2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/export.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
44 #include <linux/hashtable.h>
46 #include "workqueue_internal.h"
52 * A bound pool is either associated or disassociated with its CPU.
53 * While associated (!DISASSOCIATED), all workers are bound to the
54 * CPU and none has %WORKER_UNBOUND set and concurrency management
57 * While DISASSOCIATED, the cpu may be offline and all workers have
58 * %WORKER_UNBOUND set and concurrency management disabled, and may
59 * be executing on any CPU. The pool behaves as an unbound one.
61 * Note that DISASSOCIATED can be flipped only while holding
62 * assoc_mutex to avoid changing binding state while
63 * create_worker() is in progress.
65 POOL_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
66 POOL_MANAGING_WORKERS
= 1 << 1, /* managing workers */
67 POOL_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
68 POOL_FREEZING
= 1 << 3, /* freeze in progress */
71 WORKER_STARTED
= 1 << 0, /* started */
72 WORKER_DIE
= 1 << 1, /* die die die */
73 WORKER_IDLE
= 1 << 2, /* is idle */
74 WORKER_PREP
= 1 << 3, /* preparing to run works */
75 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
76 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
78 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_UNBOUND
|
81 NR_STD_WORKER_POOLS
= 2, /* # standard pools per cpu */
83 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
85 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
86 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
88 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100 >= 2 ? HZ
/ 100 : 2,
89 /* call for help after 10ms
91 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
92 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
95 * Rescue workers are used only on emergencies and shared by
98 RESCUER_NICE_LEVEL
= -20,
99 HIGHPRI_NICE_LEVEL
= -20,
103 * Structure fields follow one of the following exclusion rules.
105 * I: Modifiable by initialization/destruction paths and read-only for
108 * P: Preemption protected. Disabling preemption is enough and should
109 * only be modified and accessed from the local cpu.
111 * L: pool->lock protected. Access with pool->lock held.
113 * X: During normal operation, modification requires pool->lock and should
114 * be done only from local cpu. Either disabling preemption on local
115 * cpu or grabbing pool->lock is enough for read access. If
116 * POOL_DISASSOCIATED is set, it's identical to L.
118 * F: wq->flush_mutex protected.
120 * W: workqueue_lock protected.
123 /* struct worker is defined in workqueue_internal.h */
126 spinlock_t lock
; /* the pool lock */
127 unsigned int cpu
; /* I: the associated cpu */
128 int id
; /* I: pool ID */
129 unsigned int flags
; /* X: flags */
131 struct list_head worklist
; /* L: list of pending works */
132 int nr_workers
; /* L: total number of workers */
134 /* nr_idle includes the ones off idle_list for rebinding */
135 int nr_idle
; /* L: currently idle ones */
137 struct list_head idle_list
; /* X: list of idle workers */
138 struct timer_list idle_timer
; /* L: worker idle timeout */
139 struct timer_list mayday_timer
; /* L: SOS timer for workers */
141 /* workers are chained either in busy_hash or idle_list */
142 DECLARE_HASHTABLE(busy_hash
, BUSY_WORKER_HASH_ORDER
);
143 /* L: hash of busy workers */
145 struct mutex assoc_mutex
; /* protect POOL_DISASSOCIATED */
146 struct ida worker_ida
; /* L: for worker IDs */
147 } ____cacheline_aligned_in_smp
;
150 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
151 * work_struct->data are used for flags and thus cwqs need to be
152 * aligned at two's power of the number of flag bits.
154 struct cpu_workqueue_struct
{
155 struct worker_pool
*pool
; /* I: the associated pool */
156 struct workqueue_struct
*wq
; /* I: the owning workqueue */
157 int work_color
; /* L: current color */
158 int flush_color
; /* L: flushing color */
159 int nr_in_flight
[WORK_NR_COLORS
];
160 /* L: nr of in_flight works */
161 int nr_active
; /* L: nr of active works */
162 int max_active
; /* L: max active works */
163 struct list_head delayed_works
; /* L: delayed works */
167 * Structure used to wait for workqueue flush.
170 struct list_head list
; /* F: list of flushers */
171 int flush_color
; /* F: flush color waiting for */
172 struct completion done
; /* flush completion */
176 * All cpumasks are assumed to be always set on UP and thus can't be
177 * used to determine whether there's something to be done.
180 typedef cpumask_var_t mayday_mask_t
;
181 #define mayday_test_and_set_cpu(cpu, mask) \
182 cpumask_test_and_set_cpu((cpu), (mask))
183 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
184 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
185 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
186 #define free_mayday_mask(mask) free_cpumask_var((mask))
188 typedef unsigned long mayday_mask_t
;
189 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
190 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
191 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
192 #define alloc_mayday_mask(maskp, gfp) true
193 #define free_mayday_mask(mask) do { } while (0)
197 * The externally visible workqueue abstraction is an array of
198 * per-CPU workqueues:
200 struct workqueue_struct
{
201 unsigned int flags
; /* W: WQ_* flags */
203 struct cpu_workqueue_struct __percpu
*pcpu
;
204 struct cpu_workqueue_struct
*single
;
206 } cpu_wq
; /* I: cwq's */
207 struct list_head list
; /* W: list of all workqueues */
209 struct mutex flush_mutex
; /* protects wq flushing */
210 int work_color
; /* F: current work color */
211 int flush_color
; /* F: current flush color */
212 atomic_t nr_cwqs_to_flush
; /* flush in progress */
213 struct wq_flusher
*first_flusher
; /* F: first flusher */
214 struct list_head flusher_queue
; /* F: flush waiters */
215 struct list_head flusher_overflow
; /* F: flush overflow list */
217 mayday_mask_t mayday_mask
; /* cpus requesting rescue */
218 struct worker
*rescuer
; /* I: rescue worker */
220 int nr_drainers
; /* W: drain in progress */
221 int saved_max_active
; /* W: saved cwq max_active */
222 #ifdef CONFIG_LOCKDEP
223 struct lockdep_map lockdep_map
;
225 char name
[]; /* I: workqueue name */
228 struct workqueue_struct
*system_wq __read_mostly
;
229 EXPORT_SYMBOL_GPL(system_wq
);
230 struct workqueue_struct
*system_highpri_wq __read_mostly
;
231 EXPORT_SYMBOL_GPL(system_highpri_wq
);
232 struct workqueue_struct
*system_long_wq __read_mostly
;
233 EXPORT_SYMBOL_GPL(system_long_wq
);
234 struct workqueue_struct
*system_unbound_wq __read_mostly
;
235 EXPORT_SYMBOL_GPL(system_unbound_wq
);
236 struct workqueue_struct
*system_freezable_wq __read_mostly
;
237 EXPORT_SYMBOL_GPL(system_freezable_wq
);
239 #define CREATE_TRACE_POINTS
240 #include <trace/events/workqueue.h>
242 #define for_each_std_worker_pool(pool, cpu) \
243 for ((pool) = &std_worker_pools(cpu)[0]; \
244 (pool) < &std_worker_pools(cpu)[NR_STD_WORKER_POOLS]; (pool)++)
246 #define for_each_busy_worker(worker, i, pos, pool) \
247 hash_for_each(pool->busy_hash, i, pos, worker, hentry)
249 static inline int __next_wq_cpu(int cpu
, const struct cpumask
*mask
,
252 if (cpu
< nr_cpu_ids
) {
254 cpu
= cpumask_next(cpu
, mask
);
255 if (cpu
< nr_cpu_ids
)
259 return WORK_CPU_UNBOUND
;
264 static inline int __next_cwq_cpu(int cpu
, const struct cpumask
*mask
,
265 struct workqueue_struct
*wq
)
267 return __next_wq_cpu(cpu
, mask
, !(wq
->flags
& WQ_UNBOUND
) ? 1 : 2);
273 * An extra cpu number is defined using an invalid cpu number
274 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
275 * specific CPU. The following iterators are similar to for_each_*_cpu()
276 * iterators but also considers the unbound CPU.
278 * for_each_wq_cpu() : possible CPUs + WORK_CPU_UNBOUND
279 * for_each_online_wq_cpu() : online CPUs + WORK_CPU_UNBOUND
280 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
281 * WORK_CPU_UNBOUND for unbound workqueues
283 #define for_each_wq_cpu(cpu) \
284 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, 3); \
285 (cpu) < WORK_CPU_END; \
286 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, 3))
288 #define for_each_online_wq_cpu(cpu) \
289 for ((cpu) = __next_wq_cpu(-1, cpu_online_mask, 3); \
290 (cpu) < WORK_CPU_END; \
291 (cpu) = __next_wq_cpu((cpu), cpu_online_mask, 3))
293 #define for_each_cwq_cpu(cpu, wq) \
294 for ((cpu) = __next_cwq_cpu(-1, cpu_possible_mask, (wq)); \
295 (cpu) < WORK_CPU_END; \
296 (cpu) = __next_cwq_cpu((cpu), cpu_possible_mask, (wq)))
298 #ifdef CONFIG_DEBUG_OBJECTS_WORK
300 static struct debug_obj_descr work_debug_descr
;
302 static void *work_debug_hint(void *addr
)
304 return ((struct work_struct
*) addr
)->func
;
308 * fixup_init is called when:
309 * - an active object is initialized
311 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
313 struct work_struct
*work
= addr
;
316 case ODEBUG_STATE_ACTIVE
:
317 cancel_work_sync(work
);
318 debug_object_init(work
, &work_debug_descr
);
326 * fixup_activate is called when:
327 * - an active object is activated
328 * - an unknown object is activated (might be a statically initialized object)
330 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
332 struct work_struct
*work
= addr
;
336 case ODEBUG_STATE_NOTAVAILABLE
:
338 * This is not really a fixup. The work struct was
339 * statically initialized. We just make sure that it
340 * is tracked in the object tracker.
342 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
343 debug_object_init(work
, &work_debug_descr
);
344 debug_object_activate(work
, &work_debug_descr
);
350 case ODEBUG_STATE_ACTIVE
:
359 * fixup_free is called when:
360 * - an active object is freed
362 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
364 struct work_struct
*work
= addr
;
367 case ODEBUG_STATE_ACTIVE
:
368 cancel_work_sync(work
);
369 debug_object_free(work
, &work_debug_descr
);
376 static struct debug_obj_descr work_debug_descr
= {
377 .name
= "work_struct",
378 .debug_hint
= work_debug_hint
,
379 .fixup_init
= work_fixup_init
,
380 .fixup_activate
= work_fixup_activate
,
381 .fixup_free
= work_fixup_free
,
384 static inline void debug_work_activate(struct work_struct
*work
)
386 debug_object_activate(work
, &work_debug_descr
);
389 static inline void debug_work_deactivate(struct work_struct
*work
)
391 debug_object_deactivate(work
, &work_debug_descr
);
394 void __init_work(struct work_struct
*work
, int onstack
)
397 debug_object_init_on_stack(work
, &work_debug_descr
);
399 debug_object_init(work
, &work_debug_descr
);
401 EXPORT_SYMBOL_GPL(__init_work
);
403 void destroy_work_on_stack(struct work_struct
*work
)
405 debug_object_free(work
, &work_debug_descr
);
407 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
410 static inline void debug_work_activate(struct work_struct
*work
) { }
411 static inline void debug_work_deactivate(struct work_struct
*work
) { }
414 /* Serializes the accesses to the list of workqueues. */
415 static DEFINE_SPINLOCK(workqueue_lock
);
416 static LIST_HEAD(workqueues
);
417 static bool workqueue_freezing
; /* W: have wqs started freezing? */
420 * The CPU standard worker pools. nr_running is the only field which is
421 * expected to be used frequently by other cpus via try_to_wake_up(). Put
422 * it in a separate cacheline.
424 static DEFINE_PER_CPU(struct worker_pool
[NR_STD_WORKER_POOLS
],
425 cpu_std_worker_pools
);
426 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t
[NR_STD_WORKER_POOLS
],
427 cpu_std_pool_nr_running
);
430 * Standard worker pools and nr_running counter for unbound CPU. The pools
431 * have POOL_DISASSOCIATED set, and all workers have WORKER_UNBOUND set.
433 static struct worker_pool unbound_std_worker_pools
[NR_STD_WORKER_POOLS
];
434 static atomic_t unbound_std_pool_nr_running
[NR_STD_WORKER_POOLS
] = {
435 [0 ... NR_STD_WORKER_POOLS
- 1] = ATOMIC_INIT(0), /* always 0 */
438 /* idr of all pools */
439 static DEFINE_MUTEX(worker_pool_idr_mutex
);
440 static DEFINE_IDR(worker_pool_idr
);
442 static int worker_thread(void *__worker
);
444 static struct worker_pool
*std_worker_pools(int cpu
)
446 if (cpu
!= WORK_CPU_UNBOUND
)
447 return per_cpu(cpu_std_worker_pools
, cpu
);
449 return unbound_std_worker_pools
;
452 static int std_worker_pool_pri(struct worker_pool
*pool
)
454 return pool
- std_worker_pools(pool
->cpu
);
457 /* allocate ID and assign it to @pool */
458 static int worker_pool_assign_id(struct worker_pool
*pool
)
462 mutex_lock(&worker_pool_idr_mutex
);
463 idr_pre_get(&worker_pool_idr
, GFP_KERNEL
);
464 ret
= idr_get_new(&worker_pool_idr
, pool
, &pool
->id
);
465 mutex_unlock(&worker_pool_idr_mutex
);
471 * Lookup worker_pool by id. The idr currently is built during boot and
472 * never modified. Don't worry about locking for now.
474 static struct worker_pool
*worker_pool_by_id(int pool_id
)
476 return idr_find(&worker_pool_idr
, pool_id
);
479 static struct worker_pool
*get_std_worker_pool(int cpu
, bool highpri
)
481 struct worker_pool
*pools
= std_worker_pools(cpu
);
483 return &pools
[highpri
];
486 static atomic_t
*get_pool_nr_running(struct worker_pool
*pool
)
489 int idx
= std_worker_pool_pri(pool
);
491 if (cpu
!= WORK_CPU_UNBOUND
)
492 return &per_cpu(cpu_std_pool_nr_running
, cpu
)[idx
];
494 return &unbound_std_pool_nr_running
[idx
];
497 static struct cpu_workqueue_struct
*get_cwq(unsigned int cpu
,
498 struct workqueue_struct
*wq
)
500 if (!(wq
->flags
& WQ_UNBOUND
)) {
501 if (likely(cpu
< nr_cpu_ids
))
502 return per_cpu_ptr(wq
->cpu_wq
.pcpu
, cpu
);
503 } else if (likely(cpu
== WORK_CPU_UNBOUND
))
504 return wq
->cpu_wq
.single
;
508 static unsigned int work_color_to_flags(int color
)
510 return color
<< WORK_STRUCT_COLOR_SHIFT
;
513 static int get_work_color(struct work_struct
*work
)
515 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
516 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
519 static int work_next_color(int color
)
521 return (color
+ 1) % WORK_NR_COLORS
;
525 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
526 * contain the pointer to the queued cwq. Once execution starts, the flag
527 * is cleared and the high bits contain OFFQ flags and pool ID.
529 * set_work_cwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
530 * and clear_work_data() can be used to set the cwq, pool or clear
531 * work->data. These functions should only be called while the work is
532 * owned - ie. while the PENDING bit is set.
534 * get_work_pool() and get_work_cwq() can be used to obtain the pool or cwq
535 * corresponding to a work. Pool is available once the work has been
536 * queued anywhere after initialization until it is sync canceled. cwq is
537 * available only while the work item is queued.
539 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
540 * canceled. While being canceled, a work item may have its PENDING set
541 * but stay off timer and worklist for arbitrarily long and nobody should
542 * try to steal the PENDING bit.
544 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
547 BUG_ON(!work_pending(work
));
548 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
551 static void set_work_cwq(struct work_struct
*work
,
552 struct cpu_workqueue_struct
*cwq
,
553 unsigned long extra_flags
)
555 set_work_data(work
, (unsigned long)cwq
,
556 WORK_STRUCT_PENDING
| WORK_STRUCT_CWQ
| extra_flags
);
559 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
563 * The following wmb is paired with the implied mb in
564 * test_and_set_bit(PENDING) and ensures all updates to @work made
565 * here are visible to and precede any updates by the next PENDING
569 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
572 static void clear_work_data(struct work_struct
*work
)
574 smp_wmb(); /* see set_work_pool_and_clear_pending() */
575 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
578 static struct cpu_workqueue_struct
*get_work_cwq(struct work_struct
*work
)
580 unsigned long data
= atomic_long_read(&work
->data
);
582 if (data
& WORK_STRUCT_CWQ
)
583 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
589 * get_work_pool - return the worker_pool a given work was associated with
590 * @work: the work item of interest
592 * Return the worker_pool @work was last associated with. %NULL if none.
594 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
596 unsigned long data
= atomic_long_read(&work
->data
);
597 struct worker_pool
*pool
;
600 if (data
& WORK_STRUCT_CWQ
)
601 return ((struct cpu_workqueue_struct
*)
602 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
604 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
605 if (pool_id
== WORK_OFFQ_POOL_NONE
)
608 pool
= worker_pool_by_id(pool_id
);
614 * get_work_pool_id - return the worker pool ID a given work is associated with
615 * @work: the work item of interest
617 * Return the worker_pool ID @work was last associated with.
618 * %WORK_OFFQ_POOL_NONE if none.
620 static int get_work_pool_id(struct work_struct
*work
)
622 struct worker_pool
*pool
= get_work_pool(work
);
624 return pool
? pool
->id
: WORK_OFFQ_POOL_NONE
;
627 static void mark_work_canceling(struct work_struct
*work
)
629 unsigned long pool_id
= get_work_pool_id(work
);
631 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
632 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
635 static bool work_is_canceling(struct work_struct
*work
)
637 unsigned long data
= atomic_long_read(&work
->data
);
639 return !(data
& WORK_STRUCT_CWQ
) && (data
& WORK_OFFQ_CANCELING
);
643 * Policy functions. These define the policies on how the global worker
644 * pools are managed. Unless noted otherwise, these functions assume that
645 * they're being called with pool->lock held.
648 static bool __need_more_worker(struct worker_pool
*pool
)
650 return !atomic_read(get_pool_nr_running(pool
));
654 * Need to wake up a worker? Called from anything but currently
657 * Note that, because unbound workers never contribute to nr_running, this
658 * function will always return %true for unbound pools as long as the
659 * worklist isn't empty.
661 static bool need_more_worker(struct worker_pool
*pool
)
663 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
666 /* Can I start working? Called from busy but !running workers. */
667 static bool may_start_working(struct worker_pool
*pool
)
669 return pool
->nr_idle
;
672 /* Do I need to keep working? Called from currently running workers. */
673 static bool keep_working(struct worker_pool
*pool
)
675 atomic_t
*nr_running
= get_pool_nr_running(pool
);
677 return !list_empty(&pool
->worklist
) && atomic_read(nr_running
) <= 1;
680 /* Do we need a new worker? Called from manager. */
681 static bool need_to_create_worker(struct worker_pool
*pool
)
683 return need_more_worker(pool
) && !may_start_working(pool
);
686 /* Do I need to be the manager? */
687 static bool need_to_manage_workers(struct worker_pool
*pool
)
689 return need_to_create_worker(pool
) ||
690 (pool
->flags
& POOL_MANAGE_WORKERS
);
693 /* Do we have too many workers and should some go away? */
694 static bool too_many_workers(struct worker_pool
*pool
)
696 bool managing
= pool
->flags
& POOL_MANAGING_WORKERS
;
697 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
698 int nr_busy
= pool
->nr_workers
- nr_idle
;
701 * nr_idle and idle_list may disagree if idle rebinding is in
702 * progress. Never return %true if idle_list is empty.
704 if (list_empty(&pool
->idle_list
))
707 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
714 /* Return the first worker. Safe with preemption disabled */
715 static struct worker
*first_worker(struct worker_pool
*pool
)
717 if (unlikely(list_empty(&pool
->idle_list
)))
720 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
724 * wake_up_worker - wake up an idle worker
725 * @pool: worker pool to wake worker from
727 * Wake up the first idle worker of @pool.
730 * spin_lock_irq(pool->lock).
732 static void wake_up_worker(struct worker_pool
*pool
)
734 struct worker
*worker
= first_worker(pool
);
737 wake_up_process(worker
->task
);
741 * wq_worker_waking_up - a worker is waking up
742 * @task: task waking up
743 * @cpu: CPU @task is waking up to
745 * This function is called during try_to_wake_up() when a worker is
749 * spin_lock_irq(rq->lock)
751 void wq_worker_waking_up(struct task_struct
*task
, unsigned int cpu
)
753 struct worker
*worker
= kthread_data(task
);
755 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
756 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
757 atomic_inc(get_pool_nr_running(worker
->pool
));
762 * wq_worker_sleeping - a worker is going to sleep
763 * @task: task going to sleep
764 * @cpu: CPU in question, must be the current CPU number
766 * This function is called during schedule() when a busy worker is
767 * going to sleep. Worker on the same cpu can be woken up by
768 * returning pointer to its task.
771 * spin_lock_irq(rq->lock)
774 * Worker task on @cpu to wake up, %NULL if none.
776 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
,
779 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
780 struct worker_pool
*pool
;
781 atomic_t
*nr_running
;
784 * Rescuers, which may not have all the fields set up like normal
785 * workers, also reach here, let's not access anything before
786 * checking NOT_RUNNING.
788 if (worker
->flags
& WORKER_NOT_RUNNING
)
792 nr_running
= get_pool_nr_running(pool
);
794 /* this can only happen on the local cpu */
795 BUG_ON(cpu
!= raw_smp_processor_id());
798 * The counterpart of the following dec_and_test, implied mb,
799 * worklist not empty test sequence is in insert_work().
800 * Please read comment there.
802 * NOT_RUNNING is clear. This means that we're bound to and
803 * running on the local cpu w/ rq lock held and preemption
804 * disabled, which in turn means that none else could be
805 * manipulating idle_list, so dereferencing idle_list without pool
808 if (atomic_dec_and_test(nr_running
) && !list_empty(&pool
->worklist
))
809 to_wakeup
= first_worker(pool
);
810 return to_wakeup
? to_wakeup
->task
: NULL
;
814 * worker_set_flags - set worker flags and adjust nr_running accordingly
816 * @flags: flags to set
817 * @wakeup: wakeup an idle worker if necessary
819 * Set @flags in @worker->flags and adjust nr_running accordingly. If
820 * nr_running becomes zero and @wakeup is %true, an idle worker is
824 * spin_lock_irq(pool->lock)
826 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
829 struct worker_pool
*pool
= worker
->pool
;
831 WARN_ON_ONCE(worker
->task
!= current
);
834 * If transitioning into NOT_RUNNING, adjust nr_running and
835 * wake up an idle worker as necessary if requested by
838 if ((flags
& WORKER_NOT_RUNNING
) &&
839 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
840 atomic_t
*nr_running
= get_pool_nr_running(pool
);
843 if (atomic_dec_and_test(nr_running
) &&
844 !list_empty(&pool
->worklist
))
845 wake_up_worker(pool
);
847 atomic_dec(nr_running
);
850 worker
->flags
|= flags
;
854 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
856 * @flags: flags to clear
858 * Clear @flags in @worker->flags and adjust nr_running accordingly.
861 * spin_lock_irq(pool->lock)
863 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
865 struct worker_pool
*pool
= worker
->pool
;
866 unsigned int oflags
= worker
->flags
;
868 WARN_ON_ONCE(worker
->task
!= current
);
870 worker
->flags
&= ~flags
;
873 * If transitioning out of NOT_RUNNING, increment nr_running. Note
874 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
875 * of multiple flags, not a single flag.
877 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
878 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
879 atomic_inc(get_pool_nr_running(pool
));
883 * find_worker_executing_work - find worker which is executing a work
884 * @pool: pool of interest
885 * @work: work to find worker for
887 * Find a worker which is executing @work on @pool by searching
888 * @pool->busy_hash which is keyed by the address of @work. For a worker
889 * to match, its current execution should match the address of @work and
890 * its work function. This is to avoid unwanted dependency between
891 * unrelated work executions through a work item being recycled while still
894 * This is a bit tricky. A work item may be freed once its execution
895 * starts and nothing prevents the freed area from being recycled for
896 * another work item. If the same work item address ends up being reused
897 * before the original execution finishes, workqueue will identify the
898 * recycled work item as currently executing and make it wait until the
899 * current execution finishes, introducing an unwanted dependency.
901 * This function checks the work item address, work function and workqueue
902 * to avoid false positives. Note that this isn't complete as one may
903 * construct a work function which can introduce dependency onto itself
904 * through a recycled work item. Well, if somebody wants to shoot oneself
905 * in the foot that badly, there's only so much we can do, and if such
906 * deadlock actually occurs, it should be easy to locate the culprit work
910 * spin_lock_irq(pool->lock).
913 * Pointer to worker which is executing @work if found, NULL
916 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
917 struct work_struct
*work
)
919 struct worker
*worker
;
920 struct hlist_node
*tmp
;
922 hash_for_each_possible(pool
->busy_hash
, worker
, tmp
, hentry
,
924 if (worker
->current_work
== work
&&
925 worker
->current_func
== work
->func
)
932 * move_linked_works - move linked works to a list
933 * @work: start of series of works to be scheduled
934 * @head: target list to append @work to
935 * @nextp: out paramter for nested worklist walking
937 * Schedule linked works starting from @work to @head. Work series to
938 * be scheduled starts at @work and includes any consecutive work with
939 * WORK_STRUCT_LINKED set in its predecessor.
941 * If @nextp is not NULL, it's updated to point to the next work of
942 * the last scheduled work. This allows move_linked_works() to be
943 * nested inside outer list_for_each_entry_safe().
946 * spin_lock_irq(pool->lock).
948 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
949 struct work_struct
**nextp
)
951 struct work_struct
*n
;
954 * Linked worklist will always end before the end of the list,
955 * use NULL for list head.
957 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
958 list_move_tail(&work
->entry
, head
);
959 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
964 * If we're already inside safe list traversal and have moved
965 * multiple works to the scheduled queue, the next position
966 * needs to be updated.
972 static void cwq_activate_delayed_work(struct work_struct
*work
)
974 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
976 trace_workqueue_activate_work(work
);
977 move_linked_works(work
, &cwq
->pool
->worklist
, NULL
);
978 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
982 static void cwq_activate_first_delayed(struct cpu_workqueue_struct
*cwq
)
984 struct work_struct
*work
= list_first_entry(&cwq
->delayed_works
,
985 struct work_struct
, entry
);
987 cwq_activate_delayed_work(work
);
991 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
992 * @cwq: cwq of interest
993 * @color: color of work which left the queue
995 * A work either has completed or is removed from pending queue,
996 * decrement nr_in_flight of its cwq and handle workqueue flushing.
999 * spin_lock_irq(pool->lock).
1001 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct
*cwq
, int color
)
1003 /* ignore uncolored works */
1004 if (color
== WORK_NO_COLOR
)
1007 cwq
->nr_in_flight
[color
]--;
1010 if (!list_empty(&cwq
->delayed_works
)) {
1011 /* one down, submit a delayed one */
1012 if (cwq
->nr_active
< cwq
->max_active
)
1013 cwq_activate_first_delayed(cwq
);
1016 /* is flush in progress and are we at the flushing tip? */
1017 if (likely(cwq
->flush_color
!= color
))
1020 /* are there still in-flight works? */
1021 if (cwq
->nr_in_flight
[color
])
1024 /* this cwq is done, clear flush_color */
1025 cwq
->flush_color
= -1;
1028 * If this was the last cwq, wake up the first flusher. It
1029 * will handle the rest.
1031 if (atomic_dec_and_test(&cwq
->wq
->nr_cwqs_to_flush
))
1032 complete(&cwq
->wq
->first_flusher
->done
);
1036 * try_to_grab_pending - steal work item from worklist and disable irq
1037 * @work: work item to steal
1038 * @is_dwork: @work is a delayed_work
1039 * @flags: place to store irq state
1041 * Try to grab PENDING bit of @work. This function can handle @work in any
1042 * stable state - idle, on timer or on worklist. Return values are
1044 * 1 if @work was pending and we successfully stole PENDING
1045 * 0 if @work was idle and we claimed PENDING
1046 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1047 * -ENOENT if someone else is canceling @work, this state may persist
1048 * for arbitrarily long
1050 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1051 * interrupted while holding PENDING and @work off queue, irq must be
1052 * disabled on entry. This, combined with delayed_work->timer being
1053 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1055 * On successful return, >= 0, irq is disabled and the caller is
1056 * responsible for releasing it using local_irq_restore(*@flags).
1058 * This function is safe to call from any context including IRQ handler.
1060 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1061 unsigned long *flags
)
1063 struct worker_pool
*pool
;
1065 local_irq_save(*flags
);
1067 /* try to steal the timer if it exists */
1069 struct delayed_work
*dwork
= to_delayed_work(work
);
1072 * dwork->timer is irqsafe. If del_timer() fails, it's
1073 * guaranteed that the timer is not queued anywhere and not
1074 * running on the local CPU.
1076 if (likely(del_timer(&dwork
->timer
)))
1080 /* try to claim PENDING the normal way */
1081 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1085 * The queueing is in progress, or it is already queued. Try to
1086 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1088 pool
= get_work_pool(work
);
1092 spin_lock(&pool
->lock
);
1093 if (!list_empty(&work
->entry
)) {
1095 * This work is queued, but perhaps we locked the wrong
1096 * pool. In that case we must see the new value after
1097 * rmb(), see insert_work()->wmb().
1100 if (pool
== get_work_pool(work
)) {
1101 debug_work_deactivate(work
);
1104 * A delayed work item cannot be grabbed directly
1105 * because it might have linked NO_COLOR work items
1106 * which, if left on the delayed_list, will confuse
1107 * cwq->nr_active management later on and cause
1108 * stall. Make sure the work item is activated
1111 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1112 cwq_activate_delayed_work(work
);
1114 list_del_init(&work
->entry
);
1115 cwq_dec_nr_in_flight(get_work_cwq(work
),
1116 get_work_color(work
));
1118 spin_unlock(&pool
->lock
);
1122 spin_unlock(&pool
->lock
);
1124 local_irq_restore(*flags
);
1125 if (work_is_canceling(work
))
1132 * insert_work - insert a work into a pool
1133 * @cwq: cwq @work belongs to
1134 * @work: work to insert
1135 * @head: insertion point
1136 * @extra_flags: extra WORK_STRUCT_* flags to set
1138 * Insert @work which belongs to @cwq after @head. @extra_flags is or'd to
1139 * work_struct flags.
1142 * spin_lock_irq(pool->lock).
1144 static void insert_work(struct cpu_workqueue_struct
*cwq
,
1145 struct work_struct
*work
, struct list_head
*head
,
1146 unsigned int extra_flags
)
1148 struct worker_pool
*pool
= cwq
->pool
;
1150 /* we own @work, set data and link */
1151 set_work_cwq(work
, cwq
, extra_flags
);
1154 * Ensure that we get the right work->data if we see the
1155 * result of list_add() below, see try_to_grab_pending().
1159 list_add_tail(&work
->entry
, head
);
1162 * Ensure either worker_sched_deactivated() sees the above
1163 * list_add_tail() or we see zero nr_running to avoid workers
1164 * lying around lazily while there are works to be processed.
1168 if (__need_more_worker(pool
))
1169 wake_up_worker(pool
);
1173 * Test whether @work is being queued from another work executing on the
1174 * same workqueue. This is rather expensive and should only be used from
1177 static bool is_chained_work(struct workqueue_struct
*wq
)
1179 unsigned long flags
;
1182 for_each_wq_cpu(cpu
) {
1183 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
1184 struct worker_pool
*pool
= cwq
->pool
;
1185 struct worker
*worker
;
1186 struct hlist_node
*pos
;
1189 spin_lock_irqsave(&pool
->lock
, flags
);
1190 for_each_busy_worker(worker
, i
, pos
, pool
) {
1191 if (worker
->task
!= current
)
1193 spin_unlock_irqrestore(&pool
->lock
, flags
);
1195 * I'm @worker, no locking necessary. See if @work
1196 * is headed to the same workqueue.
1198 return worker
->current_cwq
->wq
== wq
;
1200 spin_unlock_irqrestore(&pool
->lock
, flags
);
1205 static void __queue_work(unsigned int cpu
, struct workqueue_struct
*wq
,
1206 struct work_struct
*work
)
1208 bool highpri
= wq
->flags
& WQ_HIGHPRI
;
1209 struct worker_pool
*pool
;
1210 struct cpu_workqueue_struct
*cwq
;
1211 struct list_head
*worklist
;
1212 unsigned int work_flags
;
1213 unsigned int req_cpu
= cpu
;
1216 * While a work item is PENDING && off queue, a task trying to
1217 * steal the PENDING will busy-loop waiting for it to either get
1218 * queued or lose PENDING. Grabbing PENDING and queueing should
1219 * happen with IRQ disabled.
1221 WARN_ON_ONCE(!irqs_disabled());
1223 debug_work_activate(work
);
1225 /* if dying, only works from the same workqueue are allowed */
1226 if (unlikely(wq
->flags
& WQ_DRAINING
) &&
1227 WARN_ON_ONCE(!is_chained_work(wq
)))
1230 /* determine pool to use */
1231 if (!(wq
->flags
& WQ_UNBOUND
)) {
1232 struct worker_pool
*last_pool
;
1234 if (cpu
== WORK_CPU_UNBOUND
)
1235 cpu
= raw_smp_processor_id();
1238 * It's multi cpu. If @work was previously on a different
1239 * cpu, it might still be running there, in which case the
1240 * work needs to be queued on that cpu to guarantee
1243 pool
= get_std_worker_pool(cpu
, highpri
);
1244 last_pool
= get_work_pool(work
);
1246 if (last_pool
&& last_pool
!= pool
) {
1247 struct worker
*worker
;
1249 spin_lock(&last_pool
->lock
);
1251 worker
= find_worker_executing_work(last_pool
, work
);
1253 if (worker
&& worker
->current_cwq
->wq
== wq
)
1256 /* meh... not running there, queue here */
1257 spin_unlock(&last_pool
->lock
);
1258 spin_lock(&pool
->lock
);
1261 spin_lock(&pool
->lock
);
1264 pool
= get_std_worker_pool(WORK_CPU_UNBOUND
, highpri
);
1265 spin_lock(&pool
->lock
);
1268 /* pool determined, get cwq and queue */
1269 cwq
= get_cwq(pool
->cpu
, wq
);
1270 trace_workqueue_queue_work(req_cpu
, cwq
, work
);
1272 if (WARN_ON(!list_empty(&work
->entry
))) {
1273 spin_unlock(&pool
->lock
);
1277 cwq
->nr_in_flight
[cwq
->work_color
]++;
1278 work_flags
= work_color_to_flags(cwq
->work_color
);
1280 if (likely(cwq
->nr_active
< cwq
->max_active
)) {
1281 trace_workqueue_activate_work(work
);
1283 worklist
= &cwq
->pool
->worklist
;
1285 work_flags
|= WORK_STRUCT_DELAYED
;
1286 worklist
= &cwq
->delayed_works
;
1289 insert_work(cwq
, work
, worklist
, work_flags
);
1291 spin_unlock(&pool
->lock
);
1295 * queue_work_on - queue work on specific cpu
1296 * @cpu: CPU number to execute work on
1297 * @wq: workqueue to use
1298 * @work: work to queue
1300 * Returns %false if @work was already on a queue, %true otherwise.
1302 * We queue the work to a specific CPU, the caller must ensure it
1305 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1306 struct work_struct
*work
)
1309 unsigned long flags
;
1311 local_irq_save(flags
);
1313 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1314 __queue_work(cpu
, wq
, work
);
1318 local_irq_restore(flags
);
1321 EXPORT_SYMBOL_GPL(queue_work_on
);
1324 * queue_work - queue work on a workqueue
1325 * @wq: workqueue to use
1326 * @work: work to queue
1328 * Returns %false if @work was already on a queue, %true otherwise.
1330 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1331 * it can be processed by another CPU.
1333 bool queue_work(struct workqueue_struct
*wq
, struct work_struct
*work
)
1335 return queue_work_on(WORK_CPU_UNBOUND
, wq
, work
);
1337 EXPORT_SYMBOL_GPL(queue_work
);
1339 void delayed_work_timer_fn(unsigned long __data
)
1341 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1343 /* should have been called from irqsafe timer with irq already off */
1344 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1346 EXPORT_SYMBOL_GPL(delayed_work_timer_fn
);
1348 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1349 struct delayed_work
*dwork
, unsigned long delay
)
1351 struct timer_list
*timer
= &dwork
->timer
;
1352 struct work_struct
*work
= &dwork
->work
;
1354 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1355 timer
->data
!= (unsigned long)dwork
);
1356 WARN_ON_ONCE(timer_pending(timer
));
1357 WARN_ON_ONCE(!list_empty(&work
->entry
));
1360 * If @delay is 0, queue @dwork->work immediately. This is for
1361 * both optimization and correctness. The earliest @timer can
1362 * expire is on the closest next tick and delayed_work users depend
1363 * on that there's no such delay when @delay is 0.
1366 __queue_work(cpu
, wq
, &dwork
->work
);
1370 timer_stats_timer_set_start_info(&dwork
->timer
);
1374 timer
->expires
= jiffies
+ delay
;
1376 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1377 add_timer_on(timer
, cpu
);
1383 * queue_delayed_work_on - queue work on specific CPU after delay
1384 * @cpu: CPU number to execute work on
1385 * @wq: workqueue to use
1386 * @dwork: work to queue
1387 * @delay: number of jiffies to wait before queueing
1389 * Returns %false if @work was already on a queue, %true otherwise. If
1390 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1393 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1394 struct delayed_work
*dwork
, unsigned long delay
)
1396 struct work_struct
*work
= &dwork
->work
;
1398 unsigned long flags
;
1400 /* read the comment in __queue_work() */
1401 local_irq_save(flags
);
1403 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1404 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1408 local_irq_restore(flags
);
1411 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1414 * queue_delayed_work - queue work on a workqueue after delay
1415 * @wq: workqueue to use
1416 * @dwork: delayable work to queue
1417 * @delay: number of jiffies to wait before queueing
1419 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1421 bool queue_delayed_work(struct workqueue_struct
*wq
,
1422 struct delayed_work
*dwork
, unsigned long delay
)
1424 return queue_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1426 EXPORT_SYMBOL_GPL(queue_delayed_work
);
1429 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1430 * @cpu: CPU number to execute work on
1431 * @wq: workqueue to use
1432 * @dwork: work to queue
1433 * @delay: number of jiffies to wait before queueing
1435 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1436 * modify @dwork's timer so that it expires after @delay. If @delay is
1437 * zero, @work is guaranteed to be scheduled immediately regardless of its
1440 * Returns %false if @dwork was idle and queued, %true if @dwork was
1441 * pending and its timer was modified.
1443 * This function is safe to call from any context including IRQ handler.
1444 * See try_to_grab_pending() for details.
1446 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1447 struct delayed_work
*dwork
, unsigned long delay
)
1449 unsigned long flags
;
1453 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1454 } while (unlikely(ret
== -EAGAIN
));
1456 if (likely(ret
>= 0)) {
1457 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1458 local_irq_restore(flags
);
1461 /* -ENOENT from try_to_grab_pending() becomes %true */
1464 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1467 * mod_delayed_work - modify delay of or queue a delayed work
1468 * @wq: workqueue to use
1469 * @dwork: work to queue
1470 * @delay: number of jiffies to wait before queueing
1472 * mod_delayed_work_on() on local CPU.
1474 bool mod_delayed_work(struct workqueue_struct
*wq
, struct delayed_work
*dwork
,
1475 unsigned long delay
)
1477 return mod_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1479 EXPORT_SYMBOL_GPL(mod_delayed_work
);
1482 * worker_enter_idle - enter idle state
1483 * @worker: worker which is entering idle state
1485 * @worker is entering idle state. Update stats and idle timer if
1489 * spin_lock_irq(pool->lock).
1491 static void worker_enter_idle(struct worker
*worker
)
1493 struct worker_pool
*pool
= worker
->pool
;
1495 BUG_ON(worker
->flags
& WORKER_IDLE
);
1496 BUG_ON(!list_empty(&worker
->entry
) &&
1497 (worker
->hentry
.next
|| worker
->hentry
.pprev
));
1499 /* can't use worker_set_flags(), also called from start_worker() */
1500 worker
->flags
|= WORKER_IDLE
;
1502 worker
->last_active
= jiffies
;
1504 /* idle_list is LIFO */
1505 list_add(&worker
->entry
, &pool
->idle_list
);
1507 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1508 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1511 * Sanity check nr_running. Because wq_unbind_fn() releases
1512 * pool->lock between setting %WORKER_UNBOUND and zapping
1513 * nr_running, the warning may trigger spuriously. Check iff
1514 * unbind is not in progress.
1516 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1517 pool
->nr_workers
== pool
->nr_idle
&&
1518 atomic_read(get_pool_nr_running(pool
)));
1522 * worker_leave_idle - leave idle state
1523 * @worker: worker which is leaving idle state
1525 * @worker is leaving idle state. Update stats.
1528 * spin_lock_irq(pool->lock).
1530 static void worker_leave_idle(struct worker
*worker
)
1532 struct worker_pool
*pool
= worker
->pool
;
1534 BUG_ON(!(worker
->flags
& WORKER_IDLE
));
1535 worker_clr_flags(worker
, WORKER_IDLE
);
1537 list_del_init(&worker
->entry
);
1541 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock pool
1544 * Works which are scheduled while the cpu is online must at least be
1545 * scheduled to a worker which is bound to the cpu so that if they are
1546 * flushed from cpu callbacks while cpu is going down, they are
1547 * guaranteed to execute on the cpu.
1549 * This function is to be used by rogue workers and rescuers to bind
1550 * themselves to the target cpu and may race with cpu going down or
1551 * coming online. kthread_bind() can't be used because it may put the
1552 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1553 * verbatim as it's best effort and blocking and pool may be
1554 * [dis]associated in the meantime.
1556 * This function tries set_cpus_allowed() and locks pool and verifies the
1557 * binding against %POOL_DISASSOCIATED which is set during
1558 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1559 * enters idle state or fetches works without dropping lock, it can
1560 * guarantee the scheduling requirement described in the first paragraph.
1563 * Might sleep. Called without any lock but returns with pool->lock
1567 * %true if the associated pool is online (@worker is successfully
1568 * bound), %false if offline.
1570 static bool worker_maybe_bind_and_lock(struct worker
*worker
)
1571 __acquires(&pool
->lock
)
1573 struct worker_pool
*pool
= worker
->pool
;
1574 struct task_struct
*task
= worker
->task
;
1578 * The following call may fail, succeed or succeed
1579 * without actually migrating the task to the cpu if
1580 * it races with cpu hotunplug operation. Verify
1581 * against POOL_DISASSOCIATED.
1583 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1584 set_cpus_allowed_ptr(task
, get_cpu_mask(pool
->cpu
));
1586 spin_lock_irq(&pool
->lock
);
1587 if (pool
->flags
& POOL_DISASSOCIATED
)
1589 if (task_cpu(task
) == pool
->cpu
&&
1590 cpumask_equal(¤t
->cpus_allowed
,
1591 get_cpu_mask(pool
->cpu
)))
1593 spin_unlock_irq(&pool
->lock
);
1596 * We've raced with CPU hot[un]plug. Give it a breather
1597 * and retry migration. cond_resched() is required here;
1598 * otherwise, we might deadlock against cpu_stop trying to
1599 * bring down the CPU on non-preemptive kernel.
1607 * Rebind an idle @worker to its CPU. worker_thread() will test
1608 * list_empty(@worker->entry) before leaving idle and call this function.
1610 static void idle_worker_rebind(struct worker
*worker
)
1612 /* CPU may go down again inbetween, clear UNBOUND only on success */
1613 if (worker_maybe_bind_and_lock(worker
))
1614 worker_clr_flags(worker
, WORKER_UNBOUND
);
1616 /* rebind complete, become available again */
1617 list_add(&worker
->entry
, &worker
->pool
->idle_list
);
1618 spin_unlock_irq(&worker
->pool
->lock
);
1622 * Function for @worker->rebind.work used to rebind unbound busy workers to
1623 * the associated cpu which is coming back online. This is scheduled by
1624 * cpu up but can race with other cpu hotplug operations and may be
1625 * executed twice without intervening cpu down.
1627 static void busy_worker_rebind_fn(struct work_struct
*work
)
1629 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1631 if (worker_maybe_bind_and_lock(worker
))
1632 worker_clr_flags(worker
, WORKER_UNBOUND
);
1634 spin_unlock_irq(&worker
->pool
->lock
);
1638 * rebind_workers - rebind all workers of a pool to the associated CPU
1639 * @pool: pool of interest
1641 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1642 * is different for idle and busy ones.
1644 * Idle ones will be removed from the idle_list and woken up. They will
1645 * add themselves back after completing rebind. This ensures that the
1646 * idle_list doesn't contain any unbound workers when re-bound busy workers
1647 * try to perform local wake-ups for concurrency management.
1649 * Busy workers can rebind after they finish their current work items.
1650 * Queueing the rebind work item at the head of the scheduled list is
1651 * enough. Note that nr_running will be properly bumped as busy workers
1654 * On return, all non-manager workers are scheduled for rebind - see
1655 * manage_workers() for the manager special case. Any idle worker
1656 * including the manager will not appear on @idle_list until rebind is
1657 * complete, making local wake-ups safe.
1659 static void rebind_workers(struct worker_pool
*pool
)
1661 struct worker
*worker
, *n
;
1662 struct hlist_node
*pos
;
1665 lockdep_assert_held(&pool
->assoc_mutex
);
1666 lockdep_assert_held(&pool
->lock
);
1668 /* dequeue and kick idle ones */
1669 list_for_each_entry_safe(worker
, n
, &pool
->idle_list
, entry
) {
1671 * idle workers should be off @pool->idle_list until rebind
1672 * is complete to avoid receiving premature local wake-ups.
1674 list_del_init(&worker
->entry
);
1677 * worker_thread() will see the above dequeuing and call
1678 * idle_worker_rebind().
1680 wake_up_process(worker
->task
);
1683 /* rebind busy workers */
1684 for_each_busy_worker(worker
, i
, pos
, pool
) {
1685 struct work_struct
*rebind_work
= &worker
->rebind_work
;
1686 struct workqueue_struct
*wq
;
1688 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
1689 work_data_bits(rebind_work
)))
1692 debug_work_activate(rebind_work
);
1695 * wq doesn't really matter but let's keep @worker->pool
1696 * and @cwq->pool consistent for sanity.
1698 if (std_worker_pool_pri(worker
->pool
))
1699 wq
= system_highpri_wq
;
1703 insert_work(get_cwq(pool
->cpu
, wq
), rebind_work
,
1704 worker
->scheduled
.next
,
1705 work_color_to_flags(WORK_NO_COLOR
));
1709 static struct worker
*alloc_worker(void)
1711 struct worker
*worker
;
1713 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1715 INIT_LIST_HEAD(&worker
->entry
);
1716 INIT_LIST_HEAD(&worker
->scheduled
);
1717 INIT_WORK(&worker
->rebind_work
, busy_worker_rebind_fn
);
1718 /* on creation a worker is in !idle && prep state */
1719 worker
->flags
= WORKER_PREP
;
1725 * create_worker - create a new workqueue worker
1726 * @pool: pool the new worker will belong to
1728 * Create a new worker which is bound to @pool. The returned worker
1729 * can be started by calling start_worker() or destroyed using
1733 * Might sleep. Does GFP_KERNEL allocations.
1736 * Pointer to the newly created worker.
1738 static struct worker
*create_worker(struct worker_pool
*pool
)
1740 const char *pri
= std_worker_pool_pri(pool
) ? "H" : "";
1741 struct worker
*worker
= NULL
;
1744 spin_lock_irq(&pool
->lock
);
1745 while (ida_get_new(&pool
->worker_ida
, &id
)) {
1746 spin_unlock_irq(&pool
->lock
);
1747 if (!ida_pre_get(&pool
->worker_ida
, GFP_KERNEL
))
1749 spin_lock_irq(&pool
->lock
);
1751 spin_unlock_irq(&pool
->lock
);
1753 worker
= alloc_worker();
1757 worker
->pool
= pool
;
1760 if (pool
->cpu
!= WORK_CPU_UNBOUND
)
1761 worker
->task
= kthread_create_on_node(worker_thread
,
1762 worker
, cpu_to_node(pool
->cpu
),
1763 "kworker/%u:%d%s", pool
->cpu
, id
, pri
);
1765 worker
->task
= kthread_create(worker_thread
, worker
,
1766 "kworker/u:%d%s", id
, pri
);
1767 if (IS_ERR(worker
->task
))
1770 if (std_worker_pool_pri(pool
))
1771 set_user_nice(worker
->task
, HIGHPRI_NICE_LEVEL
);
1774 * Determine CPU binding of the new worker depending on
1775 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1776 * flag remains stable across this function. See the comments
1777 * above the flag definition for details.
1779 * As an unbound worker may later become a regular one if CPU comes
1780 * online, make sure every worker has %PF_THREAD_BOUND set.
1782 if (!(pool
->flags
& POOL_DISASSOCIATED
)) {
1783 kthread_bind(worker
->task
, pool
->cpu
);
1785 worker
->task
->flags
|= PF_THREAD_BOUND
;
1786 worker
->flags
|= WORKER_UNBOUND
;
1792 spin_lock_irq(&pool
->lock
);
1793 ida_remove(&pool
->worker_ida
, id
);
1794 spin_unlock_irq(&pool
->lock
);
1801 * start_worker - start a newly created worker
1802 * @worker: worker to start
1804 * Make the pool aware of @worker and start it.
1807 * spin_lock_irq(pool->lock).
1809 static void start_worker(struct worker
*worker
)
1811 worker
->flags
|= WORKER_STARTED
;
1812 worker
->pool
->nr_workers
++;
1813 worker_enter_idle(worker
);
1814 wake_up_process(worker
->task
);
1818 * destroy_worker - destroy a workqueue worker
1819 * @worker: worker to be destroyed
1821 * Destroy @worker and adjust @pool stats accordingly.
1824 * spin_lock_irq(pool->lock) which is released and regrabbed.
1826 static void destroy_worker(struct worker
*worker
)
1828 struct worker_pool
*pool
= worker
->pool
;
1829 int id
= worker
->id
;
1831 /* sanity check frenzy */
1832 BUG_ON(worker
->current_work
);
1833 BUG_ON(!list_empty(&worker
->scheduled
));
1835 if (worker
->flags
& WORKER_STARTED
)
1837 if (worker
->flags
& WORKER_IDLE
)
1840 list_del_init(&worker
->entry
);
1841 worker
->flags
|= WORKER_DIE
;
1843 spin_unlock_irq(&pool
->lock
);
1845 kthread_stop(worker
->task
);
1848 spin_lock_irq(&pool
->lock
);
1849 ida_remove(&pool
->worker_ida
, id
);
1852 static void idle_worker_timeout(unsigned long __pool
)
1854 struct worker_pool
*pool
= (void *)__pool
;
1856 spin_lock_irq(&pool
->lock
);
1858 if (too_many_workers(pool
)) {
1859 struct worker
*worker
;
1860 unsigned long expires
;
1862 /* idle_list is kept in LIFO order, check the last one */
1863 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1864 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1866 if (time_before(jiffies
, expires
))
1867 mod_timer(&pool
->idle_timer
, expires
);
1869 /* it's been idle for too long, wake up manager */
1870 pool
->flags
|= POOL_MANAGE_WORKERS
;
1871 wake_up_worker(pool
);
1875 spin_unlock_irq(&pool
->lock
);
1878 static bool send_mayday(struct work_struct
*work
)
1880 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1881 struct workqueue_struct
*wq
= cwq
->wq
;
1884 if (!(wq
->flags
& WQ_RESCUER
))
1887 /* mayday mayday mayday */
1888 cpu
= cwq
->pool
->cpu
;
1889 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1890 if (cpu
== WORK_CPU_UNBOUND
)
1892 if (!mayday_test_and_set_cpu(cpu
, wq
->mayday_mask
))
1893 wake_up_process(wq
->rescuer
->task
);
1897 static void pool_mayday_timeout(unsigned long __pool
)
1899 struct worker_pool
*pool
= (void *)__pool
;
1900 struct work_struct
*work
;
1902 spin_lock_irq(&pool
->lock
);
1904 if (need_to_create_worker(pool
)) {
1906 * We've been trying to create a new worker but
1907 * haven't been successful. We might be hitting an
1908 * allocation deadlock. Send distress signals to
1911 list_for_each_entry(work
, &pool
->worklist
, entry
)
1915 spin_unlock_irq(&pool
->lock
);
1917 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1921 * maybe_create_worker - create a new worker if necessary
1922 * @pool: pool to create a new worker for
1924 * Create a new worker for @pool if necessary. @pool is guaranteed to
1925 * have at least one idle worker on return from this function. If
1926 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1927 * sent to all rescuers with works scheduled on @pool to resolve
1928 * possible allocation deadlock.
1930 * On return, need_to_create_worker() is guaranteed to be false and
1931 * may_start_working() true.
1934 * spin_lock_irq(pool->lock) which may be released and regrabbed
1935 * multiple times. Does GFP_KERNEL allocations. Called only from
1939 * false if no action was taken and pool->lock stayed locked, true
1942 static bool maybe_create_worker(struct worker_pool
*pool
)
1943 __releases(&pool
->lock
)
1944 __acquires(&pool
->lock
)
1946 if (!need_to_create_worker(pool
))
1949 spin_unlock_irq(&pool
->lock
);
1951 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1952 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1955 struct worker
*worker
;
1957 worker
= create_worker(pool
);
1959 del_timer_sync(&pool
->mayday_timer
);
1960 spin_lock_irq(&pool
->lock
);
1961 start_worker(worker
);
1962 BUG_ON(need_to_create_worker(pool
));
1966 if (!need_to_create_worker(pool
))
1969 __set_current_state(TASK_INTERRUPTIBLE
);
1970 schedule_timeout(CREATE_COOLDOWN
);
1972 if (!need_to_create_worker(pool
))
1976 del_timer_sync(&pool
->mayday_timer
);
1977 spin_lock_irq(&pool
->lock
);
1978 if (need_to_create_worker(pool
))
1984 * maybe_destroy_worker - destroy workers which have been idle for a while
1985 * @pool: pool to destroy workers for
1987 * Destroy @pool workers which have been idle for longer than
1988 * IDLE_WORKER_TIMEOUT.
1991 * spin_lock_irq(pool->lock) which may be released and regrabbed
1992 * multiple times. Called only from manager.
1995 * false if no action was taken and pool->lock stayed locked, true
1998 static bool maybe_destroy_workers(struct worker_pool
*pool
)
2002 while (too_many_workers(pool
)) {
2003 struct worker
*worker
;
2004 unsigned long expires
;
2006 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
2007 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
2009 if (time_before(jiffies
, expires
)) {
2010 mod_timer(&pool
->idle_timer
, expires
);
2014 destroy_worker(worker
);
2022 * manage_workers - manage worker pool
2025 * Assume the manager role and manage the worker pool @worker belongs
2026 * to. At any given time, there can be only zero or one manager per
2027 * pool. The exclusion is handled automatically by this function.
2029 * The caller can safely start processing works on false return. On
2030 * true return, it's guaranteed that need_to_create_worker() is false
2031 * and may_start_working() is true.
2034 * spin_lock_irq(pool->lock) which may be released and regrabbed
2035 * multiple times. Does GFP_KERNEL allocations.
2038 * spin_lock_irq(pool->lock) which may be released and regrabbed
2039 * multiple times. Does GFP_KERNEL allocations.
2041 static bool manage_workers(struct worker
*worker
)
2043 struct worker_pool
*pool
= worker
->pool
;
2046 if (pool
->flags
& POOL_MANAGING_WORKERS
)
2049 pool
->flags
|= POOL_MANAGING_WORKERS
;
2052 * To simplify both worker management and CPU hotplug, hold off
2053 * management while hotplug is in progress. CPU hotplug path can't
2054 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2055 * lead to idle worker depletion (all become busy thinking someone
2056 * else is managing) which in turn can result in deadlock under
2057 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2058 * manager against CPU hotplug.
2060 * assoc_mutex would always be free unless CPU hotplug is in
2061 * progress. trylock first without dropping @pool->lock.
2063 if (unlikely(!mutex_trylock(&pool
->assoc_mutex
))) {
2064 spin_unlock_irq(&pool
->lock
);
2065 mutex_lock(&pool
->assoc_mutex
);
2067 * CPU hotplug could have happened while we were waiting
2068 * for assoc_mutex. Hotplug itself can't handle us
2069 * because manager isn't either on idle or busy list, and
2070 * @pool's state and ours could have deviated.
2072 * As hotplug is now excluded via assoc_mutex, we can
2073 * simply try to bind. It will succeed or fail depending
2074 * on @pool's current state. Try it and adjust
2075 * %WORKER_UNBOUND accordingly.
2077 if (worker_maybe_bind_and_lock(worker
))
2078 worker
->flags
&= ~WORKER_UNBOUND
;
2080 worker
->flags
|= WORKER_UNBOUND
;
2085 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2088 * Destroy and then create so that may_start_working() is true
2091 ret
|= maybe_destroy_workers(pool
);
2092 ret
|= maybe_create_worker(pool
);
2094 pool
->flags
&= ~POOL_MANAGING_WORKERS
;
2095 mutex_unlock(&pool
->assoc_mutex
);
2100 * process_one_work - process single work
2102 * @work: work to process
2104 * Process @work. This function contains all the logics necessary to
2105 * process a single work including synchronization against and
2106 * interaction with other workers on the same cpu, queueing and
2107 * flushing. As long as context requirement is met, any worker can
2108 * call this function to process a work.
2111 * spin_lock_irq(pool->lock) which is released and regrabbed.
2113 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2114 __releases(&pool
->lock
)
2115 __acquires(&pool
->lock
)
2117 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
2118 struct worker_pool
*pool
= worker
->pool
;
2119 bool cpu_intensive
= cwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2121 struct worker
*collision
;
2122 #ifdef CONFIG_LOCKDEP
2124 * It is permissible to free the struct work_struct from
2125 * inside the function that is called from it, this we need to
2126 * take into account for lockdep too. To avoid bogus "held
2127 * lock freed" warnings as well as problems when looking into
2128 * work->lockdep_map, make a copy and use that here.
2130 struct lockdep_map lockdep_map
;
2132 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2135 * Ensure we're on the correct CPU. DISASSOCIATED test is
2136 * necessary to avoid spurious warnings from rescuers servicing the
2137 * unbound or a disassociated pool.
2139 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2140 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2141 raw_smp_processor_id() != pool
->cpu
);
2144 * A single work shouldn't be executed concurrently by
2145 * multiple workers on a single cpu. Check whether anyone is
2146 * already processing the work. If so, defer the work to the
2147 * currently executing one.
2149 collision
= find_worker_executing_work(pool
, work
);
2150 if (unlikely(collision
)) {
2151 move_linked_works(work
, &collision
->scheduled
, NULL
);
2155 /* claim and dequeue */
2156 debug_work_deactivate(work
);
2157 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2158 worker
->current_work
= work
;
2159 worker
->current_func
= work
->func
;
2160 worker
->current_cwq
= cwq
;
2161 work_color
= get_work_color(work
);
2163 list_del_init(&work
->entry
);
2166 * CPU intensive works don't participate in concurrency
2167 * management. They're the scheduler's responsibility.
2169 if (unlikely(cpu_intensive
))
2170 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2173 * Unbound pool isn't concurrency managed and work items should be
2174 * executed ASAP. Wake up another worker if necessary.
2176 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2177 wake_up_worker(pool
);
2180 * Record the last pool and clear PENDING which should be the last
2181 * update to @work. Also, do this inside @pool->lock so that
2182 * PENDING and queued state changes happen together while IRQ is
2185 set_work_pool_and_clear_pending(work
, pool
->id
);
2187 spin_unlock_irq(&pool
->lock
);
2189 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2190 lock_map_acquire(&lockdep_map
);
2191 trace_workqueue_execute_start(work
);
2192 worker
->current_func(work
);
2194 * While we must be careful to not use "work" after this, the trace
2195 * point will only record its address.
2197 trace_workqueue_execute_end(work
);
2198 lock_map_release(&lockdep_map
);
2199 lock_map_release(&cwq
->wq
->lockdep_map
);
2201 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2202 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2203 " last function: %pf\n",
2204 current
->comm
, preempt_count(), task_pid_nr(current
),
2205 worker
->current_func
);
2206 debug_show_held_locks(current
);
2210 spin_lock_irq(&pool
->lock
);
2212 /* clear cpu intensive status */
2213 if (unlikely(cpu_intensive
))
2214 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2216 /* we're done with it, release */
2217 hash_del(&worker
->hentry
);
2218 worker
->current_work
= NULL
;
2219 worker
->current_func
= NULL
;
2220 worker
->current_cwq
= NULL
;
2221 cwq_dec_nr_in_flight(cwq
, work_color
);
2225 * process_scheduled_works - process scheduled works
2228 * Process all scheduled works. Please note that the scheduled list
2229 * may change while processing a work, so this function repeatedly
2230 * fetches a work from the top and executes it.
2233 * spin_lock_irq(pool->lock) which may be released and regrabbed
2236 static void process_scheduled_works(struct worker
*worker
)
2238 while (!list_empty(&worker
->scheduled
)) {
2239 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2240 struct work_struct
, entry
);
2241 process_one_work(worker
, work
);
2246 * worker_thread - the worker thread function
2249 * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools
2250 * of these per each cpu. These workers process all works regardless of
2251 * their specific target workqueue. The only exception is works which
2252 * belong to workqueues with a rescuer which will be explained in
2255 static int worker_thread(void *__worker
)
2257 struct worker
*worker
= __worker
;
2258 struct worker_pool
*pool
= worker
->pool
;
2260 /* tell the scheduler that this is a workqueue worker */
2261 worker
->task
->flags
|= PF_WQ_WORKER
;
2263 spin_lock_irq(&pool
->lock
);
2265 /* we are off idle list if destruction or rebind is requested */
2266 if (unlikely(list_empty(&worker
->entry
))) {
2267 spin_unlock_irq(&pool
->lock
);
2269 /* if DIE is set, destruction is requested */
2270 if (worker
->flags
& WORKER_DIE
) {
2271 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2275 /* otherwise, rebind */
2276 idle_worker_rebind(worker
);
2280 worker_leave_idle(worker
);
2282 /* no more worker necessary? */
2283 if (!need_more_worker(pool
))
2286 /* do we need to manage? */
2287 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2291 * ->scheduled list can only be filled while a worker is
2292 * preparing to process a work or actually processing it.
2293 * Make sure nobody diddled with it while I was sleeping.
2295 BUG_ON(!list_empty(&worker
->scheduled
));
2298 * When control reaches this point, we're guaranteed to have
2299 * at least one idle worker or that someone else has already
2300 * assumed the manager role.
2302 worker_clr_flags(worker
, WORKER_PREP
);
2305 struct work_struct
*work
=
2306 list_first_entry(&pool
->worklist
,
2307 struct work_struct
, entry
);
2309 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2310 /* optimization path, not strictly necessary */
2311 process_one_work(worker
, work
);
2312 if (unlikely(!list_empty(&worker
->scheduled
)))
2313 process_scheduled_works(worker
);
2315 move_linked_works(work
, &worker
->scheduled
, NULL
);
2316 process_scheduled_works(worker
);
2318 } while (keep_working(pool
));
2320 worker_set_flags(worker
, WORKER_PREP
, false);
2322 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2326 * pool->lock is held and there's no work to process and no need to
2327 * manage, sleep. Workers are woken up only while holding
2328 * pool->lock or from local cpu, so setting the current state
2329 * before releasing pool->lock is enough to prevent losing any
2332 worker_enter_idle(worker
);
2333 __set_current_state(TASK_INTERRUPTIBLE
);
2334 spin_unlock_irq(&pool
->lock
);
2340 * rescuer_thread - the rescuer thread function
2343 * Workqueue rescuer thread function. There's one rescuer for each
2344 * workqueue which has WQ_RESCUER set.
2346 * Regular work processing on a pool may block trying to create a new
2347 * worker which uses GFP_KERNEL allocation which has slight chance of
2348 * developing into deadlock if some works currently on the same queue
2349 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2350 * the problem rescuer solves.
2352 * When such condition is possible, the pool summons rescuers of all
2353 * workqueues which have works queued on the pool and let them process
2354 * those works so that forward progress can be guaranteed.
2356 * This should happen rarely.
2358 static int rescuer_thread(void *__rescuer
)
2360 struct worker
*rescuer
= __rescuer
;
2361 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2362 struct list_head
*scheduled
= &rescuer
->scheduled
;
2363 bool is_unbound
= wq
->flags
& WQ_UNBOUND
;
2366 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2369 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2370 * doesn't participate in concurrency management.
2372 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2374 set_current_state(TASK_INTERRUPTIBLE
);
2376 if (kthread_should_stop()) {
2377 __set_current_state(TASK_RUNNING
);
2378 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2383 * See whether any cpu is asking for help. Unbounded
2384 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2386 for_each_mayday_cpu(cpu
, wq
->mayday_mask
) {
2387 unsigned int tcpu
= is_unbound
? WORK_CPU_UNBOUND
: cpu
;
2388 struct cpu_workqueue_struct
*cwq
= get_cwq(tcpu
, wq
);
2389 struct worker_pool
*pool
= cwq
->pool
;
2390 struct work_struct
*work
, *n
;
2392 __set_current_state(TASK_RUNNING
);
2393 mayday_clear_cpu(cpu
, wq
->mayday_mask
);
2395 /* migrate to the target cpu if possible */
2396 rescuer
->pool
= pool
;
2397 worker_maybe_bind_and_lock(rescuer
);
2400 * Slurp in all works issued via this workqueue and
2403 BUG_ON(!list_empty(&rescuer
->scheduled
));
2404 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2405 if (get_work_cwq(work
) == cwq
)
2406 move_linked_works(work
, scheduled
, &n
);
2408 process_scheduled_works(rescuer
);
2411 * Leave this pool. If keep_working() is %true, notify a
2412 * regular worker; otherwise, we end up with 0 concurrency
2413 * and stalling the execution.
2415 if (keep_working(pool
))
2416 wake_up_worker(pool
);
2418 spin_unlock_irq(&pool
->lock
);
2421 /* rescuers should never participate in concurrency management */
2422 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2428 struct work_struct work
;
2429 struct completion done
;
2432 static void wq_barrier_func(struct work_struct
*work
)
2434 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2435 complete(&barr
->done
);
2439 * insert_wq_barrier - insert a barrier work
2440 * @cwq: cwq to insert barrier into
2441 * @barr: wq_barrier to insert
2442 * @target: target work to attach @barr to
2443 * @worker: worker currently executing @target, NULL if @target is not executing
2445 * @barr is linked to @target such that @barr is completed only after
2446 * @target finishes execution. Please note that the ordering
2447 * guarantee is observed only with respect to @target and on the local
2450 * Currently, a queued barrier can't be canceled. This is because
2451 * try_to_grab_pending() can't determine whether the work to be
2452 * grabbed is at the head of the queue and thus can't clear LINKED
2453 * flag of the previous work while there must be a valid next work
2454 * after a work with LINKED flag set.
2456 * Note that when @worker is non-NULL, @target may be modified
2457 * underneath us, so we can't reliably determine cwq from @target.
2460 * spin_lock_irq(pool->lock).
2462 static void insert_wq_barrier(struct cpu_workqueue_struct
*cwq
,
2463 struct wq_barrier
*barr
,
2464 struct work_struct
*target
, struct worker
*worker
)
2466 struct list_head
*head
;
2467 unsigned int linked
= 0;
2470 * debugobject calls are safe here even with pool->lock locked
2471 * as we know for sure that this will not trigger any of the
2472 * checks and call back into the fixup functions where we
2475 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2476 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2477 init_completion(&barr
->done
);
2480 * If @target is currently being executed, schedule the
2481 * barrier to the worker; otherwise, put it after @target.
2484 head
= worker
->scheduled
.next
;
2486 unsigned long *bits
= work_data_bits(target
);
2488 head
= target
->entry
.next
;
2489 /* there can already be other linked works, inherit and set */
2490 linked
= *bits
& WORK_STRUCT_LINKED
;
2491 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2494 debug_work_activate(&barr
->work
);
2495 insert_work(cwq
, &barr
->work
, head
,
2496 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2500 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2501 * @wq: workqueue being flushed
2502 * @flush_color: new flush color, < 0 for no-op
2503 * @work_color: new work color, < 0 for no-op
2505 * Prepare cwqs for workqueue flushing.
2507 * If @flush_color is non-negative, flush_color on all cwqs should be
2508 * -1. If no cwq has in-flight commands at the specified color, all
2509 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2510 * has in flight commands, its cwq->flush_color is set to
2511 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2512 * wakeup logic is armed and %true is returned.
2514 * The caller should have initialized @wq->first_flusher prior to
2515 * calling this function with non-negative @flush_color. If
2516 * @flush_color is negative, no flush color update is done and %false
2519 * If @work_color is non-negative, all cwqs should have the same
2520 * work_color which is previous to @work_color and all will be
2521 * advanced to @work_color.
2524 * mutex_lock(wq->flush_mutex).
2527 * %true if @flush_color >= 0 and there's something to flush. %false
2530 static bool flush_workqueue_prep_cwqs(struct workqueue_struct
*wq
,
2531 int flush_color
, int work_color
)
2536 if (flush_color
>= 0) {
2537 BUG_ON(atomic_read(&wq
->nr_cwqs_to_flush
));
2538 atomic_set(&wq
->nr_cwqs_to_flush
, 1);
2541 for_each_cwq_cpu(cpu
, wq
) {
2542 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2543 struct worker_pool
*pool
= cwq
->pool
;
2545 spin_lock_irq(&pool
->lock
);
2547 if (flush_color
>= 0) {
2548 BUG_ON(cwq
->flush_color
!= -1);
2550 if (cwq
->nr_in_flight
[flush_color
]) {
2551 cwq
->flush_color
= flush_color
;
2552 atomic_inc(&wq
->nr_cwqs_to_flush
);
2557 if (work_color
>= 0) {
2558 BUG_ON(work_color
!= work_next_color(cwq
->work_color
));
2559 cwq
->work_color
= work_color
;
2562 spin_unlock_irq(&pool
->lock
);
2565 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_cwqs_to_flush
))
2566 complete(&wq
->first_flusher
->done
);
2572 * flush_workqueue - ensure that any scheduled work has run to completion.
2573 * @wq: workqueue to flush
2575 * Forces execution of the workqueue and blocks until its completion.
2576 * This is typically used in driver shutdown handlers.
2578 * We sleep until all works which were queued on entry have been handled,
2579 * but we are not livelocked by new incoming ones.
2581 void flush_workqueue(struct workqueue_struct
*wq
)
2583 struct wq_flusher this_flusher
= {
2584 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2586 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2590 lock_map_acquire(&wq
->lockdep_map
);
2591 lock_map_release(&wq
->lockdep_map
);
2593 mutex_lock(&wq
->flush_mutex
);
2596 * Start-to-wait phase
2598 next_color
= work_next_color(wq
->work_color
);
2600 if (next_color
!= wq
->flush_color
) {
2602 * Color space is not full. The current work_color
2603 * becomes our flush_color and work_color is advanced
2606 BUG_ON(!list_empty(&wq
->flusher_overflow
));
2607 this_flusher
.flush_color
= wq
->work_color
;
2608 wq
->work_color
= next_color
;
2610 if (!wq
->first_flusher
) {
2611 /* no flush in progress, become the first flusher */
2612 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2614 wq
->first_flusher
= &this_flusher
;
2616 if (!flush_workqueue_prep_cwqs(wq
, wq
->flush_color
,
2618 /* nothing to flush, done */
2619 wq
->flush_color
= next_color
;
2620 wq
->first_flusher
= NULL
;
2625 BUG_ON(wq
->flush_color
== this_flusher
.flush_color
);
2626 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2627 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2631 * Oops, color space is full, wait on overflow queue.
2632 * The next flush completion will assign us
2633 * flush_color and transfer to flusher_queue.
2635 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2638 mutex_unlock(&wq
->flush_mutex
);
2640 wait_for_completion(&this_flusher
.done
);
2643 * Wake-up-and-cascade phase
2645 * First flushers are responsible for cascading flushes and
2646 * handling overflow. Non-first flushers can simply return.
2648 if (wq
->first_flusher
!= &this_flusher
)
2651 mutex_lock(&wq
->flush_mutex
);
2653 /* we might have raced, check again with mutex held */
2654 if (wq
->first_flusher
!= &this_flusher
)
2657 wq
->first_flusher
= NULL
;
2659 BUG_ON(!list_empty(&this_flusher
.list
));
2660 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2663 struct wq_flusher
*next
, *tmp
;
2665 /* complete all the flushers sharing the current flush color */
2666 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2667 if (next
->flush_color
!= wq
->flush_color
)
2669 list_del_init(&next
->list
);
2670 complete(&next
->done
);
2673 BUG_ON(!list_empty(&wq
->flusher_overflow
) &&
2674 wq
->flush_color
!= work_next_color(wq
->work_color
));
2676 /* this flush_color is finished, advance by one */
2677 wq
->flush_color
= work_next_color(wq
->flush_color
);
2679 /* one color has been freed, handle overflow queue */
2680 if (!list_empty(&wq
->flusher_overflow
)) {
2682 * Assign the same color to all overflowed
2683 * flushers, advance work_color and append to
2684 * flusher_queue. This is the start-to-wait
2685 * phase for these overflowed flushers.
2687 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2688 tmp
->flush_color
= wq
->work_color
;
2690 wq
->work_color
= work_next_color(wq
->work_color
);
2692 list_splice_tail_init(&wq
->flusher_overflow
,
2693 &wq
->flusher_queue
);
2694 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2697 if (list_empty(&wq
->flusher_queue
)) {
2698 BUG_ON(wq
->flush_color
!= wq
->work_color
);
2703 * Need to flush more colors. Make the next flusher
2704 * the new first flusher and arm cwqs.
2706 BUG_ON(wq
->flush_color
== wq
->work_color
);
2707 BUG_ON(wq
->flush_color
!= next
->flush_color
);
2709 list_del_init(&next
->list
);
2710 wq
->first_flusher
= next
;
2712 if (flush_workqueue_prep_cwqs(wq
, wq
->flush_color
, -1))
2716 * Meh... this color is already done, clear first
2717 * flusher and repeat cascading.
2719 wq
->first_flusher
= NULL
;
2723 mutex_unlock(&wq
->flush_mutex
);
2725 EXPORT_SYMBOL_GPL(flush_workqueue
);
2728 * drain_workqueue - drain a workqueue
2729 * @wq: workqueue to drain
2731 * Wait until the workqueue becomes empty. While draining is in progress,
2732 * only chain queueing is allowed. IOW, only currently pending or running
2733 * work items on @wq can queue further work items on it. @wq is flushed
2734 * repeatedly until it becomes empty. The number of flushing is detemined
2735 * by the depth of chaining and should be relatively short. Whine if it
2738 void drain_workqueue(struct workqueue_struct
*wq
)
2740 unsigned int flush_cnt
= 0;
2744 * __queue_work() needs to test whether there are drainers, is much
2745 * hotter than drain_workqueue() and already looks at @wq->flags.
2746 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2748 spin_lock(&workqueue_lock
);
2749 if (!wq
->nr_drainers
++)
2750 wq
->flags
|= WQ_DRAINING
;
2751 spin_unlock(&workqueue_lock
);
2753 flush_workqueue(wq
);
2755 for_each_cwq_cpu(cpu
, wq
) {
2756 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2759 spin_lock_irq(&cwq
->pool
->lock
);
2760 drained
= !cwq
->nr_active
&& list_empty(&cwq
->delayed_works
);
2761 spin_unlock_irq(&cwq
->pool
->lock
);
2766 if (++flush_cnt
== 10 ||
2767 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2768 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2769 wq
->name
, flush_cnt
);
2773 spin_lock(&workqueue_lock
);
2774 if (!--wq
->nr_drainers
)
2775 wq
->flags
&= ~WQ_DRAINING
;
2776 spin_unlock(&workqueue_lock
);
2778 EXPORT_SYMBOL_GPL(drain_workqueue
);
2780 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2782 struct worker
*worker
= NULL
;
2783 struct worker_pool
*pool
;
2784 struct cpu_workqueue_struct
*cwq
;
2787 pool
= get_work_pool(work
);
2791 spin_lock_irq(&pool
->lock
);
2792 if (!list_empty(&work
->entry
)) {
2794 * See the comment near try_to_grab_pending()->smp_rmb().
2795 * If it was re-queued to a different pool under us, we
2796 * are not going to wait.
2799 cwq
= get_work_cwq(work
);
2800 if (unlikely(!cwq
|| pool
!= cwq
->pool
))
2803 worker
= find_worker_executing_work(pool
, work
);
2806 cwq
= worker
->current_cwq
;
2809 insert_wq_barrier(cwq
, barr
, work
, worker
);
2810 spin_unlock_irq(&pool
->lock
);
2813 * If @max_active is 1 or rescuer is in use, flushing another work
2814 * item on the same workqueue may lead to deadlock. Make sure the
2815 * flusher is not running on the same workqueue by verifying write
2818 if (cwq
->wq
->saved_max_active
== 1 || cwq
->wq
->flags
& WQ_RESCUER
)
2819 lock_map_acquire(&cwq
->wq
->lockdep_map
);
2821 lock_map_acquire_read(&cwq
->wq
->lockdep_map
);
2822 lock_map_release(&cwq
->wq
->lockdep_map
);
2826 spin_unlock_irq(&pool
->lock
);
2831 * flush_work - wait for a work to finish executing the last queueing instance
2832 * @work: the work to flush
2834 * Wait until @work has finished execution. @work is guaranteed to be idle
2835 * on return if it hasn't been requeued since flush started.
2838 * %true if flush_work() waited for the work to finish execution,
2839 * %false if it was already idle.
2841 bool flush_work(struct work_struct
*work
)
2843 struct wq_barrier barr
;
2845 lock_map_acquire(&work
->lockdep_map
);
2846 lock_map_release(&work
->lockdep_map
);
2848 if (start_flush_work(work
, &barr
)) {
2849 wait_for_completion(&barr
.done
);
2850 destroy_work_on_stack(&barr
.work
);
2856 EXPORT_SYMBOL_GPL(flush_work
);
2858 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2860 unsigned long flags
;
2864 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2866 * If someone else is canceling, wait for the same event it
2867 * would be waiting for before retrying.
2869 if (unlikely(ret
== -ENOENT
))
2871 } while (unlikely(ret
< 0));
2873 /* tell other tasks trying to grab @work to back off */
2874 mark_work_canceling(work
);
2875 local_irq_restore(flags
);
2878 clear_work_data(work
);
2883 * cancel_work_sync - cancel a work and wait for it to finish
2884 * @work: the work to cancel
2886 * Cancel @work and wait for its execution to finish. This function
2887 * can be used even if the work re-queues itself or migrates to
2888 * another workqueue. On return from this function, @work is
2889 * guaranteed to be not pending or executing on any CPU.
2891 * cancel_work_sync(&delayed_work->work) must not be used for
2892 * delayed_work's. Use cancel_delayed_work_sync() instead.
2894 * The caller must ensure that the workqueue on which @work was last
2895 * queued can't be destroyed before this function returns.
2898 * %true if @work was pending, %false otherwise.
2900 bool cancel_work_sync(struct work_struct
*work
)
2902 return __cancel_work_timer(work
, false);
2904 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2907 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2908 * @dwork: the delayed work to flush
2910 * Delayed timer is cancelled and the pending work is queued for
2911 * immediate execution. Like flush_work(), this function only
2912 * considers the last queueing instance of @dwork.
2915 * %true if flush_work() waited for the work to finish execution,
2916 * %false if it was already idle.
2918 bool flush_delayed_work(struct delayed_work
*dwork
)
2920 local_irq_disable();
2921 if (del_timer_sync(&dwork
->timer
))
2922 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2924 return flush_work(&dwork
->work
);
2926 EXPORT_SYMBOL(flush_delayed_work
);
2929 * cancel_delayed_work - cancel a delayed work
2930 * @dwork: delayed_work to cancel
2932 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2933 * and canceled; %false if wasn't pending. Note that the work callback
2934 * function may still be running on return, unless it returns %true and the
2935 * work doesn't re-arm itself. Explicitly flush or use
2936 * cancel_delayed_work_sync() to wait on it.
2938 * This function is safe to call from any context including IRQ handler.
2940 bool cancel_delayed_work(struct delayed_work
*dwork
)
2942 unsigned long flags
;
2946 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2947 } while (unlikely(ret
== -EAGAIN
));
2949 if (unlikely(ret
< 0))
2952 set_work_pool_and_clear_pending(&dwork
->work
,
2953 get_work_pool_id(&dwork
->work
));
2954 local_irq_restore(flags
);
2957 EXPORT_SYMBOL(cancel_delayed_work
);
2960 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2961 * @dwork: the delayed work cancel
2963 * This is cancel_work_sync() for delayed works.
2966 * %true if @dwork was pending, %false otherwise.
2968 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2970 return __cancel_work_timer(&dwork
->work
, true);
2972 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2975 * schedule_work_on - put work task on a specific cpu
2976 * @cpu: cpu to put the work task on
2977 * @work: job to be done
2979 * This puts a job on a specific cpu
2981 bool schedule_work_on(int cpu
, struct work_struct
*work
)
2983 return queue_work_on(cpu
, system_wq
, work
);
2985 EXPORT_SYMBOL(schedule_work_on
);
2988 * schedule_work - put work task in global workqueue
2989 * @work: job to be done
2991 * Returns %false if @work was already on the kernel-global workqueue and
2994 * This puts a job in the kernel-global workqueue if it was not already
2995 * queued and leaves it in the same position on the kernel-global
2996 * workqueue otherwise.
2998 bool schedule_work(struct work_struct
*work
)
3000 return queue_work(system_wq
, work
);
3002 EXPORT_SYMBOL(schedule_work
);
3005 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3007 * @dwork: job to be done
3008 * @delay: number of jiffies to wait
3010 * After waiting for a given time this puts a job in the kernel-global
3011 * workqueue on the specified CPU.
3013 bool schedule_delayed_work_on(int cpu
, struct delayed_work
*dwork
,
3014 unsigned long delay
)
3016 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
3018 EXPORT_SYMBOL(schedule_delayed_work_on
);
3021 * schedule_delayed_work - put work task in global workqueue after delay
3022 * @dwork: job to be done
3023 * @delay: number of jiffies to wait or 0 for immediate execution
3025 * After waiting for a given time this puts a job in the kernel-global
3028 bool schedule_delayed_work(struct delayed_work
*dwork
, unsigned long delay
)
3030 return queue_delayed_work(system_wq
, dwork
, delay
);
3032 EXPORT_SYMBOL(schedule_delayed_work
);
3035 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3036 * @func: the function to call
3038 * schedule_on_each_cpu() executes @func on each online CPU using the
3039 * system workqueue and blocks until all CPUs have completed.
3040 * schedule_on_each_cpu() is very slow.
3043 * 0 on success, -errno on failure.
3045 int schedule_on_each_cpu(work_func_t func
)
3048 struct work_struct __percpu
*works
;
3050 works
= alloc_percpu(struct work_struct
);
3056 for_each_online_cpu(cpu
) {
3057 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3059 INIT_WORK(work
, func
);
3060 schedule_work_on(cpu
, work
);
3063 for_each_online_cpu(cpu
)
3064 flush_work(per_cpu_ptr(works
, cpu
));
3072 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3074 * Forces execution of the kernel-global workqueue and blocks until its
3077 * Think twice before calling this function! It's very easy to get into
3078 * trouble if you don't take great care. Either of the following situations
3079 * will lead to deadlock:
3081 * One of the work items currently on the workqueue needs to acquire
3082 * a lock held by your code or its caller.
3084 * Your code is running in the context of a work routine.
3086 * They will be detected by lockdep when they occur, but the first might not
3087 * occur very often. It depends on what work items are on the workqueue and
3088 * what locks they need, which you have no control over.
3090 * In most situations flushing the entire workqueue is overkill; you merely
3091 * need to know that a particular work item isn't queued and isn't running.
3092 * In such cases you should use cancel_delayed_work_sync() or
3093 * cancel_work_sync() instead.
3095 void flush_scheduled_work(void)
3097 flush_workqueue(system_wq
);
3099 EXPORT_SYMBOL(flush_scheduled_work
);
3102 * execute_in_process_context - reliably execute the routine with user context
3103 * @fn: the function to execute
3104 * @ew: guaranteed storage for the execute work structure (must
3105 * be available when the work executes)
3107 * Executes the function immediately if process context is available,
3108 * otherwise schedules the function for delayed execution.
3110 * Returns: 0 - function was executed
3111 * 1 - function was scheduled for execution
3113 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3115 if (!in_interrupt()) {
3120 INIT_WORK(&ew
->work
, fn
);
3121 schedule_work(&ew
->work
);
3125 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3127 int keventd_up(void)
3129 return system_wq
!= NULL
;
3132 static int alloc_cwqs(struct workqueue_struct
*wq
)
3135 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3136 * Make sure that the alignment isn't lower than that of
3137 * unsigned long long.
3139 const size_t size
= sizeof(struct cpu_workqueue_struct
);
3140 const size_t align
= max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS
,
3141 __alignof__(unsigned long long));
3143 if (!(wq
->flags
& WQ_UNBOUND
))
3144 wq
->cpu_wq
.pcpu
= __alloc_percpu(size
, align
);
3149 * Allocate enough room to align cwq and put an extra
3150 * pointer at the end pointing back to the originally
3151 * allocated pointer which will be used for free.
3153 ptr
= kzalloc(size
+ align
+ sizeof(void *), GFP_KERNEL
);
3155 wq
->cpu_wq
.single
= PTR_ALIGN(ptr
, align
);
3156 *(void **)(wq
->cpu_wq
.single
+ 1) = ptr
;
3160 /* just in case, make sure it's actually aligned */
3161 BUG_ON(!IS_ALIGNED(wq
->cpu_wq
.v
, align
));
3162 return wq
->cpu_wq
.v
? 0 : -ENOMEM
;
3165 static void free_cwqs(struct workqueue_struct
*wq
)
3167 if (!(wq
->flags
& WQ_UNBOUND
))
3168 free_percpu(wq
->cpu_wq
.pcpu
);
3169 else if (wq
->cpu_wq
.single
) {
3170 /* the pointer to free is stored right after the cwq */
3171 kfree(*(void **)(wq
->cpu_wq
.single
+ 1));
3175 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3178 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3180 if (max_active
< 1 || max_active
> lim
)
3181 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3182 max_active
, name
, 1, lim
);
3184 return clamp_val(max_active
, 1, lim
);
3187 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3190 struct lock_class_key
*key
,
3191 const char *lock_name
, ...)
3193 va_list args
, args1
;
3194 struct workqueue_struct
*wq
;
3198 /* determine namelen, allocate wq and format name */
3199 va_start(args
, lock_name
);
3200 va_copy(args1
, args
);
3201 namelen
= vsnprintf(NULL
, 0, fmt
, args
) + 1;
3203 wq
= kzalloc(sizeof(*wq
) + namelen
, GFP_KERNEL
);
3207 vsnprintf(wq
->name
, namelen
, fmt
, args1
);
3212 * Workqueues which may be used during memory reclaim should
3213 * have a rescuer to guarantee forward progress.
3215 if (flags
& WQ_MEM_RECLAIM
)
3216 flags
|= WQ_RESCUER
;
3218 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3219 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3223 wq
->saved_max_active
= max_active
;
3224 mutex_init(&wq
->flush_mutex
);
3225 atomic_set(&wq
->nr_cwqs_to_flush
, 0);
3226 INIT_LIST_HEAD(&wq
->flusher_queue
);
3227 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3229 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3230 INIT_LIST_HEAD(&wq
->list
);
3232 if (alloc_cwqs(wq
) < 0)
3235 for_each_cwq_cpu(cpu
, wq
) {
3236 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3238 BUG_ON((unsigned long)cwq
& WORK_STRUCT_FLAG_MASK
);
3239 cwq
->pool
= get_std_worker_pool(cpu
, flags
& WQ_HIGHPRI
);
3241 cwq
->flush_color
= -1;
3242 cwq
->max_active
= max_active
;
3243 INIT_LIST_HEAD(&cwq
->delayed_works
);
3246 if (flags
& WQ_RESCUER
) {
3247 struct worker
*rescuer
;
3249 if (!alloc_mayday_mask(&wq
->mayday_mask
, GFP_KERNEL
))
3252 wq
->rescuer
= rescuer
= alloc_worker();
3256 rescuer
->rescue_wq
= wq
;
3257 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
3259 if (IS_ERR(rescuer
->task
))
3262 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
3263 wake_up_process(rescuer
->task
);
3267 * workqueue_lock protects global freeze state and workqueues
3268 * list. Grab it, set max_active accordingly and add the new
3269 * workqueue to workqueues list.
3271 spin_lock(&workqueue_lock
);
3273 if (workqueue_freezing
&& wq
->flags
& WQ_FREEZABLE
)
3274 for_each_cwq_cpu(cpu
, wq
)
3275 get_cwq(cpu
, wq
)->max_active
= 0;
3277 list_add(&wq
->list
, &workqueues
);
3279 spin_unlock(&workqueue_lock
);
3285 free_mayday_mask(wq
->mayday_mask
);
3291 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3294 * destroy_workqueue - safely terminate a workqueue
3295 * @wq: target workqueue
3297 * Safely destroy a workqueue. All work currently pending will be done first.
3299 void destroy_workqueue(struct workqueue_struct
*wq
)
3303 /* drain it before proceeding with destruction */
3304 drain_workqueue(wq
);
3307 * wq list is used to freeze wq, remove from list after
3308 * flushing is complete in case freeze races us.
3310 spin_lock(&workqueue_lock
);
3311 list_del(&wq
->list
);
3312 spin_unlock(&workqueue_lock
);
3315 for_each_cwq_cpu(cpu
, wq
) {
3316 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3319 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
3320 BUG_ON(cwq
->nr_in_flight
[i
]);
3321 BUG_ON(cwq
->nr_active
);
3322 BUG_ON(!list_empty(&cwq
->delayed_works
));
3325 if (wq
->flags
& WQ_RESCUER
) {
3326 kthread_stop(wq
->rescuer
->task
);
3327 free_mayday_mask(wq
->mayday_mask
);
3334 EXPORT_SYMBOL_GPL(destroy_workqueue
);
3337 * cwq_set_max_active - adjust max_active of a cwq
3338 * @cwq: target cpu_workqueue_struct
3339 * @max_active: new max_active value.
3341 * Set @cwq->max_active to @max_active and activate delayed works if
3345 * spin_lock_irq(pool->lock).
3347 static void cwq_set_max_active(struct cpu_workqueue_struct
*cwq
, int max_active
)
3349 cwq
->max_active
= max_active
;
3351 while (!list_empty(&cwq
->delayed_works
) &&
3352 cwq
->nr_active
< cwq
->max_active
)
3353 cwq_activate_first_delayed(cwq
);
3357 * workqueue_set_max_active - adjust max_active of a workqueue
3358 * @wq: target workqueue
3359 * @max_active: new max_active value.
3361 * Set max_active of @wq to @max_active.
3364 * Don't call from IRQ context.
3366 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
3370 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
3372 spin_lock(&workqueue_lock
);
3374 wq
->saved_max_active
= max_active
;
3376 for_each_cwq_cpu(cpu
, wq
) {
3377 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3378 struct worker_pool
*pool
= cwq
->pool
;
3380 spin_lock_irq(&pool
->lock
);
3382 if (!(wq
->flags
& WQ_FREEZABLE
) ||
3383 !(pool
->flags
& POOL_FREEZING
))
3384 cwq_set_max_active(cwq
, max_active
);
3386 spin_unlock_irq(&pool
->lock
);
3389 spin_unlock(&workqueue_lock
);
3391 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
3394 * workqueue_congested - test whether a workqueue is congested
3395 * @cpu: CPU in question
3396 * @wq: target workqueue
3398 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3399 * no synchronization around this function and the test result is
3400 * unreliable and only useful as advisory hints or for debugging.
3403 * %true if congested, %false otherwise.
3405 bool workqueue_congested(unsigned int cpu
, struct workqueue_struct
*wq
)
3407 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3409 return !list_empty(&cwq
->delayed_works
);
3411 EXPORT_SYMBOL_GPL(workqueue_congested
);
3414 * work_busy - test whether a work is currently pending or running
3415 * @work: the work to be tested
3417 * Test whether @work is currently pending or running. There is no
3418 * synchronization around this function and the test result is
3419 * unreliable and only useful as advisory hints or for debugging.
3422 * OR'd bitmask of WORK_BUSY_* bits.
3424 unsigned int work_busy(struct work_struct
*work
)
3426 struct worker_pool
*pool
= get_work_pool(work
);
3427 unsigned long flags
;
3428 unsigned int ret
= 0;
3430 if (work_pending(work
))
3431 ret
|= WORK_BUSY_PENDING
;
3434 spin_lock_irqsave(&pool
->lock
, flags
);
3435 if (find_worker_executing_work(pool
, work
))
3436 ret
|= WORK_BUSY_RUNNING
;
3437 spin_unlock_irqrestore(&pool
->lock
, flags
);
3442 EXPORT_SYMBOL_GPL(work_busy
);
3447 * There are two challenges in supporting CPU hotplug. Firstly, there
3448 * are a lot of assumptions on strong associations among work, cwq and
3449 * pool which make migrating pending and scheduled works very
3450 * difficult to implement without impacting hot paths. Secondly,
3451 * worker pools serve mix of short, long and very long running works making
3452 * blocked draining impractical.
3454 * This is solved by allowing the pools to be disassociated from the CPU
3455 * running as an unbound one and allowing it to be reattached later if the
3456 * cpu comes back online.
3459 static void wq_unbind_fn(struct work_struct
*work
)
3461 int cpu
= smp_processor_id();
3462 struct worker_pool
*pool
;
3463 struct worker
*worker
;
3464 struct hlist_node
*pos
;
3467 for_each_std_worker_pool(pool
, cpu
) {
3468 BUG_ON(cpu
!= smp_processor_id());
3470 mutex_lock(&pool
->assoc_mutex
);
3471 spin_lock_irq(&pool
->lock
);
3474 * We've claimed all manager positions. Make all workers
3475 * unbound and set DISASSOCIATED. Before this, all workers
3476 * except for the ones which are still executing works from
3477 * before the last CPU down must be on the cpu. After
3478 * this, they may become diasporas.
3480 list_for_each_entry(worker
, &pool
->idle_list
, entry
)
3481 worker
->flags
|= WORKER_UNBOUND
;
3483 for_each_busy_worker(worker
, i
, pos
, pool
)
3484 worker
->flags
|= WORKER_UNBOUND
;
3486 pool
->flags
|= POOL_DISASSOCIATED
;
3488 spin_unlock_irq(&pool
->lock
);
3489 mutex_unlock(&pool
->assoc_mutex
);
3493 * Call schedule() so that we cross rq->lock and thus can guarantee
3494 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3495 * as scheduler callbacks may be invoked from other cpus.
3500 * Sched callbacks are disabled now. Zap nr_running. After this,
3501 * nr_running stays zero and need_more_worker() and keep_working()
3502 * are always true as long as the worklist is not empty. Pools on
3503 * @cpu now behave as unbound (in terms of concurrency management)
3504 * pools which are served by workers tied to the CPU.
3506 * On return from this function, the current worker would trigger
3507 * unbound chain execution of pending work items if other workers
3510 for_each_std_worker_pool(pool
, cpu
)
3511 atomic_set(get_pool_nr_running(pool
), 0);
3515 * Workqueues should be brought up before normal priority CPU notifiers.
3516 * This will be registered high priority CPU notifier.
3518 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
3519 unsigned long action
,
3522 unsigned int cpu
= (unsigned long)hcpu
;
3523 struct worker_pool
*pool
;
3525 switch (action
& ~CPU_TASKS_FROZEN
) {
3526 case CPU_UP_PREPARE
:
3527 for_each_std_worker_pool(pool
, cpu
) {
3528 struct worker
*worker
;
3530 if (pool
->nr_workers
)
3533 worker
= create_worker(pool
);
3537 spin_lock_irq(&pool
->lock
);
3538 start_worker(worker
);
3539 spin_unlock_irq(&pool
->lock
);
3543 case CPU_DOWN_FAILED
:
3545 for_each_std_worker_pool(pool
, cpu
) {
3546 mutex_lock(&pool
->assoc_mutex
);
3547 spin_lock_irq(&pool
->lock
);
3549 pool
->flags
&= ~POOL_DISASSOCIATED
;
3550 rebind_workers(pool
);
3552 spin_unlock_irq(&pool
->lock
);
3553 mutex_unlock(&pool
->assoc_mutex
);
3561 * Workqueues should be brought down after normal priority CPU notifiers.
3562 * This will be registered as low priority CPU notifier.
3564 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
3565 unsigned long action
,
3568 unsigned int cpu
= (unsigned long)hcpu
;
3569 struct work_struct unbind_work
;
3571 switch (action
& ~CPU_TASKS_FROZEN
) {
3572 case CPU_DOWN_PREPARE
:
3573 /* unbinding should happen on the local CPU */
3574 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
3575 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
3576 flush_work(&unbind_work
);
3584 struct work_for_cpu
{
3585 struct work_struct work
;
3591 static void work_for_cpu_fn(struct work_struct
*work
)
3593 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
3595 wfc
->ret
= wfc
->fn(wfc
->arg
);
3599 * work_on_cpu - run a function in user context on a particular cpu
3600 * @cpu: the cpu to run on
3601 * @fn: the function to run
3602 * @arg: the function arg
3604 * This will return the value @fn returns.
3605 * It is up to the caller to ensure that the cpu doesn't go offline.
3606 * The caller must not hold any locks which would prevent @fn from completing.
3608 long work_on_cpu(unsigned int cpu
, long (*fn
)(void *), void *arg
)
3610 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
3612 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
3613 schedule_work_on(cpu
, &wfc
.work
);
3614 flush_work(&wfc
.work
);
3617 EXPORT_SYMBOL_GPL(work_on_cpu
);
3618 #endif /* CONFIG_SMP */
3620 #ifdef CONFIG_FREEZER
3623 * freeze_workqueues_begin - begin freezing workqueues
3625 * Start freezing workqueues. After this function returns, all freezable
3626 * workqueues will queue new works to their frozen_works list instead of
3630 * Grabs and releases workqueue_lock and pool->lock's.
3632 void freeze_workqueues_begin(void)
3636 spin_lock(&workqueue_lock
);
3638 BUG_ON(workqueue_freezing
);
3639 workqueue_freezing
= true;
3641 for_each_wq_cpu(cpu
) {
3642 struct worker_pool
*pool
;
3643 struct workqueue_struct
*wq
;
3645 for_each_std_worker_pool(pool
, cpu
) {
3646 spin_lock_irq(&pool
->lock
);
3648 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
3649 pool
->flags
|= POOL_FREEZING
;
3651 list_for_each_entry(wq
, &workqueues
, list
) {
3652 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3654 if (cwq
&& cwq
->pool
== pool
&&
3655 (wq
->flags
& WQ_FREEZABLE
))
3656 cwq
->max_active
= 0;
3659 spin_unlock_irq(&pool
->lock
);
3663 spin_unlock(&workqueue_lock
);
3667 * freeze_workqueues_busy - are freezable workqueues still busy?
3669 * Check whether freezing is complete. This function must be called
3670 * between freeze_workqueues_begin() and thaw_workqueues().
3673 * Grabs and releases workqueue_lock.
3676 * %true if some freezable workqueues are still busy. %false if freezing
3679 bool freeze_workqueues_busy(void)
3684 spin_lock(&workqueue_lock
);
3686 BUG_ON(!workqueue_freezing
);
3688 for_each_wq_cpu(cpu
) {
3689 struct workqueue_struct
*wq
;
3691 * nr_active is monotonically decreasing. It's safe
3692 * to peek without lock.
3694 list_for_each_entry(wq
, &workqueues
, list
) {
3695 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3697 if (!cwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3700 BUG_ON(cwq
->nr_active
< 0);
3701 if (cwq
->nr_active
) {
3708 spin_unlock(&workqueue_lock
);
3713 * thaw_workqueues - thaw workqueues
3715 * Thaw workqueues. Normal queueing is restored and all collected
3716 * frozen works are transferred to their respective pool worklists.
3719 * Grabs and releases workqueue_lock and pool->lock's.
3721 void thaw_workqueues(void)
3725 spin_lock(&workqueue_lock
);
3727 if (!workqueue_freezing
)
3730 for_each_wq_cpu(cpu
) {
3731 struct worker_pool
*pool
;
3732 struct workqueue_struct
*wq
;
3734 for_each_std_worker_pool(pool
, cpu
) {
3735 spin_lock_irq(&pool
->lock
);
3737 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
3738 pool
->flags
&= ~POOL_FREEZING
;
3740 list_for_each_entry(wq
, &workqueues
, list
) {
3741 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3743 if (!cwq
|| cwq
->pool
!= pool
||
3744 !(wq
->flags
& WQ_FREEZABLE
))
3747 /* restore max_active and repopulate worklist */
3748 cwq_set_max_active(cwq
, wq
->saved_max_active
);
3751 wake_up_worker(pool
);
3753 spin_unlock_irq(&pool
->lock
);
3757 workqueue_freezing
= false;
3759 spin_unlock(&workqueue_lock
);
3761 #endif /* CONFIG_FREEZER */
3763 static int __init
init_workqueues(void)
3767 /* make sure we have enough bits for OFFQ pool ID */
3768 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_POOL_SHIFT
)) <
3769 WORK_CPU_END
* NR_STD_WORKER_POOLS
);
3771 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
3772 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
3774 /* initialize CPU pools */
3775 for_each_wq_cpu(cpu
) {
3776 struct worker_pool
*pool
;
3778 for_each_std_worker_pool(pool
, cpu
) {
3779 spin_lock_init(&pool
->lock
);
3781 pool
->flags
|= POOL_DISASSOCIATED
;
3782 INIT_LIST_HEAD(&pool
->worklist
);
3783 INIT_LIST_HEAD(&pool
->idle_list
);
3784 hash_init(pool
->busy_hash
);
3786 init_timer_deferrable(&pool
->idle_timer
);
3787 pool
->idle_timer
.function
= idle_worker_timeout
;
3788 pool
->idle_timer
.data
= (unsigned long)pool
;
3790 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3791 (unsigned long)pool
);
3793 mutex_init(&pool
->assoc_mutex
);
3794 ida_init(&pool
->worker_ida
);
3797 BUG_ON(worker_pool_assign_id(pool
));
3801 /* create the initial worker */
3802 for_each_online_wq_cpu(cpu
) {
3803 struct worker_pool
*pool
;
3805 for_each_std_worker_pool(pool
, cpu
) {
3806 struct worker
*worker
;
3808 if (cpu
!= WORK_CPU_UNBOUND
)
3809 pool
->flags
&= ~POOL_DISASSOCIATED
;
3811 worker
= create_worker(pool
);
3813 spin_lock_irq(&pool
->lock
);
3814 start_worker(worker
);
3815 spin_unlock_irq(&pool
->lock
);
3819 system_wq
= alloc_workqueue("events", 0, 0);
3820 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
3821 system_long_wq
= alloc_workqueue("events_long", 0, 0);
3822 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
3823 WQ_UNBOUND_MAX_ACTIVE
);
3824 system_freezable_wq
= alloc_workqueue("events_freezable",
3826 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
3827 !system_unbound_wq
|| !system_freezable_wq
);
3830 early_initcall(init_workqueues
);