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, pool) \
255 hash_for_each(pool->busy_hash, i, 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
;
913 hash_for_each_possible(pool
->busy_hash
, worker
, hentry
,
915 if (worker
->current_work
== work
&&
916 worker
->current_func
== work
->func
)
923 * move_linked_works - move linked works to a list
924 * @work: start of series of works to be scheduled
925 * @head: target list to append @work to
926 * @nextp: out paramter for nested worklist walking
928 * Schedule linked works starting from @work to @head. Work series to
929 * be scheduled starts at @work and includes any consecutive work with
930 * WORK_STRUCT_LINKED set in its predecessor.
932 * If @nextp is not NULL, it's updated to point to the next work of
933 * the last scheduled work. This allows move_linked_works() to be
934 * nested inside outer list_for_each_entry_safe().
937 * spin_lock_irq(pool->lock).
939 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
940 struct work_struct
**nextp
)
942 struct work_struct
*n
;
945 * Linked worklist will always end before the end of the list,
946 * use NULL for list head.
948 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
949 list_move_tail(&work
->entry
, head
);
950 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
955 * If we're already inside safe list traversal and have moved
956 * multiple works to the scheduled queue, the next position
957 * needs to be updated.
963 static void pwq_activate_delayed_work(struct work_struct
*work
)
965 struct pool_workqueue
*pwq
= get_work_pwq(work
);
967 trace_workqueue_activate_work(work
);
968 move_linked_works(work
, &pwq
->pool
->worklist
, NULL
);
969 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
973 static void pwq_activate_first_delayed(struct pool_workqueue
*pwq
)
975 struct work_struct
*work
= list_first_entry(&pwq
->delayed_works
,
976 struct work_struct
, entry
);
978 pwq_activate_delayed_work(work
);
982 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
983 * @pwq: pwq of interest
984 * @color: color of work which left the queue
986 * A work either has completed or is removed from pending queue,
987 * decrement nr_in_flight of its pwq and handle workqueue flushing.
990 * spin_lock_irq(pool->lock).
992 static void pwq_dec_nr_in_flight(struct pool_workqueue
*pwq
, int color
)
994 /* ignore uncolored works */
995 if (color
== WORK_NO_COLOR
)
998 pwq
->nr_in_flight
[color
]--;
1001 if (!list_empty(&pwq
->delayed_works
)) {
1002 /* one down, submit a delayed one */
1003 if (pwq
->nr_active
< pwq
->max_active
)
1004 pwq_activate_first_delayed(pwq
);
1007 /* is flush in progress and are we at the flushing tip? */
1008 if (likely(pwq
->flush_color
!= color
))
1011 /* are there still in-flight works? */
1012 if (pwq
->nr_in_flight
[color
])
1015 /* this pwq is done, clear flush_color */
1016 pwq
->flush_color
= -1;
1019 * If this was the last pwq, wake up the first flusher. It
1020 * will handle the rest.
1022 if (atomic_dec_and_test(&pwq
->wq
->nr_pwqs_to_flush
))
1023 complete(&pwq
->wq
->first_flusher
->done
);
1027 * try_to_grab_pending - steal work item from worklist and disable irq
1028 * @work: work item to steal
1029 * @is_dwork: @work is a delayed_work
1030 * @flags: place to store irq state
1032 * Try to grab PENDING bit of @work. This function can handle @work in any
1033 * stable state - idle, on timer or on worklist. Return values are
1035 * 1 if @work was pending and we successfully stole PENDING
1036 * 0 if @work was idle and we claimed PENDING
1037 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1038 * -ENOENT if someone else is canceling @work, this state may persist
1039 * for arbitrarily long
1041 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1042 * interrupted while holding PENDING and @work off queue, irq must be
1043 * disabled on entry. This, combined with delayed_work->timer being
1044 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1046 * On successful return, >= 0, irq is disabled and the caller is
1047 * responsible for releasing it using local_irq_restore(*@flags).
1049 * This function is safe to call from any context including IRQ handler.
1051 static int try_to_grab_pending(struct work_struct
*work
, bool is_dwork
,
1052 unsigned long *flags
)
1054 struct worker_pool
*pool
;
1055 struct pool_workqueue
*pwq
;
1057 local_irq_save(*flags
);
1059 /* try to steal the timer if it exists */
1061 struct delayed_work
*dwork
= to_delayed_work(work
);
1064 * dwork->timer is irqsafe. If del_timer() fails, it's
1065 * guaranteed that the timer is not queued anywhere and not
1066 * running on the local CPU.
1068 if (likely(del_timer(&dwork
->timer
)))
1072 /* try to claim PENDING the normal way */
1073 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
1077 * The queueing is in progress, or it is already queued. Try to
1078 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1080 pool
= get_work_pool(work
);
1084 spin_lock(&pool
->lock
);
1086 * work->data is guaranteed to point to pwq only while the work
1087 * item is queued on pwq->wq, and both updating work->data to point
1088 * to pwq on queueing and to pool on dequeueing are done under
1089 * pwq->pool->lock. This in turn guarantees that, if work->data
1090 * points to pwq which is associated with a locked pool, the work
1091 * item is currently queued on that pool.
1093 pwq
= get_work_pwq(work
);
1094 if (pwq
&& pwq
->pool
== pool
) {
1095 debug_work_deactivate(work
);
1098 * A delayed work item cannot be grabbed directly because
1099 * it might have linked NO_COLOR work items which, if left
1100 * on the delayed_list, will confuse pwq->nr_active
1101 * management later on and cause stall. Make sure the work
1102 * item is activated before grabbing.
1104 if (*work_data_bits(work
) & WORK_STRUCT_DELAYED
)
1105 pwq_activate_delayed_work(work
);
1107 list_del_init(&work
->entry
);
1108 pwq_dec_nr_in_flight(get_work_pwq(work
), get_work_color(work
));
1110 /* work->data points to pwq iff queued, point to pool */
1111 set_work_pool_and_keep_pending(work
, pool
->id
);
1113 spin_unlock(&pool
->lock
);
1116 spin_unlock(&pool
->lock
);
1118 local_irq_restore(*flags
);
1119 if (work_is_canceling(work
))
1126 * insert_work - insert a work into a pool
1127 * @pwq: pwq @work belongs to
1128 * @work: work to insert
1129 * @head: insertion point
1130 * @extra_flags: extra WORK_STRUCT_* flags to set
1132 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1133 * work_struct flags.
1136 * spin_lock_irq(pool->lock).
1138 static void insert_work(struct pool_workqueue
*pwq
, struct work_struct
*work
,
1139 struct list_head
*head
, unsigned int extra_flags
)
1141 struct worker_pool
*pool
= pwq
->pool
;
1143 /* we own @work, set data and link */
1144 set_work_pwq(work
, pwq
, extra_flags
);
1145 list_add_tail(&work
->entry
, head
);
1148 * Ensure either worker_sched_deactivated() sees the above
1149 * list_add_tail() or we see zero nr_running to avoid workers
1150 * lying around lazily while there are works to be processed.
1154 if (__need_more_worker(pool
))
1155 wake_up_worker(pool
);
1159 * Test whether @work is being queued from another work executing on the
1162 static bool is_chained_work(struct workqueue_struct
*wq
)
1164 struct worker
*worker
;
1166 worker
= current_wq_worker();
1168 * Return %true iff I'm a worker execuing a work item on @wq. If
1169 * I'm @worker, it's safe to dereference it without locking.
1171 return worker
&& worker
->current_pwq
->wq
== wq
;
1174 static void __queue_work(unsigned int cpu
, struct workqueue_struct
*wq
,
1175 struct work_struct
*work
)
1177 struct pool_workqueue
*pwq
;
1178 struct list_head
*worklist
;
1179 unsigned int work_flags
;
1180 unsigned int req_cpu
= cpu
;
1183 * While a work item is PENDING && off queue, a task trying to
1184 * steal the PENDING will busy-loop waiting for it to either get
1185 * queued or lose PENDING. Grabbing PENDING and queueing should
1186 * happen with IRQ disabled.
1188 WARN_ON_ONCE(!irqs_disabled());
1190 debug_work_activate(work
);
1192 /* if dying, only works from the same workqueue are allowed */
1193 if (unlikely(wq
->flags
& WQ_DRAINING
) &&
1194 WARN_ON_ONCE(!is_chained_work(wq
)))
1197 /* determine the pwq to use */
1198 if (!(wq
->flags
& WQ_UNBOUND
)) {
1199 struct worker_pool
*last_pool
;
1201 if (cpu
== WORK_CPU_UNBOUND
)
1202 cpu
= raw_smp_processor_id();
1205 * It's multi cpu. If @work was previously on a different
1206 * cpu, it might still be running there, in which case the
1207 * work needs to be queued on that cpu to guarantee
1210 pwq
= get_pwq(cpu
, wq
);
1211 last_pool
= get_work_pool(work
);
1213 if (last_pool
&& last_pool
!= pwq
->pool
) {
1214 struct worker
*worker
;
1216 spin_lock(&last_pool
->lock
);
1218 worker
= find_worker_executing_work(last_pool
, work
);
1220 if (worker
&& worker
->current_pwq
->wq
== wq
) {
1221 pwq
= get_pwq(last_pool
->cpu
, wq
);
1223 /* meh... not running there, queue here */
1224 spin_unlock(&last_pool
->lock
);
1225 spin_lock(&pwq
->pool
->lock
);
1228 spin_lock(&pwq
->pool
->lock
);
1231 pwq
= get_pwq(WORK_CPU_UNBOUND
, wq
);
1232 spin_lock(&pwq
->pool
->lock
);
1235 /* pwq determined, queue */
1236 trace_workqueue_queue_work(req_cpu
, pwq
, work
);
1238 if (WARN_ON(!list_empty(&work
->entry
))) {
1239 spin_unlock(&pwq
->pool
->lock
);
1243 pwq
->nr_in_flight
[pwq
->work_color
]++;
1244 work_flags
= work_color_to_flags(pwq
->work_color
);
1246 if (likely(pwq
->nr_active
< pwq
->max_active
)) {
1247 trace_workqueue_activate_work(work
);
1249 worklist
= &pwq
->pool
->worklist
;
1251 work_flags
|= WORK_STRUCT_DELAYED
;
1252 worklist
= &pwq
->delayed_works
;
1255 insert_work(pwq
, work
, worklist
, work_flags
);
1257 spin_unlock(&pwq
->pool
->lock
);
1261 * queue_work_on - queue work on specific cpu
1262 * @cpu: CPU number to execute work on
1263 * @wq: workqueue to use
1264 * @work: work to queue
1266 * Returns %false if @work was already on a queue, %true otherwise.
1268 * We queue the work to a specific CPU, the caller must ensure it
1271 bool queue_work_on(int cpu
, struct workqueue_struct
*wq
,
1272 struct work_struct
*work
)
1275 unsigned long flags
;
1277 local_irq_save(flags
);
1279 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1280 __queue_work(cpu
, wq
, work
);
1284 local_irq_restore(flags
);
1287 EXPORT_SYMBOL_GPL(queue_work_on
);
1290 * queue_work - queue work on a workqueue
1291 * @wq: workqueue to use
1292 * @work: work to queue
1294 * Returns %false if @work was already on a queue, %true otherwise.
1296 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1297 * it can be processed by another CPU.
1299 bool queue_work(struct workqueue_struct
*wq
, struct work_struct
*work
)
1301 return queue_work_on(WORK_CPU_UNBOUND
, wq
, work
);
1303 EXPORT_SYMBOL_GPL(queue_work
);
1305 void delayed_work_timer_fn(unsigned long __data
)
1307 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1309 /* should have been called from irqsafe timer with irq already off */
1310 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
1312 EXPORT_SYMBOL(delayed_work_timer_fn
);
1314 static void __queue_delayed_work(int cpu
, struct workqueue_struct
*wq
,
1315 struct delayed_work
*dwork
, unsigned long delay
)
1317 struct timer_list
*timer
= &dwork
->timer
;
1318 struct work_struct
*work
= &dwork
->work
;
1320 WARN_ON_ONCE(timer
->function
!= delayed_work_timer_fn
||
1321 timer
->data
!= (unsigned long)dwork
);
1322 WARN_ON_ONCE(timer_pending(timer
));
1323 WARN_ON_ONCE(!list_empty(&work
->entry
));
1326 * If @delay is 0, queue @dwork->work immediately. This is for
1327 * both optimization and correctness. The earliest @timer can
1328 * expire is on the closest next tick and delayed_work users depend
1329 * on that there's no such delay when @delay is 0.
1332 __queue_work(cpu
, wq
, &dwork
->work
);
1336 timer_stats_timer_set_start_info(&dwork
->timer
);
1340 timer
->expires
= jiffies
+ delay
;
1342 if (unlikely(cpu
!= WORK_CPU_UNBOUND
))
1343 add_timer_on(timer
, cpu
);
1349 * queue_delayed_work_on - queue work on specific CPU after delay
1350 * @cpu: CPU number to execute work on
1351 * @wq: workqueue to use
1352 * @dwork: work to queue
1353 * @delay: number of jiffies to wait before queueing
1355 * Returns %false if @work was already on a queue, %true otherwise. If
1356 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1359 bool queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1360 struct delayed_work
*dwork
, unsigned long delay
)
1362 struct work_struct
*work
= &dwork
->work
;
1364 unsigned long flags
;
1366 /* read the comment in __queue_work() */
1367 local_irq_save(flags
);
1369 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1370 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1374 local_irq_restore(flags
);
1377 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1380 * queue_delayed_work - queue work on a workqueue after delay
1381 * @wq: workqueue to use
1382 * @dwork: delayable work to queue
1383 * @delay: number of jiffies to wait before queueing
1385 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1387 bool queue_delayed_work(struct workqueue_struct
*wq
,
1388 struct delayed_work
*dwork
, unsigned long delay
)
1390 return queue_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1392 EXPORT_SYMBOL_GPL(queue_delayed_work
);
1395 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1396 * @cpu: CPU number to execute work on
1397 * @wq: workqueue to use
1398 * @dwork: work to queue
1399 * @delay: number of jiffies to wait before queueing
1401 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1402 * modify @dwork's timer so that it expires after @delay. If @delay is
1403 * zero, @work is guaranteed to be scheduled immediately regardless of its
1406 * Returns %false if @dwork was idle and queued, %true if @dwork was
1407 * pending and its timer was modified.
1409 * This function is safe to call from any context including IRQ handler.
1410 * See try_to_grab_pending() for details.
1412 bool mod_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1413 struct delayed_work
*dwork
, unsigned long delay
)
1415 unsigned long flags
;
1419 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
1420 } while (unlikely(ret
== -EAGAIN
));
1422 if (likely(ret
>= 0)) {
1423 __queue_delayed_work(cpu
, wq
, dwork
, delay
);
1424 local_irq_restore(flags
);
1427 /* -ENOENT from try_to_grab_pending() becomes %true */
1430 EXPORT_SYMBOL_GPL(mod_delayed_work_on
);
1433 * mod_delayed_work - modify delay of or queue a delayed work
1434 * @wq: workqueue to use
1435 * @dwork: work to queue
1436 * @delay: number of jiffies to wait before queueing
1438 * mod_delayed_work_on() on local CPU.
1440 bool mod_delayed_work(struct workqueue_struct
*wq
, struct delayed_work
*dwork
,
1441 unsigned long delay
)
1443 return mod_delayed_work_on(WORK_CPU_UNBOUND
, wq
, dwork
, delay
);
1445 EXPORT_SYMBOL_GPL(mod_delayed_work
);
1448 * worker_enter_idle - enter idle state
1449 * @worker: worker which is entering idle state
1451 * @worker is entering idle state. Update stats and idle timer if
1455 * spin_lock_irq(pool->lock).
1457 static void worker_enter_idle(struct worker
*worker
)
1459 struct worker_pool
*pool
= worker
->pool
;
1461 BUG_ON(worker
->flags
& WORKER_IDLE
);
1462 BUG_ON(!list_empty(&worker
->entry
) &&
1463 (worker
->hentry
.next
|| worker
->hentry
.pprev
));
1465 /* can't use worker_set_flags(), also called from start_worker() */
1466 worker
->flags
|= WORKER_IDLE
;
1468 worker
->last_active
= jiffies
;
1470 /* idle_list is LIFO */
1471 list_add(&worker
->entry
, &pool
->idle_list
);
1473 if (too_many_workers(pool
) && !timer_pending(&pool
->idle_timer
))
1474 mod_timer(&pool
->idle_timer
, jiffies
+ IDLE_WORKER_TIMEOUT
);
1477 * Sanity check nr_running. Because wq_unbind_fn() releases
1478 * pool->lock between setting %WORKER_UNBOUND and zapping
1479 * nr_running, the warning may trigger spuriously. Check iff
1480 * unbind is not in progress.
1482 WARN_ON_ONCE(!(pool
->flags
& POOL_DISASSOCIATED
) &&
1483 pool
->nr_workers
== pool
->nr_idle
&&
1484 atomic_read(&pool
->nr_running
));
1488 * worker_leave_idle - leave idle state
1489 * @worker: worker which is leaving idle state
1491 * @worker is leaving idle state. Update stats.
1494 * spin_lock_irq(pool->lock).
1496 static void worker_leave_idle(struct worker
*worker
)
1498 struct worker_pool
*pool
= worker
->pool
;
1500 BUG_ON(!(worker
->flags
& WORKER_IDLE
));
1501 worker_clr_flags(worker
, WORKER_IDLE
);
1503 list_del_init(&worker
->entry
);
1507 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock pool
1510 * Works which are scheduled while the cpu is online must at least be
1511 * scheduled to a worker which is bound to the cpu so that if they are
1512 * flushed from cpu callbacks while cpu is going down, they are
1513 * guaranteed to execute on the cpu.
1515 * This function is to be used by rogue workers and rescuers to bind
1516 * themselves to the target cpu and may race with cpu going down or
1517 * coming online. kthread_bind() can't be used because it may put the
1518 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1519 * verbatim as it's best effort and blocking and pool may be
1520 * [dis]associated in the meantime.
1522 * This function tries set_cpus_allowed() and locks pool and verifies the
1523 * binding against %POOL_DISASSOCIATED which is set during
1524 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1525 * enters idle state or fetches works without dropping lock, it can
1526 * guarantee the scheduling requirement described in the first paragraph.
1529 * Might sleep. Called without any lock but returns with pool->lock
1533 * %true if the associated pool is online (@worker is successfully
1534 * bound), %false if offline.
1536 static bool worker_maybe_bind_and_lock(struct worker
*worker
)
1537 __acquires(&pool
->lock
)
1539 struct worker_pool
*pool
= worker
->pool
;
1540 struct task_struct
*task
= worker
->task
;
1544 * The following call may fail, succeed or succeed
1545 * without actually migrating the task to the cpu if
1546 * it races with cpu hotunplug operation. Verify
1547 * against POOL_DISASSOCIATED.
1549 if (!(pool
->flags
& POOL_DISASSOCIATED
))
1550 set_cpus_allowed_ptr(task
, get_cpu_mask(pool
->cpu
));
1552 spin_lock_irq(&pool
->lock
);
1553 if (pool
->flags
& POOL_DISASSOCIATED
)
1555 if (task_cpu(task
) == pool
->cpu
&&
1556 cpumask_equal(¤t
->cpus_allowed
,
1557 get_cpu_mask(pool
->cpu
)))
1559 spin_unlock_irq(&pool
->lock
);
1562 * We've raced with CPU hot[un]plug. Give it a breather
1563 * and retry migration. cond_resched() is required here;
1564 * otherwise, we might deadlock against cpu_stop trying to
1565 * bring down the CPU on non-preemptive kernel.
1573 * Rebind an idle @worker to its CPU. worker_thread() will test
1574 * list_empty(@worker->entry) before leaving idle and call this function.
1576 static void idle_worker_rebind(struct worker
*worker
)
1578 /* CPU may go down again inbetween, clear UNBOUND only on success */
1579 if (worker_maybe_bind_and_lock(worker
))
1580 worker_clr_flags(worker
, WORKER_UNBOUND
);
1582 /* rebind complete, become available again */
1583 list_add(&worker
->entry
, &worker
->pool
->idle_list
);
1584 spin_unlock_irq(&worker
->pool
->lock
);
1588 * Function for @worker->rebind.work used to rebind unbound busy workers to
1589 * the associated cpu which is coming back online. This is scheduled by
1590 * cpu up but can race with other cpu hotplug operations and may be
1591 * executed twice without intervening cpu down.
1593 static void busy_worker_rebind_fn(struct work_struct
*work
)
1595 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1597 if (worker_maybe_bind_and_lock(worker
))
1598 worker_clr_flags(worker
, WORKER_UNBOUND
);
1600 spin_unlock_irq(&worker
->pool
->lock
);
1604 * rebind_workers - rebind all workers of a pool to the associated CPU
1605 * @pool: pool of interest
1607 * @pool->cpu is coming online. Rebind all workers to the CPU. Rebinding
1608 * is different for idle and busy ones.
1610 * Idle ones will be removed from the idle_list and woken up. They will
1611 * add themselves back after completing rebind. This ensures that the
1612 * idle_list doesn't contain any unbound workers when re-bound busy workers
1613 * try to perform local wake-ups for concurrency management.
1615 * Busy workers can rebind after they finish their current work items.
1616 * Queueing the rebind work item at the head of the scheduled list is
1617 * enough. Note that nr_running will be properly bumped as busy workers
1620 * On return, all non-manager workers are scheduled for rebind - see
1621 * manage_workers() for the manager special case. Any idle worker
1622 * including the manager will not appear on @idle_list until rebind is
1623 * complete, making local wake-ups safe.
1625 static void rebind_workers(struct worker_pool
*pool
)
1627 struct worker
*worker
, *n
;
1630 lockdep_assert_held(&pool
->assoc_mutex
);
1631 lockdep_assert_held(&pool
->lock
);
1633 /* dequeue and kick idle ones */
1634 list_for_each_entry_safe(worker
, n
, &pool
->idle_list
, entry
) {
1636 * idle workers should be off @pool->idle_list until rebind
1637 * is complete to avoid receiving premature local wake-ups.
1639 list_del_init(&worker
->entry
);
1642 * worker_thread() will see the above dequeuing and call
1643 * idle_worker_rebind().
1645 wake_up_process(worker
->task
);
1648 /* rebind busy workers */
1649 for_each_busy_worker(worker
, i
, pool
) {
1650 struct work_struct
*rebind_work
= &worker
->rebind_work
;
1651 struct workqueue_struct
*wq
;
1653 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
1654 work_data_bits(rebind_work
)))
1657 debug_work_activate(rebind_work
);
1660 * wq doesn't really matter but let's keep @worker->pool
1661 * and @pwq->pool consistent for sanity.
1663 if (std_worker_pool_pri(worker
->pool
))
1664 wq
= system_highpri_wq
;
1668 insert_work(get_pwq(pool
->cpu
, wq
), rebind_work
,
1669 worker
->scheduled
.next
,
1670 work_color_to_flags(WORK_NO_COLOR
));
1674 static struct worker
*alloc_worker(void)
1676 struct worker
*worker
;
1678 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1680 INIT_LIST_HEAD(&worker
->entry
);
1681 INIT_LIST_HEAD(&worker
->scheduled
);
1682 INIT_WORK(&worker
->rebind_work
, busy_worker_rebind_fn
);
1683 /* on creation a worker is in !idle && prep state */
1684 worker
->flags
= WORKER_PREP
;
1690 * create_worker - create a new workqueue worker
1691 * @pool: pool the new worker will belong to
1693 * Create a new worker which is bound to @pool. The returned worker
1694 * can be started by calling start_worker() or destroyed using
1698 * Might sleep. Does GFP_KERNEL allocations.
1701 * Pointer to the newly created worker.
1703 static struct worker
*create_worker(struct worker_pool
*pool
)
1705 const char *pri
= std_worker_pool_pri(pool
) ? "H" : "";
1706 struct worker
*worker
= NULL
;
1709 spin_lock_irq(&pool
->lock
);
1710 while (ida_get_new(&pool
->worker_ida
, &id
)) {
1711 spin_unlock_irq(&pool
->lock
);
1712 if (!ida_pre_get(&pool
->worker_ida
, GFP_KERNEL
))
1714 spin_lock_irq(&pool
->lock
);
1716 spin_unlock_irq(&pool
->lock
);
1718 worker
= alloc_worker();
1722 worker
->pool
= pool
;
1725 if (pool
->cpu
!= WORK_CPU_UNBOUND
)
1726 worker
->task
= kthread_create_on_node(worker_thread
,
1727 worker
, cpu_to_node(pool
->cpu
),
1728 "kworker/%u:%d%s", pool
->cpu
, id
, pri
);
1730 worker
->task
= kthread_create(worker_thread
, worker
,
1731 "kworker/u:%d%s", id
, pri
);
1732 if (IS_ERR(worker
->task
))
1735 if (std_worker_pool_pri(pool
))
1736 set_user_nice(worker
->task
, HIGHPRI_NICE_LEVEL
);
1739 * Determine CPU binding of the new worker depending on
1740 * %POOL_DISASSOCIATED. The caller is responsible for ensuring the
1741 * flag remains stable across this function. See the comments
1742 * above the flag definition for details.
1744 * As an unbound worker may later become a regular one if CPU comes
1745 * online, make sure every worker has %PF_THREAD_BOUND set.
1747 if (!(pool
->flags
& POOL_DISASSOCIATED
)) {
1748 kthread_bind(worker
->task
, pool
->cpu
);
1750 worker
->task
->flags
|= PF_THREAD_BOUND
;
1751 worker
->flags
|= WORKER_UNBOUND
;
1757 spin_lock_irq(&pool
->lock
);
1758 ida_remove(&pool
->worker_ida
, id
);
1759 spin_unlock_irq(&pool
->lock
);
1766 * start_worker - start a newly created worker
1767 * @worker: worker to start
1769 * Make the pool aware of @worker and start it.
1772 * spin_lock_irq(pool->lock).
1774 static void start_worker(struct worker
*worker
)
1776 worker
->flags
|= WORKER_STARTED
;
1777 worker
->pool
->nr_workers
++;
1778 worker_enter_idle(worker
);
1779 wake_up_process(worker
->task
);
1783 * destroy_worker - destroy a workqueue worker
1784 * @worker: worker to be destroyed
1786 * Destroy @worker and adjust @pool stats accordingly.
1789 * spin_lock_irq(pool->lock) which is released and regrabbed.
1791 static void destroy_worker(struct worker
*worker
)
1793 struct worker_pool
*pool
= worker
->pool
;
1794 int id
= worker
->id
;
1796 /* sanity check frenzy */
1797 BUG_ON(worker
->current_work
);
1798 BUG_ON(!list_empty(&worker
->scheduled
));
1800 if (worker
->flags
& WORKER_STARTED
)
1802 if (worker
->flags
& WORKER_IDLE
)
1805 list_del_init(&worker
->entry
);
1806 worker
->flags
|= WORKER_DIE
;
1808 spin_unlock_irq(&pool
->lock
);
1810 kthread_stop(worker
->task
);
1813 spin_lock_irq(&pool
->lock
);
1814 ida_remove(&pool
->worker_ida
, id
);
1817 static void idle_worker_timeout(unsigned long __pool
)
1819 struct worker_pool
*pool
= (void *)__pool
;
1821 spin_lock_irq(&pool
->lock
);
1823 if (too_many_workers(pool
)) {
1824 struct worker
*worker
;
1825 unsigned long expires
;
1827 /* idle_list is kept in LIFO order, check the last one */
1828 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1829 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1831 if (time_before(jiffies
, expires
))
1832 mod_timer(&pool
->idle_timer
, expires
);
1834 /* it's been idle for too long, wake up manager */
1835 pool
->flags
|= POOL_MANAGE_WORKERS
;
1836 wake_up_worker(pool
);
1840 spin_unlock_irq(&pool
->lock
);
1843 static bool send_mayday(struct work_struct
*work
)
1845 struct pool_workqueue
*pwq
= get_work_pwq(work
);
1846 struct workqueue_struct
*wq
= pwq
->wq
;
1849 if (!(wq
->flags
& WQ_RESCUER
))
1852 /* mayday mayday mayday */
1853 cpu
= pwq
->pool
->cpu
;
1854 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1855 if (cpu
== WORK_CPU_UNBOUND
)
1857 if (!mayday_test_and_set_cpu(cpu
, wq
->mayday_mask
))
1858 wake_up_process(wq
->rescuer
->task
);
1862 static void pool_mayday_timeout(unsigned long __pool
)
1864 struct worker_pool
*pool
= (void *)__pool
;
1865 struct work_struct
*work
;
1867 spin_lock_irq(&pool
->lock
);
1869 if (need_to_create_worker(pool
)) {
1871 * We've been trying to create a new worker but
1872 * haven't been successful. We might be hitting an
1873 * allocation deadlock. Send distress signals to
1876 list_for_each_entry(work
, &pool
->worklist
, entry
)
1880 spin_unlock_irq(&pool
->lock
);
1882 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1886 * maybe_create_worker - create a new worker if necessary
1887 * @pool: pool to create a new worker for
1889 * Create a new worker for @pool if necessary. @pool is guaranteed to
1890 * have at least one idle worker on return from this function. If
1891 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1892 * sent to all rescuers with works scheduled on @pool to resolve
1893 * possible allocation deadlock.
1895 * On return, need_to_create_worker() is guaranteed to be false and
1896 * may_start_working() true.
1899 * spin_lock_irq(pool->lock) which may be released and regrabbed
1900 * multiple times. Does GFP_KERNEL allocations. Called only from
1904 * false if no action was taken and pool->lock stayed locked, true
1907 static bool maybe_create_worker(struct worker_pool
*pool
)
1908 __releases(&pool
->lock
)
1909 __acquires(&pool
->lock
)
1911 if (!need_to_create_worker(pool
))
1914 spin_unlock_irq(&pool
->lock
);
1916 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1917 mod_timer(&pool
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1920 struct worker
*worker
;
1922 worker
= create_worker(pool
);
1924 del_timer_sync(&pool
->mayday_timer
);
1925 spin_lock_irq(&pool
->lock
);
1926 start_worker(worker
);
1927 BUG_ON(need_to_create_worker(pool
));
1931 if (!need_to_create_worker(pool
))
1934 __set_current_state(TASK_INTERRUPTIBLE
);
1935 schedule_timeout(CREATE_COOLDOWN
);
1937 if (!need_to_create_worker(pool
))
1941 del_timer_sync(&pool
->mayday_timer
);
1942 spin_lock_irq(&pool
->lock
);
1943 if (need_to_create_worker(pool
))
1949 * maybe_destroy_worker - destroy workers which have been idle for a while
1950 * @pool: pool to destroy workers for
1952 * Destroy @pool workers which have been idle for longer than
1953 * IDLE_WORKER_TIMEOUT.
1956 * spin_lock_irq(pool->lock) which may be released and regrabbed
1957 * multiple times. Called only from manager.
1960 * false if no action was taken and pool->lock stayed locked, true
1963 static bool maybe_destroy_workers(struct worker_pool
*pool
)
1967 while (too_many_workers(pool
)) {
1968 struct worker
*worker
;
1969 unsigned long expires
;
1971 worker
= list_entry(pool
->idle_list
.prev
, struct worker
, entry
);
1972 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1974 if (time_before(jiffies
, expires
)) {
1975 mod_timer(&pool
->idle_timer
, expires
);
1979 destroy_worker(worker
);
1987 * manage_workers - manage worker pool
1990 * Assume the manager role and manage the worker pool @worker belongs
1991 * to. At any given time, there can be only zero or one manager per
1992 * pool. The exclusion is handled automatically by this function.
1994 * The caller can safely start processing works on false return. On
1995 * true return, it's guaranteed that need_to_create_worker() is false
1996 * and may_start_working() is true.
1999 * spin_lock_irq(pool->lock) which may be released and regrabbed
2000 * multiple times. Does GFP_KERNEL allocations.
2003 * spin_lock_irq(pool->lock) which may be released and regrabbed
2004 * multiple times. Does GFP_KERNEL allocations.
2006 static bool manage_workers(struct worker
*worker
)
2008 struct worker_pool
*pool
= worker
->pool
;
2011 if (pool
->flags
& POOL_MANAGING_WORKERS
)
2014 pool
->flags
|= POOL_MANAGING_WORKERS
;
2017 * To simplify both worker management and CPU hotplug, hold off
2018 * management while hotplug is in progress. CPU hotplug path can't
2019 * grab %POOL_MANAGING_WORKERS to achieve this because that can
2020 * lead to idle worker depletion (all become busy thinking someone
2021 * else is managing) which in turn can result in deadlock under
2022 * extreme circumstances. Use @pool->assoc_mutex to synchronize
2023 * manager against CPU hotplug.
2025 * assoc_mutex would always be free unless CPU hotplug is in
2026 * progress. trylock first without dropping @pool->lock.
2028 if (unlikely(!mutex_trylock(&pool
->assoc_mutex
))) {
2029 spin_unlock_irq(&pool
->lock
);
2030 mutex_lock(&pool
->assoc_mutex
);
2032 * CPU hotplug could have happened while we were waiting
2033 * for assoc_mutex. Hotplug itself can't handle us
2034 * because manager isn't either on idle or busy list, and
2035 * @pool's state and ours could have deviated.
2037 * As hotplug is now excluded via assoc_mutex, we can
2038 * simply try to bind. It will succeed or fail depending
2039 * on @pool's current state. Try it and adjust
2040 * %WORKER_UNBOUND accordingly.
2042 if (worker_maybe_bind_and_lock(worker
))
2043 worker
->flags
&= ~WORKER_UNBOUND
;
2045 worker
->flags
|= WORKER_UNBOUND
;
2050 pool
->flags
&= ~POOL_MANAGE_WORKERS
;
2053 * Destroy and then create so that may_start_working() is true
2056 ret
|= maybe_destroy_workers(pool
);
2057 ret
|= maybe_create_worker(pool
);
2059 pool
->flags
&= ~POOL_MANAGING_WORKERS
;
2060 mutex_unlock(&pool
->assoc_mutex
);
2065 * process_one_work - process single work
2067 * @work: work to process
2069 * Process @work. This function contains all the logics necessary to
2070 * process a single work including synchronization against and
2071 * interaction with other workers on the same cpu, queueing and
2072 * flushing. As long as context requirement is met, any worker can
2073 * call this function to process a work.
2076 * spin_lock_irq(pool->lock) which is released and regrabbed.
2078 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
2079 __releases(&pool
->lock
)
2080 __acquires(&pool
->lock
)
2082 struct pool_workqueue
*pwq
= get_work_pwq(work
);
2083 struct worker_pool
*pool
= worker
->pool
;
2084 bool cpu_intensive
= pwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
2086 struct worker
*collision
;
2087 #ifdef CONFIG_LOCKDEP
2089 * It is permissible to free the struct work_struct from
2090 * inside the function that is called from it, this we need to
2091 * take into account for lockdep too. To avoid bogus "held
2092 * lock freed" warnings as well as problems when looking into
2093 * work->lockdep_map, make a copy and use that here.
2095 struct lockdep_map lockdep_map
;
2097 lockdep_copy_map(&lockdep_map
, &work
->lockdep_map
);
2100 * Ensure we're on the correct CPU. DISASSOCIATED test is
2101 * necessary to avoid spurious warnings from rescuers servicing the
2102 * unbound or a disassociated pool.
2104 WARN_ON_ONCE(!(worker
->flags
& WORKER_UNBOUND
) &&
2105 !(pool
->flags
& POOL_DISASSOCIATED
) &&
2106 raw_smp_processor_id() != pool
->cpu
);
2109 * A single work shouldn't be executed concurrently by
2110 * multiple workers on a single cpu. Check whether anyone is
2111 * already processing the work. If so, defer the work to the
2112 * currently executing one.
2114 collision
= find_worker_executing_work(pool
, work
);
2115 if (unlikely(collision
)) {
2116 move_linked_works(work
, &collision
->scheduled
, NULL
);
2120 /* claim and dequeue */
2121 debug_work_deactivate(work
);
2122 hash_add(pool
->busy_hash
, &worker
->hentry
, (unsigned long)work
);
2123 worker
->current_work
= work
;
2124 worker
->current_func
= work
->func
;
2125 worker
->current_pwq
= pwq
;
2126 work_color
= get_work_color(work
);
2128 list_del_init(&work
->entry
);
2131 * CPU intensive works don't participate in concurrency
2132 * management. They're the scheduler's responsibility.
2134 if (unlikely(cpu_intensive
))
2135 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
2138 * Unbound pool isn't concurrency managed and work items should be
2139 * executed ASAP. Wake up another worker if necessary.
2141 if ((worker
->flags
& WORKER_UNBOUND
) && need_more_worker(pool
))
2142 wake_up_worker(pool
);
2145 * Record the last pool and clear PENDING which should be the last
2146 * update to @work. Also, do this inside @pool->lock so that
2147 * PENDING and queued state changes happen together while IRQ is
2150 set_work_pool_and_clear_pending(work
, pool
->id
);
2152 spin_unlock_irq(&pool
->lock
);
2154 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2155 lock_map_acquire(&lockdep_map
);
2156 trace_workqueue_execute_start(work
);
2157 worker
->current_func(work
);
2159 * While we must be careful to not use "work" after this, the trace
2160 * point will only record its address.
2162 trace_workqueue_execute_end(work
);
2163 lock_map_release(&lockdep_map
);
2164 lock_map_release(&pwq
->wq
->lockdep_map
);
2166 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
2167 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2168 " last function: %pf\n",
2169 current
->comm
, preempt_count(), task_pid_nr(current
),
2170 worker
->current_func
);
2171 debug_show_held_locks(current
);
2175 spin_lock_irq(&pool
->lock
);
2177 /* clear cpu intensive status */
2178 if (unlikely(cpu_intensive
))
2179 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
2181 /* we're done with it, release */
2182 hash_del(&worker
->hentry
);
2183 worker
->current_work
= NULL
;
2184 worker
->current_func
= NULL
;
2185 worker
->current_pwq
= NULL
;
2186 pwq_dec_nr_in_flight(pwq
, work_color
);
2190 * process_scheduled_works - process scheduled works
2193 * Process all scheduled works. Please note that the scheduled list
2194 * may change while processing a work, so this function repeatedly
2195 * fetches a work from the top and executes it.
2198 * spin_lock_irq(pool->lock) which may be released and regrabbed
2201 static void process_scheduled_works(struct worker
*worker
)
2203 while (!list_empty(&worker
->scheduled
)) {
2204 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
2205 struct work_struct
, entry
);
2206 process_one_work(worker
, work
);
2211 * worker_thread - the worker thread function
2214 * The worker thread function. There are NR_CPU_WORKER_POOLS dynamic pools
2215 * of these per each cpu. These workers process all works regardless of
2216 * their specific target workqueue. The only exception is works which
2217 * belong to workqueues with a rescuer which will be explained in
2220 static int worker_thread(void *__worker
)
2222 struct worker
*worker
= __worker
;
2223 struct worker_pool
*pool
= worker
->pool
;
2225 /* tell the scheduler that this is a workqueue worker */
2226 worker
->task
->flags
|= PF_WQ_WORKER
;
2228 spin_lock_irq(&pool
->lock
);
2230 /* we are off idle list if destruction or rebind is requested */
2231 if (unlikely(list_empty(&worker
->entry
))) {
2232 spin_unlock_irq(&pool
->lock
);
2234 /* if DIE is set, destruction is requested */
2235 if (worker
->flags
& WORKER_DIE
) {
2236 worker
->task
->flags
&= ~PF_WQ_WORKER
;
2240 /* otherwise, rebind */
2241 idle_worker_rebind(worker
);
2245 worker_leave_idle(worker
);
2247 /* no more worker necessary? */
2248 if (!need_more_worker(pool
))
2251 /* do we need to manage? */
2252 if (unlikely(!may_start_working(pool
)) && manage_workers(worker
))
2256 * ->scheduled list can only be filled while a worker is
2257 * preparing to process a work or actually processing it.
2258 * Make sure nobody diddled with it while I was sleeping.
2260 BUG_ON(!list_empty(&worker
->scheduled
));
2263 * When control reaches this point, we're guaranteed to have
2264 * at least one idle worker or that someone else has already
2265 * assumed the manager role.
2267 worker_clr_flags(worker
, WORKER_PREP
);
2270 struct work_struct
*work
=
2271 list_first_entry(&pool
->worklist
,
2272 struct work_struct
, entry
);
2274 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
2275 /* optimization path, not strictly necessary */
2276 process_one_work(worker
, work
);
2277 if (unlikely(!list_empty(&worker
->scheduled
)))
2278 process_scheduled_works(worker
);
2280 move_linked_works(work
, &worker
->scheduled
, NULL
);
2281 process_scheduled_works(worker
);
2283 } while (keep_working(pool
));
2285 worker_set_flags(worker
, WORKER_PREP
, false);
2287 if (unlikely(need_to_manage_workers(pool
)) && manage_workers(worker
))
2291 * pool->lock is held and there's no work to process and no need to
2292 * manage, sleep. Workers are woken up only while holding
2293 * pool->lock or from local cpu, so setting the current state
2294 * before releasing pool->lock is enough to prevent losing any
2297 worker_enter_idle(worker
);
2298 __set_current_state(TASK_INTERRUPTIBLE
);
2299 spin_unlock_irq(&pool
->lock
);
2305 * rescuer_thread - the rescuer thread function
2308 * Workqueue rescuer thread function. There's one rescuer for each
2309 * workqueue which has WQ_RESCUER set.
2311 * Regular work processing on a pool may block trying to create a new
2312 * worker which uses GFP_KERNEL allocation which has slight chance of
2313 * developing into deadlock if some works currently on the same queue
2314 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2315 * the problem rescuer solves.
2317 * When such condition is possible, the pool summons rescuers of all
2318 * workqueues which have works queued on the pool and let them process
2319 * those works so that forward progress can be guaranteed.
2321 * This should happen rarely.
2323 static int rescuer_thread(void *__rescuer
)
2325 struct worker
*rescuer
= __rescuer
;
2326 struct workqueue_struct
*wq
= rescuer
->rescue_wq
;
2327 struct list_head
*scheduled
= &rescuer
->scheduled
;
2328 bool is_unbound
= wq
->flags
& WQ_UNBOUND
;
2331 set_user_nice(current
, RESCUER_NICE_LEVEL
);
2334 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2335 * doesn't participate in concurrency management.
2337 rescuer
->task
->flags
|= PF_WQ_WORKER
;
2339 set_current_state(TASK_INTERRUPTIBLE
);
2341 if (kthread_should_stop()) {
2342 __set_current_state(TASK_RUNNING
);
2343 rescuer
->task
->flags
&= ~PF_WQ_WORKER
;
2348 * See whether any cpu is asking for help. Unbounded
2349 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2351 for_each_mayday_cpu(cpu
, wq
->mayday_mask
) {
2352 unsigned int tcpu
= is_unbound
? WORK_CPU_UNBOUND
: cpu
;
2353 struct pool_workqueue
*pwq
= get_pwq(tcpu
, wq
);
2354 struct worker_pool
*pool
= pwq
->pool
;
2355 struct work_struct
*work
, *n
;
2357 __set_current_state(TASK_RUNNING
);
2358 mayday_clear_cpu(cpu
, wq
->mayday_mask
);
2360 /* migrate to the target cpu if possible */
2361 rescuer
->pool
= pool
;
2362 worker_maybe_bind_and_lock(rescuer
);
2365 * Slurp in all works issued via this workqueue and
2368 BUG_ON(!list_empty(&rescuer
->scheduled
));
2369 list_for_each_entry_safe(work
, n
, &pool
->worklist
, entry
)
2370 if (get_work_pwq(work
) == pwq
)
2371 move_linked_works(work
, scheduled
, &n
);
2373 process_scheduled_works(rescuer
);
2376 * Leave this pool. If keep_working() is %true, notify a
2377 * regular worker; otherwise, we end up with 0 concurrency
2378 * and stalling the execution.
2380 if (keep_working(pool
))
2381 wake_up_worker(pool
);
2383 spin_unlock_irq(&pool
->lock
);
2386 /* rescuers should never participate in concurrency management */
2387 WARN_ON_ONCE(!(rescuer
->flags
& WORKER_NOT_RUNNING
));
2393 struct work_struct work
;
2394 struct completion done
;
2397 static void wq_barrier_func(struct work_struct
*work
)
2399 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2400 complete(&barr
->done
);
2404 * insert_wq_barrier - insert a barrier work
2405 * @pwq: pwq to insert barrier into
2406 * @barr: wq_barrier to insert
2407 * @target: target work to attach @barr to
2408 * @worker: worker currently executing @target, NULL if @target is not executing
2410 * @barr is linked to @target such that @barr is completed only after
2411 * @target finishes execution. Please note that the ordering
2412 * guarantee is observed only with respect to @target and on the local
2415 * Currently, a queued barrier can't be canceled. This is because
2416 * try_to_grab_pending() can't determine whether the work to be
2417 * grabbed is at the head of the queue and thus can't clear LINKED
2418 * flag of the previous work while there must be a valid next work
2419 * after a work with LINKED flag set.
2421 * Note that when @worker is non-NULL, @target may be modified
2422 * underneath us, so we can't reliably determine pwq from @target.
2425 * spin_lock_irq(pool->lock).
2427 static void insert_wq_barrier(struct pool_workqueue
*pwq
,
2428 struct wq_barrier
*barr
,
2429 struct work_struct
*target
, struct worker
*worker
)
2431 struct list_head
*head
;
2432 unsigned int linked
= 0;
2435 * debugobject calls are safe here even with pool->lock locked
2436 * as we know for sure that this will not trigger any of the
2437 * checks and call back into the fixup functions where we
2440 INIT_WORK_ONSTACK(&barr
->work
, wq_barrier_func
);
2441 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2442 init_completion(&barr
->done
);
2445 * If @target is currently being executed, schedule the
2446 * barrier to the worker; otherwise, put it after @target.
2449 head
= worker
->scheduled
.next
;
2451 unsigned long *bits
= work_data_bits(target
);
2453 head
= target
->entry
.next
;
2454 /* there can already be other linked works, inherit and set */
2455 linked
= *bits
& WORK_STRUCT_LINKED
;
2456 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2459 debug_work_activate(&barr
->work
);
2460 insert_work(pwq
, &barr
->work
, head
,
2461 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2465 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2466 * @wq: workqueue being flushed
2467 * @flush_color: new flush color, < 0 for no-op
2468 * @work_color: new work color, < 0 for no-op
2470 * Prepare pwqs for workqueue flushing.
2472 * If @flush_color is non-negative, flush_color on all pwqs should be
2473 * -1. If no pwq has in-flight commands at the specified color, all
2474 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2475 * has in flight commands, its pwq->flush_color is set to
2476 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2477 * wakeup logic is armed and %true is returned.
2479 * The caller should have initialized @wq->first_flusher prior to
2480 * calling this function with non-negative @flush_color. If
2481 * @flush_color is negative, no flush color update is done and %false
2484 * If @work_color is non-negative, all pwqs should have the same
2485 * work_color which is previous to @work_color and all will be
2486 * advanced to @work_color.
2489 * mutex_lock(wq->flush_mutex).
2492 * %true if @flush_color >= 0 and there's something to flush. %false
2495 static bool flush_workqueue_prep_pwqs(struct workqueue_struct
*wq
,
2496 int flush_color
, int work_color
)
2501 if (flush_color
>= 0) {
2502 BUG_ON(atomic_read(&wq
->nr_pwqs_to_flush
));
2503 atomic_set(&wq
->nr_pwqs_to_flush
, 1);
2506 for_each_pwq_cpu(cpu
, wq
) {
2507 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
2508 struct worker_pool
*pool
= pwq
->pool
;
2510 spin_lock_irq(&pool
->lock
);
2512 if (flush_color
>= 0) {
2513 BUG_ON(pwq
->flush_color
!= -1);
2515 if (pwq
->nr_in_flight
[flush_color
]) {
2516 pwq
->flush_color
= flush_color
;
2517 atomic_inc(&wq
->nr_pwqs_to_flush
);
2522 if (work_color
>= 0) {
2523 BUG_ON(work_color
!= work_next_color(pwq
->work_color
));
2524 pwq
->work_color
= work_color
;
2527 spin_unlock_irq(&pool
->lock
);
2530 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_pwqs_to_flush
))
2531 complete(&wq
->first_flusher
->done
);
2537 * flush_workqueue - ensure that any scheduled work has run to completion.
2538 * @wq: workqueue to flush
2540 * Forces execution of the workqueue and blocks until its completion.
2541 * This is typically used in driver shutdown handlers.
2543 * We sleep until all works which were queued on entry have been handled,
2544 * but we are not livelocked by new incoming ones.
2546 void flush_workqueue(struct workqueue_struct
*wq
)
2548 struct wq_flusher this_flusher
= {
2549 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2551 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2555 lock_map_acquire(&wq
->lockdep_map
);
2556 lock_map_release(&wq
->lockdep_map
);
2558 mutex_lock(&wq
->flush_mutex
);
2561 * Start-to-wait phase
2563 next_color
= work_next_color(wq
->work_color
);
2565 if (next_color
!= wq
->flush_color
) {
2567 * Color space is not full. The current work_color
2568 * becomes our flush_color and work_color is advanced
2571 BUG_ON(!list_empty(&wq
->flusher_overflow
));
2572 this_flusher
.flush_color
= wq
->work_color
;
2573 wq
->work_color
= next_color
;
2575 if (!wq
->first_flusher
) {
2576 /* no flush in progress, become the first flusher */
2577 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2579 wq
->first_flusher
= &this_flusher
;
2581 if (!flush_workqueue_prep_pwqs(wq
, wq
->flush_color
,
2583 /* nothing to flush, done */
2584 wq
->flush_color
= next_color
;
2585 wq
->first_flusher
= NULL
;
2590 BUG_ON(wq
->flush_color
== this_flusher
.flush_color
);
2591 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2592 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2596 * Oops, color space is full, wait on overflow queue.
2597 * The next flush completion will assign us
2598 * flush_color and transfer to flusher_queue.
2600 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2603 mutex_unlock(&wq
->flush_mutex
);
2605 wait_for_completion(&this_flusher
.done
);
2608 * Wake-up-and-cascade phase
2610 * First flushers are responsible for cascading flushes and
2611 * handling overflow. Non-first flushers can simply return.
2613 if (wq
->first_flusher
!= &this_flusher
)
2616 mutex_lock(&wq
->flush_mutex
);
2618 /* we might have raced, check again with mutex held */
2619 if (wq
->first_flusher
!= &this_flusher
)
2622 wq
->first_flusher
= NULL
;
2624 BUG_ON(!list_empty(&this_flusher
.list
));
2625 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2628 struct wq_flusher
*next
, *tmp
;
2630 /* complete all the flushers sharing the current flush color */
2631 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2632 if (next
->flush_color
!= wq
->flush_color
)
2634 list_del_init(&next
->list
);
2635 complete(&next
->done
);
2638 BUG_ON(!list_empty(&wq
->flusher_overflow
) &&
2639 wq
->flush_color
!= work_next_color(wq
->work_color
));
2641 /* this flush_color is finished, advance by one */
2642 wq
->flush_color
= work_next_color(wq
->flush_color
);
2644 /* one color has been freed, handle overflow queue */
2645 if (!list_empty(&wq
->flusher_overflow
)) {
2647 * Assign the same color to all overflowed
2648 * flushers, advance work_color and append to
2649 * flusher_queue. This is the start-to-wait
2650 * phase for these overflowed flushers.
2652 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2653 tmp
->flush_color
= wq
->work_color
;
2655 wq
->work_color
= work_next_color(wq
->work_color
);
2657 list_splice_tail_init(&wq
->flusher_overflow
,
2658 &wq
->flusher_queue
);
2659 flush_workqueue_prep_pwqs(wq
, -1, wq
->work_color
);
2662 if (list_empty(&wq
->flusher_queue
)) {
2663 BUG_ON(wq
->flush_color
!= wq
->work_color
);
2668 * Need to flush more colors. Make the next flusher
2669 * the new first flusher and arm pwqs.
2671 BUG_ON(wq
->flush_color
== wq
->work_color
);
2672 BUG_ON(wq
->flush_color
!= next
->flush_color
);
2674 list_del_init(&next
->list
);
2675 wq
->first_flusher
= next
;
2677 if (flush_workqueue_prep_pwqs(wq
, wq
->flush_color
, -1))
2681 * Meh... this color is already done, clear first
2682 * flusher and repeat cascading.
2684 wq
->first_flusher
= NULL
;
2688 mutex_unlock(&wq
->flush_mutex
);
2690 EXPORT_SYMBOL_GPL(flush_workqueue
);
2693 * drain_workqueue - drain a workqueue
2694 * @wq: workqueue to drain
2696 * Wait until the workqueue becomes empty. While draining is in progress,
2697 * only chain queueing is allowed. IOW, only currently pending or running
2698 * work items on @wq can queue further work items on it. @wq is flushed
2699 * repeatedly until it becomes empty. The number of flushing is detemined
2700 * by the depth of chaining and should be relatively short. Whine if it
2703 void drain_workqueue(struct workqueue_struct
*wq
)
2705 unsigned int flush_cnt
= 0;
2709 * __queue_work() needs to test whether there are drainers, is much
2710 * hotter than drain_workqueue() and already looks at @wq->flags.
2711 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2713 spin_lock(&workqueue_lock
);
2714 if (!wq
->nr_drainers
++)
2715 wq
->flags
|= WQ_DRAINING
;
2716 spin_unlock(&workqueue_lock
);
2718 flush_workqueue(wq
);
2720 for_each_pwq_cpu(cpu
, wq
) {
2721 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
2724 spin_lock_irq(&pwq
->pool
->lock
);
2725 drained
= !pwq
->nr_active
&& list_empty(&pwq
->delayed_works
);
2726 spin_unlock_irq(&pwq
->pool
->lock
);
2731 if (++flush_cnt
== 10 ||
2732 (flush_cnt
% 100 == 0 && flush_cnt
<= 1000))
2733 pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2734 wq
->name
, flush_cnt
);
2738 spin_lock(&workqueue_lock
);
2739 if (!--wq
->nr_drainers
)
2740 wq
->flags
&= ~WQ_DRAINING
;
2741 spin_unlock(&workqueue_lock
);
2743 EXPORT_SYMBOL_GPL(drain_workqueue
);
2745 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
)
2747 struct worker
*worker
= NULL
;
2748 struct worker_pool
*pool
;
2749 struct pool_workqueue
*pwq
;
2752 pool
= get_work_pool(work
);
2756 spin_lock_irq(&pool
->lock
);
2757 /* see the comment in try_to_grab_pending() with the same code */
2758 pwq
= get_work_pwq(work
);
2760 if (unlikely(pwq
->pool
!= pool
))
2763 worker
= find_worker_executing_work(pool
, work
);
2766 pwq
= worker
->current_pwq
;
2769 insert_wq_barrier(pwq
, barr
, work
, worker
);
2770 spin_unlock_irq(&pool
->lock
);
2773 * If @max_active is 1 or rescuer is in use, flushing another work
2774 * item on the same workqueue may lead to deadlock. Make sure the
2775 * flusher is not running on the same workqueue by verifying write
2778 if (pwq
->wq
->saved_max_active
== 1 || pwq
->wq
->flags
& WQ_RESCUER
)
2779 lock_map_acquire(&pwq
->wq
->lockdep_map
);
2781 lock_map_acquire_read(&pwq
->wq
->lockdep_map
);
2782 lock_map_release(&pwq
->wq
->lockdep_map
);
2786 spin_unlock_irq(&pool
->lock
);
2791 * flush_work - wait for a work to finish executing the last queueing instance
2792 * @work: the work to flush
2794 * Wait until @work has finished execution. @work is guaranteed to be idle
2795 * on return if it hasn't been requeued since flush started.
2798 * %true if flush_work() waited for the work to finish execution,
2799 * %false if it was already idle.
2801 bool flush_work(struct work_struct
*work
)
2803 struct wq_barrier barr
;
2805 lock_map_acquire(&work
->lockdep_map
);
2806 lock_map_release(&work
->lockdep_map
);
2808 if (start_flush_work(work
, &barr
)) {
2809 wait_for_completion(&barr
.done
);
2810 destroy_work_on_stack(&barr
.work
);
2816 EXPORT_SYMBOL_GPL(flush_work
);
2818 static bool __cancel_work_timer(struct work_struct
*work
, bool is_dwork
)
2820 unsigned long flags
;
2824 ret
= try_to_grab_pending(work
, is_dwork
, &flags
);
2826 * If someone else is canceling, wait for the same event it
2827 * would be waiting for before retrying.
2829 if (unlikely(ret
== -ENOENT
))
2831 } while (unlikely(ret
< 0));
2833 /* tell other tasks trying to grab @work to back off */
2834 mark_work_canceling(work
);
2835 local_irq_restore(flags
);
2838 clear_work_data(work
);
2843 * cancel_work_sync - cancel a work and wait for it to finish
2844 * @work: the work to cancel
2846 * Cancel @work and wait for its execution to finish. This function
2847 * can be used even if the work re-queues itself or migrates to
2848 * another workqueue. On return from this function, @work is
2849 * guaranteed to be not pending or executing on any CPU.
2851 * cancel_work_sync(&delayed_work->work) must not be used for
2852 * delayed_work's. Use cancel_delayed_work_sync() instead.
2854 * The caller must ensure that the workqueue on which @work was last
2855 * queued can't be destroyed before this function returns.
2858 * %true if @work was pending, %false otherwise.
2860 bool cancel_work_sync(struct work_struct
*work
)
2862 return __cancel_work_timer(work
, false);
2864 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2867 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2868 * @dwork: the delayed work to flush
2870 * Delayed timer is cancelled and the pending work is queued for
2871 * immediate execution. Like flush_work(), this function only
2872 * considers the last queueing instance of @dwork.
2875 * %true if flush_work() waited for the work to finish execution,
2876 * %false if it was already idle.
2878 bool flush_delayed_work(struct delayed_work
*dwork
)
2880 local_irq_disable();
2881 if (del_timer_sync(&dwork
->timer
))
2882 __queue_work(dwork
->cpu
, dwork
->wq
, &dwork
->work
);
2884 return flush_work(&dwork
->work
);
2886 EXPORT_SYMBOL(flush_delayed_work
);
2889 * cancel_delayed_work - cancel a delayed work
2890 * @dwork: delayed_work to cancel
2892 * Kill off a pending delayed_work. Returns %true if @dwork was pending
2893 * and canceled; %false if wasn't pending. Note that the work callback
2894 * function may still be running on return, unless it returns %true and the
2895 * work doesn't re-arm itself. Explicitly flush or use
2896 * cancel_delayed_work_sync() to wait on it.
2898 * This function is safe to call from any context including IRQ handler.
2900 bool cancel_delayed_work(struct delayed_work
*dwork
)
2902 unsigned long flags
;
2906 ret
= try_to_grab_pending(&dwork
->work
, true, &flags
);
2907 } while (unlikely(ret
== -EAGAIN
));
2909 if (unlikely(ret
< 0))
2912 set_work_pool_and_clear_pending(&dwork
->work
,
2913 get_work_pool_id(&dwork
->work
));
2914 local_irq_restore(flags
);
2917 EXPORT_SYMBOL(cancel_delayed_work
);
2920 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2921 * @dwork: the delayed work cancel
2923 * This is cancel_work_sync() for delayed works.
2926 * %true if @dwork was pending, %false otherwise.
2928 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2930 return __cancel_work_timer(&dwork
->work
, true);
2932 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2935 * schedule_work_on - put work task on a specific cpu
2936 * @cpu: cpu to put the work task on
2937 * @work: job to be done
2939 * This puts a job on a specific cpu
2941 bool schedule_work_on(int cpu
, struct work_struct
*work
)
2943 return queue_work_on(cpu
, system_wq
, work
);
2945 EXPORT_SYMBOL(schedule_work_on
);
2948 * schedule_work - put work task in global workqueue
2949 * @work: job to be done
2951 * Returns %false if @work was already on the kernel-global workqueue and
2954 * This puts a job in the kernel-global workqueue if it was not already
2955 * queued and leaves it in the same position on the kernel-global
2956 * workqueue otherwise.
2958 bool schedule_work(struct work_struct
*work
)
2960 return queue_work(system_wq
, work
);
2962 EXPORT_SYMBOL(schedule_work
);
2965 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2967 * @dwork: job to be done
2968 * @delay: number of jiffies to wait
2970 * After waiting for a given time this puts a job in the kernel-global
2971 * workqueue on the specified CPU.
2973 bool schedule_delayed_work_on(int cpu
, struct delayed_work
*dwork
,
2974 unsigned long delay
)
2976 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
2978 EXPORT_SYMBOL(schedule_delayed_work_on
);
2981 * schedule_delayed_work - put work task in global workqueue after delay
2982 * @dwork: job to be done
2983 * @delay: number of jiffies to wait or 0 for immediate execution
2985 * After waiting for a given time this puts a job in the kernel-global
2988 bool schedule_delayed_work(struct delayed_work
*dwork
, unsigned long delay
)
2990 return queue_delayed_work(system_wq
, dwork
, delay
);
2992 EXPORT_SYMBOL(schedule_delayed_work
);
2995 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2996 * @func: the function to call
2998 * schedule_on_each_cpu() executes @func on each online CPU using the
2999 * system workqueue and blocks until all CPUs have completed.
3000 * schedule_on_each_cpu() is very slow.
3003 * 0 on success, -errno on failure.
3005 int schedule_on_each_cpu(work_func_t func
)
3008 struct work_struct __percpu
*works
;
3010 works
= alloc_percpu(struct work_struct
);
3016 for_each_online_cpu(cpu
) {
3017 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
3019 INIT_WORK(work
, func
);
3020 schedule_work_on(cpu
, work
);
3023 for_each_online_cpu(cpu
)
3024 flush_work(per_cpu_ptr(works
, cpu
));
3032 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3034 * Forces execution of the kernel-global workqueue and blocks until its
3037 * Think twice before calling this function! It's very easy to get into
3038 * trouble if you don't take great care. Either of the following situations
3039 * will lead to deadlock:
3041 * One of the work items currently on the workqueue needs to acquire
3042 * a lock held by your code or its caller.
3044 * Your code is running in the context of a work routine.
3046 * They will be detected by lockdep when they occur, but the first might not
3047 * occur very often. It depends on what work items are on the workqueue and
3048 * what locks they need, which you have no control over.
3050 * In most situations flushing the entire workqueue is overkill; you merely
3051 * need to know that a particular work item isn't queued and isn't running.
3052 * In such cases you should use cancel_delayed_work_sync() or
3053 * cancel_work_sync() instead.
3055 void flush_scheduled_work(void)
3057 flush_workqueue(system_wq
);
3059 EXPORT_SYMBOL(flush_scheduled_work
);
3062 * execute_in_process_context - reliably execute the routine with user context
3063 * @fn: the function to execute
3064 * @ew: guaranteed storage for the execute work structure (must
3065 * be available when the work executes)
3067 * Executes the function immediately if process context is available,
3068 * otherwise schedules the function for delayed execution.
3070 * Returns: 0 - function was executed
3071 * 1 - function was scheduled for execution
3073 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
3075 if (!in_interrupt()) {
3080 INIT_WORK(&ew
->work
, fn
);
3081 schedule_work(&ew
->work
);
3085 EXPORT_SYMBOL_GPL(execute_in_process_context
);
3087 int keventd_up(void)
3089 return system_wq
!= NULL
;
3092 static int alloc_pwqs(struct workqueue_struct
*wq
)
3095 * pwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3096 * Make sure that the alignment isn't lower than that of
3097 * unsigned long long.
3099 const size_t size
= sizeof(struct pool_workqueue
);
3100 const size_t align
= max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS
,
3101 __alignof__(unsigned long long));
3103 if (!(wq
->flags
& WQ_UNBOUND
))
3104 wq
->pool_wq
.pcpu
= __alloc_percpu(size
, align
);
3109 * Allocate enough room to align pwq and put an extra
3110 * pointer at the end pointing back to the originally
3111 * allocated pointer which will be used for free.
3113 ptr
= kzalloc(size
+ align
+ sizeof(void *), GFP_KERNEL
);
3115 wq
->pool_wq
.single
= PTR_ALIGN(ptr
, align
);
3116 *(void **)(wq
->pool_wq
.single
+ 1) = ptr
;
3120 /* just in case, make sure it's actually aligned */
3121 BUG_ON(!IS_ALIGNED(wq
->pool_wq
.v
, align
));
3122 return wq
->pool_wq
.v
? 0 : -ENOMEM
;
3125 static void free_pwqs(struct workqueue_struct
*wq
)
3127 if (!(wq
->flags
& WQ_UNBOUND
))
3128 free_percpu(wq
->pool_wq
.pcpu
);
3129 else if (wq
->pool_wq
.single
) {
3130 /* the pointer to free is stored right after the pwq */
3131 kfree(*(void **)(wq
->pool_wq
.single
+ 1));
3135 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
3138 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
3140 if (max_active
< 1 || max_active
> lim
)
3141 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3142 max_active
, name
, 1, lim
);
3144 return clamp_val(max_active
, 1, lim
);
3147 struct workqueue_struct
*__alloc_workqueue_key(const char *fmt
,
3150 struct lock_class_key
*key
,
3151 const char *lock_name
, ...)
3153 va_list args
, args1
;
3154 struct workqueue_struct
*wq
;
3158 /* determine namelen, allocate wq and format name */
3159 va_start(args
, lock_name
);
3160 va_copy(args1
, args
);
3161 namelen
= vsnprintf(NULL
, 0, fmt
, args
) + 1;
3163 wq
= kzalloc(sizeof(*wq
) + namelen
, GFP_KERNEL
);
3167 vsnprintf(wq
->name
, namelen
, fmt
, args1
);
3172 * Workqueues which may be used during memory reclaim should
3173 * have a rescuer to guarantee forward progress.
3175 if (flags
& WQ_MEM_RECLAIM
)
3176 flags
|= WQ_RESCUER
;
3178 max_active
= max_active
?: WQ_DFL_ACTIVE
;
3179 max_active
= wq_clamp_max_active(max_active
, flags
, wq
->name
);
3183 wq
->saved_max_active
= max_active
;
3184 mutex_init(&wq
->flush_mutex
);
3185 atomic_set(&wq
->nr_pwqs_to_flush
, 0);
3186 INIT_LIST_HEAD(&wq
->flusher_queue
);
3187 INIT_LIST_HEAD(&wq
->flusher_overflow
);
3189 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
3190 INIT_LIST_HEAD(&wq
->list
);
3192 if (alloc_pwqs(wq
) < 0)
3195 for_each_pwq_cpu(cpu
, wq
) {
3196 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3198 BUG_ON((unsigned long)pwq
& WORK_STRUCT_FLAG_MASK
);
3199 pwq
->pool
= get_std_worker_pool(cpu
, flags
& WQ_HIGHPRI
);
3201 pwq
->flush_color
= -1;
3202 pwq
->max_active
= max_active
;
3203 INIT_LIST_HEAD(&pwq
->delayed_works
);
3206 if (flags
& WQ_RESCUER
) {
3207 struct worker
*rescuer
;
3209 if (!alloc_mayday_mask(&wq
->mayday_mask
, GFP_KERNEL
))
3212 wq
->rescuer
= rescuer
= alloc_worker();
3216 rescuer
->rescue_wq
= wq
;
3217 rescuer
->task
= kthread_create(rescuer_thread
, rescuer
, "%s",
3219 if (IS_ERR(rescuer
->task
))
3222 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
3223 wake_up_process(rescuer
->task
);
3227 * workqueue_lock protects global freeze state and workqueues
3228 * list. Grab it, set max_active accordingly and add the new
3229 * workqueue to workqueues list.
3231 spin_lock(&workqueue_lock
);
3233 if (workqueue_freezing
&& wq
->flags
& WQ_FREEZABLE
)
3234 for_each_pwq_cpu(cpu
, wq
)
3235 get_pwq(cpu
, wq
)->max_active
= 0;
3237 list_add(&wq
->list
, &workqueues
);
3239 spin_unlock(&workqueue_lock
);
3245 free_mayday_mask(wq
->mayday_mask
);
3251 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
3254 * destroy_workqueue - safely terminate a workqueue
3255 * @wq: target workqueue
3257 * Safely destroy a workqueue. All work currently pending will be done first.
3259 void destroy_workqueue(struct workqueue_struct
*wq
)
3263 /* drain it before proceeding with destruction */
3264 drain_workqueue(wq
);
3267 * wq list is used to freeze wq, remove from list after
3268 * flushing is complete in case freeze races us.
3270 spin_lock(&workqueue_lock
);
3271 list_del(&wq
->list
);
3272 spin_unlock(&workqueue_lock
);
3275 for_each_pwq_cpu(cpu
, wq
) {
3276 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3279 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
3280 BUG_ON(pwq
->nr_in_flight
[i
]);
3281 BUG_ON(pwq
->nr_active
);
3282 BUG_ON(!list_empty(&pwq
->delayed_works
));
3285 if (wq
->flags
& WQ_RESCUER
) {
3286 kthread_stop(wq
->rescuer
->task
);
3287 free_mayday_mask(wq
->mayday_mask
);
3294 EXPORT_SYMBOL_GPL(destroy_workqueue
);
3297 * pwq_set_max_active - adjust max_active of a pwq
3298 * @pwq: target pool_workqueue
3299 * @max_active: new max_active value.
3301 * Set @pwq->max_active to @max_active and activate delayed works if
3305 * spin_lock_irq(pool->lock).
3307 static void pwq_set_max_active(struct pool_workqueue
*pwq
, int max_active
)
3309 pwq
->max_active
= max_active
;
3311 while (!list_empty(&pwq
->delayed_works
) &&
3312 pwq
->nr_active
< pwq
->max_active
)
3313 pwq_activate_first_delayed(pwq
);
3317 * workqueue_set_max_active - adjust max_active of a workqueue
3318 * @wq: target workqueue
3319 * @max_active: new max_active value.
3321 * Set max_active of @wq to @max_active.
3324 * Don't call from IRQ context.
3326 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
3330 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
3332 spin_lock(&workqueue_lock
);
3334 wq
->saved_max_active
= max_active
;
3336 for_each_pwq_cpu(cpu
, wq
) {
3337 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3338 struct worker_pool
*pool
= pwq
->pool
;
3340 spin_lock_irq(&pool
->lock
);
3342 if (!(wq
->flags
& WQ_FREEZABLE
) ||
3343 !(pool
->flags
& POOL_FREEZING
))
3344 pwq_set_max_active(pwq
, max_active
);
3346 spin_unlock_irq(&pool
->lock
);
3349 spin_unlock(&workqueue_lock
);
3351 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
3354 * workqueue_congested - test whether a workqueue is congested
3355 * @cpu: CPU in question
3356 * @wq: target workqueue
3358 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3359 * no synchronization around this function and the test result is
3360 * unreliable and only useful as advisory hints or for debugging.
3363 * %true if congested, %false otherwise.
3365 bool workqueue_congested(unsigned int cpu
, struct workqueue_struct
*wq
)
3367 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3369 return !list_empty(&pwq
->delayed_works
);
3371 EXPORT_SYMBOL_GPL(workqueue_congested
);
3374 * work_busy - test whether a work is currently pending or running
3375 * @work: the work to be tested
3377 * Test whether @work is currently pending or running. There is no
3378 * synchronization around this function and the test result is
3379 * unreliable and only useful as advisory hints or for debugging.
3382 * OR'd bitmask of WORK_BUSY_* bits.
3384 unsigned int work_busy(struct work_struct
*work
)
3386 struct worker_pool
*pool
= get_work_pool(work
);
3387 unsigned long flags
;
3388 unsigned int ret
= 0;
3390 if (work_pending(work
))
3391 ret
|= WORK_BUSY_PENDING
;
3394 spin_lock_irqsave(&pool
->lock
, flags
);
3395 if (find_worker_executing_work(pool
, work
))
3396 ret
|= WORK_BUSY_RUNNING
;
3397 spin_unlock_irqrestore(&pool
->lock
, flags
);
3402 EXPORT_SYMBOL_GPL(work_busy
);
3407 * There are two challenges in supporting CPU hotplug. Firstly, there
3408 * are a lot of assumptions on strong associations among work, pwq and
3409 * pool which make migrating pending and scheduled works very
3410 * difficult to implement without impacting hot paths. Secondly,
3411 * worker pools serve mix of short, long and very long running works making
3412 * blocked draining impractical.
3414 * This is solved by allowing the pools to be disassociated from the CPU
3415 * running as an unbound one and allowing it to be reattached later if the
3416 * cpu comes back online.
3419 static void wq_unbind_fn(struct work_struct
*work
)
3421 int cpu
= smp_processor_id();
3422 struct worker_pool
*pool
;
3423 struct worker
*worker
;
3426 for_each_std_worker_pool(pool
, cpu
) {
3427 BUG_ON(cpu
!= smp_processor_id());
3429 mutex_lock(&pool
->assoc_mutex
);
3430 spin_lock_irq(&pool
->lock
);
3433 * We've claimed all manager positions. Make all workers
3434 * unbound and set DISASSOCIATED. Before this, all workers
3435 * except for the ones which are still executing works from
3436 * before the last CPU down must be on the cpu. After
3437 * this, they may become diasporas.
3439 list_for_each_entry(worker
, &pool
->idle_list
, entry
)
3440 worker
->flags
|= WORKER_UNBOUND
;
3442 for_each_busy_worker(worker
, i
, pool
)
3443 worker
->flags
|= WORKER_UNBOUND
;
3445 pool
->flags
|= POOL_DISASSOCIATED
;
3447 spin_unlock_irq(&pool
->lock
);
3448 mutex_unlock(&pool
->assoc_mutex
);
3452 * Call schedule() so that we cross rq->lock and thus can guarantee
3453 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3454 * as scheduler callbacks may be invoked from other cpus.
3459 * Sched callbacks are disabled now. Zap nr_running. After this,
3460 * nr_running stays zero and need_more_worker() and keep_working()
3461 * are always true as long as the worklist is not empty. Pools on
3462 * @cpu now behave as unbound (in terms of concurrency management)
3463 * pools which are served by workers tied to the CPU.
3465 * On return from this function, the current worker would trigger
3466 * unbound chain execution of pending work items if other workers
3469 for_each_std_worker_pool(pool
, cpu
)
3470 atomic_set(&pool
->nr_running
, 0);
3474 * Workqueues should be brought up before normal priority CPU notifiers.
3475 * This will be registered high priority CPU notifier.
3477 static int __cpuinit
workqueue_cpu_up_callback(struct notifier_block
*nfb
,
3478 unsigned long action
,
3481 unsigned int cpu
= (unsigned long)hcpu
;
3482 struct worker_pool
*pool
;
3484 switch (action
& ~CPU_TASKS_FROZEN
) {
3485 case CPU_UP_PREPARE
:
3486 for_each_std_worker_pool(pool
, cpu
) {
3487 struct worker
*worker
;
3489 if (pool
->nr_workers
)
3492 worker
= create_worker(pool
);
3496 spin_lock_irq(&pool
->lock
);
3497 start_worker(worker
);
3498 spin_unlock_irq(&pool
->lock
);
3502 case CPU_DOWN_FAILED
:
3504 for_each_std_worker_pool(pool
, cpu
) {
3505 mutex_lock(&pool
->assoc_mutex
);
3506 spin_lock_irq(&pool
->lock
);
3508 pool
->flags
&= ~POOL_DISASSOCIATED
;
3509 rebind_workers(pool
);
3511 spin_unlock_irq(&pool
->lock
);
3512 mutex_unlock(&pool
->assoc_mutex
);
3520 * Workqueues should be brought down after normal priority CPU notifiers.
3521 * This will be registered as low priority CPU notifier.
3523 static int __cpuinit
workqueue_cpu_down_callback(struct notifier_block
*nfb
,
3524 unsigned long action
,
3527 unsigned int cpu
= (unsigned long)hcpu
;
3528 struct work_struct unbind_work
;
3530 switch (action
& ~CPU_TASKS_FROZEN
) {
3531 case CPU_DOWN_PREPARE
:
3532 /* unbinding should happen on the local CPU */
3533 INIT_WORK_ONSTACK(&unbind_work
, wq_unbind_fn
);
3534 queue_work_on(cpu
, system_highpri_wq
, &unbind_work
);
3535 flush_work(&unbind_work
);
3543 struct work_for_cpu
{
3544 struct work_struct work
;
3550 static void work_for_cpu_fn(struct work_struct
*work
)
3552 struct work_for_cpu
*wfc
= container_of(work
, struct work_for_cpu
, work
);
3554 wfc
->ret
= wfc
->fn(wfc
->arg
);
3558 * work_on_cpu - run a function in user context on a particular cpu
3559 * @cpu: the cpu to run on
3560 * @fn: the function to run
3561 * @arg: the function arg
3563 * This will return the value @fn returns.
3564 * It is up to the caller to ensure that the cpu doesn't go offline.
3565 * The caller must not hold any locks which would prevent @fn from completing.
3567 long work_on_cpu(unsigned int cpu
, long (*fn
)(void *), void *arg
)
3569 struct work_for_cpu wfc
= { .fn
= fn
, .arg
= arg
};
3571 INIT_WORK_ONSTACK(&wfc
.work
, work_for_cpu_fn
);
3572 schedule_work_on(cpu
, &wfc
.work
);
3573 flush_work(&wfc
.work
);
3576 EXPORT_SYMBOL_GPL(work_on_cpu
);
3577 #endif /* CONFIG_SMP */
3579 #ifdef CONFIG_FREEZER
3582 * freeze_workqueues_begin - begin freezing workqueues
3584 * Start freezing workqueues. After this function returns, all freezable
3585 * workqueues will queue new works to their frozen_works list instead of
3589 * Grabs and releases workqueue_lock and pool->lock's.
3591 void freeze_workqueues_begin(void)
3595 spin_lock(&workqueue_lock
);
3597 BUG_ON(workqueue_freezing
);
3598 workqueue_freezing
= true;
3600 for_each_wq_cpu(cpu
) {
3601 struct worker_pool
*pool
;
3602 struct workqueue_struct
*wq
;
3604 for_each_std_worker_pool(pool
, cpu
) {
3605 spin_lock_irq(&pool
->lock
);
3607 WARN_ON_ONCE(pool
->flags
& POOL_FREEZING
);
3608 pool
->flags
|= POOL_FREEZING
;
3610 list_for_each_entry(wq
, &workqueues
, list
) {
3611 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3613 if (pwq
&& pwq
->pool
== pool
&&
3614 (wq
->flags
& WQ_FREEZABLE
))
3615 pwq
->max_active
= 0;
3618 spin_unlock_irq(&pool
->lock
);
3622 spin_unlock(&workqueue_lock
);
3626 * freeze_workqueues_busy - are freezable workqueues still busy?
3628 * Check whether freezing is complete. This function must be called
3629 * between freeze_workqueues_begin() and thaw_workqueues().
3632 * Grabs and releases workqueue_lock.
3635 * %true if some freezable workqueues are still busy. %false if freezing
3638 bool freeze_workqueues_busy(void)
3643 spin_lock(&workqueue_lock
);
3645 BUG_ON(!workqueue_freezing
);
3647 for_each_wq_cpu(cpu
) {
3648 struct workqueue_struct
*wq
;
3650 * nr_active is monotonically decreasing. It's safe
3651 * to peek without lock.
3653 list_for_each_entry(wq
, &workqueues
, list
) {
3654 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3656 if (!pwq
|| !(wq
->flags
& WQ_FREEZABLE
))
3659 BUG_ON(pwq
->nr_active
< 0);
3660 if (pwq
->nr_active
) {
3667 spin_unlock(&workqueue_lock
);
3672 * thaw_workqueues - thaw workqueues
3674 * Thaw workqueues. Normal queueing is restored and all collected
3675 * frozen works are transferred to their respective pool worklists.
3678 * Grabs and releases workqueue_lock and pool->lock's.
3680 void thaw_workqueues(void)
3684 spin_lock(&workqueue_lock
);
3686 if (!workqueue_freezing
)
3689 for_each_wq_cpu(cpu
) {
3690 struct worker_pool
*pool
;
3691 struct workqueue_struct
*wq
;
3693 for_each_std_worker_pool(pool
, cpu
) {
3694 spin_lock_irq(&pool
->lock
);
3696 WARN_ON_ONCE(!(pool
->flags
& POOL_FREEZING
));
3697 pool
->flags
&= ~POOL_FREEZING
;
3699 list_for_each_entry(wq
, &workqueues
, list
) {
3700 struct pool_workqueue
*pwq
= get_pwq(cpu
, wq
);
3702 if (!pwq
|| pwq
->pool
!= pool
||
3703 !(wq
->flags
& WQ_FREEZABLE
))
3706 /* restore max_active and repopulate worklist */
3707 pwq_set_max_active(pwq
, wq
->saved_max_active
);
3710 wake_up_worker(pool
);
3712 spin_unlock_irq(&pool
->lock
);
3716 workqueue_freezing
= false;
3718 spin_unlock(&workqueue_lock
);
3720 #endif /* CONFIG_FREEZER */
3722 static int __init
init_workqueues(void)
3726 /* make sure we have enough bits for OFFQ pool ID */
3727 BUILD_BUG_ON((1LU << (BITS_PER_LONG
- WORK_OFFQ_POOL_SHIFT
)) <
3728 WORK_CPU_END
* NR_STD_WORKER_POOLS
);
3730 cpu_notifier(workqueue_cpu_up_callback
, CPU_PRI_WORKQUEUE_UP
);
3731 hotcpu_notifier(workqueue_cpu_down_callback
, CPU_PRI_WORKQUEUE_DOWN
);
3733 /* initialize CPU pools */
3734 for_each_wq_cpu(cpu
) {
3735 struct worker_pool
*pool
;
3737 for_each_std_worker_pool(pool
, cpu
) {
3738 spin_lock_init(&pool
->lock
);
3740 pool
->flags
|= POOL_DISASSOCIATED
;
3741 INIT_LIST_HEAD(&pool
->worklist
);
3742 INIT_LIST_HEAD(&pool
->idle_list
);
3743 hash_init(pool
->busy_hash
);
3745 init_timer_deferrable(&pool
->idle_timer
);
3746 pool
->idle_timer
.function
= idle_worker_timeout
;
3747 pool
->idle_timer
.data
= (unsigned long)pool
;
3749 setup_timer(&pool
->mayday_timer
, pool_mayday_timeout
,
3750 (unsigned long)pool
);
3752 mutex_init(&pool
->assoc_mutex
);
3753 ida_init(&pool
->worker_ida
);
3756 BUG_ON(worker_pool_assign_id(pool
));
3760 /* create the initial worker */
3761 for_each_online_wq_cpu(cpu
) {
3762 struct worker_pool
*pool
;
3764 for_each_std_worker_pool(pool
, cpu
) {
3765 struct worker
*worker
;
3767 if (cpu
!= WORK_CPU_UNBOUND
)
3768 pool
->flags
&= ~POOL_DISASSOCIATED
;
3770 worker
= create_worker(pool
);
3772 spin_lock_irq(&pool
->lock
);
3773 start_worker(worker
);
3774 spin_unlock_irq(&pool
->lock
);
3778 system_wq
= alloc_workqueue("events", 0, 0);
3779 system_highpri_wq
= alloc_workqueue("events_highpri", WQ_HIGHPRI
, 0);
3780 system_long_wq
= alloc_workqueue("events_long", 0, 0);
3781 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
3782 WQ_UNBOUND_MAX_ACTIVE
);
3783 system_freezable_wq
= alloc_workqueue("events_freezable",
3785 BUG_ON(!system_wq
|| !system_highpri_wq
|| !system_long_wq
||
3786 !system_unbound_wq
|| !system_freezable_wq
);
3789 early_initcall(init_workqueues
);