2 * linux/kernel/workqueue.c
4 * Generic mechanism for defining kernel helper threads for running
5 * arbitrary tasks in process context.
7 * Started by Ingo Molnar, Copyright (C) 2002
9 * Derived from the taskqueue/keventd code by:
11 * David Woodhouse <dwmw2@infradead.org>
13 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
14 * Theodore Ts'o <tytso@mit.edu>
16 * Made to use alloc_percpu by Christoph Lameter.
19 #include <linux/module.h>
20 #include <linux/kernel.h>
21 #include <linux/sched.h>
22 #include <linux/init.h>
23 #include <linux/signal.h>
24 #include <linux/completion.h>
25 #include <linux/workqueue.h>
26 #include <linux/slab.h>
27 #include <linux/cpu.h>
28 #include <linux/notifier.h>
29 #include <linux/kthread.h>
30 #include <linux/hardirq.h>
31 #include <linux/mempolicy.h>
32 #include <linux/freezer.h>
33 #include <linux/kallsyms.h>
34 #include <linux/debug_locks.h>
35 #include <linux/lockdep.h>
36 #include <linux/idr.h>
38 #include "workqueue_sched.h"
41 /* global_cwq flags */
42 GCWQ_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
43 GCWQ_MANAGING_WORKERS
= 1 << 1, /* managing workers */
44 GCWQ_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
45 GCWQ_FREEZING
= 1 << 3, /* freeze in progress */
46 GCWQ_HIGHPRI_PENDING
= 1 << 4, /* highpri works on queue */
49 WORKER_STARTED
= 1 << 0, /* started */
50 WORKER_DIE
= 1 << 1, /* die die die */
51 WORKER_IDLE
= 1 << 2, /* is idle */
52 WORKER_PREP
= 1 << 3, /* preparing to run works */
53 WORKER_ROGUE
= 1 << 4, /* not bound to any cpu */
54 WORKER_REBIND
= 1 << 5, /* mom is home, come back */
55 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
56 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
58 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_ROGUE
| WORKER_REBIND
|
59 WORKER_CPU_INTENSIVE
| WORKER_UNBOUND
,
61 /* gcwq->trustee_state */
62 TRUSTEE_START
= 0, /* start */
63 TRUSTEE_IN_CHARGE
= 1, /* trustee in charge of gcwq */
64 TRUSTEE_BUTCHER
= 2, /* butcher workers */
65 TRUSTEE_RELEASE
= 3, /* release workers */
66 TRUSTEE_DONE
= 4, /* trustee is done */
68 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
69 BUSY_WORKER_HASH_SIZE
= 1 << BUSY_WORKER_HASH_ORDER
,
70 BUSY_WORKER_HASH_MASK
= BUSY_WORKER_HASH_SIZE
- 1,
72 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
73 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
75 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100, /* call for help after 10ms */
76 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
77 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
78 TRUSTEE_COOLDOWN
= HZ
/ 10, /* for trustee draining */
81 * Rescue workers are used only on emergencies and shared by
84 RESCUER_NICE_LEVEL
= -20,
88 * Structure fields follow one of the following exclusion rules.
90 * I: Modifiable by initialization/destruction paths and read-only for
93 * P: Preemption protected. Disabling preemption is enough and should
94 * only be modified and accessed from the local cpu.
96 * L: gcwq->lock protected. Access with gcwq->lock held.
98 * X: During normal operation, modification requires gcwq->lock and
99 * should be done only from local cpu. Either disabling preemption
100 * on local cpu or grabbing gcwq->lock is enough for read access.
101 * If GCWQ_DISASSOCIATED is set, it's identical to L.
103 * F: wq->flush_mutex protected.
105 * W: workqueue_lock protected.
111 * The poor guys doing the actual heavy lifting. All on-duty workers
112 * are either serving the manager role, on idle list or on busy hash.
115 /* on idle list while idle, on busy hash table while busy */
117 struct list_head entry
; /* L: while idle */
118 struct hlist_node hentry
; /* L: while busy */
121 struct work_struct
*current_work
; /* L: work being processed */
122 struct cpu_workqueue_struct
*current_cwq
; /* L: current_work's cwq */
123 struct list_head scheduled
; /* L: scheduled works */
124 struct task_struct
*task
; /* I: worker task */
125 struct global_cwq
*gcwq
; /* I: the associated gcwq */
126 /* 64 bytes boundary on 64bit, 32 on 32bit */
127 unsigned long last_active
; /* L: last active timestamp */
128 unsigned int flags
; /* X: flags */
129 int id
; /* I: worker id */
130 struct work_struct rebind_work
; /* L: rebind worker to cpu */
134 * Global per-cpu workqueue. There's one and only one for each cpu
135 * and all works are queued and processed here regardless of their
139 spinlock_t lock
; /* the gcwq lock */
140 struct list_head worklist
; /* L: list of pending works */
141 unsigned int cpu
; /* I: the associated cpu */
142 unsigned int flags
; /* L: GCWQ_* flags */
144 int nr_workers
; /* L: total number of workers */
145 int nr_idle
; /* L: currently idle ones */
147 /* workers are chained either in the idle_list or busy_hash */
148 struct list_head idle_list
; /* X: list of idle workers */
149 struct hlist_head busy_hash
[BUSY_WORKER_HASH_SIZE
];
150 /* L: hash of busy workers */
152 struct timer_list idle_timer
; /* L: worker idle timeout */
153 struct timer_list mayday_timer
; /* L: SOS timer for dworkers */
155 struct ida worker_ida
; /* L: for worker IDs */
157 struct task_struct
*trustee
; /* L: for gcwq shutdown */
158 unsigned int trustee_state
; /* L: trustee state */
159 wait_queue_head_t trustee_wait
; /* trustee wait */
160 struct worker
*first_idle
; /* L: first idle worker */
161 } ____cacheline_aligned_in_smp
;
164 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
165 * work_struct->data are used for flags and thus cwqs need to be
166 * aligned at two's power of the number of flag bits.
168 struct cpu_workqueue_struct
{
169 struct global_cwq
*gcwq
; /* I: the associated gcwq */
170 struct workqueue_struct
*wq
; /* I: the owning workqueue */
171 int work_color
; /* L: current color */
172 int flush_color
; /* L: flushing color */
173 int nr_in_flight
[WORK_NR_COLORS
];
174 /* L: nr of in_flight works */
175 int nr_active
; /* L: nr of active works */
176 int max_active
; /* L: max active works */
177 struct list_head delayed_works
; /* L: delayed works */
181 * Structure used to wait for workqueue flush.
184 struct list_head list
; /* F: list of flushers */
185 int flush_color
; /* F: flush color waiting for */
186 struct completion done
; /* flush completion */
190 * All cpumasks are assumed to be always set on UP and thus can't be
191 * used to determine whether there's something to be done.
194 typedef cpumask_var_t mayday_mask_t
;
195 #define mayday_test_and_set_cpu(cpu, mask) \
196 cpumask_test_and_set_cpu((cpu), (mask))
197 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
198 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
199 #define alloc_mayday_mask(maskp, gfp) alloc_cpumask_var((maskp), (gfp))
200 #define free_mayday_mask(mask) free_cpumask_var((mask))
202 typedef unsigned long mayday_mask_t
;
203 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
204 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
205 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
206 #define alloc_mayday_mask(maskp, gfp) true
207 #define free_mayday_mask(mask) do { } while (0)
211 * The externally visible workqueue abstraction is an array of
212 * per-CPU workqueues:
214 struct workqueue_struct
{
215 unsigned int flags
; /* I: WQ_* flags */
217 struct cpu_workqueue_struct __percpu
*pcpu
;
218 struct cpu_workqueue_struct
*single
;
220 } cpu_wq
; /* I: cwq's */
221 struct list_head list
; /* W: list of all workqueues */
223 struct mutex flush_mutex
; /* protects wq flushing */
224 int work_color
; /* F: current work color */
225 int flush_color
; /* F: current flush color */
226 atomic_t nr_cwqs_to_flush
; /* flush in progress */
227 struct wq_flusher
*first_flusher
; /* F: first flusher */
228 struct list_head flusher_queue
; /* F: flush waiters */
229 struct list_head flusher_overflow
; /* F: flush overflow list */
231 mayday_mask_t mayday_mask
; /* cpus requesting rescue */
232 struct worker
*rescuer
; /* I: rescue worker */
234 int saved_max_active
; /* W: saved cwq max_active */
235 const char *name
; /* I: workqueue name */
236 #ifdef CONFIG_LOCKDEP
237 struct lockdep_map lockdep_map
;
241 struct workqueue_struct
*system_wq __read_mostly
;
242 struct workqueue_struct
*system_long_wq __read_mostly
;
243 struct workqueue_struct
*system_nrt_wq __read_mostly
;
244 struct workqueue_struct
*system_unbound_wq __read_mostly
;
245 EXPORT_SYMBOL_GPL(system_wq
);
246 EXPORT_SYMBOL_GPL(system_long_wq
);
247 EXPORT_SYMBOL_GPL(system_nrt_wq
);
248 EXPORT_SYMBOL_GPL(system_unbound_wq
);
250 #define for_each_busy_worker(worker, i, pos, gcwq) \
251 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
252 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
254 static inline int __next_gcwq_cpu(int cpu
, const struct cpumask
*mask
,
257 if (cpu
< nr_cpu_ids
) {
259 cpu
= cpumask_next(cpu
, mask
);
260 if (cpu
< nr_cpu_ids
)
264 return WORK_CPU_UNBOUND
;
266 return WORK_CPU_NONE
;
269 static inline int __next_wq_cpu(int cpu
, const struct cpumask
*mask
,
270 struct workqueue_struct
*wq
)
272 return __next_gcwq_cpu(cpu
, mask
, !(wq
->flags
& WQ_UNBOUND
) ? 1 : 2);
278 * An extra gcwq is defined for an invalid cpu number
279 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
280 * specific CPU. The following iterators are similar to
281 * for_each_*_cpu() iterators but also considers the unbound gcwq.
283 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
284 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
285 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
286 * WORK_CPU_UNBOUND for unbound workqueues
288 #define for_each_gcwq_cpu(cpu) \
289 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
290 (cpu) < WORK_CPU_NONE; \
291 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
293 #define for_each_online_gcwq_cpu(cpu) \
294 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
295 (cpu) < WORK_CPU_NONE; \
296 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
298 #define for_each_cwq_cpu(cpu, wq) \
299 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
300 (cpu) < WORK_CPU_NONE; \
301 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
303 #ifdef CONFIG_LOCKDEP
305 * in_workqueue_context() - in context of specified workqueue?
306 * @wq: the workqueue of interest
308 * Checks lockdep state to see if the current task is executing from
309 * within a workqueue item. This function exists only if lockdep is
312 int in_workqueue_context(struct workqueue_struct
*wq
)
314 return lock_is_held(&wq
->lockdep_map
);
318 #ifdef CONFIG_DEBUG_OBJECTS_WORK
320 static struct debug_obj_descr work_debug_descr
;
323 * fixup_init is called when:
324 * - an active object is initialized
326 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
328 struct work_struct
*work
= addr
;
331 case ODEBUG_STATE_ACTIVE
:
332 cancel_work_sync(work
);
333 debug_object_init(work
, &work_debug_descr
);
341 * fixup_activate is called when:
342 * - an active object is activated
343 * - an unknown object is activated (might be a statically initialized object)
345 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
347 struct work_struct
*work
= addr
;
351 case ODEBUG_STATE_NOTAVAILABLE
:
353 * This is not really a fixup. The work struct was
354 * statically initialized. We just make sure that it
355 * is tracked in the object tracker.
357 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
358 debug_object_init(work
, &work_debug_descr
);
359 debug_object_activate(work
, &work_debug_descr
);
365 case ODEBUG_STATE_ACTIVE
:
374 * fixup_free is called when:
375 * - an active object is freed
377 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
379 struct work_struct
*work
= addr
;
382 case ODEBUG_STATE_ACTIVE
:
383 cancel_work_sync(work
);
384 debug_object_free(work
, &work_debug_descr
);
391 static struct debug_obj_descr work_debug_descr
= {
392 .name
= "work_struct",
393 .fixup_init
= work_fixup_init
,
394 .fixup_activate
= work_fixup_activate
,
395 .fixup_free
= work_fixup_free
,
398 static inline void debug_work_activate(struct work_struct
*work
)
400 debug_object_activate(work
, &work_debug_descr
);
403 static inline void debug_work_deactivate(struct work_struct
*work
)
405 debug_object_deactivate(work
, &work_debug_descr
);
408 void __init_work(struct work_struct
*work
, int onstack
)
411 debug_object_init_on_stack(work
, &work_debug_descr
);
413 debug_object_init(work
, &work_debug_descr
);
415 EXPORT_SYMBOL_GPL(__init_work
);
417 void destroy_work_on_stack(struct work_struct
*work
)
419 debug_object_free(work
, &work_debug_descr
);
421 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
424 static inline void debug_work_activate(struct work_struct
*work
) { }
425 static inline void debug_work_deactivate(struct work_struct
*work
) { }
428 /* Serializes the accesses to the list of workqueues. */
429 static DEFINE_SPINLOCK(workqueue_lock
);
430 static LIST_HEAD(workqueues
);
431 static bool workqueue_freezing
; /* W: have wqs started freezing? */
434 * The almighty global cpu workqueues. nr_running is the only field
435 * which is expected to be used frequently by other cpus via
436 * try_to_wake_up(). Put it in a separate cacheline.
438 static DEFINE_PER_CPU(struct global_cwq
, global_cwq
);
439 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t
, gcwq_nr_running
);
442 * Global cpu workqueue and nr_running counter for unbound gcwq. The
443 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
444 * workers have WORKER_UNBOUND set.
446 static struct global_cwq unbound_global_cwq
;
447 static atomic_t unbound_gcwq_nr_running
= ATOMIC_INIT(0); /* always 0 */
449 static int worker_thread(void *__worker
);
451 static struct global_cwq
*get_gcwq(unsigned int cpu
)
453 if (cpu
!= WORK_CPU_UNBOUND
)
454 return &per_cpu(global_cwq
, cpu
);
456 return &unbound_global_cwq
;
459 static atomic_t
*get_gcwq_nr_running(unsigned int cpu
)
461 if (cpu
!= WORK_CPU_UNBOUND
)
462 return &per_cpu(gcwq_nr_running
, cpu
);
464 return &unbound_gcwq_nr_running
;
467 static struct cpu_workqueue_struct
*get_cwq(unsigned int cpu
,
468 struct workqueue_struct
*wq
)
470 if (!(wq
->flags
& WQ_UNBOUND
)) {
471 if (likely(cpu
< nr_cpu_ids
)) {
473 return per_cpu_ptr(wq
->cpu_wq
.pcpu
, cpu
);
475 return wq
->cpu_wq
.single
;
478 } else if (likely(cpu
== WORK_CPU_UNBOUND
))
479 return wq
->cpu_wq
.single
;
483 static unsigned int work_color_to_flags(int color
)
485 return color
<< WORK_STRUCT_COLOR_SHIFT
;
488 static int get_work_color(struct work_struct
*work
)
490 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
491 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
494 static int work_next_color(int color
)
496 return (color
+ 1) % WORK_NR_COLORS
;
500 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
501 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
502 * cleared and the work data contains the cpu number it was last on.
504 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
505 * cwq, cpu or clear work->data. These functions should only be
506 * called while the work is owned - ie. while the PENDING bit is set.
508 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
509 * corresponding to a work. gcwq is available once the work has been
510 * queued anywhere after initialization. cwq is available only from
511 * queueing until execution starts.
513 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
516 BUG_ON(!work_pending(work
));
517 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
520 static void set_work_cwq(struct work_struct
*work
,
521 struct cpu_workqueue_struct
*cwq
,
522 unsigned long extra_flags
)
524 set_work_data(work
, (unsigned long)cwq
,
525 WORK_STRUCT_PENDING
| WORK_STRUCT_CWQ
| extra_flags
);
528 static void set_work_cpu(struct work_struct
*work
, unsigned int cpu
)
530 set_work_data(work
, cpu
<< WORK_STRUCT_FLAG_BITS
, WORK_STRUCT_PENDING
);
533 static void clear_work_data(struct work_struct
*work
)
535 set_work_data(work
, WORK_STRUCT_NO_CPU
, 0);
538 static struct cpu_workqueue_struct
*get_work_cwq(struct work_struct
*work
)
540 unsigned long data
= atomic_long_read(&work
->data
);
542 if (data
& WORK_STRUCT_CWQ
)
543 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
548 static struct global_cwq
*get_work_gcwq(struct work_struct
*work
)
550 unsigned long data
= atomic_long_read(&work
->data
);
553 if (data
& WORK_STRUCT_CWQ
)
554 return ((struct cpu_workqueue_struct
*)
555 (data
& WORK_STRUCT_WQ_DATA_MASK
))->gcwq
;
557 cpu
= data
>> WORK_STRUCT_FLAG_BITS
;
558 if (cpu
== WORK_CPU_NONE
)
561 BUG_ON(cpu
>= nr_cpu_ids
&& cpu
!= WORK_CPU_UNBOUND
);
562 return get_gcwq(cpu
);
566 * Policy functions. These define the policies on how the global
567 * worker pool is managed. Unless noted otherwise, these functions
568 * assume that they're being called with gcwq->lock held.
571 static bool __need_more_worker(struct global_cwq
*gcwq
)
573 return !atomic_read(get_gcwq_nr_running(gcwq
->cpu
)) ||
574 gcwq
->flags
& GCWQ_HIGHPRI_PENDING
;
578 * Need to wake up a worker? Called from anything but currently
581 static bool need_more_worker(struct global_cwq
*gcwq
)
583 return !list_empty(&gcwq
->worklist
) && __need_more_worker(gcwq
);
586 /* Can I start working? Called from busy but !running workers. */
587 static bool may_start_working(struct global_cwq
*gcwq
)
589 return gcwq
->nr_idle
;
592 /* Do I need to keep working? Called from currently running workers. */
593 static bool keep_working(struct global_cwq
*gcwq
)
595 atomic_t
*nr_running
= get_gcwq_nr_running(gcwq
->cpu
);
597 return !list_empty(&gcwq
->worklist
) && atomic_read(nr_running
) <= 1;
600 /* Do we need a new worker? Called from manager. */
601 static bool need_to_create_worker(struct global_cwq
*gcwq
)
603 return need_more_worker(gcwq
) && !may_start_working(gcwq
);
606 /* Do I need to be the manager? */
607 static bool need_to_manage_workers(struct global_cwq
*gcwq
)
609 return need_to_create_worker(gcwq
) || gcwq
->flags
& GCWQ_MANAGE_WORKERS
;
612 /* Do we have too many workers and should some go away? */
613 static bool too_many_workers(struct global_cwq
*gcwq
)
615 bool managing
= gcwq
->flags
& GCWQ_MANAGING_WORKERS
;
616 int nr_idle
= gcwq
->nr_idle
+ managing
; /* manager is considered idle */
617 int nr_busy
= gcwq
->nr_workers
- nr_idle
;
619 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
626 /* Return the first worker. Safe with preemption disabled */
627 static struct worker
*first_worker(struct global_cwq
*gcwq
)
629 if (unlikely(list_empty(&gcwq
->idle_list
)))
632 return list_first_entry(&gcwq
->idle_list
, struct worker
, entry
);
636 * wake_up_worker - wake up an idle worker
637 * @gcwq: gcwq to wake worker for
639 * Wake up the first idle worker of @gcwq.
642 * spin_lock_irq(gcwq->lock).
644 static void wake_up_worker(struct global_cwq
*gcwq
)
646 struct worker
*worker
= first_worker(gcwq
);
649 wake_up_process(worker
->task
);
653 * wq_worker_waking_up - a worker is waking up
654 * @task: task waking up
655 * @cpu: CPU @task is waking up to
657 * This function is called during try_to_wake_up() when a worker is
661 * spin_lock_irq(rq->lock)
663 void wq_worker_waking_up(struct task_struct
*task
, unsigned int cpu
)
665 struct worker
*worker
= kthread_data(task
);
667 if (likely(!(worker
->flags
& WORKER_NOT_RUNNING
)))
668 atomic_inc(get_gcwq_nr_running(cpu
));
672 * wq_worker_sleeping - a worker is going to sleep
673 * @task: task going to sleep
674 * @cpu: CPU in question, must be the current CPU number
676 * This function is called during schedule() when a busy worker is
677 * going to sleep. Worker on the same cpu can be woken up by
678 * returning pointer to its task.
681 * spin_lock_irq(rq->lock)
684 * Worker task on @cpu to wake up, %NULL if none.
686 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
,
689 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
690 struct global_cwq
*gcwq
= get_gcwq(cpu
);
691 atomic_t
*nr_running
= get_gcwq_nr_running(cpu
);
693 if (unlikely(worker
->flags
& WORKER_NOT_RUNNING
))
696 /* this can only happen on the local cpu */
697 BUG_ON(cpu
!= raw_smp_processor_id());
700 * The counterpart of the following dec_and_test, implied mb,
701 * worklist not empty test sequence is in insert_work().
702 * Please read comment there.
704 * NOT_RUNNING is clear. This means that trustee is not in
705 * charge and we're running on the local cpu w/ rq lock held
706 * and preemption disabled, which in turn means that none else
707 * could be manipulating idle_list, so dereferencing idle_list
708 * without gcwq lock is safe.
710 if (atomic_dec_and_test(nr_running
) && !list_empty(&gcwq
->worklist
))
711 to_wakeup
= first_worker(gcwq
);
712 return to_wakeup
? to_wakeup
->task
: NULL
;
716 * worker_set_flags - set worker flags and adjust nr_running accordingly
718 * @flags: flags to set
719 * @wakeup: wakeup an idle worker if necessary
721 * Set @flags in @worker->flags and adjust nr_running accordingly. If
722 * nr_running becomes zero and @wakeup is %true, an idle worker is
726 * spin_lock_irq(gcwq->lock)
728 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
731 struct global_cwq
*gcwq
= worker
->gcwq
;
733 WARN_ON_ONCE(worker
->task
!= current
);
736 * If transitioning into NOT_RUNNING, adjust nr_running and
737 * wake up an idle worker as necessary if requested by
740 if ((flags
& WORKER_NOT_RUNNING
) &&
741 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
742 atomic_t
*nr_running
= get_gcwq_nr_running(gcwq
->cpu
);
745 if (atomic_dec_and_test(nr_running
) &&
746 !list_empty(&gcwq
->worklist
))
747 wake_up_worker(gcwq
);
749 atomic_dec(nr_running
);
752 worker
->flags
|= flags
;
756 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
758 * @flags: flags to clear
760 * Clear @flags in @worker->flags and adjust nr_running accordingly.
763 * spin_lock_irq(gcwq->lock)
765 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
767 struct global_cwq
*gcwq
= worker
->gcwq
;
768 unsigned int oflags
= worker
->flags
;
770 WARN_ON_ONCE(worker
->task
!= current
);
772 worker
->flags
&= ~flags
;
774 /* if transitioning out of NOT_RUNNING, increment nr_running */
775 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
776 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
777 atomic_inc(get_gcwq_nr_running(gcwq
->cpu
));
781 * busy_worker_head - return the busy hash head for a work
782 * @gcwq: gcwq of interest
783 * @work: work to be hashed
785 * Return hash head of @gcwq for @work.
788 * spin_lock_irq(gcwq->lock).
791 * Pointer to the hash head.
793 static struct hlist_head
*busy_worker_head(struct global_cwq
*gcwq
,
794 struct work_struct
*work
)
796 const int base_shift
= ilog2(sizeof(struct work_struct
));
797 unsigned long v
= (unsigned long)work
;
799 /* simple shift and fold hash, do we need something better? */
801 v
+= v
>> BUSY_WORKER_HASH_ORDER
;
802 v
&= BUSY_WORKER_HASH_MASK
;
804 return &gcwq
->busy_hash
[v
];
808 * __find_worker_executing_work - find worker which is executing a work
809 * @gcwq: gcwq of interest
810 * @bwh: hash head as returned by busy_worker_head()
811 * @work: work to find worker for
813 * Find a worker which is executing @work on @gcwq. @bwh should be
814 * the hash head obtained by calling busy_worker_head() with the same
818 * spin_lock_irq(gcwq->lock).
821 * Pointer to worker which is executing @work if found, NULL
824 static struct worker
*__find_worker_executing_work(struct global_cwq
*gcwq
,
825 struct hlist_head
*bwh
,
826 struct work_struct
*work
)
828 struct worker
*worker
;
829 struct hlist_node
*tmp
;
831 hlist_for_each_entry(worker
, tmp
, bwh
, hentry
)
832 if (worker
->current_work
== work
)
838 * find_worker_executing_work - find worker which is executing a work
839 * @gcwq: gcwq of interest
840 * @work: work to find worker for
842 * Find a worker which is executing @work on @gcwq. This function is
843 * identical to __find_worker_executing_work() except that this
844 * function calculates @bwh itself.
847 * spin_lock_irq(gcwq->lock).
850 * Pointer to worker which is executing @work if found, NULL
853 static struct worker
*find_worker_executing_work(struct global_cwq
*gcwq
,
854 struct work_struct
*work
)
856 return __find_worker_executing_work(gcwq
, busy_worker_head(gcwq
, work
),
861 * gcwq_determine_ins_pos - find insertion position
862 * @gcwq: gcwq of interest
863 * @cwq: cwq a work is being queued for
865 * A work for @cwq is about to be queued on @gcwq, determine insertion
866 * position for the work. If @cwq is for HIGHPRI wq, the work is
867 * queued at the head of the queue but in FIFO order with respect to
868 * other HIGHPRI works; otherwise, at the end of the queue. This
869 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
870 * there are HIGHPRI works pending.
873 * spin_lock_irq(gcwq->lock).
876 * Pointer to inserstion position.
878 static inline struct list_head
*gcwq_determine_ins_pos(struct global_cwq
*gcwq
,
879 struct cpu_workqueue_struct
*cwq
)
881 struct work_struct
*twork
;
883 if (likely(!(cwq
->wq
->flags
& WQ_HIGHPRI
)))
884 return &gcwq
->worklist
;
886 list_for_each_entry(twork
, &gcwq
->worklist
, entry
) {
887 struct cpu_workqueue_struct
*tcwq
= get_work_cwq(twork
);
889 if (!(tcwq
->wq
->flags
& WQ_HIGHPRI
))
893 gcwq
->flags
|= GCWQ_HIGHPRI_PENDING
;
894 return &twork
->entry
;
898 * insert_work - insert a work into gcwq
899 * @cwq: cwq @work belongs to
900 * @work: work to insert
901 * @head: insertion point
902 * @extra_flags: extra WORK_STRUCT_* flags to set
904 * Insert @work which belongs to @cwq into @gcwq after @head.
905 * @extra_flags is or'd to work_struct flags.
908 * spin_lock_irq(gcwq->lock).
910 static void insert_work(struct cpu_workqueue_struct
*cwq
,
911 struct work_struct
*work
, struct list_head
*head
,
912 unsigned int extra_flags
)
914 struct global_cwq
*gcwq
= cwq
->gcwq
;
916 /* we own @work, set data and link */
917 set_work_cwq(work
, cwq
, extra_flags
);
920 * Ensure that we get the right work->data if we see the
921 * result of list_add() below, see try_to_grab_pending().
925 list_add_tail(&work
->entry
, head
);
928 * Ensure either worker_sched_deactivated() sees the above
929 * list_add_tail() or we see zero nr_running to avoid workers
930 * lying around lazily while there are works to be processed.
934 if (__need_more_worker(gcwq
))
935 wake_up_worker(gcwq
);
938 static void __queue_work(unsigned int cpu
, struct workqueue_struct
*wq
,
939 struct work_struct
*work
)
941 struct global_cwq
*gcwq
;
942 struct cpu_workqueue_struct
*cwq
;
943 struct list_head
*worklist
;
946 debug_work_activate(work
);
948 if (WARN_ON_ONCE(wq
->flags
& WQ_DYING
))
951 /* determine gcwq to use */
952 if (!(wq
->flags
& WQ_UNBOUND
)) {
953 struct global_cwq
*last_gcwq
;
955 if (unlikely(cpu
== WORK_CPU_UNBOUND
))
956 cpu
= raw_smp_processor_id();
959 * It's multi cpu. If @wq is non-reentrant and @work
960 * was previously on a different cpu, it might still
961 * be running there, in which case the work needs to
962 * be queued on that cpu to guarantee non-reentrance.
964 gcwq
= get_gcwq(cpu
);
965 if (wq
->flags
& WQ_NON_REENTRANT
&&
966 (last_gcwq
= get_work_gcwq(work
)) && last_gcwq
!= gcwq
) {
967 struct worker
*worker
;
969 spin_lock_irqsave(&last_gcwq
->lock
, flags
);
971 worker
= find_worker_executing_work(last_gcwq
, work
);
973 if (worker
&& worker
->current_cwq
->wq
== wq
)
976 /* meh... not running there, queue here */
977 spin_unlock_irqrestore(&last_gcwq
->lock
, flags
);
978 spin_lock_irqsave(&gcwq
->lock
, flags
);
981 spin_lock_irqsave(&gcwq
->lock
, flags
);
983 gcwq
= get_gcwq(WORK_CPU_UNBOUND
);
984 spin_lock_irqsave(&gcwq
->lock
, flags
);
987 /* gcwq determined, get cwq and queue */
988 cwq
= get_cwq(gcwq
->cpu
, wq
);
990 BUG_ON(!list_empty(&work
->entry
));
992 cwq
->nr_in_flight
[cwq
->work_color
]++;
994 if (likely(cwq
->nr_active
< cwq
->max_active
)) {
996 worklist
= gcwq_determine_ins_pos(gcwq
, cwq
);
998 worklist
= &cwq
->delayed_works
;
1000 insert_work(cwq
, work
, worklist
, work_color_to_flags(cwq
->work_color
));
1002 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1006 * queue_work - queue work on a workqueue
1007 * @wq: workqueue to use
1008 * @work: work to queue
1010 * Returns 0 if @work was already on a queue, non-zero otherwise.
1012 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1013 * it can be processed by another CPU.
1015 int queue_work(struct workqueue_struct
*wq
, struct work_struct
*work
)
1019 ret
= queue_work_on(get_cpu(), wq
, work
);
1024 EXPORT_SYMBOL_GPL(queue_work
);
1027 * queue_work_on - queue work on specific cpu
1028 * @cpu: CPU number to execute work on
1029 * @wq: workqueue to use
1030 * @work: work to queue
1032 * Returns 0 if @work was already on a queue, non-zero otherwise.
1034 * We queue the work to a specific CPU, the caller must ensure it
1038 queue_work_on(int cpu
, struct workqueue_struct
*wq
, struct work_struct
*work
)
1042 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1043 __queue_work(cpu
, wq
, work
);
1048 EXPORT_SYMBOL_GPL(queue_work_on
);
1050 static void delayed_work_timer_fn(unsigned long __data
)
1052 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1053 struct cpu_workqueue_struct
*cwq
= get_work_cwq(&dwork
->work
);
1055 __queue_work(smp_processor_id(), cwq
->wq
, &dwork
->work
);
1059 * queue_delayed_work - queue work on a workqueue after delay
1060 * @wq: workqueue to use
1061 * @dwork: delayable work to queue
1062 * @delay: number of jiffies to wait before queueing
1064 * Returns 0 if @work was already on a queue, non-zero otherwise.
1066 int queue_delayed_work(struct workqueue_struct
*wq
,
1067 struct delayed_work
*dwork
, unsigned long delay
)
1070 return queue_work(wq
, &dwork
->work
);
1072 return queue_delayed_work_on(-1, wq
, dwork
, delay
);
1074 EXPORT_SYMBOL_GPL(queue_delayed_work
);
1077 * queue_delayed_work_on - queue work on specific CPU after delay
1078 * @cpu: CPU number to execute work on
1079 * @wq: workqueue to use
1080 * @dwork: work to queue
1081 * @delay: number of jiffies to wait before queueing
1083 * Returns 0 if @work was already on a queue, non-zero otherwise.
1085 int queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1086 struct delayed_work
*dwork
, unsigned long delay
)
1089 struct timer_list
*timer
= &dwork
->timer
;
1090 struct work_struct
*work
= &dwork
->work
;
1092 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1095 BUG_ON(timer_pending(timer
));
1096 BUG_ON(!list_empty(&work
->entry
));
1098 timer_stats_timer_set_start_info(&dwork
->timer
);
1101 * This stores cwq for the moment, for the timer_fn.
1102 * Note that the work's gcwq is preserved to allow
1103 * reentrance detection for delayed works.
1105 if (!(wq
->flags
& WQ_UNBOUND
)) {
1106 struct global_cwq
*gcwq
= get_work_gcwq(work
);
1108 if (gcwq
&& gcwq
->cpu
!= WORK_CPU_UNBOUND
)
1111 lcpu
= raw_smp_processor_id();
1113 lcpu
= WORK_CPU_UNBOUND
;
1115 set_work_cwq(work
, get_cwq(lcpu
, wq
), 0);
1117 timer
->expires
= jiffies
+ delay
;
1118 timer
->data
= (unsigned long)dwork
;
1119 timer
->function
= delayed_work_timer_fn
;
1121 if (unlikely(cpu
>= 0))
1122 add_timer_on(timer
, cpu
);
1129 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1132 * worker_enter_idle - enter idle state
1133 * @worker: worker which is entering idle state
1135 * @worker is entering idle state. Update stats and idle timer if
1139 * spin_lock_irq(gcwq->lock).
1141 static void worker_enter_idle(struct worker
*worker
)
1143 struct global_cwq
*gcwq
= worker
->gcwq
;
1145 BUG_ON(worker
->flags
& WORKER_IDLE
);
1146 BUG_ON(!list_empty(&worker
->entry
) &&
1147 (worker
->hentry
.next
|| worker
->hentry
.pprev
));
1149 /* can't use worker_set_flags(), also called from start_worker() */
1150 worker
->flags
|= WORKER_IDLE
;
1152 worker
->last_active
= jiffies
;
1154 /* idle_list is LIFO */
1155 list_add(&worker
->entry
, &gcwq
->idle_list
);
1157 if (likely(!(worker
->flags
& WORKER_ROGUE
))) {
1158 if (too_many_workers(gcwq
) && !timer_pending(&gcwq
->idle_timer
))
1159 mod_timer(&gcwq
->idle_timer
,
1160 jiffies
+ IDLE_WORKER_TIMEOUT
);
1162 wake_up_all(&gcwq
->trustee_wait
);
1164 /* sanity check nr_running */
1165 WARN_ON_ONCE(gcwq
->nr_workers
== gcwq
->nr_idle
&&
1166 atomic_read(get_gcwq_nr_running(gcwq
->cpu
)));
1170 * worker_leave_idle - leave idle state
1171 * @worker: worker which is leaving idle state
1173 * @worker is leaving idle state. Update stats.
1176 * spin_lock_irq(gcwq->lock).
1178 static void worker_leave_idle(struct worker
*worker
)
1180 struct global_cwq
*gcwq
= worker
->gcwq
;
1182 BUG_ON(!(worker
->flags
& WORKER_IDLE
));
1183 worker_clr_flags(worker
, WORKER_IDLE
);
1185 list_del_init(&worker
->entry
);
1189 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1192 * Works which are scheduled while the cpu is online must at least be
1193 * scheduled to a worker which is bound to the cpu so that if they are
1194 * flushed from cpu callbacks while cpu is going down, they are
1195 * guaranteed to execute on the cpu.
1197 * This function is to be used by rogue workers and rescuers to bind
1198 * themselves to the target cpu and may race with cpu going down or
1199 * coming online. kthread_bind() can't be used because it may put the
1200 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1201 * verbatim as it's best effort and blocking and gcwq may be
1202 * [dis]associated in the meantime.
1204 * This function tries set_cpus_allowed() and locks gcwq and verifies
1205 * the binding against GCWQ_DISASSOCIATED which is set during
1206 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1207 * idle state or fetches works without dropping lock, it can guarantee
1208 * the scheduling requirement described in the first paragraph.
1211 * Might sleep. Called without any lock but returns with gcwq->lock
1215 * %true if the associated gcwq is online (@worker is successfully
1216 * bound), %false if offline.
1218 static bool worker_maybe_bind_and_lock(struct worker
*worker
)
1219 __acquires(&gcwq
->lock
)
1221 struct global_cwq
*gcwq
= worker
->gcwq
;
1222 struct task_struct
*task
= worker
->task
;
1226 * The following call may fail, succeed or succeed
1227 * without actually migrating the task to the cpu if
1228 * it races with cpu hotunplug operation. Verify
1229 * against GCWQ_DISASSOCIATED.
1231 if (!(gcwq
->flags
& GCWQ_DISASSOCIATED
))
1232 set_cpus_allowed_ptr(task
, get_cpu_mask(gcwq
->cpu
));
1234 spin_lock_irq(&gcwq
->lock
);
1235 if (gcwq
->flags
& GCWQ_DISASSOCIATED
)
1237 if (task_cpu(task
) == gcwq
->cpu
&&
1238 cpumask_equal(¤t
->cpus_allowed
,
1239 get_cpu_mask(gcwq
->cpu
)))
1241 spin_unlock_irq(&gcwq
->lock
);
1243 /* CPU has come up inbetween, retry migration */
1249 * Function for worker->rebind_work used to rebind rogue busy workers
1250 * to the associated cpu which is coming back online. This is
1251 * scheduled by cpu up but can race with other cpu hotplug operations
1252 * and may be executed twice without intervening cpu down.
1254 static void worker_rebind_fn(struct work_struct
*work
)
1256 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1257 struct global_cwq
*gcwq
= worker
->gcwq
;
1259 if (worker_maybe_bind_and_lock(worker
))
1260 worker_clr_flags(worker
, WORKER_REBIND
);
1262 spin_unlock_irq(&gcwq
->lock
);
1265 static struct worker
*alloc_worker(void)
1267 struct worker
*worker
;
1269 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1271 INIT_LIST_HEAD(&worker
->entry
);
1272 INIT_LIST_HEAD(&worker
->scheduled
);
1273 INIT_WORK(&worker
->rebind_work
, worker_rebind_fn
);
1274 /* on creation a worker is in !idle && prep state */
1275 worker
->flags
= WORKER_PREP
;
1281 * create_worker - create a new workqueue worker
1282 * @gcwq: gcwq the new worker will belong to
1283 * @bind: whether to set affinity to @cpu or not
1285 * Create a new worker which is bound to @gcwq. The returned worker
1286 * can be started by calling start_worker() or destroyed using
1290 * Might sleep. Does GFP_KERNEL allocations.
1293 * Pointer to the newly created worker.
1295 static struct worker
*create_worker(struct global_cwq
*gcwq
, bool bind
)
1297 bool on_unbound_cpu
= gcwq
->cpu
== WORK_CPU_UNBOUND
;
1298 struct worker
*worker
= NULL
;
1301 spin_lock_irq(&gcwq
->lock
);
1302 while (ida_get_new(&gcwq
->worker_ida
, &id
)) {
1303 spin_unlock_irq(&gcwq
->lock
);
1304 if (!ida_pre_get(&gcwq
->worker_ida
, GFP_KERNEL
))
1306 spin_lock_irq(&gcwq
->lock
);
1308 spin_unlock_irq(&gcwq
->lock
);
1310 worker
= alloc_worker();
1314 worker
->gcwq
= gcwq
;
1317 if (!on_unbound_cpu
)
1318 worker
->task
= kthread_create(worker_thread
, worker
,
1319 "kworker/%u:%d", gcwq
->cpu
, id
);
1321 worker
->task
= kthread_create(worker_thread
, worker
,
1322 "kworker/u:%d", id
);
1323 if (IS_ERR(worker
->task
))
1327 * A rogue worker will become a regular one if CPU comes
1328 * online later on. Make sure every worker has
1329 * PF_THREAD_BOUND set.
1331 if (bind
&& !on_unbound_cpu
)
1332 kthread_bind(worker
->task
, gcwq
->cpu
);
1334 worker
->task
->flags
|= PF_THREAD_BOUND
;
1336 worker
->flags
|= WORKER_UNBOUND
;
1342 spin_lock_irq(&gcwq
->lock
);
1343 ida_remove(&gcwq
->worker_ida
, id
);
1344 spin_unlock_irq(&gcwq
->lock
);
1351 * start_worker - start a newly created worker
1352 * @worker: worker to start
1354 * Make the gcwq aware of @worker and start it.
1357 * spin_lock_irq(gcwq->lock).
1359 static void start_worker(struct worker
*worker
)
1361 worker
->flags
|= WORKER_STARTED
;
1362 worker
->gcwq
->nr_workers
++;
1363 worker_enter_idle(worker
);
1364 wake_up_process(worker
->task
);
1368 * destroy_worker - destroy a workqueue worker
1369 * @worker: worker to be destroyed
1371 * Destroy @worker and adjust @gcwq stats accordingly.
1374 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1376 static void destroy_worker(struct worker
*worker
)
1378 struct global_cwq
*gcwq
= worker
->gcwq
;
1379 int id
= worker
->id
;
1381 /* sanity check frenzy */
1382 BUG_ON(worker
->current_work
);
1383 BUG_ON(!list_empty(&worker
->scheduled
));
1385 if (worker
->flags
& WORKER_STARTED
)
1387 if (worker
->flags
& WORKER_IDLE
)
1390 list_del_init(&worker
->entry
);
1391 worker
->flags
|= WORKER_DIE
;
1393 spin_unlock_irq(&gcwq
->lock
);
1395 kthread_stop(worker
->task
);
1398 spin_lock_irq(&gcwq
->lock
);
1399 ida_remove(&gcwq
->worker_ida
, id
);
1402 static void idle_worker_timeout(unsigned long __gcwq
)
1404 struct global_cwq
*gcwq
= (void *)__gcwq
;
1406 spin_lock_irq(&gcwq
->lock
);
1408 if (too_many_workers(gcwq
)) {
1409 struct worker
*worker
;
1410 unsigned long expires
;
1412 /* idle_list is kept in LIFO order, check the last one */
1413 worker
= list_entry(gcwq
->idle_list
.prev
, struct worker
, entry
);
1414 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1416 if (time_before(jiffies
, expires
))
1417 mod_timer(&gcwq
->idle_timer
, expires
);
1419 /* it's been idle for too long, wake up manager */
1420 gcwq
->flags
|= GCWQ_MANAGE_WORKERS
;
1421 wake_up_worker(gcwq
);
1425 spin_unlock_irq(&gcwq
->lock
);
1428 static bool send_mayday(struct work_struct
*work
)
1430 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1431 struct workqueue_struct
*wq
= cwq
->wq
;
1434 if (!(wq
->flags
& WQ_RESCUER
))
1437 /* mayday mayday mayday */
1438 cpu
= cwq
->gcwq
->cpu
;
1439 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1440 if (cpu
== WORK_CPU_UNBOUND
)
1442 if (!mayday_test_and_set_cpu(cpu
, wq
->mayday_mask
))
1443 wake_up_process(wq
->rescuer
->task
);
1447 static void gcwq_mayday_timeout(unsigned long __gcwq
)
1449 struct global_cwq
*gcwq
= (void *)__gcwq
;
1450 struct work_struct
*work
;
1452 spin_lock_irq(&gcwq
->lock
);
1454 if (need_to_create_worker(gcwq
)) {
1456 * We've been trying to create a new worker but
1457 * haven't been successful. We might be hitting an
1458 * allocation deadlock. Send distress signals to
1461 list_for_each_entry(work
, &gcwq
->worklist
, entry
)
1465 spin_unlock_irq(&gcwq
->lock
);
1467 mod_timer(&gcwq
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1471 * maybe_create_worker - create a new worker if necessary
1472 * @gcwq: gcwq to create a new worker for
1474 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1475 * have at least one idle worker on return from this function. If
1476 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1477 * sent to all rescuers with works scheduled on @gcwq to resolve
1478 * possible allocation deadlock.
1480 * On return, need_to_create_worker() is guaranteed to be false and
1481 * may_start_working() true.
1484 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1485 * multiple times. Does GFP_KERNEL allocations. Called only from
1489 * false if no action was taken and gcwq->lock stayed locked, true
1492 static bool maybe_create_worker(struct global_cwq
*gcwq
)
1493 __releases(&gcwq
->lock
)
1494 __acquires(&gcwq
->lock
)
1496 if (!need_to_create_worker(gcwq
))
1499 spin_unlock_irq(&gcwq
->lock
);
1501 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1502 mod_timer(&gcwq
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1505 struct worker
*worker
;
1507 worker
= create_worker(gcwq
, true);
1509 del_timer_sync(&gcwq
->mayday_timer
);
1510 spin_lock_irq(&gcwq
->lock
);
1511 start_worker(worker
);
1512 BUG_ON(need_to_create_worker(gcwq
));
1516 if (!need_to_create_worker(gcwq
))
1519 __set_current_state(TASK_INTERRUPTIBLE
);
1520 schedule_timeout(CREATE_COOLDOWN
);
1522 if (!need_to_create_worker(gcwq
))
1526 del_timer_sync(&gcwq
->mayday_timer
);
1527 spin_lock_irq(&gcwq
->lock
);
1528 if (need_to_create_worker(gcwq
))
1534 * maybe_destroy_worker - destroy workers which have been idle for a while
1535 * @gcwq: gcwq to destroy workers for
1537 * Destroy @gcwq workers which have been idle for longer than
1538 * IDLE_WORKER_TIMEOUT.
1541 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1542 * multiple times. Called only from manager.
1545 * false if no action was taken and gcwq->lock stayed locked, true
1548 static bool maybe_destroy_workers(struct global_cwq
*gcwq
)
1552 while (too_many_workers(gcwq
)) {
1553 struct worker
*worker
;
1554 unsigned long expires
;
1556 worker
= list_entry(gcwq
->idle_list
.prev
, struct worker
, entry
);
1557 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1559 if (time_before(jiffies
, expires
)) {
1560 mod_timer(&gcwq
->idle_timer
, expires
);
1564 destroy_worker(worker
);
1572 * manage_workers - manage worker pool
1575 * Assume the manager role and manage gcwq worker pool @worker belongs
1576 * to. At any given time, there can be only zero or one manager per
1577 * gcwq. The exclusion is handled automatically by this function.
1579 * The caller can safely start processing works on false return. On
1580 * true return, it's guaranteed that need_to_create_worker() is false
1581 * and may_start_working() is true.
1584 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1585 * multiple times. Does GFP_KERNEL allocations.
1588 * false if no action was taken and gcwq->lock stayed locked, true if
1589 * some action was taken.
1591 static bool manage_workers(struct worker
*worker
)
1593 struct global_cwq
*gcwq
= worker
->gcwq
;
1596 if (gcwq
->flags
& GCWQ_MANAGING_WORKERS
)
1599 gcwq
->flags
&= ~GCWQ_MANAGE_WORKERS
;
1600 gcwq
->flags
|= GCWQ_MANAGING_WORKERS
;
1603 * Destroy and then create so that may_start_working() is true
1606 ret
|= maybe_destroy_workers(gcwq
);
1607 ret
|= maybe_create_worker(gcwq
);
1609 gcwq
->flags
&= ~GCWQ_MANAGING_WORKERS
;
1612 * The trustee might be waiting to take over the manager
1613 * position, tell it we're done.
1615 if (unlikely(gcwq
->trustee
))
1616 wake_up_all(&gcwq
->trustee_wait
);
1622 * move_linked_works - move linked works to a list
1623 * @work: start of series of works to be scheduled
1624 * @head: target list to append @work to
1625 * @nextp: out paramter for nested worklist walking
1627 * Schedule linked works starting from @work to @head. Work series to
1628 * be scheduled starts at @work and includes any consecutive work with
1629 * WORK_STRUCT_LINKED set in its predecessor.
1631 * If @nextp is not NULL, it's updated to point to the next work of
1632 * the last scheduled work. This allows move_linked_works() to be
1633 * nested inside outer list_for_each_entry_safe().
1636 * spin_lock_irq(gcwq->lock).
1638 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1639 struct work_struct
**nextp
)
1641 struct work_struct
*n
;
1644 * Linked worklist will always end before the end of the list,
1645 * use NULL for list head.
1647 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1648 list_move_tail(&work
->entry
, head
);
1649 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1654 * If we're already inside safe list traversal and have moved
1655 * multiple works to the scheduled queue, the next position
1656 * needs to be updated.
1662 static void cwq_activate_first_delayed(struct cpu_workqueue_struct
*cwq
)
1664 struct work_struct
*work
= list_first_entry(&cwq
->delayed_works
,
1665 struct work_struct
, entry
);
1666 struct list_head
*pos
= gcwq_determine_ins_pos(cwq
->gcwq
, cwq
);
1668 move_linked_works(work
, pos
, NULL
);
1673 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1674 * @cwq: cwq of interest
1675 * @color: color of work which left the queue
1677 * A work either has completed or is removed from pending queue,
1678 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1681 * spin_lock_irq(gcwq->lock).
1683 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct
*cwq
, int color
)
1685 /* ignore uncolored works */
1686 if (color
== WORK_NO_COLOR
)
1689 cwq
->nr_in_flight
[color
]--;
1692 if (!list_empty(&cwq
->delayed_works
)) {
1693 /* one down, submit a delayed one */
1694 if (cwq
->nr_active
< cwq
->max_active
)
1695 cwq_activate_first_delayed(cwq
);
1698 /* is flush in progress and are we at the flushing tip? */
1699 if (likely(cwq
->flush_color
!= color
))
1702 /* are there still in-flight works? */
1703 if (cwq
->nr_in_flight
[color
])
1706 /* this cwq is done, clear flush_color */
1707 cwq
->flush_color
= -1;
1710 * If this was the last cwq, wake up the first flusher. It
1711 * will handle the rest.
1713 if (atomic_dec_and_test(&cwq
->wq
->nr_cwqs_to_flush
))
1714 complete(&cwq
->wq
->first_flusher
->done
);
1718 * process_one_work - process single work
1720 * @work: work to process
1722 * Process @work. This function contains all the logics necessary to
1723 * process a single work including synchronization against and
1724 * interaction with other workers on the same cpu, queueing and
1725 * flushing. As long as context requirement is met, any worker can
1726 * call this function to process a work.
1729 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1731 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
1732 __releases(&gcwq
->lock
)
1733 __acquires(&gcwq
->lock
)
1735 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1736 struct global_cwq
*gcwq
= cwq
->gcwq
;
1737 struct hlist_head
*bwh
= busy_worker_head(gcwq
, work
);
1738 bool cpu_intensive
= cwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
1739 work_func_t f
= work
->func
;
1741 struct worker
*collision
;
1742 #ifdef CONFIG_LOCKDEP
1744 * It is permissible to free the struct work_struct from
1745 * inside the function that is called from it, this we need to
1746 * take into account for lockdep too. To avoid bogus "held
1747 * lock freed" warnings as well as problems when looking into
1748 * work->lockdep_map, make a copy and use that here.
1750 struct lockdep_map lockdep_map
= work
->lockdep_map
;
1753 * A single work shouldn't be executed concurrently by
1754 * multiple workers on a single cpu. Check whether anyone is
1755 * already processing the work. If so, defer the work to the
1756 * currently executing one.
1758 collision
= __find_worker_executing_work(gcwq
, bwh
, work
);
1759 if (unlikely(collision
)) {
1760 move_linked_works(work
, &collision
->scheduled
, NULL
);
1764 /* claim and process */
1765 debug_work_deactivate(work
);
1766 hlist_add_head(&worker
->hentry
, bwh
);
1767 worker
->current_work
= work
;
1768 worker
->current_cwq
= cwq
;
1769 work_color
= get_work_color(work
);
1771 /* record the current cpu number in the work data and dequeue */
1772 set_work_cpu(work
, gcwq
->cpu
);
1773 list_del_init(&work
->entry
);
1776 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1777 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1779 if (unlikely(gcwq
->flags
& GCWQ_HIGHPRI_PENDING
)) {
1780 struct work_struct
*nwork
= list_first_entry(&gcwq
->worklist
,
1781 struct work_struct
, entry
);
1783 if (!list_empty(&gcwq
->worklist
) &&
1784 get_work_cwq(nwork
)->wq
->flags
& WQ_HIGHPRI
)
1785 wake_up_worker(gcwq
);
1787 gcwq
->flags
&= ~GCWQ_HIGHPRI_PENDING
;
1791 * CPU intensive works don't participate in concurrency
1792 * management. They're the scheduler's responsibility.
1794 if (unlikely(cpu_intensive
))
1795 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
1797 spin_unlock_irq(&gcwq
->lock
);
1799 work_clear_pending(work
);
1800 lock_map_acquire(&cwq
->wq
->lockdep_map
);
1801 lock_map_acquire(&lockdep_map
);
1803 lock_map_release(&lockdep_map
);
1804 lock_map_release(&cwq
->wq
->lockdep_map
);
1806 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
1807 printk(KERN_ERR
"BUG: workqueue leaked lock or atomic: "
1809 current
->comm
, preempt_count(), task_pid_nr(current
));
1810 printk(KERN_ERR
" last function: ");
1811 print_symbol("%s\n", (unsigned long)f
);
1812 debug_show_held_locks(current
);
1816 spin_lock_irq(&gcwq
->lock
);
1818 /* clear cpu intensive status */
1819 if (unlikely(cpu_intensive
))
1820 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
1822 /* we're done with it, release */
1823 hlist_del_init(&worker
->hentry
);
1824 worker
->current_work
= NULL
;
1825 worker
->current_cwq
= NULL
;
1826 cwq_dec_nr_in_flight(cwq
, work_color
);
1830 * process_scheduled_works - process scheduled works
1833 * Process all scheduled works. Please note that the scheduled list
1834 * may change while processing a work, so this function repeatedly
1835 * fetches a work from the top and executes it.
1838 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1841 static void process_scheduled_works(struct worker
*worker
)
1843 while (!list_empty(&worker
->scheduled
)) {
1844 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
1845 struct work_struct
, entry
);
1846 process_one_work(worker
, work
);
1851 * worker_thread - the worker thread function
1854 * The gcwq worker thread function. There's a single dynamic pool of
1855 * these per each cpu. These workers process all works regardless of
1856 * their specific target workqueue. The only exception is works which
1857 * belong to workqueues with a rescuer which will be explained in
1860 static int worker_thread(void *__worker
)
1862 struct worker
*worker
= __worker
;
1863 struct global_cwq
*gcwq
= worker
->gcwq
;
1865 /* tell the scheduler that this is a workqueue worker */
1866 worker
->task
->flags
|= PF_WQ_WORKER
;
1868 spin_lock_irq(&gcwq
->lock
);
1870 /* DIE can be set only while we're idle, checking here is enough */
1871 if (worker
->flags
& WORKER_DIE
) {
1872 spin_unlock_irq(&gcwq
->lock
);
1873 worker
->task
->flags
&= ~PF_WQ_WORKER
;
1877 worker_leave_idle(worker
);
1879 /* no more worker necessary? */
1880 if (!need_more_worker(gcwq
))
1883 /* do we need to manage? */
1884 if (unlikely(!may_start_working(gcwq
)) && manage_workers(worker
))
1888 * ->scheduled list can only be filled while a worker is
1889 * preparing to process a work or actually processing it.
1890 * Make sure nobody diddled with it while I was sleeping.
1892 BUG_ON(!list_empty(&worker
->scheduled
));
1895 * When control reaches this point, we're guaranteed to have
1896 * at least one idle worker or that someone else has already
1897 * assumed the manager role.
1899 worker_clr_flags(worker
, WORKER_PREP
);
1902 struct work_struct
*work
=
1903 list_first_entry(&gcwq
->worklist
,
1904 struct work_struct
, entry
);
1906 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
1907 /* optimization path, not strictly necessary */
1908 process_one_work(worker
, work
);
1909 if (unlikely(!list_empty(&worker
->scheduled
)))
1910 process_scheduled_works(worker
);
1912 move_linked_works(work
, &worker
->scheduled
, NULL
);
1913 process_scheduled_works(worker
);
1915 } while (keep_working(gcwq
));
1917 worker_set_flags(worker
, WORKER_PREP
, false);
1919 if (unlikely(need_to_manage_workers(gcwq
)) && manage_workers(worker
))
1923 * gcwq->lock is held and there's no work to process and no
1924 * need to manage, sleep. Workers are woken up only while
1925 * holding gcwq->lock or from local cpu, so setting the
1926 * current state before releasing gcwq->lock is enough to
1927 * prevent losing any event.
1929 worker_enter_idle(worker
);
1930 __set_current_state(TASK_INTERRUPTIBLE
);
1931 spin_unlock_irq(&gcwq
->lock
);
1937 * rescuer_thread - the rescuer thread function
1938 * @__wq: the associated workqueue
1940 * Workqueue rescuer thread function. There's one rescuer for each
1941 * workqueue which has WQ_RESCUER set.
1943 * Regular work processing on a gcwq may block trying to create a new
1944 * worker which uses GFP_KERNEL allocation which has slight chance of
1945 * developing into deadlock if some works currently on the same queue
1946 * need to be processed to satisfy the GFP_KERNEL allocation. This is
1947 * the problem rescuer solves.
1949 * When such condition is possible, the gcwq summons rescuers of all
1950 * workqueues which have works queued on the gcwq and let them process
1951 * those works so that forward progress can be guaranteed.
1953 * This should happen rarely.
1955 static int rescuer_thread(void *__wq
)
1957 struct workqueue_struct
*wq
= __wq
;
1958 struct worker
*rescuer
= wq
->rescuer
;
1959 struct list_head
*scheduled
= &rescuer
->scheduled
;
1960 bool is_unbound
= wq
->flags
& WQ_UNBOUND
;
1963 set_user_nice(current
, RESCUER_NICE_LEVEL
);
1965 set_current_state(TASK_INTERRUPTIBLE
);
1967 if (kthread_should_stop())
1971 * See whether any cpu is asking for help. Unbounded
1972 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
1974 for_each_mayday_cpu(cpu
, wq
->mayday_mask
) {
1975 unsigned int tcpu
= is_unbound
? WORK_CPU_UNBOUND
: cpu
;
1976 struct cpu_workqueue_struct
*cwq
= get_cwq(tcpu
, wq
);
1977 struct global_cwq
*gcwq
= cwq
->gcwq
;
1978 struct work_struct
*work
, *n
;
1980 __set_current_state(TASK_RUNNING
);
1981 mayday_clear_cpu(cpu
, wq
->mayday_mask
);
1983 /* migrate to the target cpu if possible */
1984 rescuer
->gcwq
= gcwq
;
1985 worker_maybe_bind_and_lock(rescuer
);
1988 * Slurp in all works issued via this workqueue and
1991 BUG_ON(!list_empty(&rescuer
->scheduled
));
1992 list_for_each_entry_safe(work
, n
, &gcwq
->worklist
, entry
)
1993 if (get_work_cwq(work
) == cwq
)
1994 move_linked_works(work
, scheduled
, &n
);
1996 process_scheduled_works(rescuer
);
1997 spin_unlock_irq(&gcwq
->lock
);
2005 struct work_struct work
;
2006 struct completion done
;
2009 static void wq_barrier_func(struct work_struct
*work
)
2011 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2012 complete(&barr
->done
);
2016 * insert_wq_barrier - insert a barrier work
2017 * @cwq: cwq to insert barrier into
2018 * @barr: wq_barrier to insert
2019 * @target: target work to attach @barr to
2020 * @worker: worker currently executing @target, NULL if @target is not executing
2022 * @barr is linked to @target such that @barr is completed only after
2023 * @target finishes execution. Please note that the ordering
2024 * guarantee is observed only with respect to @target and on the local
2027 * Currently, a queued barrier can't be canceled. This is because
2028 * try_to_grab_pending() can't determine whether the work to be
2029 * grabbed is at the head of the queue and thus can't clear LINKED
2030 * flag of the previous work while there must be a valid next work
2031 * after a work with LINKED flag set.
2033 * Note that when @worker is non-NULL, @target may be modified
2034 * underneath us, so we can't reliably determine cwq from @target.
2037 * spin_lock_irq(gcwq->lock).
2039 static void insert_wq_barrier(struct cpu_workqueue_struct
*cwq
,
2040 struct wq_barrier
*barr
,
2041 struct work_struct
*target
, struct worker
*worker
)
2043 struct list_head
*head
;
2044 unsigned int linked
= 0;
2047 * debugobject calls are safe here even with gcwq->lock locked
2048 * as we know for sure that this will not trigger any of the
2049 * checks and call back into the fixup functions where we
2052 INIT_WORK_ON_STACK(&barr
->work
, wq_barrier_func
);
2053 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2054 init_completion(&barr
->done
);
2057 * If @target is currently being executed, schedule the
2058 * barrier to the worker; otherwise, put it after @target.
2061 head
= worker
->scheduled
.next
;
2063 unsigned long *bits
= work_data_bits(target
);
2065 head
= target
->entry
.next
;
2066 /* there can already be other linked works, inherit and set */
2067 linked
= *bits
& WORK_STRUCT_LINKED
;
2068 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2071 debug_work_activate(&barr
->work
);
2072 insert_work(cwq
, &barr
->work
, head
,
2073 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2077 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2078 * @wq: workqueue being flushed
2079 * @flush_color: new flush color, < 0 for no-op
2080 * @work_color: new work color, < 0 for no-op
2082 * Prepare cwqs for workqueue flushing.
2084 * If @flush_color is non-negative, flush_color on all cwqs should be
2085 * -1. If no cwq has in-flight commands at the specified color, all
2086 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2087 * has in flight commands, its cwq->flush_color is set to
2088 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2089 * wakeup logic is armed and %true is returned.
2091 * The caller should have initialized @wq->first_flusher prior to
2092 * calling this function with non-negative @flush_color. If
2093 * @flush_color is negative, no flush color update is done and %false
2096 * If @work_color is non-negative, all cwqs should have the same
2097 * work_color which is previous to @work_color and all will be
2098 * advanced to @work_color.
2101 * mutex_lock(wq->flush_mutex).
2104 * %true if @flush_color >= 0 and there's something to flush. %false
2107 static bool flush_workqueue_prep_cwqs(struct workqueue_struct
*wq
,
2108 int flush_color
, int work_color
)
2113 if (flush_color
>= 0) {
2114 BUG_ON(atomic_read(&wq
->nr_cwqs_to_flush
));
2115 atomic_set(&wq
->nr_cwqs_to_flush
, 1);
2118 for_each_cwq_cpu(cpu
, wq
) {
2119 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2120 struct global_cwq
*gcwq
= cwq
->gcwq
;
2122 spin_lock_irq(&gcwq
->lock
);
2124 if (flush_color
>= 0) {
2125 BUG_ON(cwq
->flush_color
!= -1);
2127 if (cwq
->nr_in_flight
[flush_color
]) {
2128 cwq
->flush_color
= flush_color
;
2129 atomic_inc(&wq
->nr_cwqs_to_flush
);
2134 if (work_color
>= 0) {
2135 BUG_ON(work_color
!= work_next_color(cwq
->work_color
));
2136 cwq
->work_color
= work_color
;
2139 spin_unlock_irq(&gcwq
->lock
);
2142 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_cwqs_to_flush
))
2143 complete(&wq
->first_flusher
->done
);
2149 * flush_workqueue - ensure that any scheduled work has run to completion.
2150 * @wq: workqueue to flush
2152 * Forces execution of the workqueue and blocks until its completion.
2153 * This is typically used in driver shutdown handlers.
2155 * We sleep until all works which were queued on entry have been handled,
2156 * but we are not livelocked by new incoming ones.
2158 void flush_workqueue(struct workqueue_struct
*wq
)
2160 struct wq_flusher this_flusher
= {
2161 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2163 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2167 lock_map_acquire(&wq
->lockdep_map
);
2168 lock_map_release(&wq
->lockdep_map
);
2170 mutex_lock(&wq
->flush_mutex
);
2173 * Start-to-wait phase
2175 next_color
= work_next_color(wq
->work_color
);
2177 if (next_color
!= wq
->flush_color
) {
2179 * Color space is not full. The current work_color
2180 * becomes our flush_color and work_color is advanced
2183 BUG_ON(!list_empty(&wq
->flusher_overflow
));
2184 this_flusher
.flush_color
= wq
->work_color
;
2185 wq
->work_color
= next_color
;
2187 if (!wq
->first_flusher
) {
2188 /* no flush in progress, become the first flusher */
2189 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2191 wq
->first_flusher
= &this_flusher
;
2193 if (!flush_workqueue_prep_cwqs(wq
, wq
->flush_color
,
2195 /* nothing to flush, done */
2196 wq
->flush_color
= next_color
;
2197 wq
->first_flusher
= NULL
;
2202 BUG_ON(wq
->flush_color
== this_flusher
.flush_color
);
2203 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2204 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2208 * Oops, color space is full, wait on overflow queue.
2209 * The next flush completion will assign us
2210 * flush_color and transfer to flusher_queue.
2212 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2215 mutex_unlock(&wq
->flush_mutex
);
2217 wait_for_completion(&this_flusher
.done
);
2220 * Wake-up-and-cascade phase
2222 * First flushers are responsible for cascading flushes and
2223 * handling overflow. Non-first flushers can simply return.
2225 if (wq
->first_flusher
!= &this_flusher
)
2228 mutex_lock(&wq
->flush_mutex
);
2230 /* we might have raced, check again with mutex held */
2231 if (wq
->first_flusher
!= &this_flusher
)
2234 wq
->first_flusher
= NULL
;
2236 BUG_ON(!list_empty(&this_flusher
.list
));
2237 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2240 struct wq_flusher
*next
, *tmp
;
2242 /* complete all the flushers sharing the current flush color */
2243 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2244 if (next
->flush_color
!= wq
->flush_color
)
2246 list_del_init(&next
->list
);
2247 complete(&next
->done
);
2250 BUG_ON(!list_empty(&wq
->flusher_overflow
) &&
2251 wq
->flush_color
!= work_next_color(wq
->work_color
));
2253 /* this flush_color is finished, advance by one */
2254 wq
->flush_color
= work_next_color(wq
->flush_color
);
2256 /* one color has been freed, handle overflow queue */
2257 if (!list_empty(&wq
->flusher_overflow
)) {
2259 * Assign the same color to all overflowed
2260 * flushers, advance work_color and append to
2261 * flusher_queue. This is the start-to-wait
2262 * phase for these overflowed flushers.
2264 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2265 tmp
->flush_color
= wq
->work_color
;
2267 wq
->work_color
= work_next_color(wq
->work_color
);
2269 list_splice_tail_init(&wq
->flusher_overflow
,
2270 &wq
->flusher_queue
);
2271 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2274 if (list_empty(&wq
->flusher_queue
)) {
2275 BUG_ON(wq
->flush_color
!= wq
->work_color
);
2280 * Need to flush more colors. Make the next flusher
2281 * the new first flusher and arm cwqs.
2283 BUG_ON(wq
->flush_color
== wq
->work_color
);
2284 BUG_ON(wq
->flush_color
!= next
->flush_color
);
2286 list_del_init(&next
->list
);
2287 wq
->first_flusher
= next
;
2289 if (flush_workqueue_prep_cwqs(wq
, wq
->flush_color
, -1))
2293 * Meh... this color is already done, clear first
2294 * flusher and repeat cascading.
2296 wq
->first_flusher
= NULL
;
2300 mutex_unlock(&wq
->flush_mutex
);
2302 EXPORT_SYMBOL_GPL(flush_workqueue
);
2305 * flush_work - block until a work_struct's callback has terminated
2306 * @work: the work which is to be flushed
2308 * Returns false if @work has already terminated.
2310 * It is expected that, prior to calling flush_work(), the caller has
2311 * arranged for the work to not be requeued, otherwise it doesn't make
2312 * sense to use this function.
2314 int flush_work(struct work_struct
*work
)
2316 struct worker
*worker
= NULL
;
2317 struct global_cwq
*gcwq
;
2318 struct cpu_workqueue_struct
*cwq
;
2319 struct wq_barrier barr
;
2322 gcwq
= get_work_gcwq(work
);
2326 spin_lock_irq(&gcwq
->lock
);
2327 if (!list_empty(&work
->entry
)) {
2329 * See the comment near try_to_grab_pending()->smp_rmb().
2330 * If it was re-queued to a different gcwq under us, we
2331 * are not going to wait.
2334 cwq
= get_work_cwq(work
);
2335 if (unlikely(!cwq
|| gcwq
!= cwq
->gcwq
))
2338 worker
= find_worker_executing_work(gcwq
, work
);
2341 cwq
= worker
->current_cwq
;
2344 insert_wq_barrier(cwq
, &barr
, work
, worker
);
2345 spin_unlock_irq(&gcwq
->lock
);
2347 lock_map_acquire(&cwq
->wq
->lockdep_map
);
2348 lock_map_release(&cwq
->wq
->lockdep_map
);
2350 wait_for_completion(&barr
.done
);
2351 destroy_work_on_stack(&barr
.work
);
2354 spin_unlock_irq(&gcwq
->lock
);
2357 EXPORT_SYMBOL_GPL(flush_work
);
2360 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2361 * so this work can't be re-armed in any way.
2363 static int try_to_grab_pending(struct work_struct
*work
)
2365 struct global_cwq
*gcwq
;
2368 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
2372 * The queueing is in progress, or it is already queued. Try to
2373 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2375 gcwq
= get_work_gcwq(work
);
2379 spin_lock_irq(&gcwq
->lock
);
2380 if (!list_empty(&work
->entry
)) {
2382 * This work is queued, but perhaps we locked the wrong gcwq.
2383 * In that case we must see the new value after rmb(), see
2384 * insert_work()->wmb().
2387 if (gcwq
== get_work_gcwq(work
)) {
2388 debug_work_deactivate(work
);
2389 list_del_init(&work
->entry
);
2390 cwq_dec_nr_in_flight(get_work_cwq(work
),
2391 get_work_color(work
));
2395 spin_unlock_irq(&gcwq
->lock
);
2400 static void wait_on_cpu_work(struct global_cwq
*gcwq
, struct work_struct
*work
)
2402 struct wq_barrier barr
;
2403 struct worker
*worker
;
2405 spin_lock_irq(&gcwq
->lock
);
2407 worker
= find_worker_executing_work(gcwq
, work
);
2408 if (unlikely(worker
))
2409 insert_wq_barrier(worker
->current_cwq
, &barr
, work
, worker
);
2411 spin_unlock_irq(&gcwq
->lock
);
2413 if (unlikely(worker
)) {
2414 wait_for_completion(&barr
.done
);
2415 destroy_work_on_stack(&barr
.work
);
2419 static void wait_on_work(struct work_struct
*work
)
2425 lock_map_acquire(&work
->lockdep_map
);
2426 lock_map_release(&work
->lockdep_map
);
2428 for_each_gcwq_cpu(cpu
)
2429 wait_on_cpu_work(get_gcwq(cpu
), work
);
2432 static int __cancel_work_timer(struct work_struct
*work
,
2433 struct timer_list
* timer
)
2438 ret
= (timer
&& likely(del_timer(timer
)));
2440 ret
= try_to_grab_pending(work
);
2442 } while (unlikely(ret
< 0));
2444 clear_work_data(work
);
2449 * cancel_work_sync - block until a work_struct's callback has terminated
2450 * @work: the work which is to be flushed
2452 * Returns true if @work was pending.
2454 * cancel_work_sync() will cancel the work if it is queued. If the work's
2455 * callback appears to be running, cancel_work_sync() will block until it
2458 * It is possible to use this function if the work re-queues itself. It can
2459 * cancel the work even if it migrates to another workqueue, however in that
2460 * case it only guarantees that work->func() has completed on the last queued
2463 * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not
2464 * pending, otherwise it goes into a busy-wait loop until the timer expires.
2466 * The caller must ensure that workqueue_struct on which this work was last
2467 * queued can't be destroyed before this function returns.
2469 int cancel_work_sync(struct work_struct
*work
)
2471 return __cancel_work_timer(work
, NULL
);
2473 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2476 * cancel_delayed_work_sync - reliably kill off a delayed work.
2477 * @dwork: the delayed work struct
2479 * Returns true if @dwork was pending.
2481 * It is possible to use this function if @dwork rearms itself via queue_work()
2482 * or queue_delayed_work(). See also the comment for cancel_work_sync().
2484 int cancel_delayed_work_sync(struct delayed_work
*dwork
)
2486 return __cancel_work_timer(&dwork
->work
, &dwork
->timer
);
2488 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2491 * schedule_work - put work task in global workqueue
2492 * @work: job to be done
2494 * Returns zero if @work was already on the kernel-global workqueue and
2495 * non-zero otherwise.
2497 * This puts a job in the kernel-global workqueue if it was not already
2498 * queued and leaves it in the same position on the kernel-global
2499 * workqueue otherwise.
2501 int schedule_work(struct work_struct
*work
)
2503 return queue_work(system_wq
, work
);
2505 EXPORT_SYMBOL(schedule_work
);
2508 * schedule_work_on - put work task on a specific cpu
2509 * @cpu: cpu to put the work task on
2510 * @work: job to be done
2512 * This puts a job on a specific cpu
2514 int schedule_work_on(int cpu
, struct work_struct
*work
)
2516 return queue_work_on(cpu
, system_wq
, work
);
2518 EXPORT_SYMBOL(schedule_work_on
);
2521 * schedule_delayed_work - put work task in global workqueue after delay
2522 * @dwork: job to be done
2523 * @delay: number of jiffies to wait or 0 for immediate execution
2525 * After waiting for a given time this puts a job in the kernel-global
2528 int schedule_delayed_work(struct delayed_work
*dwork
,
2529 unsigned long delay
)
2531 return queue_delayed_work(system_wq
, dwork
, delay
);
2533 EXPORT_SYMBOL(schedule_delayed_work
);
2536 * flush_delayed_work - block until a dwork_struct's callback has terminated
2537 * @dwork: the delayed work which is to be flushed
2539 * Any timeout is cancelled, and any pending work is run immediately.
2541 void flush_delayed_work(struct delayed_work
*dwork
)
2543 if (del_timer_sync(&dwork
->timer
)) {
2544 __queue_work(get_cpu(), get_work_cwq(&dwork
->work
)->wq
,
2548 flush_work(&dwork
->work
);
2550 EXPORT_SYMBOL(flush_delayed_work
);
2553 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2555 * @dwork: job to be done
2556 * @delay: number of jiffies to wait
2558 * After waiting for a given time this puts a job in the kernel-global
2559 * workqueue on the specified CPU.
2561 int schedule_delayed_work_on(int cpu
,
2562 struct delayed_work
*dwork
, unsigned long delay
)
2564 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
2566 EXPORT_SYMBOL(schedule_delayed_work_on
);
2569 * schedule_on_each_cpu - call a function on each online CPU from keventd
2570 * @func: the function to call
2572 * Returns zero on success.
2573 * Returns -ve errno on failure.
2575 * schedule_on_each_cpu() is very slow.
2577 int schedule_on_each_cpu(work_func_t func
)
2580 struct work_struct __percpu
*works
;
2582 works
= alloc_percpu(struct work_struct
);
2588 for_each_online_cpu(cpu
) {
2589 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
2591 INIT_WORK(work
, func
);
2592 schedule_work_on(cpu
, work
);
2595 for_each_online_cpu(cpu
)
2596 flush_work(per_cpu_ptr(works
, cpu
));
2604 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2606 * Forces execution of the kernel-global workqueue and blocks until its
2609 * Think twice before calling this function! It's very easy to get into
2610 * trouble if you don't take great care. Either of the following situations
2611 * will lead to deadlock:
2613 * One of the work items currently on the workqueue needs to acquire
2614 * a lock held by your code or its caller.
2616 * Your code is running in the context of a work routine.
2618 * They will be detected by lockdep when they occur, but the first might not
2619 * occur very often. It depends on what work items are on the workqueue and
2620 * what locks they need, which you have no control over.
2622 * In most situations flushing the entire workqueue is overkill; you merely
2623 * need to know that a particular work item isn't queued and isn't running.
2624 * In such cases you should use cancel_delayed_work_sync() or
2625 * cancel_work_sync() instead.
2627 void flush_scheduled_work(void)
2629 flush_workqueue(system_wq
);
2631 EXPORT_SYMBOL(flush_scheduled_work
);
2634 * execute_in_process_context - reliably execute the routine with user context
2635 * @fn: the function to execute
2636 * @ew: guaranteed storage for the execute work structure (must
2637 * be available when the work executes)
2639 * Executes the function immediately if process context is available,
2640 * otherwise schedules the function for delayed execution.
2642 * Returns: 0 - function was executed
2643 * 1 - function was scheduled for execution
2645 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
2647 if (!in_interrupt()) {
2652 INIT_WORK(&ew
->work
, fn
);
2653 schedule_work(&ew
->work
);
2657 EXPORT_SYMBOL_GPL(execute_in_process_context
);
2659 int keventd_up(void)
2661 return system_wq
!= NULL
;
2664 static int alloc_cwqs(struct workqueue_struct
*wq
)
2667 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2668 * Make sure that the alignment isn't lower than that of
2669 * unsigned long long.
2671 const size_t size
= sizeof(struct cpu_workqueue_struct
);
2672 const size_t align
= max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS
,
2673 __alignof__(unsigned long long));
2675 bool percpu
= !(wq
->flags
& WQ_UNBOUND
);
2677 bool percpu
= false;
2681 wq
->cpu_wq
.pcpu
= __alloc_percpu(size
, align
);
2686 * Allocate enough room to align cwq and put an extra
2687 * pointer at the end pointing back to the originally
2688 * allocated pointer which will be used for free.
2690 ptr
= kzalloc(size
+ align
+ sizeof(void *), GFP_KERNEL
);
2692 wq
->cpu_wq
.single
= PTR_ALIGN(ptr
, align
);
2693 *(void **)(wq
->cpu_wq
.single
+ 1) = ptr
;
2697 /* just in case, make sure it's actually aligned */
2698 BUG_ON(!IS_ALIGNED(wq
->cpu_wq
.v
, align
));
2699 return wq
->cpu_wq
.v
? 0 : -ENOMEM
;
2702 static void free_cwqs(struct workqueue_struct
*wq
)
2705 bool percpu
= !(wq
->flags
& WQ_UNBOUND
);
2707 bool percpu
= false;
2711 free_percpu(wq
->cpu_wq
.pcpu
);
2712 else if (wq
->cpu_wq
.single
) {
2713 /* the pointer to free is stored right after the cwq */
2714 kfree(*(void **)(wq
->cpu_wq
.single
+ 1));
2718 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
2721 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
2723 if (max_active
< 1 || max_active
> lim
)
2724 printk(KERN_WARNING
"workqueue: max_active %d requested for %s "
2725 "is out of range, clamping between %d and %d\n",
2726 max_active
, name
, 1, lim
);
2728 return clamp_val(max_active
, 1, lim
);
2731 struct workqueue_struct
*__alloc_workqueue_key(const char *name
,
2734 struct lock_class_key
*key
,
2735 const char *lock_name
)
2737 struct workqueue_struct
*wq
;
2741 * Unbound workqueues aren't concurrency managed and should be
2742 * dispatched to workers immediately.
2744 if (flags
& WQ_UNBOUND
)
2745 flags
|= WQ_HIGHPRI
;
2747 max_active
= max_active
?: WQ_DFL_ACTIVE
;
2748 max_active
= wq_clamp_max_active(max_active
, flags
, name
);
2750 wq
= kzalloc(sizeof(*wq
), GFP_KERNEL
);
2755 wq
->saved_max_active
= max_active
;
2756 mutex_init(&wq
->flush_mutex
);
2757 atomic_set(&wq
->nr_cwqs_to_flush
, 0);
2758 INIT_LIST_HEAD(&wq
->flusher_queue
);
2759 INIT_LIST_HEAD(&wq
->flusher_overflow
);
2762 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
2763 INIT_LIST_HEAD(&wq
->list
);
2765 if (alloc_cwqs(wq
) < 0)
2768 for_each_cwq_cpu(cpu
, wq
) {
2769 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2770 struct global_cwq
*gcwq
= get_gcwq(cpu
);
2772 BUG_ON((unsigned long)cwq
& WORK_STRUCT_FLAG_MASK
);
2775 cwq
->flush_color
= -1;
2776 cwq
->max_active
= max_active
;
2777 INIT_LIST_HEAD(&cwq
->delayed_works
);
2780 if (flags
& WQ_RESCUER
) {
2781 struct worker
*rescuer
;
2783 if (!alloc_mayday_mask(&wq
->mayday_mask
, GFP_KERNEL
))
2786 wq
->rescuer
= rescuer
= alloc_worker();
2790 rescuer
->task
= kthread_create(rescuer_thread
, wq
, "%s", name
);
2791 if (IS_ERR(rescuer
->task
))
2794 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
2795 wake_up_process(rescuer
->task
);
2799 * workqueue_lock protects global freeze state and workqueues
2800 * list. Grab it, set max_active accordingly and add the new
2801 * workqueue to workqueues list.
2803 spin_lock(&workqueue_lock
);
2805 if (workqueue_freezing
&& wq
->flags
& WQ_FREEZEABLE
)
2806 for_each_cwq_cpu(cpu
, wq
)
2807 get_cwq(cpu
, wq
)->max_active
= 0;
2809 list_add(&wq
->list
, &workqueues
);
2811 spin_unlock(&workqueue_lock
);
2817 free_mayday_mask(wq
->mayday_mask
);
2823 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
2826 * destroy_workqueue - safely terminate a workqueue
2827 * @wq: target workqueue
2829 * Safely destroy a workqueue. All work currently pending will be done first.
2831 void destroy_workqueue(struct workqueue_struct
*wq
)
2835 wq
->flags
|= WQ_DYING
;
2836 flush_workqueue(wq
);
2839 * wq list is used to freeze wq, remove from list after
2840 * flushing is complete in case freeze races us.
2842 spin_lock(&workqueue_lock
);
2843 list_del(&wq
->list
);
2844 spin_unlock(&workqueue_lock
);
2847 for_each_cwq_cpu(cpu
, wq
) {
2848 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2851 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
2852 BUG_ON(cwq
->nr_in_flight
[i
]);
2853 BUG_ON(cwq
->nr_active
);
2854 BUG_ON(!list_empty(&cwq
->delayed_works
));
2857 if (wq
->flags
& WQ_RESCUER
) {
2858 kthread_stop(wq
->rescuer
->task
);
2859 free_mayday_mask(wq
->mayday_mask
);
2866 EXPORT_SYMBOL_GPL(destroy_workqueue
);
2869 * workqueue_set_max_active - adjust max_active of a workqueue
2870 * @wq: target workqueue
2871 * @max_active: new max_active value.
2873 * Set max_active of @wq to @max_active.
2876 * Don't call from IRQ context.
2878 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
2882 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
2884 spin_lock(&workqueue_lock
);
2886 wq
->saved_max_active
= max_active
;
2888 for_each_cwq_cpu(cpu
, wq
) {
2889 struct global_cwq
*gcwq
= get_gcwq(cpu
);
2891 spin_lock_irq(&gcwq
->lock
);
2893 if (!(wq
->flags
& WQ_FREEZEABLE
) ||
2894 !(gcwq
->flags
& GCWQ_FREEZING
))
2895 get_cwq(gcwq
->cpu
, wq
)->max_active
= max_active
;
2897 spin_unlock_irq(&gcwq
->lock
);
2900 spin_unlock(&workqueue_lock
);
2902 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
2905 * workqueue_congested - test whether a workqueue is congested
2906 * @cpu: CPU in question
2907 * @wq: target workqueue
2909 * Test whether @wq's cpu workqueue for @cpu is congested. There is
2910 * no synchronization around this function and the test result is
2911 * unreliable and only useful as advisory hints or for debugging.
2914 * %true if congested, %false otherwise.
2916 bool workqueue_congested(unsigned int cpu
, struct workqueue_struct
*wq
)
2918 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2920 return !list_empty(&cwq
->delayed_works
);
2922 EXPORT_SYMBOL_GPL(workqueue_congested
);
2925 * work_cpu - return the last known associated cpu for @work
2926 * @work: the work of interest
2929 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
2931 unsigned int work_cpu(struct work_struct
*work
)
2933 struct global_cwq
*gcwq
= get_work_gcwq(work
);
2935 return gcwq
? gcwq
->cpu
: WORK_CPU_NONE
;
2937 EXPORT_SYMBOL_GPL(work_cpu
);
2940 * work_busy - test whether a work is currently pending or running
2941 * @work: the work to be tested
2943 * Test whether @work is currently pending or running. There is no
2944 * synchronization around this function and the test result is
2945 * unreliable and only useful as advisory hints or for debugging.
2946 * Especially for reentrant wqs, the pending state might hide the
2950 * OR'd bitmask of WORK_BUSY_* bits.
2952 unsigned int work_busy(struct work_struct
*work
)
2954 struct global_cwq
*gcwq
= get_work_gcwq(work
);
2955 unsigned long flags
;
2956 unsigned int ret
= 0;
2961 spin_lock_irqsave(&gcwq
->lock
, flags
);
2963 if (work_pending(work
))
2964 ret
|= WORK_BUSY_PENDING
;
2965 if (find_worker_executing_work(gcwq
, work
))
2966 ret
|= WORK_BUSY_RUNNING
;
2968 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
2972 EXPORT_SYMBOL_GPL(work_busy
);
2977 * There are two challenges in supporting CPU hotplug. Firstly, there
2978 * are a lot of assumptions on strong associations among work, cwq and
2979 * gcwq which make migrating pending and scheduled works very
2980 * difficult to implement without impacting hot paths. Secondly,
2981 * gcwqs serve mix of short, long and very long running works making
2982 * blocked draining impractical.
2984 * This is solved by allowing a gcwq to be detached from CPU, running
2985 * it with unbound (rogue) workers and allowing it to be reattached
2986 * later if the cpu comes back online. A separate thread is created
2987 * to govern a gcwq in such state and is called the trustee of the
2990 * Trustee states and their descriptions.
2992 * START Command state used on startup. On CPU_DOWN_PREPARE, a
2993 * new trustee is started with this state.
2995 * IN_CHARGE Once started, trustee will enter this state after
2996 * assuming the manager role and making all existing
2997 * workers rogue. DOWN_PREPARE waits for trustee to
2998 * enter this state. After reaching IN_CHARGE, trustee
2999 * tries to execute the pending worklist until it's empty
3000 * and the state is set to BUTCHER, or the state is set
3003 * BUTCHER Command state which is set by the cpu callback after
3004 * the cpu has went down. Once this state is set trustee
3005 * knows that there will be no new works on the worklist
3006 * and once the worklist is empty it can proceed to
3007 * killing idle workers.
3009 * RELEASE Command state which is set by the cpu callback if the
3010 * cpu down has been canceled or it has come online
3011 * again. After recognizing this state, trustee stops
3012 * trying to drain or butcher and clears ROGUE, rebinds
3013 * all remaining workers back to the cpu and releases
3016 * DONE Trustee will enter this state after BUTCHER or RELEASE
3019 * trustee CPU draining
3020 * took over down complete
3021 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3023 * | CPU is back online v return workers |
3024 * ----------------> RELEASE --------------
3028 * trustee_wait_event_timeout - timed event wait for trustee
3029 * @cond: condition to wait for
3030 * @timeout: timeout in jiffies
3032 * wait_event_timeout() for trustee to use. Handles locking and
3033 * checks for RELEASE request.
3036 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3037 * multiple times. To be used by trustee.
3040 * Positive indicating left time if @cond is satisfied, 0 if timed
3041 * out, -1 if canceled.
3043 #define trustee_wait_event_timeout(cond, timeout) ({ \
3044 long __ret = (timeout); \
3045 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3047 spin_unlock_irq(&gcwq->lock); \
3048 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3049 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3051 spin_lock_irq(&gcwq->lock); \
3053 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3057 * trustee_wait_event - event wait for trustee
3058 * @cond: condition to wait for
3060 * wait_event() for trustee to use. Automatically handles locking and
3061 * checks for CANCEL request.
3064 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3065 * multiple times. To be used by trustee.
3068 * 0 if @cond is satisfied, -1 if canceled.
3070 #define trustee_wait_event(cond) ({ \
3072 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3073 __ret1 < 0 ? -1 : 0; \
3076 static int __cpuinit
trustee_thread(void *__gcwq
)
3078 struct global_cwq
*gcwq
= __gcwq
;
3079 struct worker
*worker
;
3080 struct work_struct
*work
;
3081 struct hlist_node
*pos
;
3085 BUG_ON(gcwq
->cpu
!= smp_processor_id());
3087 spin_lock_irq(&gcwq
->lock
);
3089 * Claim the manager position and make all workers rogue.
3090 * Trustee must be bound to the target cpu and can't be
3093 BUG_ON(gcwq
->cpu
!= smp_processor_id());
3094 rc
= trustee_wait_event(!(gcwq
->flags
& GCWQ_MANAGING_WORKERS
));
3097 gcwq
->flags
|= GCWQ_MANAGING_WORKERS
;
3099 list_for_each_entry(worker
, &gcwq
->idle_list
, entry
)
3100 worker
->flags
|= WORKER_ROGUE
;
3102 for_each_busy_worker(worker
, i
, pos
, gcwq
)
3103 worker
->flags
|= WORKER_ROGUE
;
3106 * Call schedule() so that we cross rq->lock and thus can
3107 * guarantee sched callbacks see the rogue flag. This is
3108 * necessary as scheduler callbacks may be invoked from other
3111 spin_unlock_irq(&gcwq
->lock
);
3113 spin_lock_irq(&gcwq
->lock
);
3116 * Sched callbacks are disabled now. Zap nr_running. After
3117 * this, nr_running stays zero and need_more_worker() and
3118 * keep_working() are always true as long as the worklist is
3121 atomic_set(get_gcwq_nr_running(gcwq
->cpu
), 0);
3123 spin_unlock_irq(&gcwq
->lock
);
3124 del_timer_sync(&gcwq
->idle_timer
);
3125 spin_lock_irq(&gcwq
->lock
);
3128 * We're now in charge. Notify and proceed to drain. We need
3129 * to keep the gcwq running during the whole CPU down
3130 * procedure as other cpu hotunplug callbacks may need to
3131 * flush currently running tasks.
3133 gcwq
->trustee_state
= TRUSTEE_IN_CHARGE
;
3134 wake_up_all(&gcwq
->trustee_wait
);
3137 * The original cpu is in the process of dying and may go away
3138 * anytime now. When that happens, we and all workers would
3139 * be migrated to other cpus. Try draining any left work. We
3140 * want to get it over with ASAP - spam rescuers, wake up as
3141 * many idlers as necessary and create new ones till the
3142 * worklist is empty. Note that if the gcwq is frozen, there
3143 * may be frozen works in freezeable cwqs. Don't declare
3144 * completion while frozen.
3146 while (gcwq
->nr_workers
!= gcwq
->nr_idle
||
3147 gcwq
->flags
& GCWQ_FREEZING
||
3148 gcwq
->trustee_state
== TRUSTEE_IN_CHARGE
) {
3151 list_for_each_entry(work
, &gcwq
->worklist
, entry
) {
3156 list_for_each_entry(worker
, &gcwq
->idle_list
, entry
) {
3159 wake_up_process(worker
->task
);
3162 if (need_to_create_worker(gcwq
)) {
3163 spin_unlock_irq(&gcwq
->lock
);
3164 worker
= create_worker(gcwq
, false);
3165 spin_lock_irq(&gcwq
->lock
);
3167 worker
->flags
|= WORKER_ROGUE
;
3168 start_worker(worker
);
3172 /* give a breather */
3173 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN
) < 0)
3178 * Either all works have been scheduled and cpu is down, or
3179 * cpu down has already been canceled. Wait for and butcher
3180 * all workers till we're canceled.
3183 rc
= trustee_wait_event(!list_empty(&gcwq
->idle_list
));
3184 while (!list_empty(&gcwq
->idle_list
))
3185 destroy_worker(list_first_entry(&gcwq
->idle_list
,
3186 struct worker
, entry
));
3187 } while (gcwq
->nr_workers
&& rc
>= 0);
3190 * At this point, either draining has completed and no worker
3191 * is left, or cpu down has been canceled or the cpu is being
3192 * brought back up. There shouldn't be any idle one left.
3193 * Tell the remaining busy ones to rebind once it finishes the
3194 * currently scheduled works by scheduling the rebind_work.
3196 WARN_ON(!list_empty(&gcwq
->idle_list
));
3198 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
3199 struct work_struct
*rebind_work
= &worker
->rebind_work
;
3202 * Rebind_work may race with future cpu hotplug
3203 * operations. Use a separate flag to mark that
3204 * rebinding is scheduled.
3206 worker
->flags
|= WORKER_REBIND
;
3207 worker
->flags
&= ~WORKER_ROGUE
;
3209 /* queue rebind_work, wq doesn't matter, use the default one */
3210 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
3211 work_data_bits(rebind_work
)))
3214 debug_work_activate(rebind_work
);
3215 insert_work(get_cwq(gcwq
->cpu
, system_wq
), rebind_work
,
3216 worker
->scheduled
.next
,
3217 work_color_to_flags(WORK_NO_COLOR
));
3220 /* relinquish manager role */
3221 gcwq
->flags
&= ~GCWQ_MANAGING_WORKERS
;
3223 /* notify completion */
3224 gcwq
->trustee
= NULL
;
3225 gcwq
->trustee_state
= TRUSTEE_DONE
;
3226 wake_up_all(&gcwq
->trustee_wait
);
3227 spin_unlock_irq(&gcwq
->lock
);
3232 * wait_trustee_state - wait for trustee to enter the specified state
3233 * @gcwq: gcwq the trustee of interest belongs to
3234 * @state: target state to wait for
3236 * Wait for the trustee to reach @state. DONE is already matched.
3239 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3240 * multiple times. To be used by cpu_callback.
3242 static void __cpuinit
wait_trustee_state(struct global_cwq
*gcwq
, int state
)
3243 __releases(&gcwq
->lock
)
3244 __acquires(&gcwq
->lock
)
3246 if (!(gcwq
->trustee_state
== state
||
3247 gcwq
->trustee_state
== TRUSTEE_DONE
)) {
3248 spin_unlock_irq(&gcwq
->lock
);
3249 __wait_event(gcwq
->trustee_wait
,
3250 gcwq
->trustee_state
== state
||
3251 gcwq
->trustee_state
== TRUSTEE_DONE
);
3252 spin_lock_irq(&gcwq
->lock
);
3256 static int __devinit
workqueue_cpu_callback(struct notifier_block
*nfb
,
3257 unsigned long action
,
3260 unsigned int cpu
= (unsigned long)hcpu
;
3261 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3262 struct task_struct
*new_trustee
= NULL
;
3263 struct worker
*uninitialized_var(new_worker
);
3264 unsigned long flags
;
3266 action
&= ~CPU_TASKS_FROZEN
;
3269 case CPU_DOWN_PREPARE
:
3270 new_trustee
= kthread_create(trustee_thread
, gcwq
,
3271 "workqueue_trustee/%d\n", cpu
);
3272 if (IS_ERR(new_trustee
))
3273 return notifier_from_errno(PTR_ERR(new_trustee
));
3274 kthread_bind(new_trustee
, cpu
);
3276 case CPU_UP_PREPARE
:
3277 BUG_ON(gcwq
->first_idle
);
3278 new_worker
= create_worker(gcwq
, false);
3281 kthread_stop(new_trustee
);
3286 /* some are called w/ irq disabled, don't disturb irq status */
3287 spin_lock_irqsave(&gcwq
->lock
, flags
);
3290 case CPU_DOWN_PREPARE
:
3291 /* initialize trustee and tell it to acquire the gcwq */
3292 BUG_ON(gcwq
->trustee
|| gcwq
->trustee_state
!= TRUSTEE_DONE
);
3293 gcwq
->trustee
= new_trustee
;
3294 gcwq
->trustee_state
= TRUSTEE_START
;
3295 wake_up_process(gcwq
->trustee
);
3296 wait_trustee_state(gcwq
, TRUSTEE_IN_CHARGE
);
3298 case CPU_UP_PREPARE
:
3299 BUG_ON(gcwq
->first_idle
);
3300 gcwq
->first_idle
= new_worker
;
3305 * Before this, the trustee and all workers except for
3306 * the ones which are still executing works from
3307 * before the last CPU down must be on the cpu. After
3308 * this, they'll all be diasporas.
3310 gcwq
->flags
|= GCWQ_DISASSOCIATED
;
3314 gcwq
->trustee_state
= TRUSTEE_BUTCHER
;
3316 case CPU_UP_CANCELED
:
3317 destroy_worker(gcwq
->first_idle
);
3318 gcwq
->first_idle
= NULL
;
3321 case CPU_DOWN_FAILED
:
3323 gcwq
->flags
&= ~GCWQ_DISASSOCIATED
;
3324 if (gcwq
->trustee_state
!= TRUSTEE_DONE
) {
3325 gcwq
->trustee_state
= TRUSTEE_RELEASE
;
3326 wake_up_process(gcwq
->trustee
);
3327 wait_trustee_state(gcwq
, TRUSTEE_DONE
);
3331 * Trustee is done and there might be no worker left.
3332 * Put the first_idle in and request a real manager to
3335 spin_unlock_irq(&gcwq
->lock
);
3336 kthread_bind(gcwq
->first_idle
->task
, cpu
);
3337 spin_lock_irq(&gcwq
->lock
);
3338 gcwq
->flags
|= GCWQ_MANAGE_WORKERS
;
3339 start_worker(gcwq
->first_idle
);
3340 gcwq
->first_idle
= NULL
;
3344 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
3346 return notifier_from_errno(0);
3351 struct work_for_cpu
{
3352 struct completion completion
;
3358 static int do_work_for_cpu(void *_wfc
)
3360 struct work_for_cpu
*wfc
= _wfc
;
3361 wfc
->ret
= wfc
->fn(wfc
->arg
);
3362 complete(&wfc
->completion
);
3367 * work_on_cpu - run a function in user context on a particular cpu
3368 * @cpu: the cpu to run on
3369 * @fn: the function to run
3370 * @arg: the function arg
3372 * This will return the value @fn returns.
3373 * It is up to the caller to ensure that the cpu doesn't go offline.
3374 * The caller must not hold any locks which would prevent @fn from completing.
3376 long work_on_cpu(unsigned int cpu
, long (*fn
)(void *), void *arg
)
3378 struct task_struct
*sub_thread
;
3379 struct work_for_cpu wfc
= {
3380 .completion
= COMPLETION_INITIALIZER_ONSTACK(wfc
.completion
),
3385 sub_thread
= kthread_create(do_work_for_cpu
, &wfc
, "work_for_cpu");
3386 if (IS_ERR(sub_thread
))
3387 return PTR_ERR(sub_thread
);
3388 kthread_bind(sub_thread
, cpu
);
3389 wake_up_process(sub_thread
);
3390 wait_for_completion(&wfc
.completion
);
3393 EXPORT_SYMBOL_GPL(work_on_cpu
);
3394 #endif /* CONFIG_SMP */
3396 #ifdef CONFIG_FREEZER
3399 * freeze_workqueues_begin - begin freezing workqueues
3401 * Start freezing workqueues. After this function returns, all
3402 * freezeable workqueues will queue new works to their frozen_works
3403 * list instead of gcwq->worklist.
3406 * Grabs and releases workqueue_lock and gcwq->lock's.
3408 void freeze_workqueues_begin(void)
3412 spin_lock(&workqueue_lock
);
3414 BUG_ON(workqueue_freezing
);
3415 workqueue_freezing
= true;
3417 for_each_gcwq_cpu(cpu
) {
3418 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3419 struct workqueue_struct
*wq
;
3421 spin_lock_irq(&gcwq
->lock
);
3423 BUG_ON(gcwq
->flags
& GCWQ_FREEZING
);
3424 gcwq
->flags
|= GCWQ_FREEZING
;
3426 list_for_each_entry(wq
, &workqueues
, list
) {
3427 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3429 if (cwq
&& wq
->flags
& WQ_FREEZEABLE
)
3430 cwq
->max_active
= 0;
3433 spin_unlock_irq(&gcwq
->lock
);
3436 spin_unlock(&workqueue_lock
);
3440 * freeze_workqueues_busy - are freezeable workqueues still busy?
3442 * Check whether freezing is complete. This function must be called
3443 * between freeze_workqueues_begin() and thaw_workqueues().
3446 * Grabs and releases workqueue_lock.
3449 * %true if some freezeable workqueues are still busy. %false if
3450 * freezing is complete.
3452 bool freeze_workqueues_busy(void)
3457 spin_lock(&workqueue_lock
);
3459 BUG_ON(!workqueue_freezing
);
3461 for_each_gcwq_cpu(cpu
) {
3462 struct workqueue_struct
*wq
;
3464 * nr_active is monotonically decreasing. It's safe
3465 * to peek without lock.
3467 list_for_each_entry(wq
, &workqueues
, list
) {
3468 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3470 if (!cwq
|| !(wq
->flags
& WQ_FREEZEABLE
))
3473 BUG_ON(cwq
->nr_active
< 0);
3474 if (cwq
->nr_active
) {
3481 spin_unlock(&workqueue_lock
);
3486 * thaw_workqueues - thaw workqueues
3488 * Thaw workqueues. Normal queueing is restored and all collected
3489 * frozen works are transferred to their respective gcwq worklists.
3492 * Grabs and releases workqueue_lock and gcwq->lock's.
3494 void thaw_workqueues(void)
3498 spin_lock(&workqueue_lock
);
3500 if (!workqueue_freezing
)
3503 for_each_gcwq_cpu(cpu
) {
3504 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3505 struct workqueue_struct
*wq
;
3507 spin_lock_irq(&gcwq
->lock
);
3509 BUG_ON(!(gcwq
->flags
& GCWQ_FREEZING
));
3510 gcwq
->flags
&= ~GCWQ_FREEZING
;
3512 list_for_each_entry(wq
, &workqueues
, list
) {
3513 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3515 if (!cwq
|| !(wq
->flags
& WQ_FREEZEABLE
))
3518 /* restore max_active and repopulate worklist */
3519 cwq
->max_active
= wq
->saved_max_active
;
3521 while (!list_empty(&cwq
->delayed_works
) &&
3522 cwq
->nr_active
< cwq
->max_active
)
3523 cwq_activate_first_delayed(cwq
);
3526 wake_up_worker(gcwq
);
3528 spin_unlock_irq(&gcwq
->lock
);
3531 workqueue_freezing
= false;
3533 spin_unlock(&workqueue_lock
);
3535 #endif /* CONFIG_FREEZER */
3537 static int __init
init_workqueues(void)
3542 cpu_notifier(workqueue_cpu_callback
, CPU_PRI_WORKQUEUE
);
3544 /* initialize gcwqs */
3545 for_each_gcwq_cpu(cpu
) {
3546 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3548 spin_lock_init(&gcwq
->lock
);
3549 INIT_LIST_HEAD(&gcwq
->worklist
);
3551 if (cpu
== WORK_CPU_UNBOUND
)
3552 gcwq
->flags
|= GCWQ_DISASSOCIATED
;
3554 INIT_LIST_HEAD(&gcwq
->idle_list
);
3555 for (i
= 0; i
< BUSY_WORKER_HASH_SIZE
; i
++)
3556 INIT_HLIST_HEAD(&gcwq
->busy_hash
[i
]);
3558 init_timer_deferrable(&gcwq
->idle_timer
);
3559 gcwq
->idle_timer
.function
= idle_worker_timeout
;
3560 gcwq
->idle_timer
.data
= (unsigned long)gcwq
;
3562 setup_timer(&gcwq
->mayday_timer
, gcwq_mayday_timeout
,
3563 (unsigned long)gcwq
);
3565 ida_init(&gcwq
->worker_ida
);
3567 gcwq
->trustee_state
= TRUSTEE_DONE
;
3568 init_waitqueue_head(&gcwq
->trustee_wait
);
3571 /* create the initial worker */
3572 for_each_online_gcwq_cpu(cpu
) {
3573 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3574 struct worker
*worker
;
3576 worker
= create_worker(gcwq
, true);
3578 spin_lock_irq(&gcwq
->lock
);
3579 start_worker(worker
);
3580 spin_unlock_irq(&gcwq
->lock
);
3583 system_wq
= alloc_workqueue("events", 0, 0);
3584 system_long_wq
= alloc_workqueue("events_long", 0, 0);
3585 system_nrt_wq
= alloc_workqueue("events_nrt", WQ_NON_REENTRANT
, 0);
3586 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
3587 WQ_UNBOUND_MAX_ACTIVE
);
3588 BUG_ON(!system_wq
|| !system_long_wq
|| !system_nrt_wq
);
3591 early_initcall(init_workqueues
);