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 */
149 * The current concurrency level. As it's likely to be accessed
150 * from other CPUs during try_to_wake_up(), put it in a separate
153 atomic_t nr_running ____cacheline_aligned_in_smp
;
154 } ____cacheline_aligned_in_smp
;
157 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
158 * of work_struct->data are used for flags and the remaining high bits
159 * point to the pwq; thus, pwqs need to be aligned at two's power of the
160 * number of flag bits.
162 struct pool_workqueue
{
163 struct worker_pool
*pool
; /* I: the associated pool */
164 struct workqueue_struct
*wq
; /* I: the owning workqueue */
165 int work_color
; /* L: current color */
166 int flush_color
; /* L: flushing color */
167 int nr_in_flight
[WORK_NR_COLORS
];
168 /* L: nr of in_flight works */
169 int nr_active
; /* L: nr of active works */
170 int max_active
; /* L: max active works */
171 struct list_head delayed_works
; /* L: delayed works */
175 * Structure used to wait for workqueue flush.
178 struct list_head list
; /* F: list of flushers */
179 int flush_color
; /* F: flush color waiting for */
180 struct completion done
; /* flush completion */
184 * All cpumasks are assumed to be always set on UP and thus can't be
185 * used to determine whether there's something to be done.
188 typedef cpumask_var_t mayday_mask_t
;
189 #define mayday_test_and_set_cpu(cpu, mask) \
190 cpumask_test_and_set_cpu((cpu), (mask))
191 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
192 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
193 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
194 #define free_mayday_mask(mask) free_cpumask_var((mask))
196 typedef unsigned long mayday_mask_t
;
197 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
198 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
199 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
200 #define alloc_mayday_mask(maskp, gfp) true
201 #define free_mayday_mask(mask) do { } while (0)
205 * The externally visible workqueue abstraction is an array of
206 * per-CPU workqueues:
208 struct workqueue_struct
{
209 unsigned int flags
; /* W: WQ_* flags */
211 struct pool_workqueue __percpu
*pcpu
;
212 struct pool_workqueue
*single
;
214 } pool_wq
; /* I: pwq's */
215 struct list_head list
; /* W: list of all workqueues */
217 struct mutex flush_mutex
; /* protects wq flushing */
218 int work_color
; /* F: current work color */
219 int flush_color
; /* F: current flush color */
220 atomic_t nr_pwqs_to_flush
; /* flush in progress */
221 struct wq_flusher
*first_flusher
; /* F: first flusher */
222 struct list_head flusher_queue
; /* F: flush waiters */
223 struct list_head flusher_overflow
; /* F: flush overflow list */
225 mayday_mask_t mayday_mask
; /* cpus requesting rescue */
226 struct worker
*rescuer
; /* I: rescue worker */
228 int nr_drainers
; /* W: drain in progress */
229 int saved_max_active
; /* W: saved pwq max_active */
230 #ifdef CONFIG_LOCKDEP
231 struct lockdep_map lockdep_map
;
233 char name
[]; /* I: workqueue name */
236 struct workqueue_struct
*system_wq __read_mostly
;
237 EXPORT_SYMBOL_GPL(system_wq
);
238 struct workqueue_struct
*system_highpri_wq __read_mostly
;
239 EXPORT_SYMBOL_GPL(system_highpri_wq
);
240 struct workqueue_struct
*system_long_wq __read_mostly
;
241 EXPORT_SYMBOL_GPL(system_long_wq
);
242 struct workqueue_struct
*system_unbound_wq __read_mostly
;
243 EXPORT_SYMBOL_GPL(system_unbound_wq
);
244 struct workqueue_struct
*system_freezable_wq __read_mostly
;
245 EXPORT_SYMBOL_GPL(system_freezable_wq
);
247 #define CREATE_TRACE_POINTS
248 #include <trace/events/workqueue.h>
250 #define for_each_std_worker_pool(pool, cpu) \
251 for ((pool) = &std_worker_pools(cpu)[0]; \
252 (pool) < &std_worker_pools(cpu)[NR_STD_WORKER_POOLS]; (pool)++)
254 #define for_each_busy_worker(worker, i, pos, pool) \
255 hash_for_each(pool->busy_hash, i, pos, worker, hentry)
257 static inline int __next_wq_cpu(int cpu
, const struct cpumask
*mask
,
260 if (cpu
< nr_cpu_ids
) {
262 cpu
= cpumask_next(cpu
, mask
);
263 if (cpu
< nr_cpu_ids
)
267 return WORK_CPU_UNBOUND
;
272 static inline int __next_pwq_cpu(int cpu
, const struct cpumask
*mask
,
273 struct workqueue_struct
*wq
)
275 return __next_wq_cpu(cpu
, mask
, !(wq
->flags
& WQ_UNBOUND
) ? 1 : 2);
281 * An extra cpu number is defined using an invalid cpu number
282 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
283 * specific CPU. The following iterators are similar to for_each_*_cpu()
284 * iterators but also considers the unbound CPU.
286 * for_each_wq_cpu() : possible CPUs + WORK_CPU_UNBOUND
287 * for_each_online_wq_cpu() : online CPUs + WORK_CPU_UNBOUND
288 * for_each_pwq_cpu() : possible CPUs for bound workqueues,
289 * WORK_CPU_UNBOUND for unbound workqueues
291 #define for_each_wq_cpu(cpu) \
292 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, 3); \
293 (cpu) < WORK_CPU_END; \
294 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, 3))
296 #define for_each_online_wq_cpu(cpu) \
297 for ((cpu) = __next_wq_cpu(-1, cpu_online_mask, 3); \
298 (cpu) < WORK_CPU_END; \
299 (cpu) = __next_wq_cpu((cpu), cpu_online_mask, 3))
301 #define for_each_pwq_cpu(cpu, wq) \
302 for ((cpu) = __next_pwq_cpu(-1, cpu_possible_mask, (wq)); \
303 (cpu) < WORK_CPU_END; \
304 (cpu) = __next_pwq_cpu((cpu), cpu_possible_mask, (wq)))
306 #ifdef CONFIG_DEBUG_OBJECTS_WORK
308 static struct debug_obj_descr work_debug_descr
;
310 static void *work_debug_hint(void *addr
)
312 return ((struct work_struct
*) addr
)->func
;
316 * fixup_init is called when:
317 * - an active object is initialized
319 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
321 struct work_struct
*work
= addr
;
324 case ODEBUG_STATE_ACTIVE
:
325 cancel_work_sync(work
);
326 debug_object_init(work
, &work_debug_descr
);
334 * fixup_activate is called when:
335 * - an active object is activated
336 * - an unknown object is activated (might be a statically initialized object)
338 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
340 struct work_struct
*work
= addr
;
344 case ODEBUG_STATE_NOTAVAILABLE
:
346 * This is not really a fixup. The work struct was
347 * statically initialized. We just make sure that it
348 * is tracked in the object tracker.
350 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
351 debug_object_init(work
, &work_debug_descr
);
352 debug_object_activate(work
, &work_debug_descr
);
358 case ODEBUG_STATE_ACTIVE
:
367 * fixup_free is called when:
368 * - an active object is freed
370 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
372 struct work_struct
*work
= addr
;
375 case ODEBUG_STATE_ACTIVE
:
376 cancel_work_sync(work
);
377 debug_object_free(work
, &work_debug_descr
);
384 static struct debug_obj_descr work_debug_descr
= {
385 .name
= "work_struct",
386 .debug_hint
= work_debug_hint
,
387 .fixup_init
= work_fixup_init
,
388 .fixup_activate
= work_fixup_activate
,
389 .fixup_free
= work_fixup_free
,
392 static inline void debug_work_activate(struct work_struct
*work
)
394 debug_object_activate(work
, &work_debug_descr
);
397 static inline void debug_work_deactivate(struct work_struct
*work
)
399 debug_object_deactivate(work
, &work_debug_descr
);
402 void __init_work(struct work_struct
*work
, int onstack
)
405 debug_object_init_on_stack(work
, &work_debug_descr
);
407 debug_object_init(work
, &work_debug_descr
);
409 EXPORT_SYMBOL_GPL(__init_work
);
411 void destroy_work_on_stack(struct work_struct
*work
)
413 debug_object_free(work
, &work_debug_descr
);
415 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
418 static inline void debug_work_activate(struct work_struct
*work
) { }
419 static inline void debug_work_deactivate(struct work_struct
*work
) { }
422 /* Serializes the accesses to the list of workqueues. */
423 static DEFINE_SPINLOCK(workqueue_lock
);
424 static LIST_HEAD(workqueues
);
425 static bool workqueue_freezing
; /* W: have wqs started freezing? */
428 * The CPU and unbound standard worker pools. The unbound ones have
429 * POOL_DISASSOCIATED set, and their workers have WORKER_UNBOUND set.
431 static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool
[NR_STD_WORKER_POOLS
],
432 cpu_std_worker_pools
);
433 static struct worker_pool unbound_std_worker_pools
[NR_STD_WORKER_POOLS
];
435 /* idr of all pools */
436 static DEFINE_MUTEX(worker_pool_idr_mutex
);
437 static DEFINE_IDR(worker_pool_idr
);
439 static int worker_thread(void *__worker
);
441 static struct worker_pool
*std_worker_pools(int cpu
)
443 if (cpu
!= WORK_CPU_UNBOUND
)
444 return per_cpu(cpu_std_worker_pools
, cpu
);
446 return unbound_std_worker_pools
;
449 static int std_worker_pool_pri(struct worker_pool
*pool
)
451 return pool
- std_worker_pools(pool
->cpu
);
454 /* allocate ID and assign it to @pool */
455 static int worker_pool_assign_id(struct worker_pool
*pool
)
459 mutex_lock(&worker_pool_idr_mutex
);
460 idr_pre_get(&worker_pool_idr
, GFP_KERNEL
);
461 ret
= idr_get_new(&worker_pool_idr
, pool
, &pool
->id
);
462 mutex_unlock(&worker_pool_idr_mutex
);
468 * Lookup worker_pool by id. The idr currently is built during boot and
469 * never modified. Don't worry about locking for now.
471 static struct worker_pool
*worker_pool_by_id(int pool_id
)
473 return idr_find(&worker_pool_idr
, pool_id
);
476 static struct worker_pool
*get_std_worker_pool(int cpu
, bool highpri
)
478 struct worker_pool
*pools
= std_worker_pools(cpu
);
480 return &pools
[highpri
];
483 static struct pool_workqueue
*get_pwq(unsigned int cpu
,
484 struct workqueue_struct
*wq
)
486 if (!(wq
->flags
& WQ_UNBOUND
)) {
487 if (likely(cpu
< nr_cpu_ids
))
488 return per_cpu_ptr(wq
->pool_wq
.pcpu
, cpu
);
489 } else if (likely(cpu
== WORK_CPU_UNBOUND
))
490 return wq
->pool_wq
.single
;
494 static unsigned int work_color_to_flags(int color
)
496 return color
<< WORK_STRUCT_COLOR_SHIFT
;
499 static int get_work_color(struct work_struct
*work
)
501 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
502 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
505 static int work_next_color(int color
)
507 return (color
+ 1) % WORK_NR_COLORS
;
511 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
512 * contain the pointer to the queued pwq. Once execution starts, the flag
513 * is cleared and the high bits contain OFFQ flags and pool ID.
515 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
516 * and clear_work_data() can be used to set the pwq, pool or clear
517 * work->data. These functions should only be called while the work is
518 * owned - ie. while the PENDING bit is set.
520 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
521 * corresponding to a work. Pool is available once the work has been
522 * queued anywhere after initialization until it is sync canceled. pwq is
523 * available only while the work item is queued.
525 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
526 * canceled. While being canceled, a work item may have its PENDING set
527 * but stay off timer and worklist for arbitrarily long and nobody should
528 * try to steal the PENDING bit.
530 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
533 BUG_ON(!work_pending(work
));
534 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
537 static void set_work_pwq(struct work_struct
*work
, struct pool_workqueue
*pwq
,
538 unsigned long extra_flags
)
540 set_work_data(work
, (unsigned long)pwq
,
541 WORK_STRUCT_PENDING
| WORK_STRUCT_PWQ
| extra_flags
);
544 static void set_work_pool_and_keep_pending(struct work_struct
*work
,
547 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
,
548 WORK_STRUCT_PENDING
);
551 static void set_work_pool_and_clear_pending(struct work_struct
*work
,
555 * The following wmb is paired with the implied mb in
556 * test_and_set_bit(PENDING) and ensures all updates to @work made
557 * here are visible to and precede any updates by the next PENDING
561 set_work_data(work
, (unsigned long)pool_id
<< WORK_OFFQ_POOL_SHIFT
, 0);
564 static void clear_work_data(struct work_struct
*work
)
566 smp_wmb(); /* see set_work_pool_and_clear_pending() */
567 set_work_data(work
, WORK_STRUCT_NO_POOL
, 0);
570 static struct pool_workqueue
*get_work_pwq(struct work_struct
*work
)
572 unsigned long data
= atomic_long_read(&work
->data
);
574 if (data
& WORK_STRUCT_PWQ
)
575 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
581 * get_work_pool - return the worker_pool a given work was associated with
582 * @work: the work item of interest
584 * Return the worker_pool @work was last associated with. %NULL if none.
586 static struct worker_pool
*get_work_pool(struct work_struct
*work
)
588 unsigned long data
= atomic_long_read(&work
->data
);
589 struct worker_pool
*pool
;
592 if (data
& WORK_STRUCT_PWQ
)
593 return ((struct pool_workqueue
*)
594 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
;
596 pool_id
= data
>> WORK_OFFQ_POOL_SHIFT
;
597 if (pool_id
== WORK_OFFQ_POOL_NONE
)
600 pool
= worker_pool_by_id(pool_id
);
606 * get_work_pool_id - return the worker pool ID a given work is associated with
607 * @work: the work item of interest
609 * Return the worker_pool ID @work was last associated with.
610 * %WORK_OFFQ_POOL_NONE if none.
612 static int get_work_pool_id(struct work_struct
*work
)
614 unsigned long data
= atomic_long_read(&work
->data
);
616 if (data
& WORK_STRUCT_PWQ
)
617 return ((struct pool_workqueue
*)
618 (data
& WORK_STRUCT_WQ_DATA_MASK
))->pool
->id
;
620 return data
>> WORK_OFFQ_POOL_SHIFT
;
623 static void mark_work_canceling(struct work_struct
*work
)
625 unsigned long pool_id
= get_work_pool_id(work
);
627 pool_id
<<= WORK_OFFQ_POOL_SHIFT
;
628 set_work_data(work
, pool_id
| WORK_OFFQ_CANCELING
, WORK_STRUCT_PENDING
);
631 static bool work_is_canceling(struct work_struct
*work
)
633 unsigned long data
= atomic_long_read(&work
->data
);
635 return !(data
& WORK_STRUCT_PWQ
) && (data
& WORK_OFFQ_CANCELING
);
639 * Policy functions. These define the policies on how the global worker
640 * pools are managed. Unless noted otherwise, these functions assume that
641 * they're being called with pool->lock held.
644 static bool __need_more_worker(struct worker_pool
*pool
)
646 return !atomic_read(&pool
->nr_running
);
650 * Need to wake up a worker? Called from anything but currently
653 * Note that, because unbound workers never contribute to nr_running, this
654 * function will always return %true for unbound pools as long as the
655 * worklist isn't empty.
657 static bool need_more_worker(struct worker_pool
*pool
)
659 return !list_empty(&pool
->worklist
) && __need_more_worker(pool
);
662 /* Can I start working? Called from busy but !running workers. */
663 static bool may_start_working(struct worker_pool
*pool
)
665 return pool
->nr_idle
;
668 /* Do I need to keep working? Called from currently running workers. */
669 static bool keep_working(struct worker_pool
*pool
)
671 return !list_empty(&pool
->worklist
) &&
672 atomic_read(&pool
->nr_running
) <= 1;
675 /* Do we need a new worker? Called from manager. */
676 static bool need_to_create_worker(struct worker_pool
*pool
)
678 return need_more_worker(pool
) && !may_start_working(pool
);
681 /* Do I need to be the manager? */
682 static bool need_to_manage_workers(struct worker_pool
*pool
)
684 return need_to_create_worker(pool
) ||
685 (pool
->flags
& POOL_MANAGE_WORKERS
);
688 /* Do we have too many workers and should some go away? */
689 static bool too_many_workers(struct worker_pool
*pool
)
691 bool managing
= pool
->flags
& POOL_MANAGING_WORKERS
;
692 int nr_idle
= pool
->nr_idle
+ managing
; /* manager is considered idle */
693 int nr_busy
= pool
->nr_workers
- nr_idle
;
696 * nr_idle and idle_list may disagree if idle rebinding is in
697 * progress. Never return %true if idle_list is empty.
699 if (list_empty(&pool
->idle_list
))
702 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
709 /* Return the first worker. Safe with preemption disabled */
710 static struct worker
*first_worker(struct worker_pool
*pool
)
712 if (unlikely(list_empty(&pool
->idle_list
)))
715 return list_first_entry(&pool
->idle_list
, struct worker
, entry
);
719 * wake_up_worker - wake up an idle worker
720 * @pool: worker pool to wake worker from
722 * Wake up the first idle worker of @pool.
725 * spin_lock_irq(pool->lock).
727 static void wake_up_worker(struct worker_pool
*pool
)
729 struct worker
*worker
= first_worker(pool
);
732 wake_up_process(worker
->task
);
736 * wq_worker_waking_up - a worker is waking up
737 * @task: task waking up
738 * @cpu: CPU @task is waking up to
740 * This function is called during try_to_wake_up() when a worker is
744 * spin_lock_irq(rq->lock)
746 void wq_worker_waking_up(struct task_struct
*task
, unsigned int cpu
)
748 struct worker
*worker
= kthread_data(task
);
750 if (!(worker
->flags
& WORKER_NOT_RUNNING
)) {
751 WARN_ON_ONCE(worker
->pool
->cpu
!= cpu
);
752 atomic_inc(&worker
->pool
->nr_running
);
757 * wq_worker_sleeping - a worker is going to sleep
758 * @task: task going to sleep
759 * @cpu: CPU in question, must be the current CPU number
761 * This function is called during schedule() when a busy worker is
762 * going to sleep. Worker on the same cpu can be woken up by
763 * returning pointer to its task.
766 * spin_lock_irq(rq->lock)
769 * Worker task on @cpu to wake up, %NULL if none.
771 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
,
774 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
775 struct worker_pool
*pool
;
778 * Rescuers, which may not have all the fields set up like normal
779 * workers, also reach here, let's not access anything before
780 * checking NOT_RUNNING.
782 if (worker
->flags
& WORKER_NOT_RUNNING
)
787 /* this can only happen on the local cpu */
788 BUG_ON(cpu
!= raw_smp_processor_id());
791 * The counterpart of the following dec_and_test, implied mb,
792 * worklist not empty test sequence is in insert_work().
793 * Please read comment there.
795 * NOT_RUNNING is clear. This means that we're bound to and
796 * running on the local cpu w/ rq lock held and preemption
797 * disabled, which in turn means that none else could be
798 * manipulating idle_list, so dereferencing idle_list without pool
801 if (atomic_dec_and_test(&pool
->nr_running
) &&
802 !list_empty(&pool
->worklist
))
803 to_wakeup
= first_worker(pool
);
804 return to_wakeup
? to_wakeup
->task
: NULL
;
808 * worker_set_flags - set worker flags and adjust nr_running accordingly
810 * @flags: flags to set
811 * @wakeup: wakeup an idle worker if necessary
813 * Set @flags in @worker->flags and adjust nr_running accordingly. If
814 * nr_running becomes zero and @wakeup is %true, an idle worker is
818 * spin_lock_irq(pool->lock)
820 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
823 struct worker_pool
*pool
= worker
->pool
;
825 WARN_ON_ONCE(worker
->task
!= current
);
828 * If transitioning into NOT_RUNNING, adjust nr_running and
829 * wake up an idle worker as necessary if requested by
832 if ((flags
& WORKER_NOT_RUNNING
) &&
833 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
835 if (atomic_dec_and_test(&pool
->nr_running
) &&
836 !list_empty(&pool
->worklist
))
837 wake_up_worker(pool
);
839 atomic_dec(&pool
->nr_running
);
842 worker
->flags
|= flags
;
846 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
848 * @flags: flags to clear
850 * Clear @flags in @worker->flags and adjust nr_running accordingly.
853 * spin_lock_irq(pool->lock)
855 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
857 struct worker_pool
*pool
= worker
->pool
;
858 unsigned int oflags
= worker
->flags
;
860 WARN_ON_ONCE(worker
->task
!= current
);
862 worker
->flags
&= ~flags
;
865 * If transitioning out of NOT_RUNNING, increment nr_running. Note
866 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
867 * of multiple flags, not a single flag.
869 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
870 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
871 atomic_inc(&pool
->nr_running
);
875 * find_worker_executing_work - find worker which is executing a work
876 * @pool: pool of interest
877 * @work: work to find worker for
879 * Find a worker which is executing @work on @pool by searching
880 * @pool->busy_hash which is keyed by the address of @work. For a worker
881 * to match, its current execution should match the address of @work and
882 * its work function. This is to avoid unwanted dependency between
883 * unrelated work executions through a work item being recycled while still
886 * This is a bit tricky. A work item may be freed once its execution
887 * starts and nothing prevents the freed area from being recycled for
888 * another work item. If the same work item address ends up being reused
889 * before the original execution finishes, workqueue will identify the
890 * recycled work item as currently executing and make it wait until the
891 * current execution finishes, introducing an unwanted dependency.
893 * This function checks the work item address, work function and workqueue
894 * to avoid false positives. Note that this isn't complete as one may
895 * construct a work function which can introduce dependency onto itself
896 * through a recycled work item. Well, if somebody wants to shoot oneself
897 * in the foot that badly, there's only so much we can do, and if such
898 * deadlock actually occurs, it should be easy to locate the culprit work
902 * spin_lock_irq(pool->lock).
905 * Pointer to worker which is executing @work if found, NULL
908 static struct worker
*find_worker_executing_work(struct worker_pool
*pool
,
909 struct work_struct
*work
)
911 struct worker
*worker
;
912 struct hlist_node
*tmp
;
914 hash_for_each_possible(pool
->busy_hash
, worker
, tmp
, hentry
,
916 if (worker
->current_work
== work
&&
917 worker
->current_func
== work
->func
)
924 * move_linked_works - move linked works to a list
925 * @work: start of series of works to be scheduled
926 * @head: target list to append @work to
927 * @nextp: out paramter for nested worklist walking
929 * Schedule linked works starting from @work to @head. Work series to
930 * be scheduled starts at @work and includes any consecutive work with
931 * WORK_STRUCT_LINKED set in its predecessor.
933 * If @nextp is not NULL, it's updated to point to the next work of
934 * the last scheduled work. This allows move_linked_works() to be
935 * nested inside outer list_for_each_entry_safe().
938 * spin_lock_irq(pool->lock).
940 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
941 struct work_struct
**nextp
)
943 struct work_struct
*n
;
946 * Linked worklist will always end before the end of the list,
947 * use NULL for list head.
949 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
950 list_move_tail(&work
->entry
, head
);
951 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
956 * If we're already inside safe list traversal and have moved
957 * multiple works to the scheduled queue, the next position
958 * needs to be updated.
964 static void pwq_activate_delayed_work(struct work_struct
*work
)
966 struct pool_workqueue
*pwq
= get_work_pwq(work
);
968 trace_workqueue_activate_work(work
);
969 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
970 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
974 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
976 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
977 struct work_struct
, entry
);
979 pwq_activate_delayed_work(work
);
983 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
984 * @pwq: pwq of interest
985 * @color: color of work which left the queue
987 * A work either has completed or is removed from pending queue,
988 * decrement nr_in_flight of its pwq and handle workqueue flushing.
991 * spin_lock_irq(pool->lock).
993 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
995 /* ignore uncolored works */
996 if (color
== WORK_NO_COLOR
)
999 pwq
->nr_in_flight
[color
]--;
1002 if (!list_empty(&pwq
->delayed_works
)) {
1003 /* one down, submit a delayed one */
1004 if (pwq
->nr_active
< pwq
->max_active
)
1005 pwq_activate_first_delayed(pwq
);
1008 /* is flush in progress and are we at the flushing tip? */
1009 if (likely(pwq
->flush_color
!= color
))
1012 /* are there still in-flight works? */
1013 if (pwq
->nr_in_flight
[color
])
1016 /* this pwq is done, clear flush_color */
1017 pwq
->flush_color
= -1;
1020 * If this was the last pwq, wake up the first flusher. It
1021 * will handle the rest.
1023 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1024 complete(&pwq
->wq
->first_flusher
->done
);
1028 * try_to_grab_pending - steal work item from worklist and disable irq
1029 * @work: work item to steal
1030 * @is_dwork: @work is a delayed_work
1031 * @flags: place to store irq state
1033 * Try to grab PENDING bit of @work. This function can handle @work in any
1034 * stable state - idle, on timer or on worklist. Return values are
1036 * 1 if @work was pending and we successfully stole PENDING
1037 * 0 if @work was idle and we claimed PENDING
1038 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1039 * -ENOENT if someone else is canceling @work, this state may persist
1040 * for arbitrarily long
1042 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1043 * interrupted while holding PENDING and @work off queue, irq must be
1044 * disabled on entry. This, combined with delayed_work->timer being
1045 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1047 * On successful return, >= 0, irq is disabled and the caller is
1048 * responsible for releasing it using local_irq_restore(*@flags).
1050 * This function is safe to call from any context including IRQ handler.
1052 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1053 unsigned long *flags
)
1055 struct worker_pool
*pool
;
1056 struct pool_workqueue
*pwq
;
1058 local_irq_save(*flags
);
1060 /* try to steal the timer if it exists */
1062 struct delayed_work
*dwork
= to_delayed_work(work
);
1065 * dwork->timer is irqsafe. If del_timer() fails, it's
1066 * guaranteed that the timer is not queued anywhere and not
1067 * running on the local CPU.
1069 if (likely(del_timer(&dwork
->timer
)))
1073 /* try to claim PENDING the normal way */
1074 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1078 * The queueing is in progress, or it is already queued. Try to
1079 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1081 pool
= get_work_pool(work
);
1085 spin_lock(&pool
->lock
);
1087 * work->data is guaranteed to point to pwq only while the work
1088 * item is queued on pwq->wq, and both updating work->data to point
1089 * to pwq on queueing and to pool on dequeueing are done under
1090 * pwq->pool->lock. This in turn guarantees that, if work->data
1091 * points to pwq which is associated with a locked pool, the work
1092 * item is currently queued on that pool.
1094 pwq
= get_work_pwq(work
);
1095 if (pwq
&& pwq
->pool
== pool
) {
1096 debug_work_deactivate(work
);
1099 * A delayed work item cannot be grabbed directly because
1100 * it might have linked NO_COLOR work items which, if left
1101 * on the delayed_list, will confuse pwq->nr_active
1102 * management later on and cause stall. Make sure the work
1103 * item is activated before grabbing.
1105 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1106 pwq_activate_delayed_work(work
);
1108 list_del_init(&work
->entry
);
1109 pwq_dec_nr_in_flight(get_work_pwq(work
), get_work_color(work
));
1111 /* work->data points to pwq iff queued, point to pool */
1112 set_work_pool_and_keep_pending(work
, pool
->id
);
1114 spin_unlock(&pool
->lock
);
1117 spin_unlock(&pool
->lock
);
1119 local_irq_restore(*flags
);
1120 if (work_is_canceling(work
))
1127 * insert_work - insert a work into a pool
1128 * @pwq: pwq @work belongs to
1129 * @work: work to insert
1130 * @head: insertion point
1131 * @extra_flags: extra WORK_STRUCT_* flags to set
1133 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1134 * work_struct flags.
1137 * spin_lock_irq(pool->lock).
1139 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1140 struct list_head
*head
, unsigned int extra_flags
)
1142 struct worker_pool
*pool
= pwq
->pool
;
1144 /* we own @work, set data and link */
1145 set_work_pwq(work
, pwq
, extra_flags
);
1146 list_add_tail(&work
->entry
, head
);
1149 * Ensure either worker_sched_deactivated() sees the above
1150 * list_add_tail() or we see zero nr_running to avoid workers
1151 * lying around lazily while there are works to be processed.
1155 if (__need_more_worker(pool
))
1156 wake_up_worker(pool
);
1160 * Test whether @work is being queued from another work executing on the
1163 static bool is_chained_work(struct workqueue_struct
*wq
)
1165 struct worker
*worker
;
1167 worker
= current_wq_worker();
1169 * Return %true iff I'm a worker execuing a work item on @wq. If
1170 * I'm @worker, it's safe to dereference it without locking.
1172 return worker
&& worker
->current_pwq
->wq
== wq
;
1175 static void __queue_work(unsigned int cpu
, struct workqueue_struct
*wq
,
1176 struct work_struct
*work
)
1178 struct pool_workqueue
*pwq
;
1179 struct list_head
*worklist
;
1180 unsigned int work_flags
;
1181 unsigned int req_cpu
= cpu
;
1184 * While a work item is PENDING && off queue, a task trying to
1185 * steal the PENDING will busy-loop waiting for it to either get
1186 * queued or lose PENDING. Grabbing PENDING and queueing should
1187 * happen with IRQ disabled.
1189 WARN_ON_ONCE(!irqs_disabled());
1191 debug_work_activate(work
);
1193 /* if dying, only works from the same workqueue are allowed */
1194 if (unlikely(wq
->flags
& WQ_DRAINING
) &&
1195 WARN_ON_ONCE(!is_chained_work(wq
)))
1198 /* determine the pwq to use */
1199 if (!(wq
->flags
& WQ_UNBOUND
)) {
1200 struct worker_pool
*last_pool
;
1202 if (cpu
== WORK_CPU_UNBOUND
)
1203 cpu
= raw_smp_processor_id();
1206 * It's multi cpu. If @work was previously on a different
1207 * cpu, it might still be running there, in which case the
1208 * work needs to be queued on that cpu to guarantee
1211 pwq
= get_pwq(cpu
, wq
);
1212 last_pool
= get_work_pool(work
);
1214 if (last_pool
&& last_pool
!= pwq
->pool
) {
1215 struct worker
*worker
;
1217 spin_lock(&last_pool
->lock
);
1219 worker
= find_worker_executing_work(last_pool
, work
);
1221 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1222 pwq
= get_pwq(last_pool
->cpu
, wq
);
1224 /* meh... not running there, queue here */
1225 spin_unlock(&last_pool
->lock
);
1226 spin_lock(&pwq
->pool
->lock
);
1229 spin_lock(&pwq
->pool
->lock
);
1232 pwq
= get_pwq(WORK_CPU_UNBOUND
, wq
);
1233 spin_lock(&pwq
->pool
->lock
);
1236 /* pwq determined, queue */
1237 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1239 if (WARN_ON(!list_empty(&work
->entry
))) {
1240 spin_unlock(&pwq
->pool
->lock
);
1244 pwq
->nr_in_flight
[pwq
->work_color
]++;
1245 work_flags
= work_color_to_flags(pwq
->work_color
);
1247 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1248 trace_workqueue_activate_work(work
);
1250 worklist
= &pwq
->pool
->worklist
;
1252 work_flags
|= WORK_STRUCT_DELAYED
;
1253 worklist
= &pwq
->delayed_works
;
1256 insert_work(pwq
, work
, worklist
, work_flags
);
1258 spin_unlock(&pwq
->pool
->lock
);
1262 * queue_work_on - queue work on specific cpu
1263 * @cpu: CPU number to execute work on
1264 * @wq: workqueue to use
1265 * @work: work to queue
1267 * Returns %false if @work was already on a queue, %true otherwise.
1269 * We queue the work to a specific CPU, the caller must ensure it
1272 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1273 struct work_struct
*work
)
1276 unsigned long flags
;
1278 local_irq_save(flags
);
1280 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1281 __queue_work(cpu
, wq
, work
);
1285 local_irq_restore(flags
);
1288 EXPORT_SYMBOL_GPL(queue_work_on
);
1291 * queue_work - queue work on a workqueue
1292 * @wq: workqueue to use
1293 * @work: work to queue
1295 * Returns %false if @work was already on a queue, %true otherwise.
1297 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1298 * it can be processed by another CPU.
1300 bool queue_work(struct workqueue_struct
*wq
, struct work_struct
*work
)
1302 return queue_work_on(WORK_CPU_UNBOUND
, wq
, work
);
1304 EXPORT_SYMBOL_GPL(queue_work
);
1306 void delayed_work_timer_fn(unsigned long __data
)
1308 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1310 /* should have been called from irqsafe timer with irq already off */
1311 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1313 EXPORT_SYMBOL(delayed_work_timer_fn
);
1315 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1316 struct delayed_work
*dwork
, unsigned long delay
)
1318 struct timer_list
*timer
= &dwork
->timer
;
1319 struct work_struct
*work
= &dwork
->work
;
1321 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1322 timer
->data
!= (unsigned long)dwork
);
1323 WARN_ON_ONCE(timer_pending(timer
));
1324 WARN_ON_ONCE(!list_empty(&work
->entry
));
1327 * If @delay is 0, queue @dwork->work immediately. This is for
1328 * both optimization and correctness. The earliest @timer can
1329 * expire is on the closest next tick and delayed_work users depend
1330 * on that there's no such delay when @delay is 0.
1333 __queue_work(cpu
, wq
, &dwork
->work
);
1337 timer_stats_timer_set_start_info(&dwork
->timer
);
1341 timer
->expires
= jiffies
+ delay
;
1343 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1344 add_timer_on(timer
, cpu
);
1350 * queue_delayed_work_on - queue work on specific CPU after delay
1351 * @cpu: CPU number to execute work on
1352 * @wq: workqueue to use
1353 * @dwork: work to queue
1354 * @delay: number of jiffies to wait before queueing
1356 * Returns %false if @work was already on a queue, %true otherwise. If
1357 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1360 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1361 struct delayed_work
*dwork
, unsigned long delay
)
1363 struct work_struct
*work
= &dwork
->work
;
1365 unsigned long flags
;
1367 /* read the comment in __queue_work() */
1368 local_irq_save(flags
);
1370 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1371 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1375 local_irq_restore(flags
);
1378 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1381 * queue_delayed_work - queue work on a workqueue after delay
1382 * @wq: workqueue to use
1383 * @dwork: delayable work to queue
1384 * @delay: number of jiffies to wait before queueing
1386 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1388 bool queue_delayed_work(struct workqueue_struct
*wq
,
1389 struct delayed_work
*dwork
, unsigned long delay
)
1391 return queue_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1393 EXPORT_SYMBOL_GPL(queue_delayed_work
);
1396 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1397 * @cpu: CPU number to execute work on
1398 * @wq: workqueue to use
1399 * @dwork: work to queue
1400 * @delay: number of jiffies to wait before queueing
1402 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1403 * modify @dwork's timer so that it expires after @delay. If @delay is
1404 * zero, @work is guaranteed to be scheduled immediately regardless of its
1407 * Returns %false if @dwork was idle and queued, %true if @dwork was
1408 * pending and its timer was modified.
1410 * This function is safe to call from any context including IRQ handler.
1411 * See try_to_grab_pending() for details.
1413 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1414 struct delayed_work
*dwork
, unsigned long delay
)
1416 unsigned long flags
;
1420 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1421 } while (unlikely(ret
== -EAGAIN
));
1423 if (likely(ret
>= 0)) {
1424 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1425 local_irq_restore(flags
);
1428 /* -ENOENT from try_to_grab_pending() becomes %true */
1431 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1434 * mod_delayed_work - modify delay of or queue a delayed work
1435 * @wq: workqueue to use
1436 * @dwork: work to queue
1437 * @delay: number of jiffies to wait before queueing
1439 * mod_delayed_work_on() on local CPU.
1441 bool mod_delayed_work(struct workqueue_struct
*wq
, struct delayed_work
*dwork
,
1442 unsigned long delay
)
1444 return mod_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1446 EXPORT_SYMBOL_GPL(mod_delayed_work
);
1449 * worker_enter_idle - enter idle state
1450 * @worker: worker which is entering idle state
1452 * @worker is entering idle state. Update stats and idle timer if
1456 * spin_lock_irq(pool->lock).
1458 static void worker_enter_idle(struct worker
*worker
)
1460 struct worker_pool
*pool
= worker
->pool
;
1462 BUG_ON(worker
->flags
& WORKER_IDLE
);
1463 BUG_ON(!list_empty(&worker
->entry
) &&
1464 (worker
->hentry
.next
|| worker
->hentry
.pprev
));
1466 /* can't use worker_set_flags(), also called from start_worker() */
1467 worker
->flags
|= WORKER_IDLE
;
1469 worker
->last_active
= jiffies
;
1471 /* idle_list is LIFO */
1472 list_add(&worker
->entry
, &pool
->idle_list
);
1474 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1475 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1478 * Sanity check nr_running. Because wq_unbind_fn() releases
1479 * pool->lock between setting %WORKER_UNBOUND and zapping
1480 * nr_running, the warning may trigger spuriously. Check iff
1481 * unbind is not in progress.
1483 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1484 pool
->nr_workers
== pool
->nr_idle
&&
1485 atomic_read(&pool
->nr_running
));
1489 * worker_leave_idle - leave idle state
1490 * @worker: worker which is leaving idle state
1492 * @worker is leaving idle state. Update stats.
1495 * spin_lock_irq(pool->lock).
1497 static void worker_leave_idle(struct worker
*worker
)
1499 struct worker_pool
*pool
= worker
->pool
;
1501 BUG_ON(!(worker
->flags
& WORKER_IDLE
));
1502 worker_clr_flags(worker
, WORKER_IDLE
);
1504 list_del_init(&worker
->entry
);
1508 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock pool
1511 * Works which are scheduled while the cpu is online must at least be
1512 * scheduled to a worker which is bound to the cpu so that if they are
1513 * flushed from cpu callbacks while cpu is going down, they are
1514 * guaranteed to execute on the cpu.
1516 * This function is to be used by rogue workers and rescuers to bind
1517 * themselves to the target cpu and may race with cpu going down or
1518 * coming online. kthread_bind() can't be used because it may put the
1519 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1520 * verbatim as it's best effort and blocking and pool may be
1521 * [dis]associated in the meantime.
1523 * This function tries set_cpus_allowed() and locks pool and verifies the
1524 * binding against %POOL_DISASSOCIATED which is set during
1525 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1526 * enters idle state or fetches works without dropping lock, it can
1527 * guarantee the scheduling requirement described in the first paragraph.
1530 * Might sleep. Called without any lock but returns with pool->lock
1534 * %true if the associated pool is online (@worker is successfully
1535 * bound), %false if offline.
1537 static bool worker_maybe_bind_and_lock(struct worker
*worker
)
1538 __acquires(&pool
->lock
)
1540 struct worker_pool
*pool
= worker
->pool
;
1541 struct task_struct
*task
= worker
->task
;
1545 * The following call may fail, succeed or succeed
1546 * without actually migrating the task to the cpu if
1547 * it races with cpu hotunplug operation. Verify
1548 * against POOL_DISASSOCIATED.
1550 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1551 set_cpus_allowed_ptr(task
, get_cpu_mask(pool
->cpu
));
1553 spin_lock_irq(&pool
->lock
);
1554 if (pool
->flags
& POOL_DISASSOCIATED
)
1556 if (task_cpu(task
) == pool
->cpu
&&
1557 cpumask_equal(¤t
->cpus_allowed
,
1558 get_cpu_mask(pool
->cpu
)))
1560 spin_unlock_irq(&pool
->lock
);
1563 * We've raced with CPU hot[un]plug. Give it a breather
1564 * and retry migration. cond_resched() is required here;
1565 * otherwise, we might deadlock against cpu_stop trying to
1566 * bring down the CPU on non-preemptive kernel.
1574 * Rebind an idle @worker to its CPU. worker_thread() will test
1575 * list_empty(@worker->entry) before leaving idle and call this function.
1577 static void idle_worker_rebind(struct worker
*worker
)
1579 /* CPU may go down again inbetween, clear UNBOUND only on success */
1580 if (worker_maybe_bind_and_lock(worker
))
1581 worker_clr_flags(worker
, WORKER_UNBOUND
);
1583 /* rebind complete, become available again */
1584 list_add(&worker
->entry
, &worker
->pool
->idle_list
);
1585 spin_unlock_irq(&worker
->pool
->lock
);
1589 * Function for @worker->rebind.work used to rebind unbound busy workers to
1590 * the associated cpu which is coming back online. This is scheduled by
1591 * cpu up but can race with other cpu hotplug operations and may be
1592 * executed twice without intervening cpu down.
1594 static void busy_worker_rebind_fn(struct work_struct
*work
)
1596 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1598 if (worker_maybe_bind_and_lock(worker
))
1599 worker_clr_flags(worker
, WORKER_UNBOUND
);
1601 spin_unlock_irq(&worker
->pool
->lock
);
1605 * rebind_workers - rebind all workers of a pool to the associated CPU
1606 * @pool: pool of interest
1608 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1609 * is different for idle and busy ones.
1611 * Idle ones will be removed from the idle_list and woken up. They will
1612 * add themselves back after completing rebind. This ensures that the
1613 * idle_list doesn't contain any unbound workers when re-bound busy workers
1614 * try to perform local wake-ups for concurrency management.
1616 * Busy workers can rebind after they finish their current work items.
1617 * Queueing the rebind work item at the head of the scheduled list is
1618 * enough. Note that nr_running will be properly bumped as busy workers
1621 * On return, all non-manager workers are scheduled for rebind - see
1622 * manage_workers() for the manager special case. Any idle worker
1623 * including the manager will not appear on @idle_list until rebind is
1624 * complete, making local wake-ups safe.
1626 static void rebind_workers(struct worker_pool
*pool
)
1628 struct worker
*worker
, *n
;
1629 struct hlist_node
*pos
;
1632 lockdep_assert_held(&pool
->assoc_mutex
);
1633 lockdep_assert_held(&pool
->lock
);
1635 /* dequeue and kick idle ones */
1636 list_for_each_entry_safe(worker
, n
, &pool
->idle_list
, entry
) {
1638 * idle workers should be off @pool->idle_list until rebind
1639 * is complete to avoid receiving premature local wake-ups.
1641 list_del_init(&worker
->entry
);
1644 * worker_thread() will see the above dequeuing and call
1645 * idle_worker_rebind().
1647 wake_up_process(worker
->task
);
1650 /* rebind busy workers */
1651 for_each_busy_worker(worker
, i
, pos
, pool
) {
1652 struct work_struct
*rebind_work
= &worker
->rebind_work
;
1653 struct workqueue_struct
*wq
;
1655 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
1656 work_data_bits(rebind_work
)))
1659 debug_work_activate(rebind_work
);
1662 * wq doesn't really matter but let's keep @worker->pool
1663 * and @pwq->pool consistent for sanity.
1665 if (std_worker_pool_pri(worker
->pool
))
1666 wq
= system_highpri_wq
;
1670 insert_work(get_pwq(pool
->cpu
, wq
), rebind_work
,
1671 worker
->scheduled
.next
,
1672 work_color_to_flags(WORK_NO_COLOR
));
1676 static struct worker
*alloc_worker(void)
1678 struct worker
*worker
;
1680 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1682 INIT_LIST_HEAD(&worker
->entry
);
1683 INIT_LIST_HEAD(&worker
->scheduled
);
1684 INIT_WORK(&worker
->rebind_work
, busy_worker_rebind_fn
);
1685 /* on creation a worker is in !idle && prep state */
1686 worker
->flags
= WORKER_PREP
;
1692 * create_worker - create a new workqueue worker
1693 * @pool: pool the new worker will belong to
1695 * Create a new worker which is bound to @pool. The returned worker
1696 * can be started by calling start_worker() or destroyed using
1700 * Might sleep. Does GFP_KERNEL allocations.
1703 * Pointer to the newly created worker.
1705 static struct worker
*create_worker(struct worker_pool
*pool
)
1707 const char *pri
= std_worker_pool_pri(pool
) ? "H" : "";
1708 struct worker
*worker
= NULL
;
1711 spin_lock_irq(&pool
->lock
);
1712 while (ida_get_new(&pool
->worker_ida
, &id
)) {
1713 spin_unlock_irq(&pool
->lock
);
1714 if (!ida_pre_get(&pool
->worker_ida
, GFP_KERNEL
))
1716 spin_lock_irq(&pool
->lock
);
1718 spin_unlock_irq(&pool
->lock
);
1720 worker
= alloc_worker();
1724 worker
->pool
= pool
;
1727 if (pool
->cpu
!= WORK_CPU_UNBOUND
)
1728 worker
->task
= kthread_create_on_node(worker_thread
,
1729 worker
, cpu_to_node(pool
->cpu
),
1730 "kworker/%u:%d%s", pool
->cpu
, id
, pri
);
1732 worker
->task
= kthread_create(worker_thread
, worker
,
1733 "kworker/u:%d%s", id
, pri
);
1734 if (IS_ERR(worker
->task
))
1737 if (std_worker_pool_pri(pool
))
1738 set_user_nice(worker
->task
, HIGHPRI_NICE_LEVEL
);
1741 * Determine CPU binding of the new worker depending on
1742 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1743 * flag remains stable across this function. See the comments
1744 * above the flag definition for details.
1746 * As an unbound worker may later become a regular one if CPU comes
1747 * online, make sure every worker has %PF_THREAD_BOUND set.
1749 if (!(pool
->flags
& POOL_DISASSOCIATED
)) {
1750 kthread_bind(worker
->task
, pool
->cpu
);
1752 worker
->task
->flags
|= PF_THREAD_BOUND
;
1753 worker
->flags
|= WORKER_UNBOUND
;
1759 spin_lock_irq(&pool
->lock
);
1760 ida_remove(&pool
->worker_ida
, id
);
1761 spin_unlock_irq(&pool
->lock
);
1768 * start_worker - start a newly created worker
1769 * @worker: worker to start
1771 * Make the pool aware of @worker and start it.
1774 * spin_lock_irq(pool->lock).
1776 static void start_worker(struct worker
*worker
)
1778 worker
->flags
|= WORKER_STARTED
;
1779 worker
->pool
->nr_workers
++;
1780 worker_enter_idle(worker
);
1781 wake_up_process(worker
->task
);
1785 * destroy_worker - destroy a workqueue worker
1786 * @worker: worker to be destroyed
1788 * Destroy @worker and adjust @pool stats accordingly.
1791 * spin_lock_irq(pool->lock) which is released and regrabbed.
1793 static void destroy_worker(struct worker
*worker
)
1795 struct worker_pool
*pool
= worker
->pool
;
1796 int id
= worker
->id
;
1798 /* sanity check frenzy */
1799 BUG_ON(worker
->current_work
);
1800 BUG_ON(!list_empty(&worker
->scheduled
));
1802 if (worker
->flags
& WORKER_STARTED
)
1804 if (worker
->flags
& WORKER_IDLE
)
1807 list_del_init(&worker
->entry
);
1808 worker
->flags
|= WORKER_DIE
;
1810 spin_unlock_irq(&pool
->lock
);
1812 kthread_stop(worker
->task
);
1815 spin_lock_irq(&pool
->lock
);
1816 ida_remove(&pool
->worker_ida
, id
);
1819 static void idle_worker_timeout(unsigned long __pool
)
1821 struct worker_pool
*pool
= (void *)__pool
;
1823 spin_lock_irq(&pool
->lock
);
1825 if (too_many_workers(pool
)) {
1826 struct worker
*worker
;
1827 unsigned long expires
;
1829 /* idle_list is kept in LIFO order, check the last one */
1830 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1831 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1833 if (time_before(jiffies
, expires
))
1834 mod_timer(&pool
->idle_timer
, expires
);
1836 /* it's been idle for too long, wake up manager */
1837 pool
->flags
|= POOL_MANAGE_WORKERS
;
1838 wake_up_worker(pool
);
1842 spin_unlock_irq(&pool
->lock
);
1845 static bool send_mayday(struct work_struct
*work
)
1847 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1848 struct workqueue_struct
*wq
= pwq
->wq
;
1851 if (!(wq
->flags
& WQ_RESCUER
))
1854 /* mayday mayday mayday */
1855 cpu
= pwq
->pool
->cpu
;
1856 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1857 if (cpu
== WORK_CPU_UNBOUND
)
1859 if (!mayday_test_and_set_cpu(cpu
, wq
->mayday_mask
))
1860 wake_up_process(wq
->rescuer
->task
);
1864 static void pool_mayday_timeout(unsigned long __pool
)
1866 struct worker_pool
*pool
= (void *)__pool
;
1867 struct work_struct
*work
;
1869 spin_lock_irq(&pool
->lock
);
1871 if (need_to_create_worker(pool
)) {
1873 * We've been trying to create a new worker but
1874 * haven't been successful. We might be hitting an
1875 * allocation deadlock. Send distress signals to
1878 list_for_each_entry(work
, &pool
->worklist
, entry
)
1882 spin_unlock_irq(&pool
->lock
);
1884 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1888 * maybe_create_worker - create a new worker if necessary
1889 * @pool: pool to create a new worker for
1891 * Create a new worker for @pool if necessary. @pool is guaranteed to
1892 * have at least one idle worker on return from this function. If
1893 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1894 * sent to all rescuers with works scheduled on @pool to resolve
1895 * possible allocation deadlock.
1897 * On return, need_to_create_worker() is guaranteed to be false and
1898 * may_start_working() true.
1901 * spin_lock_irq(pool->lock) which may be released and regrabbed
1902 * multiple times. Does GFP_KERNEL allocations. Called only from
1906 * false if no action was taken and pool->lock stayed locked, true
1909 static bool maybe_create_worker(struct worker_pool
*pool
)
1910 __releases(&pool
->lock
)
1911 __acquires(&pool
->lock
)
1913 if (!need_to_create_worker(pool
))
1916 spin_unlock_irq(&pool
->lock
);
1918 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1919 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1922 struct worker
*worker
;
1924 worker
= create_worker(pool
);
1926 del_timer_sync(&pool
->mayday_timer
);
1927 spin_lock_irq(&pool
->lock
);
1928 start_worker(worker
);
1929 BUG_ON(need_to_create_worker(pool
));
1933 if (!need_to_create_worker(pool
))
1936 __set_current_state(TASK_INTERRUPTIBLE
);
1937 schedule_timeout(CREATE_COOLDOWN
);
1939 if (!need_to_create_worker(pool
))
1943 del_timer_sync(&pool
->mayday_timer
);
1944 spin_lock_irq(&pool
->lock
);
1945 if (need_to_create_worker(pool
))
1951 * maybe_destroy_worker - destroy workers which have been idle for a while
1952 * @pool: pool to destroy workers for
1954 * Destroy @pool workers which have been idle for longer than
1955 * IDLE_WORKER_TIMEOUT.
1958 * spin_lock_irq(pool->lock) which may be released and regrabbed
1959 * multiple times. Called only from manager.
1962 * false if no action was taken and pool->lock stayed locked, true
1965 static bool maybe_destroy_workers(struct worker_pool
*pool
)
1969 while (too_many_workers(pool
)) {
1970 struct worker
*worker
;
1971 unsigned long expires
;
1973 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1974 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1976 if (time_before(jiffies
, expires
)) {
1977 mod_timer(&pool
->idle_timer
, expires
);
1981 destroy_worker(worker
);
1989 * manage_workers - manage worker pool
1992 * Assume the manager role and manage the worker pool @worker belongs
1993 * to. At any given time, there can be only zero or one manager per
1994 * pool. The exclusion is handled automatically by this function.
1996 * The caller can safely start processing works on false return. On
1997 * true return, it's guaranteed that need_to_create_worker() is false
1998 * and may_start_working() is true.
2001 * spin_lock_irq(pool->lock) which may be released and regrabbed
2002 * multiple times. Does GFP_KERNEL allocations.
2005 * spin_lock_irq(pool->lock) which may be released and regrabbed
2006 * multiple times. Does GFP_KERNEL allocations.
2008 static bool manage_workers(struct worker
*worker
)
2010 struct worker_pool
*pool
= worker
->pool
;
2013 if (pool
->flags
& POOL_MANAGING_WORKERS
)
2016 pool
->flags
|= POOL_MANAGING_WORKERS
;
2019 * To simplify both worker management and CPU hotplug, hold off
2020 * management while hotplug is in progress. CPU hotplug path can't
2021 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2022 * lead to idle worker depletion (all become busy thinking someone
2023 * else is managing) which in turn can result in deadlock under
2024 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2025 * manager against CPU hotplug.
2027 * assoc_mutex would always be free unless CPU hotplug is in
2028 * progress. trylock first without dropping @pool->lock.
2030 if (unlikely(!mutex_trylock(&pool
->assoc_mutex
))) {
2031 spin_unlock_irq(&pool
->lock
);
2032 mutex_lock(&pool
->assoc_mutex
);
2034 * CPU hotplug could have happened while we were waiting
2035 * for assoc_mutex. Hotplug itself can't handle us
2036 * because manager isn't either on idle or busy list, and
2037 * @pool's state and ours could have deviated.
2039 * As hotplug is now excluded via assoc_mutex, we can
2040 * simply try to bind. It will succeed or fail depending
2041 * on @pool's current state. Try it and adjust
2042 * %WORKER_UNBOUND accordingly.
2044 if (worker_maybe_bind_and_lock(worker
))
2045 worker
->flags
&= ~WORKER_UNBOUND
;
2047 worker
->flags
|= WORKER_UNBOUND
;
2052 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2055 * Destroy and then create so that may_start_working() is true
2058 ret
|= maybe_destroy_workers(pool
);
2059 ret
|= maybe_create_worker(pool
);
2061 pool
->flags
&= ~POOL_MANAGING_WORKERS
;
2062 mutex_unlock(&pool
->assoc_mutex
);
2067 * process_one_work - process single work
2069 * @work: work to process
2071 * Process @work. This function contains all the logics necessary to
2072 * process a single work including synchronization against and
2073 * interaction with other workers on the same cpu, queueing and
2074 * flushing. As long as context requirement is met, any worker can
2075 * call this function to process a work.
2078 * spin_lock_irq(pool->lock) which is released and regrabbed.
2080 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2081 __releases(&pool
->lock
)
2082 __acquires(&pool
->lock
)
2084 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2085 struct worker_pool
*pool
= worker
->pool
;
2086 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2088 struct worker
*collision
;
2089 #ifdef CONFIG_LOCKDEP
2091 * It is permissible to free the struct work_struct from
2092 * inside the function that is called from it, this we need to
2093 * take into account for lockdep too. To avoid bogus "held
2094 * lock freed" warnings as well as problems when looking into
2095 * work->lockdep_map, make a copy and use that here.
2097 struct lockdep_map lockdep_map
;
2099 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2102 * Ensure we're on the correct CPU. DISASSOCIATED test is
2103 * necessary to avoid spurious warnings from rescuers servicing the
2104 * unbound or a disassociated pool.
2106 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2107 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2108 raw_smp_processor_id() != pool
->cpu
);
2111 * A single work shouldn't be executed concurrently by
2112 * multiple workers on a single cpu. Check whether anyone is
2113 * already processing the work. If so, defer the work to the
2114 * currently executing one.
2116 collision
= find_worker_executing_work(pool
, work
);
2117 if (unlikely(collision
)) {
2118 move_linked_works(work
, &collision
->scheduled
, NULL
);
2122 /* claim and dequeue */
2123 debug_work_deactivate(work
);
2124 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2125 worker
->current_work
= work
;
2126 worker
->current_func
= work
->func
;
2127 worker
->current_pwq
= pwq
;
2128 work_color
= get_work_color(work
);
2130 list_del_init(&work
->entry
);
2133 * CPU intensive works don't participate in concurrency
2134 * management. They're the scheduler's responsibility.
2136 if (unlikely(cpu_intensive
))
2137 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2140 * Unbound pool isn't concurrency managed and work items should be
2141 * executed ASAP. Wake up another worker if necessary.
2143 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2144 wake_up_worker(pool
);
2147 * Record the last pool and clear PENDING which should be the last
2148 * update to @work. Also, do this inside @pool->lock so that
2149 * PENDING and queued state changes happen together while IRQ is
2152 set_work_pool_and_clear_pending(work
, pool
->id
);
2154 spin_unlock_irq(&pool
->lock
);
2156 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2157 lock_map_acquire(&lockdep_map
);
2158 trace_workqueue_execute_start(work
);
2159 worker
->current_func(work
);
2161 * While we must be careful to not use "work" after this, the trace
2162 * point will only record its address.
2164 trace_workqueue_execute_end(work
);
2165 lock_map_release(&lockdep_map
);
2166 lock_map_release(&pwq
->wq
->lockdep_map
);
2168 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2169 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2170 " last function: %pf\n",
2171 current
->comm
, preempt_count(), task_pid_nr(current
),
2172 worker
->current_func
);
2173 debug_show_held_locks(current
);
2177 spin_lock_irq(&pool
->lock
);
2179 /* clear cpu intensive status */
2180 if (unlikely(cpu_intensive
))
2181 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2183 /* we're done with it, release */
2184 hash_del(&worker
->hentry
);
2185 worker
->current_work
= NULL
;
2186 worker
->current_func
= NULL
;
2187 worker
->current_pwq
= NULL
;
2188 pwq_dec_nr_in_flight(pwq
, work_color
);
2192 * process_scheduled_works - process scheduled works
2195 * Process all scheduled works. Please note that the scheduled list
2196 * may change while processing a work, so this function repeatedly
2197 * fetches a work from the top and executes it.
2200 * spin_lock_irq(pool->lock) which may be released and regrabbed
2203 static void process_scheduled_works(struct worker
*worker
)
2205 while (!list_empty(&worker
->scheduled
)) {
2206 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2207 struct work_struct
, entry
);
2208 process_one_work(worker
, work
);
2213 * worker_thread - the worker thread function
2216 * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools
2217 * of these per each cpu. These workers process all works regardless of
2218 * their specific target workqueue. The only exception is works which
2219 * belong to workqueues with a rescuer which will be explained in
2222 static int worker_thread(void *__worker
)
2224 struct worker
*worker
= __worker
;
2225 struct worker_pool
*pool
= worker
->pool
;
2227 /* tell the scheduler that this is a workqueue worker */
2228 worker
->task
->flags
|= PF_WQ_WORKER
;
2230 spin_lock_irq(&pool
->lock
);
2232 /* we are off idle list if destruction or rebind is requested */
2233 if (unlikely(list_empty(&worker
->entry
))) {
2234 spin_unlock_irq(&pool
->lock
);
2236 /* if DIE is set, destruction is requested */
2237 if (worker
->flags
& WORKER_DIE
) {
2238 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2242 /* otherwise, rebind */
2243 idle_worker_rebind(worker
);
2247 worker_leave_idle(worker
);
2249 /* no more worker necessary? */
2250 if (!need_more_worker(pool
))
2253 /* do we need to manage? */
2254 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2258 * ->scheduled list can only be filled while a worker is
2259 * preparing to process a work or actually processing it.
2260 * Make sure nobody diddled with it while I was sleeping.
2262 BUG_ON(!list_empty(&worker
->scheduled
));
2265 * When control reaches this point, we're guaranteed to have
2266 * at least one idle worker or that someone else has already
2267 * assumed the manager role.
2269 worker_clr_flags(worker
, WORKER_PREP
);
2272 struct work_struct
*work
=
2273 list_first_entry(&pool
->worklist
,
2274 struct work_struct
, entry
);
2276 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2277 /* optimization path, not strictly necessary */
2278 process_one_work(worker
, work
);
2279 if (unlikely(!list_empty(&worker
->scheduled
)))
2280 process_scheduled_works(worker
);
2282 move_linked_works(work
, &worker
->scheduled
, NULL
);
2283 process_scheduled_works(worker
);
2285 } while (keep_working(pool
));
2287 worker_set_flags(worker
, WORKER_PREP
, false);
2289 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2293 * pool->lock is held and there's no work to process and no need to
2294 * manage, sleep. Workers are woken up only while holding
2295 * pool->lock or from local cpu, so setting the current state
2296 * before releasing pool->lock is enough to prevent losing any
2299 worker_enter_idle(worker
);
2300 __set_current_state(TASK_INTERRUPTIBLE
);
2301 spin_unlock_irq(&pool
->lock
);
2307 * rescuer_thread - the rescuer thread function
2310 * Workqueue rescuer thread function. There's one rescuer for each
2311 * workqueue which has WQ_RESCUER set.
2313 * Regular work processing on a pool may block trying to create a new
2314 * worker which uses GFP_KERNEL allocation which has slight chance of
2315 * developing into deadlock if some works currently on the same queue
2316 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2317 * the problem rescuer solves.
2319 * When such condition is possible, the pool summons rescuers of all
2320 * workqueues which have works queued on the pool and let them process
2321 * those works so that forward progress can be guaranteed.
2323 * This should happen rarely.
2325 static int rescuer_thread(void *__rescuer
)
2327 struct worker
*rescuer
= __rescuer
;
2328 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2329 struct list_head
*scheduled
= &rescuer
->scheduled
;
2330 bool is_unbound
= wq
->flags
& WQ_UNBOUND
;
2333 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2336 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2337 * doesn't participate in concurrency management.
2339 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2341 set_current_state(TASK_INTERRUPTIBLE
);
2343 if (kthread_should_stop()) {
2344 __set_current_state(TASK_RUNNING
);
2345 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2350 * See whether any cpu is asking for help. Unbounded
2351 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2353 for_each_mayday_cpu(cpu
, wq
->mayday_mask
) {
2354 unsigned int tcpu
= is_unbound
? WORK_CPU_UNBOUND
: cpu
;
2355 struct pool_workqueue
*pwq
= get_pwq(tcpu
, wq
);
2356 struct worker_pool
*pool
= pwq
->pool
;
2357 struct work_struct
*work
, *n
;
2359 __set_current_state(TASK_RUNNING
);
2360 mayday_clear_cpu(cpu
, wq
->mayday_mask
);
2362 /* migrate to the target cpu if possible */
2363 rescuer
->pool
= pool
;
2364 worker_maybe_bind_and_lock(rescuer
);
2367 * Slurp in all works issued via this workqueue and
2370 BUG_ON(!list_empty(&rescuer
->scheduled
));
2371 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2372 if (get_work_pwq(work
) == pwq
)
2373 move_linked_works(work
, scheduled
, &n
);
2375 process_scheduled_works(rescuer
);
2378 * Leave this pool. If keep_working() is %true, notify a
2379 * regular worker; otherwise, we end up with 0 concurrency
2380 * and stalling the execution.
2382 if (keep_working(pool
))
2383 wake_up_worker(pool
);
2385 spin_unlock_irq(&pool
->lock
);
2388 /* rescuers should never participate in concurrency management */
2389 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2395 struct work_struct work
;
2396 struct completion done
;
2399 static void wq_barrier_func(struct work_struct
*work
)
2401 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2402 complete(&barr
->done
);
2406 * insert_wq_barrier - insert a barrier work
2407 * @pwq: pwq to insert barrier into
2408 * @barr: wq_barrier to insert
2409 * @target: target work to attach @barr to
2410 * @worker: worker currently executing @target, NULL if @target is not executing
2412 * @barr is linked to @target such that @barr is completed only after
2413 * @target finishes execution. Please note that the ordering
2414 * guarantee is observed only with respect to @target and on the local
2417 * Currently, a queued barrier can't be canceled. This is because
2418 * try_to_grab_pending() can't determine whether the work to be
2419 * grabbed is at the head of the queue and thus can't clear LINKED
2420 * flag of the previous work while there must be a valid next work
2421 * after a work with LINKED flag set.
2423 * Note that when @worker is non-NULL, @target may be modified
2424 * underneath us, so we can't reliably determine pwq from @target.
2427 * spin_lock_irq(pool->lock).
2429 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2430 struct wq_barrier
*barr
,
2431 struct work_struct
*target
, struct worker
*worker
)
2433 struct list_head
*head
;
2434 unsigned int linked
= 0;
2437 * debugobject calls are safe here even with pool->lock locked
2438 * as we know for sure that this will not trigger any of the
2439 * checks and call back into the fixup functions where we
2442 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2443 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2444 init_completion(&barr
->done
);
2447 * If @target is currently being executed, schedule the
2448 * barrier to the worker; otherwise, put it after @target.
2451 head
= worker
->scheduled
.next
;
2453 unsigned long *bits
= work_data_bits(target
);
2455 head
= target
->entry
.next
;
2456 /* there can already be other linked works, inherit and set */
2457 linked
= *bits
& WORK_STRUCT_LINKED
;
2458 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2461 debug_work_activate(&barr
->work
);
2462 insert_work(pwq
, &barr
->work
, head
,
2463 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2467 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2468 * @wq: workqueue being flushed
2469 * @flush_color: new flush color, < 0 for no-op
2470 * @work_color: new work color, < 0 for no-op
2472 * Prepare pwqs for workqueue flushing.
2474 * If @flush_color is non-negative, flush_color on all pwqs should be
2475 * -1. If no pwq has in-flight commands at the specified color, all
2476 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2477 * has in flight commands, its pwq->flush_color is set to
2478 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2479 * wakeup logic is armed and %true is returned.
2481 * The caller should have initialized @wq->first_flusher prior to
2482 * calling this function with non-negative @flush_color. If
2483 * @flush_color is negative, no flush color update is done and %false
2486 * If @work_color is non-negative, all pwqs should have the same
2487 * work_color which is previous to @work_color and all will be
2488 * advanced to @work_color.
2491 * mutex_lock(wq->flush_mutex).
2494 * %true if @flush_color >= 0 and there's something to flush. %false
2497 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2498 int flush_color
, int work_color
)
2503 if (flush_color
>= 0) {
2504 BUG_ON(atomic_read(&wq
->nr_pwqs_to_flush
));
2505 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2508 for_each_pwq_cpu(cpu
, wq
) {
2509 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
2510 struct worker_pool
*pool
= pwq
->pool
;
2512 spin_lock_irq(&pool
->lock
);
2514 if (flush_color
>= 0) {
2515 BUG_ON(pwq
->flush_color
!= -1);
2517 if (pwq
->nr_in_flight
[flush_color
]) {
2518 pwq
->flush_color
= flush_color
;
2519 atomic_inc(&wq
->nr_pwqs_to_flush
);
2524 if (work_color
>= 0) {
2525 BUG_ON(work_color
!= work_next_color(pwq
->work_color
));
2526 pwq
->work_color
= work_color
;
2529 spin_unlock_irq(&pool
->lock
);
2532 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2533 complete(&wq
->first_flusher
->done
);
2539 * flush_workqueue - ensure that any scheduled work has run to completion.
2540 * @wq: workqueue to flush
2542 * Forces execution of the workqueue and blocks until its completion.
2543 * This is typically used in driver shutdown handlers.
2545 * We sleep until all works which were queued on entry have been handled,
2546 * but we are not livelocked by new incoming ones.
2548 void flush_workqueue(struct workqueue_struct
*wq
)
2550 struct wq_flusher this_flusher
= {
2551 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2553 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2557 lock_map_acquire(&wq
->lockdep_map
);
2558 lock_map_release(&wq
->lockdep_map
);
2560 mutex_lock(&wq
->flush_mutex
);
2563 * Start-to-wait phase
2565 next_color
= work_next_color(wq
->work_color
);
2567 if (next_color
!= wq
->flush_color
) {
2569 * Color space is not full. The current work_color
2570 * becomes our flush_color and work_color is advanced
2573 BUG_ON(!list_empty(&wq
->flusher_overflow
));
2574 this_flusher
.flush_color
= wq
->work_color
;
2575 wq
->work_color
= next_color
;
2577 if (!wq
->first_flusher
) {
2578 /* no flush in progress, become the first flusher */
2579 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2581 wq
->first_flusher
= &this_flusher
;
2583 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2585 /* nothing to flush, done */
2586 wq
->flush_color
= next_color
;
2587 wq
->first_flusher
= NULL
;
2592 BUG_ON(wq
->flush_color
== this_flusher
.flush_color
);
2593 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2594 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2598 * Oops, color space is full, wait on overflow queue.
2599 * The next flush completion will assign us
2600 * flush_color and transfer to flusher_queue.
2602 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2605 mutex_unlock(&wq
->flush_mutex
);
2607 wait_for_completion(&this_flusher
.done
);
2610 * Wake-up-and-cascade phase
2612 * First flushers are responsible for cascading flushes and
2613 * handling overflow. Non-first flushers can simply return.
2615 if (wq
->first_flusher
!= &this_flusher
)
2618 mutex_lock(&wq
->flush_mutex
);
2620 /* we might have raced, check again with mutex held */
2621 if (wq
->first_flusher
!= &this_flusher
)
2624 wq
->first_flusher
= NULL
;
2626 BUG_ON(!list_empty(&this_flusher
.list
));
2627 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2630 struct wq_flusher
*next
, *tmp
;
2632 /* complete all the flushers sharing the current flush color */
2633 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2634 if (next
->flush_color
!= wq
->flush_color
)
2636 list_del_init(&next
->list
);
2637 complete(&next
->done
);
2640 BUG_ON(!list_empty(&wq
->flusher_overflow
) &&
2641 wq
->flush_color
!= work_next_color(wq
->work_color
));
2643 /* this flush_color is finished, advance by one */
2644 wq
->flush_color
= work_next_color(wq
->flush_color
);
2646 /* one color has been freed, handle overflow queue */
2647 if (!list_empty(&wq
->flusher_overflow
)) {
2649 * Assign the same color to all overflowed
2650 * flushers, advance work_color and append to
2651 * flusher_queue. This is the start-to-wait
2652 * phase for these overflowed flushers.
2654 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2655 tmp
->flush_color
= wq
->work_color
;
2657 wq
->work_color
= work_next_color(wq
->work_color
);
2659 list_splice_tail_init(&wq
->flusher_overflow
,
2660 &wq
->flusher_queue
);
2661 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2664 if (list_empty(&wq
->flusher_queue
)) {
2665 BUG_ON(wq
->flush_color
!= wq
->work_color
);
2670 * Need to flush more colors. Make the next flusher
2671 * the new first flusher and arm pwqs.
2673 BUG_ON(wq
->flush_color
== wq
->work_color
);
2674 BUG_ON(wq
->flush_color
!= next
->flush_color
);
2676 list_del_init(&next
->list
);
2677 wq
->first_flusher
= next
;
2679 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2683 * Meh... this color is already done, clear first
2684 * flusher and repeat cascading.
2686 wq
->first_flusher
= NULL
;
2690 mutex_unlock(&wq
->flush_mutex
);
2692 EXPORT_SYMBOL_GPL(flush_workqueue
);
2695 * drain_workqueue - drain a workqueue
2696 * @wq: workqueue to drain
2698 * Wait until the workqueue becomes empty. While draining is in progress,
2699 * only chain queueing is allowed. IOW, only currently pending or running
2700 * work items on @wq can queue further work items on it. @wq is flushed
2701 * repeatedly until it becomes empty. The number of flushing is detemined
2702 * by the depth of chaining and should be relatively short. Whine if it
2705 void drain_workqueue(struct workqueue_struct
*wq
)
2707 unsigned int flush_cnt
= 0;
2711 * __queue_work() needs to test whether there are drainers, is much
2712 * hotter than drain_workqueue() and already looks at @wq->flags.
2713 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2715 spin_lock(&workqueue_lock
);
2716 if (!wq
->nr_drainers
++)
2717 wq
->flags
|= WQ_DRAINING
;
2718 spin_unlock(&workqueue_lock
);
2720 flush_workqueue(wq
);
2722 for_each_pwq_cpu(cpu
, wq
) {
2723 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
2726 spin_lock_irq(&pwq
->pool
->lock
);
2727 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2728 spin_unlock_irq(&pwq
->pool
->lock
);
2733 if (++flush_cnt
== 10 ||
2734 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2735 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2736 wq
->name
, flush_cnt
);
2740 spin_lock(&workqueue_lock
);
2741 if (!--wq
->nr_drainers
)
2742 wq
->flags
&= ~WQ_DRAINING
;
2743 spin_unlock(&workqueue_lock
);
2745 EXPORT_SYMBOL_GPL(drain_workqueue
);
2747 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2749 struct worker
*worker
= NULL
;
2750 struct worker_pool
*pool
;
2751 struct pool_workqueue
*pwq
;
2754 pool
= get_work_pool(work
);
2758 spin_lock_irq(&pool
->lock
);
2759 /* see the comment in try_to_grab_pending() with the same code */
2760 pwq
= get_work_pwq(work
);
2762 if (unlikely(pwq
->pool
!= pool
))
2765 worker
= find_worker_executing_work(pool
, work
);
2768 pwq
= worker
->current_pwq
;
2771 insert_wq_barrier(pwq
, barr
, work
, worker
);
2772 spin_unlock_irq(&pool
->lock
);
2775 * If @max_active is 1 or rescuer is in use, flushing another work
2776 * item on the same workqueue may lead to deadlock. Make sure the
2777 * flusher is not running on the same workqueue by verifying write
2780 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->flags
& WQ_RESCUER
)
2781 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2783 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2784 lock_map_release(&pwq
->wq
->lockdep_map
);
2788 spin_unlock_irq(&pool
->lock
);
2793 * flush_work - wait for a work to finish executing the last queueing instance
2794 * @work: the work to flush
2796 * Wait until @work has finished execution. @work is guaranteed to be idle
2797 * on return if it hasn't been requeued since flush started.
2800 * %true if flush_work() waited for the work to finish execution,
2801 * %false if it was already idle.
2803 bool flush_work(struct work_struct
*work
)
2805 struct wq_barrier barr
;
2807 lock_map_acquire(&work
->lockdep_map
);
2808 lock_map_release(&work
->lockdep_map
);
2810 if (start_flush_work(work
, &barr
)) {
2811 wait_for_completion(&barr
.done
);
2812 destroy_work_on_stack(&barr
.work
);
2818 EXPORT_SYMBOL_GPL(flush_work
);
2820 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2822 unsigned long flags
;
2826 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2828 * If someone else is canceling, wait for the same event it
2829 * would be waiting for before retrying.
2831 if (unlikely(ret
== -ENOENT
))
2833 } while (unlikely(ret
< 0));
2835 /* tell other tasks trying to grab @work to back off */
2836 mark_work_canceling(work
);
2837 local_irq_restore(flags
);
2840 clear_work_data(work
);
2845 * cancel_work_sync - cancel a work and wait for it to finish
2846 * @work: the work to cancel
2848 * Cancel @work and wait for its execution to finish. This function
2849 * can be used even if the work re-queues itself or migrates to
2850 * another workqueue. On return from this function, @work is
2851 * guaranteed to be not pending or executing on any CPU.
2853 * cancel_work_sync(&delayed_work->work) must not be used for
2854 * delayed_work's. Use cancel_delayed_work_sync() instead.
2856 * The caller must ensure that the workqueue on which @work was last
2857 * queued can't be destroyed before this function returns.
2860 * %true if @work was pending, %false otherwise.
2862 bool cancel_work_sync(struct work_struct
*work
)
2864 return __cancel_work_timer(work
, false);
2866 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2869 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2870 * @dwork: the delayed work to flush
2872 * Delayed timer is cancelled and the pending work is queued for
2873 * immediate execution. Like flush_work(), this function only
2874 * considers the last queueing instance of @dwork.
2877 * %true if flush_work() waited for the work to finish execution,
2878 * %false if it was already idle.
2880 bool flush_delayed_work(struct delayed_work
*dwork
)
2882 local_irq_disable();
2883 if (del_timer_sync(&dwork
->timer
))
2884 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2886 return flush_work(&dwork
->work
);
2888 EXPORT_SYMBOL(flush_delayed_work
);
2891 * cancel_delayed_work - cancel a delayed work
2892 * @dwork: delayed_work to cancel
2894 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2895 * and canceled; %false if wasn't pending. Note that the work callback
2896 * function may still be running on return, unless it returns %true and the
2897 * work doesn't re-arm itself. Explicitly flush or use
2898 * cancel_delayed_work_sync() to wait on it.
2900 * This function is safe to call from any context including IRQ handler.
2902 bool cancel_delayed_work(struct delayed_work
*dwork
)
2904 unsigned long flags
;
2908 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2909 } while (unlikely(ret
== -EAGAIN
));
2911 if (unlikely(ret
< 0))
2914 set_work_pool_and_clear_pending(&dwork
->work
,
2915 get_work_pool_id(&dwork
->work
));
2916 local_irq_restore(flags
);
2919 EXPORT_SYMBOL(cancel_delayed_work
);
2922 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2923 * @dwork: the delayed work cancel
2925 * This is cancel_work_sync() for delayed works.
2928 * %true if @dwork was pending, %false otherwise.
2930 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2932 return __cancel_work_timer(&dwork
->work
, true);
2934 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2937 * schedule_work_on - put work task on a specific cpu
2938 * @cpu: cpu to put the work task on
2939 * @work: job to be done
2941 * This puts a job on a specific cpu
2943 bool schedule_work_on(int cpu
, struct work_struct
*work
)
2945 return queue_work_on(cpu
, system_wq
, work
);
2947 EXPORT_SYMBOL(schedule_work_on
);
2950 * schedule_work - put work task in global workqueue
2951 * @work: job to be done
2953 * Returns %false if @work was already on the kernel-global workqueue and
2956 * This puts a job in the kernel-global workqueue if it was not already
2957 * queued and leaves it in the same position on the kernel-global
2958 * workqueue otherwise.
2960 bool schedule_work(struct work_struct
*work
)
2962 return queue_work(system_wq
, work
);
2964 EXPORT_SYMBOL(schedule_work
);
2967 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2969 * @dwork: job to be done
2970 * @delay: number of jiffies to wait
2972 * After waiting for a given time this puts a job in the kernel-global
2973 * workqueue on the specified CPU.
2975 bool schedule_delayed_work_on(int cpu
, struct delayed_work
*dwork
,
2976 unsigned long delay
)
2978 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
2980 EXPORT_SYMBOL(schedule_delayed_work_on
);
2983 * schedule_delayed_work - put work task in global workqueue after delay
2984 * @dwork: job to be done
2985 * @delay: number of jiffies to wait or 0 for immediate execution
2987 * After waiting for a given time this puts a job in the kernel-global
2990 bool schedule_delayed_work(struct delayed_work
*dwork
, unsigned long delay
)
2992 return queue_delayed_work(system_wq
, dwork
, delay
);
2994 EXPORT_SYMBOL(schedule_delayed_work
);
2997 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2998 * @func: the function to call
3000 * schedule_on_each_cpu() executes @func on each online CPU using the
3001 * system workqueue and blocks until all CPUs have completed.
3002 * schedule_on_each_cpu() is very slow.
3005 * 0 on success, -errno on failure.
3007 int schedule_on_each_cpu(work_func_t func
)
3010 struct work_struct __percpu
*works
;
3012 works
= alloc_percpu(struct work_struct
);
3018 for_each_online_cpu(cpu
) {
3019 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3021 INIT_WORK(work
, func
);
3022 schedule_work_on(cpu
, work
);
3025 for_each_online_cpu(cpu
)
3026 flush_work(per_cpu_ptr(works
, cpu
));
3034 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3036 * Forces execution of the kernel-global workqueue and blocks until its
3039 * Think twice before calling this function! It's very easy to get into
3040 * trouble if you don't take great care. Either of the following situations
3041 * will lead to deadlock:
3043 * One of the work items currently on the workqueue needs to acquire
3044 * a lock held by your code or its caller.
3046 * Your code is running in the context of a work routine.
3048 * They will be detected by lockdep when they occur, but the first might not
3049 * occur very often. It depends on what work items are on the workqueue and
3050 * what locks they need, which you have no control over.
3052 * In most situations flushing the entire workqueue is overkill; you merely
3053 * need to know that a particular work item isn't queued and isn't running.
3054 * In such cases you should use cancel_delayed_work_sync() or
3055 * cancel_work_sync() instead.
3057 void flush_scheduled_work(void)
3059 flush_workqueue(system_wq
);
3061 EXPORT_SYMBOL(flush_scheduled_work
);
3064 * execute_in_process_context - reliably execute the routine with user context
3065 * @fn: the function to execute
3066 * @ew: guaranteed storage for the execute work structure (must
3067 * be available when the work executes)
3069 * Executes the function immediately if process context is available,
3070 * otherwise schedules the function for delayed execution.
3072 * Returns: 0 - function was executed
3073 * 1 - function was scheduled for execution
3075 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3077 if (!in_interrupt()) {
3082 INIT_WORK(&ew
->work
, fn
);
3083 schedule_work(&ew
->work
);
3087 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3089 int keventd_up(void)
3091 return system_wq
!= NULL
;
3094 static int alloc_pwqs(struct workqueue_struct
*wq
)
3097 * pwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3098 * Make sure that the alignment isn't lower than that of
3099 * unsigned long long.
3101 const size_t size
= sizeof(struct pool_workqueue
);
3102 const size_t align
= max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS
,
3103 __alignof__(unsigned long long));
3105 if (!(wq
->flags
& WQ_UNBOUND
))
3106 wq
->pool_wq
.pcpu
= __alloc_percpu(size
, align
);
3111 * Allocate enough room to align pwq and put an extra
3112 * pointer at the end pointing back to the originally
3113 * allocated pointer which will be used for free.
3115 ptr
= kzalloc(size
+ align
+ sizeof(void *), GFP_KERNEL
);
3117 wq
->pool_wq
.single
= PTR_ALIGN(ptr
, align
);
3118 *(void **)(wq
->pool_wq
.single
+ 1) = ptr
;
3122 /* just in case, make sure it's actually aligned */
3123 BUG_ON(!IS_ALIGNED(wq
->pool_wq
.v
, align
));
3124 return wq
->pool_wq
.v
? 0 : -ENOMEM
;
3127 static void free_pwqs(struct workqueue_struct
*wq
)
3129 if (!(wq
->flags
& WQ_UNBOUND
))
3130 free_percpu(wq
->pool_wq
.pcpu
);
3131 else if (wq
->pool_wq
.single
) {
3132 /* the pointer to free is stored right after the pwq */
3133 kfree(*(void **)(wq
->pool_wq
.single
+ 1));
3137 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3140 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3142 if (max_active
< 1 || max_active
> lim
)
3143 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3144 max_active
, name
, 1, lim
);
3146 return clamp_val(max_active
, 1, lim
);
3149 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3152 struct lock_class_key
*key
,
3153 const char *lock_name
, ...)
3155 va_list args
, args1
;
3156 struct workqueue_struct
*wq
;
3160 /* determine namelen, allocate wq and format name */
3161 va_start(args
, lock_name
);
3162 va_copy(args1
, args
);
3163 namelen
= vsnprintf(NULL
, 0, fmt
, args
) + 1;
3165 wq
= kzalloc(sizeof(*wq
) + namelen
, GFP_KERNEL
);
3169 vsnprintf(wq
->name
, namelen
, fmt
, args1
);
3174 * Workqueues which may be used during memory reclaim should
3175 * have a rescuer to guarantee forward progress.
3177 if (flags
& WQ_MEM_RECLAIM
)
3178 flags
|= WQ_RESCUER
;
3180 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3181 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3185 wq
->saved_max_active
= max_active
;
3186 mutex_init(&wq
->flush_mutex
);
3187 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
3188 INIT_LIST_HEAD(&wq
->flusher_queue
);
3189 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3191 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3192 INIT_LIST_HEAD(&wq
->list
);
3194 if (alloc_pwqs(wq
) < 0)
3197 for_each_pwq_cpu(cpu
, wq
) {
3198 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3200 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3201 pwq
->pool
= get_std_worker_pool(cpu
, flags
& WQ_HIGHPRI
);
3203 pwq
->flush_color
= -1;
3204 pwq
->max_active
= max_active
;
3205 INIT_LIST_HEAD(&pwq
->delayed_works
);
3208 if (flags
& WQ_RESCUER
) {
3209 struct worker
*rescuer
;
3211 if (!alloc_mayday_mask(&wq
->mayday_mask
, GFP_KERNEL
))
3214 wq
->rescuer
= rescuer
= alloc_worker();
3218 rescuer
->rescue_wq
= wq
;
3219 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
3221 if (IS_ERR(rescuer
->task
))
3224 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
3225 wake_up_process(rescuer
->task
);
3229 * workqueue_lock protects global freeze state and workqueues
3230 * list. Grab it, set max_active accordingly and add the new
3231 * workqueue to workqueues list.
3233 spin_lock(&workqueue_lock
);
3235 if (workqueue_freezing
&& wq
->flags
& WQ_FREEZABLE
)
3236 for_each_pwq_cpu(cpu
, wq
)
3237 get_pwq(cpu
, wq
)->max_active
= 0;
3239 list_add(&wq
->list
, &workqueues
);
3241 spin_unlock(&workqueue_lock
);
3247 free_mayday_mask(wq
->mayday_mask
);
3253 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3256 * destroy_workqueue - safely terminate a workqueue
3257 * @wq: target workqueue
3259 * Safely destroy a workqueue. All work currently pending will be done first.
3261 void destroy_workqueue(struct workqueue_struct
*wq
)
3265 /* drain it before proceeding with destruction */
3266 drain_workqueue(wq
);
3269 * wq list is used to freeze wq, remove from list after
3270 * flushing is complete in case freeze races us.
3272 spin_lock(&workqueue_lock
);
3273 list_del(&wq
->list
);
3274 spin_unlock(&workqueue_lock
);
3277 for_each_pwq_cpu(cpu
, wq
) {
3278 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3281 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
3282 BUG_ON(pwq
->nr_in_flight
[i
]);
3283 BUG_ON(pwq
->nr_active
);
3284 BUG_ON(!list_empty(&pwq
->delayed_works
));
3287 if (wq
->flags
& WQ_RESCUER
) {
3288 kthread_stop(wq
->rescuer
->task
);
3289 free_mayday_mask(wq
->mayday_mask
);
3296 EXPORT_SYMBOL_GPL(destroy_workqueue
);
3299 * pwq_set_max_active - adjust max_active of a pwq
3300 * @pwq: target pool_workqueue
3301 * @max_active: new max_active value.
3303 * Set @pwq->max_active to @max_active and activate delayed works if
3307 * spin_lock_irq(pool->lock).
3309 static void pwq_set_max_active(struct pool_workqueue
*pwq
, int max_active
)
3311 pwq
->max_active
= max_active
;
3313 while (!list_empty(&pwq
->delayed_works
) &&
3314 pwq
->nr_active
< pwq
->max_active
)
3315 pwq_activate_first_delayed(pwq
);
3319 * workqueue_set_max_active - adjust max_active of a workqueue
3320 * @wq: target workqueue
3321 * @max_active: new max_active value.
3323 * Set max_active of @wq to @max_active.
3326 * Don't call from IRQ context.
3328 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
3332 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
3334 spin_lock(&workqueue_lock
);
3336 wq
->saved_max_active
= max_active
;
3338 for_each_pwq_cpu(cpu
, wq
) {
3339 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3340 struct worker_pool
*pool
= pwq
->pool
;
3342 spin_lock_irq(&pool
->lock
);
3344 if (!(wq
->flags
& WQ_FREEZABLE
) ||
3345 !(pool
->flags
& POOL_FREEZING
))
3346 pwq_set_max_active(pwq
, max_active
);
3348 spin_unlock_irq(&pool
->lock
);
3351 spin_unlock(&workqueue_lock
);
3353 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
3356 * workqueue_congested - test whether a workqueue is congested
3357 * @cpu: CPU in question
3358 * @wq: target workqueue
3360 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3361 * no synchronization around this function and the test result is
3362 * unreliable and only useful as advisory hints or for debugging.
3365 * %true if congested, %false otherwise.
3367 bool workqueue_congested(unsigned int cpu
, struct workqueue_struct
*wq
)
3369 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3371 return !list_empty(&pwq
->delayed_works
);
3373 EXPORT_SYMBOL_GPL(workqueue_congested
);
3376 * work_busy - test whether a work is currently pending or running
3377 * @work: the work to be tested
3379 * Test whether @work is currently pending or running. There is no
3380 * synchronization around this function and the test result is
3381 * unreliable and only useful as advisory hints or for debugging.
3384 * OR'd bitmask of WORK_BUSY_* bits.
3386 unsigned int work_busy(struct work_struct
*work
)
3388 struct worker_pool
*pool
= get_work_pool(work
);
3389 unsigned long flags
;
3390 unsigned int ret
= 0;
3392 if (work_pending(work
))
3393 ret
|= WORK_BUSY_PENDING
;
3396 spin_lock_irqsave(&pool
->lock
, flags
);
3397 if (find_worker_executing_work(pool
, work
))
3398 ret
|= WORK_BUSY_RUNNING
;
3399 spin_unlock_irqrestore(&pool
->lock
, flags
);
3404 EXPORT_SYMBOL_GPL(work_busy
);
3409 * There are two challenges in supporting CPU hotplug. Firstly, there
3410 * are a lot of assumptions on strong associations among work, pwq and
3411 * pool which make migrating pending and scheduled works very
3412 * difficult to implement without impacting hot paths. Secondly,
3413 * worker pools serve mix of short, long and very long running works making
3414 * blocked draining impractical.
3416 * This is solved by allowing the pools to be disassociated from the CPU
3417 * running as an unbound one and allowing it to be reattached later if the
3418 * cpu comes back online.
3421 static void wq_unbind_fn(struct work_struct
*work
)
3423 int cpu
= smp_processor_id();
3424 struct worker_pool
*pool
;
3425 struct worker
*worker
;
3426 struct hlist_node
*pos
;
3429 for_each_std_worker_pool(pool
, cpu
) {
3430 BUG_ON(cpu
!= smp_processor_id());
3432 mutex_lock(&pool
->assoc_mutex
);
3433 spin_lock_irq(&pool
->lock
);
3436 * We've claimed all manager positions. Make all workers
3437 * unbound and set DISASSOCIATED. Before this, all workers
3438 * except for the ones which are still executing works from
3439 * before the last CPU down must be on the cpu. After
3440 * this, they may become diasporas.
3442 list_for_each_entry(worker
, &pool
->idle_list
, entry
)
3443 worker
->flags
|= WORKER_UNBOUND
;
3445 for_each_busy_worker(worker
, i
, pos
, pool
)
3446 worker
->flags
|= WORKER_UNBOUND
;
3448 pool
->flags
|= POOL_DISASSOCIATED
;
3450 spin_unlock_irq(&pool
->lock
);
3451 mutex_unlock(&pool
->assoc_mutex
);
3455 * Call schedule() so that we cross rq->lock and thus can guarantee
3456 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3457 * as scheduler callbacks may be invoked from other cpus.
3462 * Sched callbacks are disabled now. Zap nr_running. After this,
3463 * nr_running stays zero and need_more_worker() and keep_working()
3464 * are always true as long as the worklist is not empty. Pools on
3465 * @cpu now behave as unbound (in terms of concurrency management)
3466 * pools which are served by workers tied to the CPU.
3468 * On return from this function, the current worker would trigger
3469 * unbound chain execution of pending work items if other workers
3472 for_each_std_worker_pool(pool
, cpu
)
3473 atomic_set(&pool
->nr_running
, 0);
3477 * Workqueues should be brought up before normal priority CPU notifiers.
3478 * This will be registered high priority CPU notifier.
3480 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
3481 unsigned long action
,
3484 unsigned int cpu
= (unsigned long)hcpu
;
3485 struct worker_pool
*pool
;
3487 switch (action
& ~CPU_TASKS_FROZEN
) {
3488 case CPU_UP_PREPARE
:
3489 for_each_std_worker_pool(pool
, cpu
) {
3490 struct worker
*worker
;
3492 if (pool
->nr_workers
)
3495 worker
= create_worker(pool
);
3499 spin_lock_irq(&pool
->lock
);
3500 start_worker(worker
);
3501 spin_unlock_irq(&pool
->lock
);
3505 case CPU_DOWN_FAILED
:
3507 for_each_std_worker_pool(pool
, cpu
) {
3508 mutex_lock(&pool
->assoc_mutex
);
3509 spin_lock_irq(&pool
->lock
);
3511 pool
->flags
&= ~POOL_DISASSOCIATED
;
3512 rebind_workers(pool
);
3514 spin_unlock_irq(&pool
->lock
);
3515 mutex_unlock(&pool
->assoc_mutex
);
3523 * Workqueues should be brought down after normal priority CPU notifiers.
3524 * This will be registered as low priority CPU notifier.
3526 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
3527 unsigned long action
,
3530 unsigned int cpu
= (unsigned long)hcpu
;
3531 struct work_struct unbind_work
;
3533 switch (action
& ~CPU_TASKS_FROZEN
) {
3534 case CPU_DOWN_PREPARE
:
3535 /* unbinding should happen on the local CPU */
3536 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
3537 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
3538 flush_work(&unbind_work
);
3546 struct work_for_cpu
{
3547 struct work_struct work
;
3553 static void work_for_cpu_fn(struct work_struct
*work
)
3555 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
3557 wfc
->ret
= wfc
->fn(wfc
->arg
);
3561 * work_on_cpu - run a function in user context on a particular cpu
3562 * @cpu: the cpu to run on
3563 * @fn: the function to run
3564 * @arg: the function arg
3566 * This will return the value @fn returns.
3567 * It is up to the caller to ensure that the cpu doesn't go offline.
3568 * The caller must not hold any locks which would prevent @fn from completing.
3570 long work_on_cpu(unsigned int cpu
, long (*fn
)(void *), void *arg
)
3572 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
3574 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
3575 schedule_work_on(cpu
, &wfc
.work
);
3576 flush_work(&wfc
.work
);
3579 EXPORT_SYMBOL_GPL(work_on_cpu
);
3580 #endif /* CONFIG_SMP */
3582 #ifdef CONFIG_FREEZER
3585 * freeze_workqueues_begin - begin freezing workqueues
3587 * Start freezing workqueues. After this function returns, all freezable
3588 * workqueues will queue new works to their frozen_works list instead of
3592 * Grabs and releases workqueue_lock and pool->lock's.
3594 void freeze_workqueues_begin(void)
3598 spin_lock(&workqueue_lock
);
3600 BUG_ON(workqueue_freezing
);
3601 workqueue_freezing
= true;
3603 for_each_wq_cpu(cpu
) {
3604 struct worker_pool
*pool
;
3605 struct workqueue_struct
*wq
;
3607 for_each_std_worker_pool(pool
, cpu
) {
3608 spin_lock_irq(&pool
->lock
);
3610 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
3611 pool
->flags
|= POOL_FREEZING
;
3613 list_for_each_entry(wq
, &workqueues
, list
) {
3614 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3616 if (pwq
&& pwq
->pool
== pool
&&
3617 (wq
->flags
& WQ_FREEZABLE
))
3618 pwq
->max_active
= 0;
3621 spin_unlock_irq(&pool
->lock
);
3625 spin_unlock(&workqueue_lock
);
3629 * freeze_workqueues_busy - are freezable workqueues still busy?
3631 * Check whether freezing is complete. This function must be called
3632 * between freeze_workqueues_begin() and thaw_workqueues().
3635 * Grabs and releases workqueue_lock.
3638 * %true if some freezable workqueues are still busy. %false if freezing
3641 bool freeze_workqueues_busy(void)
3646 spin_lock(&workqueue_lock
);
3648 BUG_ON(!workqueue_freezing
);
3650 for_each_wq_cpu(cpu
) {
3651 struct workqueue_struct
*wq
;
3653 * nr_active is monotonically decreasing. It's safe
3654 * to peek without lock.
3656 list_for_each_entry(wq
, &workqueues
, list
) {
3657 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3659 if (!pwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3662 BUG_ON(pwq
->nr_active
< 0);
3663 if (pwq
->nr_active
) {
3670 spin_unlock(&workqueue_lock
);
3675 * thaw_workqueues - thaw workqueues
3677 * Thaw workqueues. Normal queueing is restored and all collected
3678 * frozen works are transferred to their respective pool worklists.
3681 * Grabs and releases workqueue_lock and pool->lock's.
3683 void thaw_workqueues(void)
3687 spin_lock(&workqueue_lock
);
3689 if (!workqueue_freezing
)
3692 for_each_wq_cpu(cpu
) {
3693 struct worker_pool
*pool
;
3694 struct workqueue_struct
*wq
;
3696 for_each_std_worker_pool(pool
, cpu
) {
3697 spin_lock_irq(&pool
->lock
);
3699 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
3700 pool
->flags
&= ~POOL_FREEZING
;
3702 list_for_each_entry(wq
, &workqueues
, list
) {
3703 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3705 if (!pwq
|| pwq
->pool
!= pool
||
3706 !(wq
->flags
& WQ_FREEZABLE
))
3709 /* restore max_active and repopulate worklist */
3710 pwq_set_max_active(pwq
, wq
->saved_max_active
);
3713 wake_up_worker(pool
);
3715 spin_unlock_irq(&pool
->lock
);
3719 workqueue_freezing
= false;
3721 spin_unlock(&workqueue_lock
);
3723 #endif /* CONFIG_FREEZER */
3725 static int __init
init_workqueues(void)
3729 /* make sure we have enough bits for OFFQ pool ID */
3730 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_POOL_SHIFT
)) <
3731 WORK_CPU_END
* NR_STD_WORKER_POOLS
);
3733 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
3734 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
3736 /* initialize CPU pools */
3737 for_each_wq_cpu(cpu
) {
3738 struct worker_pool
*pool
;
3740 for_each_std_worker_pool(pool
, cpu
) {
3741 spin_lock_init(&pool
->lock
);
3743 pool
->flags
|= POOL_DISASSOCIATED
;
3744 INIT_LIST_HEAD(&pool
->worklist
);
3745 INIT_LIST_HEAD(&pool
->idle_list
);
3746 hash_init(pool
->busy_hash
);
3748 init_timer_deferrable(&pool
->idle_timer
);
3749 pool
->idle_timer
.function
= idle_worker_timeout
;
3750 pool
->idle_timer
.data
= (unsigned long)pool
;
3752 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3753 (unsigned long)pool
);
3755 mutex_init(&pool
->assoc_mutex
);
3756 ida_init(&pool
->worker_ida
);
3759 BUG_ON(worker_pool_assign_id(pool
));
3763 /* create the initial worker */
3764 for_each_online_wq_cpu(cpu
) {
3765 struct worker_pool
*pool
;
3767 for_each_std_worker_pool(pool
, cpu
) {
3768 struct worker
*worker
;
3770 if (cpu
!= WORK_CPU_UNBOUND
)
3771 pool
->flags
&= ~POOL_DISASSOCIATED
;
3773 worker
= create_worker(pool
);
3775 spin_lock_irq(&pool
->lock
);
3776 start_worker(worker
);
3777 spin_unlock_irq(&pool
->lock
);
3781 system_wq
= alloc_workqueue("events", 0, 0);
3782 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
3783 system_long_wq
= alloc_workqueue("events_long", 0, 0);
3784 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
3785 WQ_UNBOUND_MAX_ACTIVE
);
3786 system_freezable_wq
= alloc_workqueue("events_freezable",
3788 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
3789 !system_unbound_wq
|| !system_freezable_wq
);
3792 early_initcall(init_workqueues
);