2105 * lead to idle worker depletion (all become busy thinking someone
2106 * else is managing) which in turn can result in deadlock under
2107 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2108 * manager against CPU hotplug.
2110 * assoc_mutex would always be free unless CPU hotplug is in
2111 * progress. trylock first without dropping @gcwq->lock.
2113 if (unlikely(!mutex_trylock(&pool
->assoc_mutex
))) {
2114 spin_unlock_irq(&pool
->gcwq
->lock
);
2115 mutex_lock(&pool
->assoc_mutex
);
2117 * CPU hotplug could have happened while we were waiting
2118 * for assoc_mutex. Hotplug itself can't handle us
2119 * because manager isn't either on idle or busy list, and
2120 * @gcwq's state and ours could have deviated.
2122 * As hotplug is now excluded via assoc_mutex, we can
2123 * simply try to bind. It will succeed or fail depending
2124 * on @gcwq's current state. Try it and adjust
2125 * %WORKER_UNBOUND accordingly.
2127 if (worker_maybe_bind_and_lock(worker
))
2128 worker
->flags
&= ~WORKER_UNBOUND
;
2130 worker
->flags
|= WORKER_UNBOUND
;
2135 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2138 * Destroy and then create so that may_start_working() is true
2141 ret
|= maybe_destroy_workers(pool
);
2142 ret
|= maybe_create_worker(pool
);
2144 pool
->flags
&= ~POOL_MANAGING_WORKERS
;
2145 mutex_unlock(&pool
->assoc_mutex
);
2150 * process_one_work - process single work
2152 * @work: work to process
2154 * Process @work. This function contains all the logics necessary to
2155 * process a single work including synchronization against and
2156 * interaction with other workers on the same cpu, queueing and
2157 * flushing. As long as context requirement is met, any worker can
2158 * call this function to process a work.
2161 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2163 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2164 __releases(&gcwq
->lock
)
2165 __acquires(&gcwq
->lock
)
2167 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
2168 struct worker_pool
*pool
= worker
->pool
;
2169 struct global_cwq
*gcwq
= pool
->gcwq
;
2170 bool cpu_intensive
= cwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2172 struct worker
*collision
;
2173 #ifdef CONFIG_LOCKDEP
2175 * It is permissible to free the struct work_struct from
2176 * inside the function that is called from it, this we need to
2177 * take into account for lockdep too. To avoid bogus "held
2178 * lock freed" warnings as well as problems when looking into
2179 * work->lockdep_map, make a copy and use that here.
2181 struct lockdep_map lockdep_map
;
2183 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2186 * Ensure we're on the correct CPU. DISASSOCIATED test is
2187 * necessary to avoid spurious warnings from rescuers servicing the
2188 * unbound or a disassociated pool.
2190 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2191 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2192 raw_smp_processor_id() != gcwq
->cpu
);
2195 * A single work shouldn't be executed concurrently by
2196 * multiple workers on a single cpu. Check whether anyone is
2197 * already processing the work. If so, defer the work to the
2198 * currently executing one.
2200 collision
= find_worker_executing_work(gcwq
, work
);
2201 if (unlikely(collision
)) {
2202 move_linked_works(work
, &collision
->scheduled
, NULL
);
2206 /* claim and dequeue */
2207 debug_work_deactivate(work
);
2208 hash_add(gcwq
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2209 worker
->current_work
= work
;
2210 worker
->current_func
= work
->func
;
2211 worker
->current_cwq
= cwq
;
2212 work_color
= get_work_color(work
);
2214 list_del_init(&work
->entry
);
2217 * CPU intensive works don't participate in concurrency
2218 * management. They're the scheduler's responsibility.
2220 if (unlikely(cpu_intensive
))
2221 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2224 * Unbound gcwq isn't concurrency managed and work items should be
2225 * executed ASAP. Wake up another worker if necessary.
2227 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2228 wake_up_worker(pool
);
2231 * Record the last pool and clear PENDING which should be the last
2232 * update to @work. Also, do this inside @gcwq->lock so that
2233 * PENDING and queued state changes happen together while IRQ is
2236 set_work_pool_and_clear_pending(work
, pool
->id
);
2238 spin_unlock_irq(&gcwq
->lock
);
2240 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2241 lock_map_acquire(&lockdep_map
);
2242 trace_workqueue_execute_start(work
);
2243 worker
->current_func(work
);
2245 * While we must be careful to not use "work" after this, the trace
2246 * point will only record its address.
2248 trace_workqueue_execute_end(work
);
2249 lock_map_release(&lockdep_map
);
2250 lock_map_release(&cwq
->wq
->lockdep_map
);
2252 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2253 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2254 " last function: %pf\n",
2255 current
->comm
, preempt_count(), task_pid_nr(current
),
2256 worker
->current_func
);
2257 debug_show_held_locks(current
);
2261 spin_lock_irq(&gcwq
->lock
);
2263 /* clear cpu intensive status */
2264 if (unlikely(cpu_intensive
))
2265 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2267 /* we're done with it, release */
2268 hash_del(&worker
->hentry
);
2269 worker
->current_work
= NULL
;
2270 worker
->current_func
= NULL
;
2271 worker
->current_cwq
= NULL
;
2272 cwq_dec_nr_in_flight(cwq
, work_color
);
2276 * process_scheduled_works - process scheduled works
2279 * Process all scheduled works. Please note that the scheduled list
2280 * may change while processing a work, so this function repeatedly
2281 * fetches a work from the top and executes it.
2284 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2287 static void process_scheduled_works(struct worker
*worker
)
2289 while (!list_empty(&worker
->scheduled
)) {
2290 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2291 struct work_struct
, entry
);
2292 process_one_work(worker
, work
);
2297 * worker_thread - the worker thread function
2300 * The gcwq worker thread function. There's a single dynamic pool of
2301 * these per each cpu. These workers process all works regardless of
2302 * their specific target workqueue. The only exception is works which
2303 * belong to workqueues with a rescuer which will be explained in
2306 static int worker_thread(void *__worker
)
2308 struct worker
*worker
= __worker
;
2309 struct worker_pool
*pool
= worker
->pool
;
2310 struct global_cwq
*gcwq
= pool
->gcwq
;
2312 /* tell the scheduler that this is a workqueue worker */
2313 worker
->task
->flags
|= PF_WQ_WORKER
;
2315 spin_lock_irq(&gcwq
->lock
);
2317 /* we are off idle list if destruction or rebind is requested */
2318 if (unlikely(list_empty(&worker
->entry
))) {
2319 spin_unlock_irq(&gcwq
->lock
);
2321 /* if DIE is set, destruction is requested */
2322 if (worker
->flags
& WORKER_DIE
) {
2323 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2327 /* otherwise, rebind */
2328 idle_worker_rebind(worker
);
2332 worker_leave_idle(worker
);
2334 /* no more worker necessary? */
2335 if (!need_more_worker(pool
))
2338 /* do we need to manage? */
2339 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2343 * ->scheduled list can only be filled while a worker is
2344 * preparing to process a work or actually processing it.
2345 * Make sure nobody diddled with it while I was sleeping.
2347 BUG_ON(!list_empty(&worker
->scheduled
));
2350 * When control reaches this point, we're guaranteed to have
2351 * at least one idle worker or that someone else has already
2352 * assumed the manager role.
2354 worker_clr_flags(worker
, WORKER_PREP
);
2357 struct work_struct
*work
=
2358 list_first_entry(&pool
->worklist
,
2359 struct work_struct
, entry
);
2361 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2362 /* optimization path, not strictly necessary */
2363 process_one_work(worker
, work
);
2364 if (unlikely(!list_empty(&worker
->scheduled
)))
2365 process_scheduled_works(worker
);
2367 move_linked_works(work
, &worker
->scheduled
, NULL
);
2368 process_scheduled_works(worker
);
2370 } while (keep_working(pool
));
2372 worker_set_flags(worker
, WORKER_PREP
, false);
2374 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2378 * gcwq->lock is held and there's no work to process and no
2379 * need to manage, sleep. Workers are woken up only while
2380 * holding gcwq->lock or from local cpu, so setting the
2381 * current state before releasing gcwq->lock is enough to
2382 * prevent losing any event.
2384 worker_enter_idle(worker
);
2385 __set_current_state(TASK_INTERRUPTIBLE
);
2386 spin_unlock_irq(&gcwq
->lock
);
2392 * rescuer_thread - the rescuer thread function
2395 * Workqueue rescuer thread function. There's one rescuer for each
2396 * workqueue which has WQ_RESCUER set.
2398 * Regular work processing on a gcwq may block trying to create a new
2399 * worker which uses GFP_KERNEL allocation which has slight chance of
2400 * developing into deadlock if some works currently on the same queue
2401 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2402 * the problem rescuer solves.
2404 * When such condition is possible, the gcwq summons rescuers of all
2405 * workqueues which have works queued on the gcwq and let them process
2406 * those works so that forward progress can be guaranteed.
2408 * This should happen rarely.
2410 static int rescuer_thread(void *__rescuer
)
2412 struct worker
*rescuer
= __rescuer
;
2413 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2414 struct list_head
*scheduled
= &rescuer
->scheduled
;
2415 bool is_unbound
= wq
->flags
& WQ_UNBOUND
;
2418 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2421 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2422 * doesn't participate in concurrency management.
2424 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2426 set_current_state(TASK_INTERRUPTIBLE
);
2428 if (kthread_should_stop()) {
2429 __set_current_state(TASK_RUNNING
);
2430 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2435 * See whether any cpu is asking for help. Unbounded
2436 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2438 for_each_mayday_cpu(cpu
, wq
->mayday_mask
) {
2439 unsigned int tcpu
= is_unbound
? WORK_CPU_UNBOUND
: cpu
;
2440 struct cpu_workqueue_struct
*cwq
= get_cwq(tcpu
, wq
);
2441 struct worker_pool
*pool
= cwq
->pool
;
2442 struct global_cwq
*gcwq
= pool
->gcwq
;
2443 struct work_struct
*work
, *n
;
2445 __set_current_state(TASK_RUNNING
);
2446 mayday_clear_cpu(cpu
, wq
->mayday_mask
);
2448 /* migrate to the target cpu if possible */
2449 rescuer
->pool
= pool
;
2450 worker_maybe_bind_and_lock(rescuer
);
2453 * Slurp in all works issued via this workqueue and
2456 BUG_ON(!list_empty(&rescuer
->scheduled
));
2457 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2458 if (get_work_cwq(work
) == cwq
)
2459 move_linked_works(work
, scheduled
, &n
);
2461 process_scheduled_works(rescuer
);
2464 * Leave this gcwq. If keep_working() is %true, notify a
2465 * regular worker; otherwise, we end up with 0 concurrency
2466 * and stalling the execution.
2468 if (keep_working(pool
))
2469 wake_up_worker(pool
);
2471 spin_unlock_irq(&gcwq
->lock
);
2474 /* rescuers should never participate in concurrency management */
2475 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2481 struct work_struct work
;
2482 struct completion done
;
2485 static void wq_barrier_func(struct work_struct
*work
)
2487 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2488 complete(&barr
->done
);
2492 * insert_wq_barrier - insert a barrier work
2493 * @cwq: cwq to insert barrier into
2494 * @barr: wq_barrier to insert
2495 * @target: target work to attach @barr to
2496 * @worker: worker currently executing @target, NULL if @target is not executing
2498 * @barr is linked to @target such that @barr is completed only after
2499 * @target finishes execution. Please note that the ordering
2500 * guarantee is observed only with respect to @target and on the local
2503 * Currently, a queued barrier can't be canceled. This is because
2504 * try_to_grab_pending() can't determine whether the work to be
2505 * grabbed is at the head of the queue and thus can't clear LINKED
2506 * flag of the previous work while there must be a valid next work
2507 * after a work with LINKED flag set.
2509 * Note that when @worker is non-NULL, @target may be modified
2510 * underneath us, so we can't reliably determine cwq from @target.
2513 * spin_lock_irq(gcwq->lock).
2515 static void insert_wq_barrier(struct cpu_workqueue_struct
*cwq
,
2516 struct wq_barrier
*barr
,
2517 struct work_struct
*target
, struct worker
*worker
)
2519 struct list_head
*head
;
2520 unsigned int linked
= 0;
2523 * debugobject calls are safe here even with gcwq->lock locked
2524 * as we know for sure that this will not trigger any of the
2525 * checks and call back into the fixup functions where we
2528 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2529 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2530 init_completion(&barr
->done
);
2533 * If @target is currently being executed, schedule the
2534 * barrier to the worker; otherwise, put it after @target.
2537 head
= worker
->scheduled
.next
;
2539 unsigned long *bits
= work_data_bits(target
);
2541 head
= target
->entry
.next
;
2542 /* there can already be other linked works, inherit and set */
2543 linked
= *bits
& WORK_STRUCT_LINKED
;
2544 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2547 debug_work_activate(&barr
->work
);
2548 insert_work(cwq
, &barr
->work
, head
,
2549 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2553 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2554 * @wq: workqueue being flushed
2555 * @flush_color: new flush color, < 0 for no-op
2556 * @work_color: new work color, < 0 for no-op
2558 * Prepare cwqs for workqueue flushing.
2560 * If @flush_color is non-negative, flush_color on all cwqs should be
2561 * -1. If no cwq has in-flight commands at the specified color, all
2562 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2563 * has in flight commands, its cwq->flush_color is set to
2564 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2565 * wakeup logic is armed and %true is returned.
2567 * The caller should have initialized @wq->first_flusher prior to
2568 * calling this function with non-negative @flush_color. If
2569 * @flush_color is negative, no flush color update is done and %false
2572 * If @work_color is non-negative, all cwqs should have the same
2573 * work_color which is previous to @work_color and all will be
2574 * advanced to @work_color.
2577 * mutex_lock(wq->flush_mutex).
2580 * %true if @flush_color >= 0 and there's something to flush. %false
2583 static bool flush_workqueue_prep_cwqs(struct workqueue_struct
*wq
,
2584 int flush_color
, int work_color
)
2589 if (flush_color
>= 0) {
2590 BUG_ON(atomic_read(&wq
->nr_cwqs_to_flush
));
2591 atomic_set(&wq
->nr_cwqs_to_flush
, 1);
2594 for_each_cwq_cpu(cpu
, wq
) {
2595 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2596 struct global_cwq
*gcwq
= cwq
->pool
->gcwq
;
2598 spin_lock_irq(&gcwq
->lock
);
2600 if (flush_color
>= 0) {
2601 BUG_ON(cwq
->flush_color
!= -1);
2603 if (cwq
->nr_in_flight
[flush_color
]) {
2604 cwq
->flush_color
= flush_color
;
2605 atomic_inc(&wq
->nr_cwqs_to_flush
);
2610 if (work_color
>= 0) {
2611 BUG_ON(work_color
!= work_next_color(cwq
->work_color
));
2612 cwq
->work_color
= work_color
;
2615 spin_unlock_irq(&gcwq
->lock
);
2618 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_cwqs_to_flush
))
2619 complete(&wq
->first_flusher
->done
);
2625 * flush_workqueue - ensure that any scheduled work has run to completion.
2626 * @wq: workqueue to flush
2628 * Forces execution of the workqueue and blocks until its completion.
2629 * This is typically used in driver shutdown handlers.
2631 * We sleep until all works which were queued on entry have been handled,
2632 * but we are not livelocked by new incoming ones.
2634 void flush_workqueue(struct workqueue_struct
*wq
)
2636 struct wq_flusher this_flusher
= {
2637 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2639 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2643 lock_map_acquire(&wq
->lockdep_map
);
2644 lock_map_release(&wq
->lockdep_map
);
2646 mutex_lock(&wq
->flush_mutex
);
2649 * Start-to-wait phase
2651 next_color
= work_next_color(wq
->work_color
);
2653 if (next_color
!= wq
->flush_color
) {
2655 * Color space is not full. The current work_color
2656 * becomes our flush_color and work_color is advanced
2659 BUG_ON(!list_empty(&wq
->flusher_overflow
));
2660 this_flusher
.flush_color
= wq
->work_color
;
2661 wq
->work_color
= next_color
;
2663 if (!wq
->first_flusher
) {
2664 /* no flush in progress, become the first flusher */
2665 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2667 wq
->first_flusher
= &this_flusher
;
2669 if (!flush_workqueue_prep_cwqs(wq
, wq
->flush_color
,
2671 /* nothing to flush, done */
2672 wq
->flush_color
= next_color
;
2673 wq
->first_flusher
= NULL
;
2678 BUG_ON(wq
->flush_color
== this_flusher
.flush_color
);
2679 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2680 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2684 * Oops, color space is full, wait on overflow queue.
2685 * The next flush completion will assign us
2686 * flush_color and transfer to flusher_queue.
2688 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2691 mutex_unlock(&wq
->flush_mutex
);
2693 wait_for_completion(&this_flusher
.done
);
2696 * Wake-up-and-cascade phase
2698 * First flushers are responsible for cascading flushes and
2699 * handling overflow. Non-first flushers can simply return.
2701 if (wq
->first_flusher
!= &this_flusher
)
2704 mutex_lock(&wq
->flush_mutex
);
2706 /* we might have raced, check again with mutex held */
2707 if (wq
->first_flusher
!= &this_flusher
)
2710 wq
->first_flusher
= NULL
;
2712 BUG_ON(!list_empty(&this_flusher
.list
));
2713 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2716 struct wq_flusher
*next
, *tmp
;
2718 /* complete all the flushers sharing the current flush color */
2719 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2720 if (next
->flush_color
!= wq
->flush_color
)
2722 list_del_init(&next
->list
);
2723 complete(&next
->done
);
2726 BUG_ON(!list_empty(&wq
->flusher_overflow
) &&
2727 wq
->flush_color
!= work_next_color(wq
->work_color
));
2729 /* this flush_color is finished, advance by one */
2730 wq
->flush_color
= work_next_color(wq
->flush_color
);
2732 /* one color has been freed, handle overflow queue */
2733 if (!list_empty(&wq
->flusher_overflow
)) {
2735 * Assign the same color to all overflowed
2736 * flushers, advance work_color and append to
2737 * flusher_queue. This is the start-to-wait
2738 * phase for these overflowed flushers.
2740 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2741 tmp
->flush_color
= wq
->work_color
;
2743 wq
->work_color
= work_next_color(wq
->work_color
);
2745 list_splice_tail_init(&wq
->flusher_overflow
,
2746 &wq
->flusher_queue
);
2747 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2750 if (list_empty(&wq
->flusher_queue
)) {
2751 BUG_ON(wq
->flush_color
!= wq
->work_color
);
2756 * Need to flush more colors. Make the next flusher
2757 * the new first flusher and arm cwqs.
2759 BUG_ON(wq
->flush_color
== wq
->work_color
);
2760 BUG_ON(wq
->flush_color
!= next
->flush_color
);
2762 list_del_init(&next
->list
);
2763 wq
->first_flusher
= next
;
2765 if (flush_workqueue_prep_cwqs(wq
, wq
->flush_color
, -1))
2769 * Meh... this color is already done, clear first
2770 * flusher and repeat cascading.
2772 wq
->first_flusher
= NULL
;
2776 mutex_unlock(&wq
->flush_mutex
);
2778 EXPORT_SYMBOL_GPL(flush_workqueue
);
2781 * drain_workqueue - drain a workqueue
2782 * @wq: workqueue to drain
2784 * Wait until the workqueue becomes empty. While draining is in progress,
2785 * only chain queueing is allowed. IOW, only currently pending or running
2786 * work items on @wq can queue further work items on it. @wq is flushed
2787 * repeatedly until it becomes empty. The number of flushing is detemined
2788 * by the depth of chaining and should be relatively short. Whine if it
2791 void drain_workqueue(struct workqueue_struct
*wq
)
2793 unsigned int flush_cnt
= 0;
2797 * __queue_work() needs to test whether there are drainers, is much
2798 * hotter than drain_workqueue() and already looks at @wq->flags.
2799 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2801 spin_lock(&workqueue_lock
);
2802 if (!wq
->nr_drainers
++)
2803 wq
->flags
|= WQ_DRAINING
;
2804 spin_unlock(&workqueue_lock
);
2806 flush_workqueue(wq
);
2808 for_each_cwq_cpu(cpu
, wq
) {
2809 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2812 spin_lock_irq(&cwq
->pool
->gcwq
->lock
);
2813 drained
= !cwq
->nr_active
&& list_empty(&cwq
->delayed_works
);
2814 spin_unlock_irq(&cwq
->pool
->gcwq
->lock
);
2819 if (++flush_cnt
== 10 ||
2820 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2821 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2822 wq
->name
, flush_cnt
);
2826 spin_lock(&workqueue_lock
);
2827 if (!--wq
->nr_drainers
)
2828 wq
->flags
&= ~WQ_DRAINING
;
2829 spin_unlock(&workqueue_lock
);
2831 EXPORT_SYMBOL_GPL(drain_workqueue
);
2833 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2835 struct worker
*worker
= NULL
;
2836 struct global_cwq
*gcwq
;
2837 struct cpu_workqueue_struct
*cwq
;
2840 gcwq
= get_work_gcwq(work
);
2844 spin_lock_irq(&gcwq
->lock
);
2845 if (!list_empty(&work
->entry
)) {
2847 * See the comment near try_to_grab_pending()->smp_rmb().
2848 * If it was re-queued to a different gcwq under us, we
2849 * are not going to wait.
2852 cwq
= get_work_cwq(work
);
2853 if (unlikely(!cwq
|| gcwq
!= cwq
->pool
->gcwq
))
2856 worker
= find_worker_executing_work(gcwq
, work
);
2859 cwq
= worker
->current_cwq
;
2862 insert_wq_barrier(cwq
, barr
, work
, worker
);
2863 spin_unlock_irq(&gcwq
->lock
);
2866 * If @max_active is 1 or rescuer is in use, flushing another work
2867 * item on the same workqueue may lead to deadlock. Make sure the
2868 * flusher is not running on the same workqueue by verifying write
2871 if (cwq
->wq
->saved_max_active
== 1 || cwq
->wq
->flags
& WQ_RESCUER
)
2872 lock_map_acquire(&cwq
->wq
->lockdep_map
);
2874 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2875 lock_map_release(&cwq
->wq
->lockdep_map
);
2879 spin_unlock_irq(&gcwq
->lock
);
2884 * flush_work - wait for a work to finish executing the last queueing instance
2885 * @work: the work to flush
2887 * Wait until @work has finished execution. @work is guaranteed to be idle
2888 * on return if it hasn't been requeued since flush started.
2891 * %true if flush_work() waited for the work to finish execution,
2892 * %false if it was already idle.
2894 bool flush_work(struct work_struct
*work
)
2896 struct wq_barrier barr
;
2898 lock_map_acquire(&work
->lockdep_map
);
2899 lock_map_release(&work
->lockdep_map
);
2901 if (start_flush_work(work
, &barr
)) {
2902 wait_for_completion(&barr
.done
);
2903 destroy_work_on_stack(&barr
.work
);
2909 EXPORT_SYMBOL_GPL(flush_work
);
2911 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2913 unsigned long flags
;
2917 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2919 * If someone else is canceling, wait for the same event it
2920 * would be waiting for before retrying.
2922 if (unlikely(ret
== -ENOENT
))
2924 } while (unlikely(ret
< 0));
2926 /* tell other tasks trying to grab @work to back off */
2927 mark_work_canceling(work
);
2928 local_irq_restore(flags
);
2931 clear_work_data(work
);
2936 * cancel_work_sync - cancel a work and wait for it to finish
2937 * @work: the work to cancel
2939 * Cancel @work and wait for its execution to finish. This function
2940 * can be used even if the work re-queues itself or migrates to
2941 * another workqueue. On return from this function, @work is
2942 * guaranteed to be not pending or executing on any CPU.
2944 * cancel_work_sync(&delayed_work->work) must not be used for
2945 * delayed_work's. Use cancel_delayed_work_sync() instead.
2947 * The caller must ensure that the workqueue on which @work was last
2948 * queued can't be destroyed before this function returns.
2951 * %true if @work was pending, %false otherwise.
2953 bool cancel_work_sync(struct work_struct
*work
)
2955 return __cancel_work_timer(work
, false);
2957 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2960 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2961 * @dwork: the delayed work to flush
2963 * Delayed timer is cancelled and the pending work is queued for
2964 * immediate execution. Like flush_work(), this function only
2965 * considers the last queueing instance of @dwork.
2968 * %true if flush_work() waited for the work to finish execution,
2969 * %false if it was already idle.
2971 bool flush_delayed_work(struct delayed_work
*dwork
)
2973 local_irq_disable();
2974 if (del_timer_sync(&dwork
->timer
))
2975 __queue_work(dwork
->cpu
,
2976 get_work_cwq(&dwork
->work
)->wq
, &dwork
->work
);
2978 return flush_work(&dwork
->work
);
2980 EXPORT_SYMBOL(flush_delayed_work
);
2983 * cancel_delayed_work - cancel a delayed work
2984 * @dwork: delayed_work to cancel
2986 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2987 * and canceled; %false if wasn't pending. Note that the work callback
2988 * function may still be running on return, unless it returns %true and the
2989 * work doesn't re-arm itself. Explicitly flush or use
2990 * cancel_delayed_work_sync() to wait on it.
2992 * This function is safe to call from any context including IRQ handler.
2994 bool cancel_delayed_work(struct delayed_work
*dwork
)
2996 unsigned long flags
;
3000 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
3001 } while (unlikely(ret
== -EAGAIN
));
3003 if (unlikely(ret
< 0))
3006 set_work_pool_and_clear_pending(&dwork
->work
,
3007 get_work_pool_id(&dwork
->work
));
3008 local_irq_restore(flags
);
3011 EXPORT_SYMBOL(cancel_delayed_work
);
3014 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3015 * @dwork: the delayed work cancel
3017 * This is cancel_work_sync() for delayed works.
3020 * %true if @dwork was pending, %false otherwise.
3022 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
3024 return __cancel_work_timer(&dwork
->work
, true);
3026 EXPORT_SYMBOL(cancel_delayed_work_sync
);
3029 * schedule_work_on - put work task on a specific cpu
3030 * @cpu: cpu to put the work task on
3031 * @work: job to be done
3033 * This puts a job on a specific cpu
3035 bool schedule_work_on(int cpu
, struct work_struct
*work
)
3037 return queue_work_on(cpu
, system_wq
, work
);
3039 EXPORT_SYMBOL(schedule_work_on
);
3042 * schedule_work - put work task in global workqueue
3043 * @work: job to be done
3045 * Returns %false if @work was already on the kernel-global workqueue and
3048 * This puts a job in the kernel-global workqueue if it was not already
3049 * queued and leaves it in the same position on the kernel-global
3050 * workqueue otherwise.
3052 bool schedule_work(struct work_struct
*work
)
3054 return queue_work(system_wq
, work
);
3056 EXPORT_SYMBOL(schedule_work
);
3059 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3061 * @dwork: job to be done
3062 * @delay: number of jiffies to wait
3064 * After waiting for a given time this puts a job in the kernel-global
3065 * workqueue on the specified CPU.
3067 bool schedule_delayed_work_on(int cpu
, struct delayed_work
*dwork
,
3068 unsigned long delay
)
3070 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
3072 EXPORT_SYMBOL(schedule_delayed_work_on
);
3075 * schedule_delayed_work - put work task in global workqueue after delay
3076 * @dwork: job to be done
3077 * @delay: number of jiffies to wait or 0 for immediate execution
3079 * After waiting for a given time this puts a job in the kernel-global
3082 bool schedule_delayed_work(struct delayed_work
*dwork
, unsigned long delay
)
3084 return queue_delayed_work(system_wq
, dwork
, delay
);
3086 EXPORT_SYMBOL(schedule_delayed_work
);
3089 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3090 * @func: the function to call
3092 * schedule_on_each_cpu() executes @func on each online CPU using the
3093 * system workqueue and blocks until all CPUs have completed.
3094 * schedule_on_each_cpu() is very slow.
3097 * 0 on success, -errno on failure.
3099 int schedule_on_each_cpu(work_func_t func
)
3102 struct work_struct __percpu
*works
;
3104 works
= alloc_percpu(struct work_struct
);
3110 for_each_online_cpu(cpu
) {
3111 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3113 INIT_WORK(work
, func
);
3114 schedule_work_on(cpu
, work
);
3117 for_each_online_cpu(cpu
)
3118 flush_work(per_cpu_ptr(works
, cpu
));
3126 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3128 * Forces execution of the kernel-global workqueue and blocks until its
3131 * Think twice before calling this function! It's very easy to get into
3132 * trouble if you don't take great care. Either of the following situations
3133 * will lead to deadlock:
3135 * One of the work items currently on the workqueue needs to acquire
3136 * a lock held by your code or its caller.
3138 * Your code is running in the context of a work routine.
3140 * They will be detected by lockdep when they occur, but the first might not
3141 * occur very often. It depends on what work items are on the workqueue and
3142 * what locks they need, which you have no control over.
3144 * In most situations flushing the entire workqueue is overkill; you merely
3145 * need to know that a particular work item isn't queued and isn't running.
3146 * In such cases you should use cancel_delayed_work_sync() or
3147 * cancel_work_sync() instead.
3149 void flush_scheduled_work(void)
3151 flush_workqueue(system_wq
);
3153 EXPORT_SYMBOL(flush_scheduled_work
);
3156 * execute_in_process_context - reliably execute the routine with user context
3157 * @fn: the function to execute
3158 * @ew: guaranteed storage for the execute work structure (must
3159 * be available when the work executes)
3161 * Executes the function immediately if process context is available,
3162 * otherwise schedules the function for delayed execution.
3164 * Returns: 0 - function was executed
3165 * 1 - function was scheduled for execution
3167 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3169 if (!in_interrupt()) {
3174 INIT_WORK(&ew
->work
, fn
);
3175 schedule_work(&ew
->work
);
3179 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3181 int keventd_up(void)
3183 return system_wq
!= NULL
;
3186 static int alloc_cwqs(struct workqueue_struct
*wq
)
3189 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3190 * Make sure that the alignment isn't lower than that of
3191 * unsigned long long.
3193 const size_t size
= sizeof(struct cpu_workqueue_struct
);
3194 const size_t align
= max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS
,
3195 __alignof__(unsigned long long));
3197 if (!(wq
->flags
& WQ_UNBOUND
))
3198 wq
->cpu_wq
.pcpu
= __alloc_percpu(size
, align
);
3203 * Allocate enough room to align cwq and put an extra
3204 * pointer at the end pointing back to the originally
3205 * allocated pointer which will be used for free.
3207 ptr
= kzalloc(size
+ align
+ sizeof(void *), GFP_KERNEL
);
3209 wq
->cpu_wq
.single
= PTR_ALIGN(ptr
, align
);
3210 *(void **)(wq
->cpu_wq
.single
+ 1) = ptr
;
3214 /* just in case, make sure it's actually aligned */
3215 BUG_ON(!IS_ALIGNED(wq
->cpu_wq
.v
, align
));
3216 return wq
->cpu_wq
.v
? 0 : -ENOMEM
;
3219 static void free_cwqs(struct workqueue_struct
*wq
)
3221 if (!(wq
->flags
& WQ_UNBOUND
))
3222 free_percpu(wq
->cpu_wq
.pcpu
);
3223 else if (wq
->cpu_wq
.single
) {
3224 /* the pointer to free is stored right after the cwq */
3225 kfree(*(void **)(wq
->cpu_wq
.single
+ 1));
3229 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3232 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3234 if (max_active
< 1 || max_active
> lim
)
3235 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3236 max_active
, name
, 1, lim
);
3238 return clamp_val(max_active
, 1, lim
);
3241 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3244 struct lock_class_key
*key
,
3245 const char *lock_name
, ...)
3247 va_list args
, args1
;
3248 struct workqueue_struct
*wq
;
3252 /* determine namelen, allocate wq and format name */
3253 va_start(args
, lock_name
);
3254 va_copy(args1
, args
);
3255 namelen
= vsnprintf(NULL
, 0, fmt
, args
) + 1;
3257 wq
= kzalloc(sizeof(*wq
) + namelen
, GFP_KERNEL
);
3261 vsnprintf(wq
->name
, namelen
, fmt
, args1
);
3266 * Workqueues which may be used during memory reclaim should
3267 * have a rescuer to guarantee forward progress.
3269 if (flags
& WQ_MEM_RECLAIM
)
3270 flags
|= WQ_RESCUER
;
3272 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3273 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3277 wq
->saved_max_active
= max_active
;
3278 mutex_init(&wq
->flush_mutex
);
3279 atomic_set(&wq
->nr_cwqs_to_flush
, 0);
3280 INIT_LIST_HEAD(&wq
->flusher_queue
);
3281 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3283 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3284 INIT_LIST_HEAD(&wq
->list
);
3286 if (alloc_cwqs(wq
) < 0)
3289 for_each_cwq_cpu(cpu
, wq
) {
3290 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3291 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3292 int pool_idx
= (bool)(flags
& WQ_HIGHPRI
);
3294 BUG_ON((unsigned long)cwq
& WORK_STRUCT_FLAG_MASK
);
3295 cwq
->pool
= &gcwq
->pools
[pool_idx
];
3297 cwq
->flush_color
= -1;
3298 cwq
->max_active
= max_active
;
3299 INIT_LIST_HEAD(&cwq
->delayed_works
);
3302 if (flags
& WQ_RESCUER
) {
3303 struct worker
*rescuer
;
3305 if (!alloc_mayday_mask(&wq
->mayday_mask
, GFP_KERNEL
))
3308 wq
->rescuer
= rescuer
= alloc_worker();
3312 rescuer
->rescue_wq
= wq
;
3313 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
3315 if (IS_ERR(rescuer
->task
))
3318 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
3319 wake_up_process(rescuer
->task
);
3323 * workqueue_lock protects global freeze state and workqueues
3324 * list. Grab it, set max_active accordingly and add the new
3325 * workqueue to workqueues list.
3327 spin_lock(&workqueue_lock
);
3329 if (workqueue_freezing
&& wq
->flags
& WQ_FREEZABLE
)
3330 for_each_cwq_cpu(cpu
, wq
)
3331 get_cwq(cpu
, wq
)->max_active
= 0;
3333 list_add(&wq
->list
, &workqueues
);
3335 spin_unlock(&workqueue_lock
);
3341 free_mayday_mask(wq
->mayday_mask
);
3347 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3350 * destroy_workqueue - safely terminate a workqueue
3351 * @wq: target workqueue
3353 * Safely destroy a workqueue. All work currently pending will be done first.
3355 void destroy_workqueue(struct workqueue_struct
*wq
)
3359 /* drain it before proceeding with destruction */
3360 drain_workqueue(wq
);
3363 * wq list is used to freeze wq, remove from list after
3364 * flushing is complete in case freeze races us.
3366 spin_lock(&workqueue_lock
);
3367 list_del(&wq
->list
);
3368 spin_unlock(&workqueue_lock
);
3371 for_each_cwq_cpu(cpu
, wq
) {
3372 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3375 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
3376 BUG_ON(cwq
->nr_in_flight
[i
]);
3377 BUG_ON(cwq
->nr_active
);
3378 BUG_ON(!list_empty(&cwq
->delayed_works
));
3381 if (wq
->flags
& WQ_RESCUER
) {
3382 kthread_stop(wq
->rescuer
->task
);
3383 free_mayday_mask(wq
->mayday_mask
);
3390 EXPORT_SYMBOL_GPL(destroy_workqueue
);
3393 * cwq_set_max_active - adjust max_active of a cwq
3394 * @cwq: target cpu_workqueue_struct
3395 * @max_active: new max_active value.
3397 * Set @cwq->max_active to @max_active and activate delayed works if
3401 * spin_lock_irq(gcwq->lock).
3403 static void cwq_set_max_active(struct cpu_workqueue_struct
*cwq
, int max_active
)
3405 cwq
->max_active
= max_active
;
3407 while (!list_empty(&cwq
->delayed_works
) &&
3408 cwq
->nr_active
< cwq
->max_active
)
3409 cwq_activate_first_delayed(cwq
);
3413 * workqueue_set_max_active - adjust max_active of a workqueue
3414 * @wq: target workqueue
3415 * @max_active: new max_active value.
3417 * Set max_active of @wq to @max_active.
3420 * Don't call from IRQ context.
3422 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
3426 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
3428 spin_lock(&workqueue_lock
);
3430 wq
->saved_max_active
= max_active
;
3432 for_each_cwq_cpu(cpu
, wq
) {
3433 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3434 struct worker_pool
*pool
= cwq
->pool
;
3435 struct global_cwq
*gcwq
= pool
->gcwq
;
3437 spin_lock_irq(&gcwq
->lock
);
3439 if (!(wq
->flags
& WQ_FREEZABLE
) ||
3440 !(pool
->flags
& POOL_FREEZING
))
3441 cwq_set_max_active(cwq
, max_active
);
3443 spin_unlock_irq(&gcwq
->lock
);
3446 spin_unlock(&workqueue_lock
);
3448 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
3451 * workqueue_congested - test whether a workqueue is congested
3452 * @cpu: CPU in question
3453 * @wq: target workqueue
3455 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3456 * no synchronization around this function and the test result is
3457 * unreliable and only useful as advisory hints or for debugging.
3460 * %true if congested, %false otherwise.
3462 bool workqueue_congested(unsigned int cpu
, struct workqueue_struct
*wq
)
3464 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3466 return !list_empty(&cwq
->delayed_works
);
3468 EXPORT_SYMBOL_GPL(workqueue_congested
);
3471 * work_busy - test whether a work is currently pending or running
3472 * @work: the work to be tested
3474 * Test whether @work is currently pending or running. There is no
3475 * synchronization around this function and the test result is
3476 * unreliable and only useful as advisory hints or for debugging.
3477 * Especially for reentrant wqs, the pending state might hide the
3481 * OR'd bitmask of WORK_BUSY_* bits.
3483 unsigned int work_busy(struct work_struct
*work
)
3485 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3486 unsigned long flags
;
3487 unsigned int ret
= 0;
3492 spin_lock_irqsave(&gcwq
->lock
, flags
);
3494 if (work_pending(work
))
3495 ret
|= WORK_BUSY_PENDING
;
3496 if (find_worker_executing_work(gcwq
, work
))
3497 ret
|= WORK_BUSY_RUNNING
;
3499 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
3503 EXPORT_SYMBOL_GPL(work_busy
);
3508 * There are two challenges in supporting CPU hotplug. Firstly, there
3509 * are a lot of assumptions on strong associations among work, cwq and
3510 * gcwq which make migrating pending and scheduled works very
3511 * difficult to implement without impacting hot paths. Secondly,
3512 * gcwqs serve mix of short, long and very long running works making
3513 * blocked draining impractical.
3515 * This is solved by allowing the pools to be disassociated from the CPU
3516 * running as an unbound one and allowing it to be reattached later if the
3517 * cpu comes back online.
3520 /* claim manager positions of all pools */
3521 static void gcwq_claim_assoc_and_lock(struct global_cwq
*gcwq
)
3523 struct worker_pool
*pool
;
3525 for_each_worker_pool(pool
, gcwq
)
3526 mutex_lock_nested(&pool
->assoc_mutex
, pool
- gcwq
->pools
);
3527 spin_lock_irq(&gcwq
->lock
);
3530 /* release manager positions */
3531 static void gcwq_release_assoc_and_unlock(struct global_cwq
*gcwq
)
3533 struct worker_pool
*pool
;
3535 spin_unlock_irq(&gcwq
->lock
);
3536 for_each_worker_pool(pool
, gcwq
)
3537 mutex_unlock(&pool
->assoc_mutex
);
3540 static void gcwq_unbind_fn(struct work_struct
*work
)
3542 struct global_cwq
*gcwq
= get_gcwq(smp_processor_id());
3543 struct worker_pool
*pool
;
3544 struct worker
*worker
;
3545 struct hlist_node
*pos
;
3548 BUG_ON(gcwq
->cpu
!= smp_processor_id());
3550 gcwq_claim_assoc_and_lock(gcwq
);
3553 * We've claimed all manager positions. Make all workers unbound
3554 * and set DISASSOCIATED. Before this, all workers except for the
3555 * ones which are still executing works from before the last CPU
3556 * down must be on the cpu. After this, they may become diasporas.
3558 for_each_worker_pool(pool
, gcwq
)
3559 list_for_each_entry(worker
, &pool
->idle_list
, entry
)
3560 worker
->flags
|= WORKER_UNBOUND
;
3562 for_each_busy_worker(worker
, i
, pos
, gcwq
)
3563 worker
->flags
|= WORKER_UNBOUND
;
3565 for_each_worker_pool(pool
, gcwq
)
3566 pool
->flags
|= POOL_DISASSOCIATED
;
3568 gcwq_release_assoc_and_unlock(gcwq
);
3571 * Call schedule() so that we cross rq->lock and thus can guarantee
3572 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3573 * as scheduler callbacks may be invoked from other cpus.
3578 * Sched callbacks are disabled now. Zap nr_running. After this,
3579 * nr_running stays zero and need_more_worker() and keep_working()
3580 * are always true as long as the worklist is not empty. @gcwq now
3581 * behaves as unbound (in terms of concurrency management) gcwq
3582 * which is served by workers tied to the CPU.
3584 * On return from this function, the current worker would trigger
3585 * unbound chain execution of pending work items if other workers
3588 for_each_worker_pool(pool
, gcwq
)
3589 atomic_set(get_pool_nr_running(pool
), 0);
3593 * Workqueues should be brought up before normal priority CPU notifiers.
3594 * This will be registered high priority CPU notifier.
3596 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
3597 unsigned long action
,
3600 unsigned int cpu
= (unsigned long)hcpu
;
3601 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3602 struct worker_pool
*pool
;
3604 switch (action
& ~CPU_TASKS_FROZEN
) {
3605 case CPU_UP_PREPARE
:
3606 for_each_worker_pool(pool
, gcwq
) {
3607 struct worker
*worker
;
3609 if (pool
->nr_workers
)
3612 worker
= create_worker(pool
);
3616 spin_lock_irq(&gcwq
->lock
);
3617 start_worker(worker
);
3618 spin_unlock_irq(&gcwq
->lock
);
3622 case CPU_DOWN_FAILED
:
3624 gcwq_claim_assoc_and_lock(gcwq
);
3625 for_each_worker_pool(pool
, gcwq
)
3626 pool
->flags
&= ~POOL_DISASSOCIATED
;
3627 rebind_workers(gcwq
);
3628 gcwq_release_assoc_and_unlock(gcwq
);
3635 * Workqueues should be brought down after normal priority CPU notifiers.
3636 * This will be registered as low priority CPU notifier.
3638 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
3639 unsigned long action
,
3642 unsigned int cpu
= (unsigned long)hcpu
;
3643 struct work_struct unbind_work
;
3645 switch (action
& ~CPU_TASKS_FROZEN
) {
3646 case CPU_DOWN_PREPARE
:
3647 /* unbinding should happen on the local CPU */
3648 INIT_WORK_ONSTACK(&unbind_work
, gcwq_unbind_fn
);
3649 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
3650 flush_work(&unbind_work
);
3658 struct work_for_cpu
{
3659 struct work_struct work
;
3665 static void work_for_cpu_fn(struct work_struct
*work
)
3667 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
3669 wfc
->ret
= wfc
->fn(wfc
->arg
);
3673 * work_on_cpu - run a function in user context on a particular cpu
3674 * @cpu: the cpu to run on
3675 * @fn: the function to run
3676 * @arg: the function arg
3678 * This will return the value @fn returns.
3679 * It is up to the caller to ensure that the cpu doesn't go offline.
3680 * The caller must not hold any locks which would prevent @fn from completing.
3682 long work_on_cpu(unsigned int cpu
, long (*fn
)(void *), void *arg
)
3684 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
3686 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
3687 schedule_work_on(cpu
, &wfc
.work
);
3688 flush_work(&wfc
.work
);
3691 EXPORT_SYMBOL_GPL(work_on_cpu
);
3692 #endif /* CONFIG_SMP */
3694 #ifdef CONFIG_FREEZER
3697 * freeze_workqueues_begin - begin freezing workqueues
3699 * Start freezing workqueues. After this function returns, all freezable
3700 * workqueues will queue new works to their frozen_works list instead of
3704 * Grabs and releases workqueue_lock and gcwq->lock's.
3706 void freeze_workqueues_begin(void)
3710 spin_lock(&workqueue_lock
);
3712 BUG_ON(workqueue_freezing
);
3713 workqueue_freezing
= true;
3715 for_each_gcwq_cpu(cpu
) {
3716 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3717 struct worker_pool
*pool
;
3718 struct workqueue_struct
*wq
;
3720 spin_lock_irq(&gcwq
->lock
);
3722 for_each_worker_pool(pool
, gcwq
) {
3723 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
3724 pool
->flags
|= POOL_FREEZING
;
3727 list_for_each_entry(wq
, &workqueues
, list
) {
3728 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3730 if (cwq
&& wq
->flags
& WQ_FREEZABLE
)
3731 cwq
->max_active
= 0;
3734 spin_unlock_irq(&gcwq
->lock
);
3737 spin_unlock(&workqueue_lock
);
3741 * freeze_workqueues_busy - are freezable workqueues still busy?
3743 * Check whether freezing is complete. This function must be called
3744 * between freeze_workqueues_begin() and thaw_workqueues().
3747 * Grabs and releases workqueue_lock.
3750 * %true if some freezable workqueues are still busy. %false if freezing
3753 bool freeze_workqueues_busy(void)
3758 spin_lock(&workqueue_lock
);
3760 BUG_ON(!workqueue_freezing
);
3762 for_each_gcwq_cpu(cpu
) {
3763 struct workqueue_struct
*wq
;
3765 * nr_active is monotonically decreasing. It's safe
3766 * to peek without lock.
3768 list_for_each_entry(wq
, &workqueues
, list
) {
3769 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3771 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3774 BUG_ON(cwq
->nr_active
< 0);
3775 if (cwq
->nr_active
) {
3782 spin_unlock(&workqueue_lock
);
3787 * thaw_workqueues - thaw workqueues
3789 * Thaw workqueues. Normal queueing is restored and all collected
3790 * frozen works are transferred to their respective gcwq worklists.
3793 * Grabs and releases workqueue_lock and gcwq->lock's.
3795 void thaw_workqueues(void)
3799 spin_lock(&workqueue_lock
);
3801 if (!workqueue_freezing
)
3804 for_each_gcwq_cpu(cpu
) {
3805 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3806 struct worker_pool
*pool
;
3807 struct workqueue_struct
*wq
;
3809 spin_lock_irq(&gcwq
->lock
);
3811 for_each_worker_pool(pool
, gcwq
) {
3812 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
3813 pool
->flags
&= ~POOL_FREEZING
;
3816 list_for_each_entry(wq
, &workqueues
, list
) {
3817 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3819 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3822 /* restore max_active and repopulate worklist */
3823 cwq_set_max_active(cwq
, wq
->saved_max_active
);
3826 for_each_worker_pool(pool
, gcwq
)
3827 wake_up_worker(pool
);
3829 spin_unlock_irq(&gcwq
->lock
);
3832 workqueue_freezing
= false;
3834 spin_unlock(&workqueue_lock
);
3836 #endif /* CONFIG_FREEZER */
3838 static int __init
init_workqueues(void)
3842 /* make sure we have enough bits for OFFQ pool ID */
3843 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_POOL_SHIFT
)) <
3844 WORK_CPU_LAST
* NR_STD_WORKER_POOLS
);
3846 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
3847 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
3849 /* initialize gcwqs */
3850 for_each_gcwq_cpu(cpu
) {
3851 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3852 struct worker_pool
*pool
;
3854 spin_lock_init(&gcwq
->lock
);
3857 hash_init(gcwq
->busy_hash
);
3859 for_each_worker_pool(pool
, gcwq
) {
3861 pool
->flags
|= POOL_DISASSOCIATED
;
3862 INIT_LIST_HEAD(&pool
->worklist
);
3863 INIT_LIST_HEAD(&pool
->idle_list
);
3865 init_timer_deferrable(&pool
->idle_timer
);
3866 pool
->idle_timer
.function
= idle_worker_timeout
;
3867 pool
->idle_timer
.data
= (unsigned long)pool
;
3869 setup_timer(&pool
->mayday_timer
, gcwq_mayday_timeout
,
3870 (unsigned long)pool
);
3872 mutex_init(&pool
->assoc_mutex
);
3873 ida_init(&pool
->worker_ida
);
3876 BUG_ON(worker_pool_assign_id(pool
));
3880 /* create the initial worker */
3881 for_each_online_gcwq_cpu(cpu
) {
3882 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3883 struct worker_pool
*pool
;
3885 for_each_worker_pool(pool
, gcwq
) {
3886 struct worker
*worker
;
3888 if (cpu
!= WORK_CPU_UNBOUND
)
3889 pool
->flags
&= ~POOL_DISASSOCIATED
;
3891 worker
= create_worker(pool
);
3893 spin_lock_irq(&gcwq
->lock
);
3894 start_worker(worker
);
3895 spin_unlock_irq(&gcwq
->lock
);
3899 system_wq
= alloc_workqueue("events", 0, 0);
3900 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
3901 system_long_wq
= alloc_workqueue("events_long", 0, 0);
3902 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
3903 WQ_UNBOUND_MAX_ACTIVE
);
3904 system_freezable_wq
= alloc_workqueue("events_freezable",
3906 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
3907 !system_unbound_wq
|| !system_freezable_wq
);
3910 early_initcall(init_workqueues
);