2 * kernel/workqueue.c - generic async execution with shared worker pool
4 * Copyright (C) 2002 Ingo Molnar
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
12 * Made to use alloc_percpu by Christoph Lameter.
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
23 * Please read Documentation/workqueue.txt for details.
26 #include <linux/module.h>
27 #include <linux/kernel.h>
28 #include <linux/sched.h>
29 #include <linux/init.h>
30 #include <linux/signal.h>
31 #include <linux/completion.h>
32 #include <linux/workqueue.h>
33 #include <linux/slab.h>
34 #include <linux/cpu.h>
35 #include <linux/notifier.h>
36 #include <linux/kthread.h>
37 #include <linux/hardirq.h>
38 #include <linux/mempolicy.h>
39 #include <linux/freezer.h>
40 #include <linux/kallsyms.h>
41 #include <linux/debug_locks.h>
42 #include <linux/lockdep.h>
43 #include <linux/idr.h>
45 #include "workqueue_sched.h"
48 /* global_cwq flags */
49 GCWQ_MANAGE_WORKERS
= 1 << 0, /* need to manage workers */
50 GCWQ_MANAGING_WORKERS
= 1 << 1, /* managing workers */
51 GCWQ_DISASSOCIATED
= 1 << 2, /* cpu can't serve workers */
52 GCWQ_FREEZING
= 1 << 3, /* freeze in progress */
53 GCWQ_HIGHPRI_PENDING
= 1 << 4, /* highpri works on queue */
56 WORKER_STARTED
= 1 << 0, /* started */
57 WORKER_DIE
= 1 << 1, /* die die die */
58 WORKER_IDLE
= 1 << 2, /* is idle */
59 WORKER_PREP
= 1 << 3, /* preparing to run works */
60 WORKER_ROGUE
= 1 << 4, /* not bound to any cpu */
61 WORKER_REBIND
= 1 << 5, /* mom is home, come back */
62 WORKER_CPU_INTENSIVE
= 1 << 6, /* cpu intensive */
63 WORKER_UNBOUND
= 1 << 7, /* worker is unbound */
65 WORKER_NOT_RUNNING
= WORKER_PREP
| WORKER_ROGUE
| WORKER_REBIND
|
66 WORKER_CPU_INTENSIVE
| WORKER_UNBOUND
,
68 /* gcwq->trustee_state */
69 TRUSTEE_START
= 0, /* start */
70 TRUSTEE_IN_CHARGE
= 1, /* trustee in charge of gcwq */
71 TRUSTEE_BUTCHER
= 2, /* butcher workers */
72 TRUSTEE_RELEASE
= 3, /* release workers */
73 TRUSTEE_DONE
= 4, /* trustee is done */
75 BUSY_WORKER_HASH_ORDER
= 6, /* 64 pointers */
76 BUSY_WORKER_HASH_SIZE
= 1 << BUSY_WORKER_HASH_ORDER
,
77 BUSY_WORKER_HASH_MASK
= BUSY_WORKER_HASH_SIZE
- 1,
79 MAX_IDLE_WORKERS_RATIO
= 4, /* 1/4 of busy can be idle */
80 IDLE_WORKER_TIMEOUT
= 300 * HZ
, /* keep idle ones for 5 mins */
82 MAYDAY_INITIAL_TIMEOUT
= HZ
/ 100, /* call for help after 10ms */
83 MAYDAY_INTERVAL
= HZ
/ 10, /* and then every 100ms */
84 CREATE_COOLDOWN
= HZ
, /* time to breath after fail */
85 TRUSTEE_COOLDOWN
= HZ
/ 10, /* for trustee draining */
88 * Rescue workers are used only on emergencies and shared by
91 RESCUER_NICE_LEVEL
= -20,
95 * Structure fields follow one of the following exclusion rules.
97 * I: Modifiable by initialization/destruction paths and read-only for
100 * P: Preemption protected. Disabling preemption is enough and should
101 * only be modified and accessed from the local cpu.
103 * L: gcwq->lock protected. Access with gcwq->lock held.
105 * X: During normal operation, modification requires gcwq->lock and
106 * should be done only from local cpu. Either disabling preemption
107 * on local cpu or grabbing gcwq->lock is enough for read access.
108 * If GCWQ_DISASSOCIATED is set, it's identical to L.
110 * F: wq->flush_mutex protected.
112 * W: workqueue_lock protected.
118 * The poor guys doing the actual heavy lifting. All on-duty workers
119 * are either serving the manager role, on idle list or on busy hash.
122 /* on idle list while idle, on busy hash table while busy */
124 struct list_head entry
; /* L: while idle */
125 struct hlist_node hentry
; /* L: while busy */
128 struct work_struct
*current_work
; /* L: work being processed */
129 struct cpu_workqueue_struct
*current_cwq
; /* L: current_work's cwq */
130 struct list_head scheduled
; /* L: scheduled works */
131 struct task_struct
*task
; /* I: worker task */
132 struct global_cwq
*gcwq
; /* I: the associated gcwq */
133 /* 64 bytes boundary on 64bit, 32 on 32bit */
134 unsigned long last_active
; /* L: last active timestamp */
135 unsigned int flags
; /* X: flags */
136 int id
; /* I: worker id */
137 struct work_struct rebind_work
; /* L: rebind worker to cpu */
141 * Global per-cpu workqueue. There's one and only one for each cpu
142 * and all works are queued and processed here regardless of their
146 spinlock_t lock
; /* the gcwq lock */
147 struct list_head worklist
; /* L: list of pending works */
148 unsigned int cpu
; /* I: the associated cpu */
149 unsigned int flags
; /* L: GCWQ_* flags */
151 int nr_workers
; /* L: total number of workers */
152 int nr_idle
; /* L: currently idle ones */
154 /* workers are chained either in the idle_list or busy_hash */
155 struct list_head idle_list
; /* X: list of idle workers */
156 struct hlist_head busy_hash
[BUSY_WORKER_HASH_SIZE
];
157 /* L: hash of busy workers */
159 struct timer_list idle_timer
; /* L: worker idle timeout */
160 struct timer_list mayday_timer
; /* L: SOS timer for dworkers */
162 struct ida worker_ida
; /* L: for worker IDs */
164 struct task_struct
*trustee
; /* L: for gcwq shutdown */
165 unsigned int trustee_state
; /* L: trustee state */
166 wait_queue_head_t trustee_wait
; /* trustee wait */
167 struct worker
*first_idle
; /* L: first idle worker */
168 } ____cacheline_aligned_in_smp
;
171 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
172 * work_struct->data are used for flags and thus cwqs need to be
173 * aligned at two's power of the number of flag bits.
175 struct cpu_workqueue_struct
{
176 struct global_cwq
*gcwq
; /* I: the associated gcwq */
177 struct workqueue_struct
*wq
; /* I: the owning workqueue */
178 int work_color
; /* L: current color */
179 int flush_color
; /* L: flushing color */
180 int nr_in_flight
[WORK_NR_COLORS
];
181 /* L: nr of in_flight works */
182 int nr_active
; /* L: nr of active works */
183 int max_active
; /* L: max active works */
184 struct list_head delayed_works
; /* L: delayed works */
188 * Structure used to wait for workqueue flush.
191 struct list_head list
; /* F: list of flushers */
192 int flush_color
; /* F: flush color waiting for */
193 struct completion done
; /* flush completion */
197 * All cpumasks are assumed to be always set on UP and thus can't be
198 * used to determine whether there's something to be done.
201 typedef cpumask_var_t mayday_mask_t
;
202 #define mayday_test_and_set_cpu(cpu, mask) \
203 cpumask_test_and_set_cpu((cpu), (mask))
204 #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
205 #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
206 #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
207 #define free_mayday_mask(mask) free_cpumask_var((mask))
209 typedef unsigned long mayday_mask_t
;
210 #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
211 #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
212 #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
213 #define alloc_mayday_mask(maskp, gfp) true
214 #define free_mayday_mask(mask) do { } while (0)
218 * The externally visible workqueue abstraction is an array of
219 * per-CPU workqueues:
221 struct workqueue_struct
{
222 unsigned int flags
; /* I: WQ_* flags */
224 struct cpu_workqueue_struct __percpu
*pcpu
;
225 struct cpu_workqueue_struct
*single
;
227 } cpu_wq
; /* I: cwq's */
228 struct list_head list
; /* W: list of all workqueues */
230 struct mutex flush_mutex
; /* protects wq flushing */
231 int work_color
; /* F: current work color */
232 int flush_color
; /* F: current flush color */
233 atomic_t nr_cwqs_to_flush
; /* flush in progress */
234 struct wq_flusher
*first_flusher
; /* F: first flusher */
235 struct list_head flusher_queue
; /* F: flush waiters */
236 struct list_head flusher_overflow
; /* F: flush overflow list */
238 mayday_mask_t mayday_mask
; /* cpus requesting rescue */
239 struct worker
*rescuer
; /* I: rescue worker */
241 int saved_max_active
; /* W: saved cwq max_active */
242 const char *name
; /* I: workqueue name */
243 #ifdef CONFIG_LOCKDEP
244 struct lockdep_map lockdep_map
;
248 struct workqueue_struct
*system_wq __read_mostly
;
249 struct workqueue_struct
*system_long_wq __read_mostly
;
250 struct workqueue_struct
*system_nrt_wq __read_mostly
;
251 struct workqueue_struct
*system_unbound_wq __read_mostly
;
252 EXPORT_SYMBOL_GPL(system_wq
);
253 EXPORT_SYMBOL_GPL(system_long_wq
);
254 EXPORT_SYMBOL_GPL(system_nrt_wq
);
255 EXPORT_SYMBOL_GPL(system_unbound_wq
);
257 #define CREATE_TRACE_POINTS
258 #include <trace/events/workqueue.h>
260 #define for_each_busy_worker(worker, i, pos, gcwq) \
261 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
262 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
264 static inline int __next_gcwq_cpu(int cpu
, const struct cpumask
*mask
,
267 if (cpu
< nr_cpu_ids
) {
269 cpu
= cpumask_next(cpu
, mask
);
270 if (cpu
< nr_cpu_ids
)
274 return WORK_CPU_UNBOUND
;
276 return WORK_CPU_NONE
;
279 static inline int __next_wq_cpu(int cpu
, const struct cpumask
*mask
,
280 struct workqueue_struct
*wq
)
282 return __next_gcwq_cpu(cpu
, mask
, !(wq
->flags
& WQ_UNBOUND
) ? 1 : 2);
288 * An extra gcwq is defined for an invalid cpu number
289 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
290 * specific CPU. The following iterators are similar to
291 * for_each_*_cpu() iterators but also considers the unbound gcwq.
293 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
294 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
295 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
296 * WORK_CPU_UNBOUND for unbound workqueues
298 #define for_each_gcwq_cpu(cpu) \
299 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
300 (cpu) < WORK_CPU_NONE; \
301 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
303 #define for_each_online_gcwq_cpu(cpu) \
304 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
305 (cpu) < WORK_CPU_NONE; \
306 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
308 #define for_each_cwq_cpu(cpu, wq) \
309 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
310 (cpu) < WORK_CPU_NONE; \
311 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
313 #ifdef CONFIG_LOCKDEP
315 * in_workqueue_context() - in context of specified workqueue?
316 * @wq: the workqueue of interest
318 * Checks lockdep state to see if the current task is executing from
319 * within a workqueue item. This function exists only if lockdep is
322 int in_workqueue_context(struct workqueue_struct
*wq
)
324 return lock_is_held(&wq
->lockdep_map
);
328 #ifdef CONFIG_DEBUG_OBJECTS_WORK
330 static struct debug_obj_descr work_debug_descr
;
333 * fixup_init is called when:
334 * - an active object is initialized
336 static int work_fixup_init(void *addr
, enum debug_obj_state state
)
338 struct work_struct
*work
= addr
;
341 case ODEBUG_STATE_ACTIVE
:
342 cancel_work_sync(work
);
343 debug_object_init(work
, &work_debug_descr
);
351 * fixup_activate is called when:
352 * - an active object is activated
353 * - an unknown object is activated (might be a statically initialized object)
355 static int work_fixup_activate(void *addr
, enum debug_obj_state state
)
357 struct work_struct
*work
= addr
;
361 case ODEBUG_STATE_NOTAVAILABLE
:
363 * This is not really a fixup. The work struct was
364 * statically initialized. We just make sure that it
365 * is tracked in the object tracker.
367 if (test_bit(WORK_STRUCT_STATIC_BIT
, work_data_bits(work
))) {
368 debug_object_init(work
, &work_debug_descr
);
369 debug_object_activate(work
, &work_debug_descr
);
375 case ODEBUG_STATE_ACTIVE
:
384 * fixup_free is called when:
385 * - an active object is freed
387 static int work_fixup_free(void *addr
, enum debug_obj_state state
)
389 struct work_struct
*work
= addr
;
392 case ODEBUG_STATE_ACTIVE
:
393 cancel_work_sync(work
);
394 debug_object_free(work
, &work_debug_descr
);
401 static struct debug_obj_descr work_debug_descr
= {
402 .name
= "work_struct",
403 .fixup_init
= work_fixup_init
,
404 .fixup_activate
= work_fixup_activate
,
405 .fixup_free
= work_fixup_free
,
408 static inline void debug_work_activate(struct work_struct
*work
)
410 debug_object_activate(work
, &work_debug_descr
);
413 static inline void debug_work_deactivate(struct work_struct
*work
)
415 debug_object_deactivate(work
, &work_debug_descr
);
418 void __init_work(struct work_struct
*work
, int onstack
)
421 debug_object_init_on_stack(work
, &work_debug_descr
);
423 debug_object_init(work
, &work_debug_descr
);
425 EXPORT_SYMBOL_GPL(__init_work
);
427 void destroy_work_on_stack(struct work_struct
*work
)
429 debug_object_free(work
, &work_debug_descr
);
431 EXPORT_SYMBOL_GPL(destroy_work_on_stack
);
434 static inline void debug_work_activate(struct work_struct
*work
) { }
435 static inline void debug_work_deactivate(struct work_struct
*work
) { }
438 /* Serializes the accesses to the list of workqueues. */
439 static DEFINE_SPINLOCK(workqueue_lock
);
440 static LIST_HEAD(workqueues
);
441 static bool workqueue_freezing
; /* W: have wqs started freezing? */
444 * The almighty global cpu workqueues. nr_running is the only field
445 * which is expected to be used frequently by other cpus via
446 * try_to_wake_up(). Put it in a separate cacheline.
448 static DEFINE_PER_CPU(struct global_cwq
, global_cwq
);
449 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t
, gcwq_nr_running
);
452 * Global cpu workqueue and nr_running counter for unbound gcwq. The
453 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
454 * workers have WORKER_UNBOUND set.
456 static struct global_cwq unbound_global_cwq
;
457 static atomic_t unbound_gcwq_nr_running
= ATOMIC_INIT(0); /* always 0 */
459 static int worker_thread(void *__worker
);
461 static struct global_cwq
*get_gcwq(unsigned int cpu
)
463 if (cpu
!= WORK_CPU_UNBOUND
)
464 return &per_cpu(global_cwq
, cpu
);
466 return &unbound_global_cwq
;
469 static atomic_t
*get_gcwq_nr_running(unsigned int cpu
)
471 if (cpu
!= WORK_CPU_UNBOUND
)
472 return &per_cpu(gcwq_nr_running
, cpu
);
474 return &unbound_gcwq_nr_running
;
477 static struct cpu_workqueue_struct
*get_cwq(unsigned int cpu
,
478 struct workqueue_struct
*wq
)
480 if (!(wq
->flags
& WQ_UNBOUND
)) {
481 if (likely(cpu
< nr_cpu_ids
)) {
483 return per_cpu_ptr(wq
->cpu_wq
.pcpu
, cpu
);
485 return wq
->cpu_wq
.single
;
488 } else if (likely(cpu
== WORK_CPU_UNBOUND
))
489 return wq
->cpu_wq
.single
;
493 static unsigned int work_color_to_flags(int color
)
495 return color
<< WORK_STRUCT_COLOR_SHIFT
;
498 static int get_work_color(struct work_struct
*work
)
500 return (*work_data_bits(work
) >> WORK_STRUCT_COLOR_SHIFT
) &
501 ((1 << WORK_STRUCT_COLOR_BITS
) - 1);
504 static int work_next_color(int color
)
506 return (color
+ 1) % WORK_NR_COLORS
;
510 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
511 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
512 * cleared and the work data contains the cpu number it was last on.
514 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
515 * cwq, cpu or clear work->data. These functions should only be
516 * called while the work is owned - ie. while the PENDING bit is set.
518 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
519 * corresponding to a work. gcwq is available once the work has been
520 * queued anywhere after initialization. cwq is available only from
521 * queueing until execution starts.
523 static inline void set_work_data(struct work_struct
*work
, unsigned long data
,
526 BUG_ON(!work_pending(work
));
527 atomic_long_set(&work
->data
, data
| flags
| work_static(work
));
530 static void set_work_cwq(struct work_struct
*work
,
531 struct cpu_workqueue_struct
*cwq
,
532 unsigned long extra_flags
)
534 set_work_data(work
, (unsigned long)cwq
,
535 WORK_STRUCT_PENDING
| WORK_STRUCT_CWQ
| extra_flags
);
538 static void set_work_cpu(struct work_struct
*work
, unsigned int cpu
)
540 set_work_data(work
, cpu
<< WORK_STRUCT_FLAG_BITS
, WORK_STRUCT_PENDING
);
543 static void clear_work_data(struct work_struct
*work
)
545 set_work_data(work
, WORK_STRUCT_NO_CPU
, 0);
548 static struct cpu_workqueue_struct
*get_work_cwq(struct work_struct
*work
)
550 unsigned long data
= atomic_long_read(&work
->data
);
552 if (data
& WORK_STRUCT_CWQ
)
553 return (void *)(data
& WORK_STRUCT_WQ_DATA_MASK
);
558 static struct global_cwq
*get_work_gcwq(struct work_struct
*work
)
560 unsigned long data
= atomic_long_read(&work
->data
);
563 if (data
& WORK_STRUCT_CWQ
)
564 return ((struct cpu_workqueue_struct
*)
565 (data
& WORK_STRUCT_WQ_DATA_MASK
))->gcwq
;
567 cpu
= data
>> WORK_STRUCT_FLAG_BITS
;
568 if (cpu
== WORK_CPU_NONE
)
571 BUG_ON(cpu
>= nr_cpu_ids
&& cpu
!= WORK_CPU_UNBOUND
);
572 return get_gcwq(cpu
);
576 * Policy functions. These define the policies on how the global
577 * worker pool is managed. Unless noted otherwise, these functions
578 * assume that they're being called with gcwq->lock held.
581 static bool __need_more_worker(struct global_cwq
*gcwq
)
583 return !atomic_read(get_gcwq_nr_running(gcwq
->cpu
)) ||
584 gcwq
->flags
& GCWQ_HIGHPRI_PENDING
;
588 * Need to wake up a worker? Called from anything but currently
591 static bool need_more_worker(struct global_cwq
*gcwq
)
593 return !list_empty(&gcwq
->worklist
) && __need_more_worker(gcwq
);
596 /* Can I start working? Called from busy but !running workers. */
597 static bool may_start_working(struct global_cwq
*gcwq
)
599 return gcwq
->nr_idle
;
602 /* Do I need to keep working? Called from currently running workers. */
603 static bool keep_working(struct global_cwq
*gcwq
)
605 atomic_t
*nr_running
= get_gcwq_nr_running(gcwq
->cpu
);
607 return !list_empty(&gcwq
->worklist
) && atomic_read(nr_running
) <= 1;
610 /* Do we need a new worker? Called from manager. */
611 static bool need_to_create_worker(struct global_cwq
*gcwq
)
613 return need_more_worker(gcwq
) && !may_start_working(gcwq
);
616 /* Do I need to be the manager? */
617 static bool need_to_manage_workers(struct global_cwq
*gcwq
)
619 return need_to_create_worker(gcwq
) || gcwq
->flags
& GCWQ_MANAGE_WORKERS
;
622 /* Do we have too many workers and should some go away? */
623 static bool too_many_workers(struct global_cwq
*gcwq
)
625 bool managing
= gcwq
->flags
& GCWQ_MANAGING_WORKERS
;
626 int nr_idle
= gcwq
->nr_idle
+ managing
; /* manager is considered idle */
627 int nr_busy
= gcwq
->nr_workers
- nr_idle
;
629 return nr_idle
> 2 && (nr_idle
- 2) * MAX_IDLE_WORKERS_RATIO
>= nr_busy
;
636 /* Return the first worker. Safe with preemption disabled */
637 static struct worker
*first_worker(struct global_cwq
*gcwq
)
639 if (unlikely(list_empty(&gcwq
->idle_list
)))
642 return list_first_entry(&gcwq
->idle_list
, struct worker
, entry
);
646 * wake_up_worker - wake up an idle worker
647 * @gcwq: gcwq to wake worker for
649 * Wake up the first idle worker of @gcwq.
652 * spin_lock_irq(gcwq->lock).
654 static void wake_up_worker(struct global_cwq
*gcwq
)
656 struct worker
*worker
= first_worker(gcwq
);
659 wake_up_process(worker
->task
);
663 * wq_worker_waking_up - a worker is waking up
664 * @task: task waking up
665 * @cpu: CPU @task is waking up to
667 * This function is called during try_to_wake_up() when a worker is
671 * spin_lock_irq(rq->lock)
673 void wq_worker_waking_up(struct task_struct
*task
, unsigned int cpu
)
675 struct worker
*worker
= kthread_data(task
);
677 if (likely(!(worker
->flags
& WORKER_NOT_RUNNING
)))
678 atomic_inc(get_gcwq_nr_running(cpu
));
682 * wq_worker_sleeping - a worker is going to sleep
683 * @task: task going to sleep
684 * @cpu: CPU in question, must be the current CPU number
686 * This function is called during schedule() when a busy worker is
687 * going to sleep. Worker on the same cpu can be woken up by
688 * returning pointer to its task.
691 * spin_lock_irq(rq->lock)
694 * Worker task on @cpu to wake up, %NULL if none.
696 struct task_struct
*wq_worker_sleeping(struct task_struct
*task
,
699 struct worker
*worker
= kthread_data(task
), *to_wakeup
= NULL
;
700 struct global_cwq
*gcwq
= get_gcwq(cpu
);
701 atomic_t
*nr_running
= get_gcwq_nr_running(cpu
);
703 if (unlikely(worker
->flags
& WORKER_NOT_RUNNING
))
706 /* this can only happen on the local cpu */
707 BUG_ON(cpu
!= raw_smp_processor_id());
710 * The counterpart of the following dec_and_test, implied mb,
711 * worklist not empty test sequence is in insert_work().
712 * Please read comment there.
714 * NOT_RUNNING is clear. This means that trustee is not in
715 * charge and we're running on the local cpu w/ rq lock held
716 * and preemption disabled, which in turn means that none else
717 * could be manipulating idle_list, so dereferencing idle_list
718 * without gcwq lock is safe.
720 if (atomic_dec_and_test(nr_running
) && !list_empty(&gcwq
->worklist
))
721 to_wakeup
= first_worker(gcwq
);
722 return to_wakeup
? to_wakeup
->task
: NULL
;
726 * worker_set_flags - set worker flags and adjust nr_running accordingly
728 * @flags: flags to set
729 * @wakeup: wakeup an idle worker if necessary
731 * Set @flags in @worker->flags and adjust nr_running accordingly. If
732 * nr_running becomes zero and @wakeup is %true, an idle worker is
736 * spin_lock_irq(gcwq->lock)
738 static inline void worker_set_flags(struct worker
*worker
, unsigned int flags
,
741 struct global_cwq
*gcwq
= worker
->gcwq
;
743 WARN_ON_ONCE(worker
->task
!= current
);
746 * If transitioning into NOT_RUNNING, adjust nr_running and
747 * wake up an idle worker as necessary if requested by
750 if ((flags
& WORKER_NOT_RUNNING
) &&
751 !(worker
->flags
& WORKER_NOT_RUNNING
)) {
752 atomic_t
*nr_running
= get_gcwq_nr_running(gcwq
->cpu
);
755 if (atomic_dec_and_test(nr_running
) &&
756 !list_empty(&gcwq
->worklist
))
757 wake_up_worker(gcwq
);
759 atomic_dec(nr_running
);
762 worker
->flags
|= flags
;
766 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
768 * @flags: flags to clear
770 * Clear @flags in @worker->flags and adjust nr_running accordingly.
773 * spin_lock_irq(gcwq->lock)
775 static inline void worker_clr_flags(struct worker
*worker
, unsigned int flags
)
777 struct global_cwq
*gcwq
= worker
->gcwq
;
778 unsigned int oflags
= worker
->flags
;
780 WARN_ON_ONCE(worker
->task
!= current
);
782 worker
->flags
&= ~flags
;
784 /* if transitioning out of NOT_RUNNING, increment nr_running */
785 if ((flags
& WORKER_NOT_RUNNING
) && (oflags
& WORKER_NOT_RUNNING
))
786 if (!(worker
->flags
& WORKER_NOT_RUNNING
))
787 atomic_inc(get_gcwq_nr_running(gcwq
->cpu
));
791 * busy_worker_head - return the busy hash head for a work
792 * @gcwq: gcwq of interest
793 * @work: work to be hashed
795 * Return hash head of @gcwq for @work.
798 * spin_lock_irq(gcwq->lock).
801 * Pointer to the hash head.
803 static struct hlist_head
*busy_worker_head(struct global_cwq
*gcwq
,
804 struct work_struct
*work
)
806 const int base_shift
= ilog2(sizeof(struct work_struct
));
807 unsigned long v
= (unsigned long)work
;
809 /* simple shift and fold hash, do we need something better? */
811 v
+= v
>> BUSY_WORKER_HASH_ORDER
;
812 v
&= BUSY_WORKER_HASH_MASK
;
814 return &gcwq
->busy_hash
[v
];
818 * __find_worker_executing_work - find worker which is executing a work
819 * @gcwq: gcwq of interest
820 * @bwh: hash head as returned by busy_worker_head()
821 * @work: work to find worker for
823 * Find a worker which is executing @work on @gcwq. @bwh should be
824 * the hash head obtained by calling busy_worker_head() with the same
828 * spin_lock_irq(gcwq->lock).
831 * Pointer to worker which is executing @work if found, NULL
834 static struct worker
*__find_worker_executing_work(struct global_cwq
*gcwq
,
835 struct hlist_head
*bwh
,
836 struct work_struct
*work
)
838 struct worker
*worker
;
839 struct hlist_node
*tmp
;
841 hlist_for_each_entry(worker
, tmp
, bwh
, hentry
)
842 if (worker
->current_work
== work
)
848 * find_worker_executing_work - find worker which is executing a work
849 * @gcwq: gcwq of interest
850 * @work: work to find worker for
852 * Find a worker which is executing @work on @gcwq. This function is
853 * identical to __find_worker_executing_work() except that this
854 * function calculates @bwh itself.
857 * spin_lock_irq(gcwq->lock).
860 * Pointer to worker which is executing @work if found, NULL
863 static struct worker
*find_worker_executing_work(struct global_cwq
*gcwq
,
864 struct work_struct
*work
)
866 return __find_worker_executing_work(gcwq
, busy_worker_head(gcwq
, work
),
871 * gcwq_determine_ins_pos - find insertion position
872 * @gcwq: gcwq of interest
873 * @cwq: cwq a work is being queued for
875 * A work for @cwq is about to be queued on @gcwq, determine insertion
876 * position for the work. If @cwq is for HIGHPRI wq, the work is
877 * queued at the head of the queue but in FIFO order with respect to
878 * other HIGHPRI works; otherwise, at the end of the queue. This
879 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
880 * there are HIGHPRI works pending.
883 * spin_lock_irq(gcwq->lock).
886 * Pointer to inserstion position.
888 static inline struct list_head
*gcwq_determine_ins_pos(struct global_cwq
*gcwq
,
889 struct cpu_workqueue_struct
*cwq
)
891 struct work_struct
*twork
;
893 if (likely(!(cwq
->wq
->flags
& WQ_HIGHPRI
)))
894 return &gcwq
->worklist
;
896 list_for_each_entry(twork
, &gcwq
->worklist
, entry
) {
897 struct cpu_workqueue_struct
*tcwq
= get_work_cwq(twork
);
899 if (!(tcwq
->wq
->flags
& WQ_HIGHPRI
))
903 gcwq
->flags
|= GCWQ_HIGHPRI_PENDING
;
904 return &twork
->entry
;
908 * insert_work - insert a work into gcwq
909 * @cwq: cwq @work belongs to
910 * @work: work to insert
911 * @head: insertion point
912 * @extra_flags: extra WORK_STRUCT_* flags to set
914 * Insert @work which belongs to @cwq into @gcwq after @head.
915 * @extra_flags is or'd to work_struct flags.
918 * spin_lock_irq(gcwq->lock).
920 static void insert_work(struct cpu_workqueue_struct
*cwq
,
921 struct work_struct
*work
, struct list_head
*head
,
922 unsigned int extra_flags
)
924 struct global_cwq
*gcwq
= cwq
->gcwq
;
926 /* we own @work, set data and link */
927 set_work_cwq(work
, cwq
, extra_flags
);
930 * Ensure that we get the right work->data if we see the
931 * result of list_add() below, see try_to_grab_pending().
935 list_add_tail(&work
->entry
, head
);
938 * Ensure either worker_sched_deactivated() sees the above
939 * list_add_tail() or we see zero nr_running to avoid workers
940 * lying around lazily while there are works to be processed.
944 if (__need_more_worker(gcwq
))
945 wake_up_worker(gcwq
);
948 static void __queue_work(unsigned int cpu
, struct workqueue_struct
*wq
,
949 struct work_struct
*work
)
951 struct global_cwq
*gcwq
;
952 struct cpu_workqueue_struct
*cwq
;
953 struct list_head
*worklist
;
954 unsigned int work_flags
;
957 debug_work_activate(work
);
959 if (WARN_ON_ONCE(wq
->flags
& WQ_DYING
))
962 /* determine gcwq to use */
963 if (!(wq
->flags
& WQ_UNBOUND
)) {
964 struct global_cwq
*last_gcwq
;
966 if (unlikely(cpu
== WORK_CPU_UNBOUND
))
967 cpu
= raw_smp_processor_id();
970 * It's multi cpu. If @wq is non-reentrant and @work
971 * was previously on a different cpu, it might still
972 * be running there, in which case the work needs to
973 * be queued on that cpu to guarantee non-reentrance.
975 gcwq
= get_gcwq(cpu
);
976 if (wq
->flags
& WQ_NON_REENTRANT
&&
977 (last_gcwq
= get_work_gcwq(work
)) && last_gcwq
!= gcwq
) {
978 struct worker
*worker
;
980 spin_lock_irqsave(&last_gcwq
->lock
, flags
);
982 worker
= find_worker_executing_work(last_gcwq
, work
);
984 if (worker
&& worker
->current_cwq
->wq
== wq
)
987 /* meh... not running there, queue here */
988 spin_unlock_irqrestore(&last_gcwq
->lock
, flags
);
989 spin_lock_irqsave(&gcwq
->lock
, flags
);
992 spin_lock_irqsave(&gcwq
->lock
, flags
);
994 gcwq
= get_gcwq(WORK_CPU_UNBOUND
);
995 spin_lock_irqsave(&gcwq
->lock
, flags
);
998 /* gcwq determined, get cwq and queue */
999 cwq
= get_cwq(gcwq
->cpu
, wq
);
1000 trace_workqueue_queue_work(cpu
, cwq
, work
);
1002 BUG_ON(!list_empty(&work
->entry
));
1004 cwq
->nr_in_flight
[cwq
->work_color
]++;
1005 work_flags
= work_color_to_flags(cwq
->work_color
);
1007 if (likely(cwq
->nr_active
< cwq
->max_active
)) {
1008 trace_workqueue_activate_work(work
);
1010 worklist
= gcwq_determine_ins_pos(gcwq
, cwq
);
1012 work_flags
|= WORK_STRUCT_DELAYED
;
1013 worklist
= &cwq
->delayed_works
;
1016 insert_work(cwq
, work
, worklist
, work_flags
);
1018 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
1022 * queue_work - queue work on a workqueue
1023 * @wq: workqueue to use
1024 * @work: work to queue
1026 * Returns 0 if @work was already on a queue, non-zero otherwise.
1028 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1029 * it can be processed by another CPU.
1031 int queue_work(struct workqueue_struct
*wq
, struct work_struct
*work
)
1035 ret
= queue_work_on(get_cpu(), wq
, work
);
1040 EXPORT_SYMBOL_GPL(queue_work
);
1043 * queue_work_on - queue work on specific cpu
1044 * @cpu: CPU number to execute work on
1045 * @wq: workqueue to use
1046 * @work: work to queue
1048 * Returns 0 if @work was already on a queue, non-zero otherwise.
1050 * We queue the work to a specific CPU, the caller must ensure it
1054 queue_work_on(int cpu
, struct workqueue_struct
*wq
, struct work_struct
*work
)
1058 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1059 __queue_work(cpu
, wq
, work
);
1064 EXPORT_SYMBOL_GPL(queue_work_on
);
1066 static void delayed_work_timer_fn(unsigned long __data
)
1068 struct delayed_work
*dwork
= (struct delayed_work
*)__data
;
1069 struct cpu_workqueue_struct
*cwq
= get_work_cwq(&dwork
->work
);
1071 __queue_work(smp_processor_id(), cwq
->wq
, &dwork
->work
);
1075 * queue_delayed_work - queue work on a workqueue after delay
1076 * @wq: workqueue to use
1077 * @dwork: delayable work to queue
1078 * @delay: number of jiffies to wait before queueing
1080 * Returns 0 if @work was already on a queue, non-zero otherwise.
1082 int queue_delayed_work(struct workqueue_struct
*wq
,
1083 struct delayed_work
*dwork
, unsigned long delay
)
1086 return queue_work(wq
, &dwork
->work
);
1088 return queue_delayed_work_on(-1, wq
, dwork
, delay
);
1090 EXPORT_SYMBOL_GPL(queue_delayed_work
);
1093 * queue_delayed_work_on - queue work on specific CPU after delay
1094 * @cpu: CPU number to execute work on
1095 * @wq: workqueue to use
1096 * @dwork: work to queue
1097 * @delay: number of jiffies to wait before queueing
1099 * Returns 0 if @work was already on a queue, non-zero otherwise.
1101 int queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
1102 struct delayed_work
*dwork
, unsigned long delay
)
1105 struct timer_list
*timer
= &dwork
->timer
;
1106 struct work_struct
*work
= &dwork
->work
;
1108 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
))) {
1111 BUG_ON(timer_pending(timer
));
1112 BUG_ON(!list_empty(&work
->entry
));
1114 timer_stats_timer_set_start_info(&dwork
->timer
);
1117 * This stores cwq for the moment, for the timer_fn.
1118 * Note that the work's gcwq is preserved to allow
1119 * reentrance detection for delayed works.
1121 if (!(wq
->flags
& WQ_UNBOUND
)) {
1122 struct global_cwq
*gcwq
= get_work_gcwq(work
);
1124 if (gcwq
&& gcwq
->cpu
!= WORK_CPU_UNBOUND
)
1127 lcpu
= raw_smp_processor_id();
1129 lcpu
= WORK_CPU_UNBOUND
;
1131 set_work_cwq(work
, get_cwq(lcpu
, wq
), 0);
1133 timer
->expires
= jiffies
+ delay
;
1134 timer
->data
= (unsigned long)dwork
;
1135 timer
->function
= delayed_work_timer_fn
;
1137 if (unlikely(cpu
>= 0))
1138 add_timer_on(timer
, cpu
);
1145 EXPORT_SYMBOL_GPL(queue_delayed_work_on
);
1148 * worker_enter_idle - enter idle state
1149 * @worker: worker which is entering idle state
1151 * @worker is entering idle state. Update stats and idle timer if
1155 * spin_lock_irq(gcwq->lock).
1157 static void worker_enter_idle(struct worker
*worker
)
1159 struct global_cwq
*gcwq
= worker
->gcwq
;
1161 BUG_ON(worker
->flags
& WORKER_IDLE
);
1162 BUG_ON(!list_empty(&worker
->entry
) &&
1163 (worker
->hentry
.next
|| worker
->hentry
.pprev
));
1165 /* can't use worker_set_flags(), also called from start_worker() */
1166 worker
->flags
|= WORKER_IDLE
;
1168 worker
->last_active
= jiffies
;
1170 /* idle_list is LIFO */
1171 list_add(&worker
->entry
, &gcwq
->idle_list
);
1173 if (likely(!(worker
->flags
& WORKER_ROGUE
))) {
1174 if (too_many_workers(gcwq
) && !timer_pending(&gcwq
->idle_timer
))
1175 mod_timer(&gcwq
->idle_timer
,
1176 jiffies
+ IDLE_WORKER_TIMEOUT
);
1178 wake_up_all(&gcwq
->trustee_wait
);
1180 /* sanity check nr_running */
1181 WARN_ON_ONCE(gcwq
->nr_workers
== gcwq
->nr_idle
&&
1182 atomic_read(get_gcwq_nr_running(gcwq
->cpu
)));
1186 * worker_leave_idle - leave idle state
1187 * @worker: worker which is leaving idle state
1189 * @worker is leaving idle state. Update stats.
1192 * spin_lock_irq(gcwq->lock).
1194 static void worker_leave_idle(struct worker
*worker
)
1196 struct global_cwq
*gcwq
= worker
->gcwq
;
1198 BUG_ON(!(worker
->flags
& WORKER_IDLE
));
1199 worker_clr_flags(worker
, WORKER_IDLE
);
1201 list_del_init(&worker
->entry
);
1205 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1208 * Works which are scheduled while the cpu is online must at least be
1209 * scheduled to a worker which is bound to the cpu so that if they are
1210 * flushed from cpu callbacks while cpu is going down, they are
1211 * guaranteed to execute on the cpu.
1213 * This function is to be used by rogue workers and rescuers to bind
1214 * themselves to the target cpu and may race with cpu going down or
1215 * coming online. kthread_bind() can't be used because it may put the
1216 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1217 * verbatim as it's best effort and blocking and gcwq may be
1218 * [dis]associated in the meantime.
1220 * This function tries set_cpus_allowed() and locks gcwq and verifies
1221 * the binding against GCWQ_DISASSOCIATED which is set during
1222 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1223 * idle state or fetches works without dropping lock, it can guarantee
1224 * the scheduling requirement described in the first paragraph.
1227 * Might sleep. Called without any lock but returns with gcwq->lock
1231 * %true if the associated gcwq is online (@worker is successfully
1232 * bound), %false if offline.
1234 static bool worker_maybe_bind_and_lock(struct worker
*worker
)
1235 __acquires(&gcwq
->lock
)
1237 struct global_cwq
*gcwq
= worker
->gcwq
;
1238 struct task_struct
*task
= worker
->task
;
1242 * The following call may fail, succeed or succeed
1243 * without actually migrating the task to the cpu if
1244 * it races with cpu hotunplug operation. Verify
1245 * against GCWQ_DISASSOCIATED.
1247 if (!(gcwq
->flags
& GCWQ_DISASSOCIATED
))
1248 set_cpus_allowed_ptr(task
, get_cpu_mask(gcwq
->cpu
));
1250 spin_lock_irq(&gcwq
->lock
);
1251 if (gcwq
->flags
& GCWQ_DISASSOCIATED
)
1253 if (task_cpu(task
) == gcwq
->cpu
&&
1254 cpumask_equal(¤t
->cpus_allowed
,
1255 get_cpu_mask(gcwq
->cpu
)))
1257 spin_unlock_irq(&gcwq
->lock
);
1259 /* CPU has come up inbetween, retry migration */
1265 * Function for worker->rebind_work used to rebind rogue busy workers
1266 * to the associated cpu which is coming back online. This is
1267 * scheduled by cpu up but can race with other cpu hotplug operations
1268 * and may be executed twice without intervening cpu down.
1270 static void worker_rebind_fn(struct work_struct
*work
)
1272 struct worker
*worker
= container_of(work
, struct worker
, rebind_work
);
1273 struct global_cwq
*gcwq
= worker
->gcwq
;
1275 if (worker_maybe_bind_and_lock(worker
))
1276 worker_clr_flags(worker
, WORKER_REBIND
);
1278 spin_unlock_irq(&gcwq
->lock
);
1281 static struct worker
*alloc_worker(void)
1283 struct worker
*worker
;
1285 worker
= kzalloc(sizeof(*worker
), GFP_KERNEL
);
1287 INIT_LIST_HEAD(&worker
->entry
);
1288 INIT_LIST_HEAD(&worker
->scheduled
);
1289 INIT_WORK(&worker
->rebind_work
, worker_rebind_fn
);
1290 /* on creation a worker is in !idle && prep state */
1291 worker
->flags
= WORKER_PREP
;
1297 * create_worker - create a new workqueue worker
1298 * @gcwq: gcwq the new worker will belong to
1299 * @bind: whether to set affinity to @cpu or not
1301 * Create a new worker which is bound to @gcwq. The returned worker
1302 * can be started by calling start_worker() or destroyed using
1306 * Might sleep. Does GFP_KERNEL allocations.
1309 * Pointer to the newly created worker.
1311 static struct worker
*create_worker(struct global_cwq
*gcwq
, bool bind
)
1313 bool on_unbound_cpu
= gcwq
->cpu
== WORK_CPU_UNBOUND
;
1314 struct worker
*worker
= NULL
;
1317 spin_lock_irq(&gcwq
->lock
);
1318 while (ida_get_new(&gcwq
->worker_ida
, &id
)) {
1319 spin_unlock_irq(&gcwq
->lock
);
1320 if (!ida_pre_get(&gcwq
->worker_ida
, GFP_KERNEL
))
1322 spin_lock_irq(&gcwq
->lock
);
1324 spin_unlock_irq(&gcwq
->lock
);
1326 worker
= alloc_worker();
1330 worker
->gcwq
= gcwq
;
1333 if (!on_unbound_cpu
)
1334 worker
->task
= kthread_create(worker_thread
, worker
,
1335 "kworker/%u:%d", gcwq
->cpu
, id
);
1337 worker
->task
= kthread_create(worker_thread
, worker
,
1338 "kworker/u:%d", id
);
1339 if (IS_ERR(worker
->task
))
1343 * A rogue worker will become a regular one if CPU comes
1344 * online later on. Make sure every worker has
1345 * PF_THREAD_BOUND set.
1347 if (bind
&& !on_unbound_cpu
)
1348 kthread_bind(worker
->task
, gcwq
->cpu
);
1350 worker
->task
->flags
|= PF_THREAD_BOUND
;
1352 worker
->flags
|= WORKER_UNBOUND
;
1358 spin_lock_irq(&gcwq
->lock
);
1359 ida_remove(&gcwq
->worker_ida
, id
);
1360 spin_unlock_irq(&gcwq
->lock
);
1367 * start_worker - start a newly created worker
1368 * @worker: worker to start
1370 * Make the gcwq aware of @worker and start it.
1373 * spin_lock_irq(gcwq->lock).
1375 static void start_worker(struct worker
*worker
)
1377 worker
->flags
|= WORKER_STARTED
;
1378 worker
->gcwq
->nr_workers
++;
1379 worker_enter_idle(worker
);
1380 wake_up_process(worker
->task
);
1384 * destroy_worker - destroy a workqueue worker
1385 * @worker: worker to be destroyed
1387 * Destroy @worker and adjust @gcwq stats accordingly.
1390 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1392 static void destroy_worker(struct worker
*worker
)
1394 struct global_cwq
*gcwq
= worker
->gcwq
;
1395 int id
= worker
->id
;
1397 /* sanity check frenzy */
1398 BUG_ON(worker
->current_work
);
1399 BUG_ON(!list_empty(&worker
->scheduled
));
1401 if (worker
->flags
& WORKER_STARTED
)
1403 if (worker
->flags
& WORKER_IDLE
)
1406 list_del_init(&worker
->entry
);
1407 worker
->flags
|= WORKER_DIE
;
1409 spin_unlock_irq(&gcwq
->lock
);
1411 kthread_stop(worker
->task
);
1414 spin_lock_irq(&gcwq
->lock
);
1415 ida_remove(&gcwq
->worker_ida
, id
);
1418 static void idle_worker_timeout(unsigned long __gcwq
)
1420 struct global_cwq
*gcwq
= (void *)__gcwq
;
1422 spin_lock_irq(&gcwq
->lock
);
1424 if (too_many_workers(gcwq
)) {
1425 struct worker
*worker
;
1426 unsigned long expires
;
1428 /* idle_list is kept in LIFO order, check the last one */
1429 worker
= list_entry(gcwq
->idle_list
.prev
, struct worker
, entry
);
1430 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1432 if (time_before(jiffies
, expires
))
1433 mod_timer(&gcwq
->idle_timer
, expires
);
1435 /* it's been idle for too long, wake up manager */
1436 gcwq
->flags
|= GCWQ_MANAGE_WORKERS
;
1437 wake_up_worker(gcwq
);
1441 spin_unlock_irq(&gcwq
->lock
);
1444 static bool send_mayday(struct work_struct
*work
)
1446 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1447 struct workqueue_struct
*wq
= cwq
->wq
;
1450 if (!(wq
->flags
& WQ_RESCUER
))
1453 /* mayday mayday mayday */
1454 cpu
= cwq
->gcwq
->cpu
;
1455 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1456 if (cpu
== WORK_CPU_UNBOUND
)
1458 if (!mayday_test_and_set_cpu(cpu
, wq
->mayday_mask
))
1459 wake_up_process(wq
->rescuer
->task
);
1463 static void gcwq_mayday_timeout(unsigned long __gcwq
)
1465 struct global_cwq
*gcwq
= (void *)__gcwq
;
1466 struct work_struct
*work
;
1468 spin_lock_irq(&gcwq
->lock
);
1470 if (need_to_create_worker(gcwq
)) {
1472 * We've been trying to create a new worker but
1473 * haven't been successful. We might be hitting an
1474 * allocation deadlock. Send distress signals to
1477 list_for_each_entry(work
, &gcwq
->worklist
, entry
)
1481 spin_unlock_irq(&gcwq
->lock
);
1483 mod_timer(&gcwq
->mayday_timer
, jiffies
+ MAYDAY_INTERVAL
);
1487 * maybe_create_worker - create a new worker if necessary
1488 * @gcwq: gcwq to create a new worker for
1490 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1491 * have at least one idle worker on return from this function. If
1492 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1493 * sent to all rescuers with works scheduled on @gcwq to resolve
1494 * possible allocation deadlock.
1496 * On return, need_to_create_worker() is guaranteed to be false and
1497 * may_start_working() true.
1500 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1501 * multiple times. Does GFP_KERNEL allocations. Called only from
1505 * false if no action was taken and gcwq->lock stayed locked, true
1508 static bool maybe_create_worker(struct global_cwq
*gcwq
)
1509 __releases(&gcwq
->lock
)
1510 __acquires(&gcwq
->lock
)
1512 if (!need_to_create_worker(gcwq
))
1515 spin_unlock_irq(&gcwq
->lock
);
1517 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1518 mod_timer(&gcwq
->mayday_timer
, jiffies
+ MAYDAY_INITIAL_TIMEOUT
);
1521 struct worker
*worker
;
1523 worker
= create_worker(gcwq
, true);
1525 del_timer_sync(&gcwq
->mayday_timer
);
1526 spin_lock_irq(&gcwq
->lock
);
1527 start_worker(worker
);
1528 BUG_ON(need_to_create_worker(gcwq
));
1532 if (!need_to_create_worker(gcwq
))
1535 __set_current_state(TASK_INTERRUPTIBLE
);
1536 schedule_timeout(CREATE_COOLDOWN
);
1538 if (!need_to_create_worker(gcwq
))
1542 del_timer_sync(&gcwq
->mayday_timer
);
1543 spin_lock_irq(&gcwq
->lock
);
1544 if (need_to_create_worker(gcwq
))
1550 * maybe_destroy_worker - destroy workers which have been idle for a while
1551 * @gcwq: gcwq to destroy workers for
1553 * Destroy @gcwq workers which have been idle for longer than
1554 * IDLE_WORKER_TIMEOUT.
1557 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1558 * multiple times. Called only from manager.
1561 * false if no action was taken and gcwq->lock stayed locked, true
1564 static bool maybe_destroy_workers(struct global_cwq
*gcwq
)
1568 while (too_many_workers(gcwq
)) {
1569 struct worker
*worker
;
1570 unsigned long expires
;
1572 worker
= list_entry(gcwq
->idle_list
.prev
, struct worker
, entry
);
1573 expires
= worker
->last_active
+ IDLE_WORKER_TIMEOUT
;
1575 if (time_before(jiffies
, expires
)) {
1576 mod_timer(&gcwq
->idle_timer
, expires
);
1580 destroy_worker(worker
);
1588 * manage_workers - manage worker pool
1591 * Assume the manager role and manage gcwq worker pool @worker belongs
1592 * to. At any given time, there can be only zero or one manager per
1593 * gcwq. The exclusion is handled automatically by this function.
1595 * The caller can safely start processing works on false return. On
1596 * true return, it's guaranteed that need_to_create_worker() is false
1597 * and may_start_working() is true.
1600 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1601 * multiple times. Does GFP_KERNEL allocations.
1604 * false if no action was taken and gcwq->lock stayed locked, true if
1605 * some action was taken.
1607 static bool manage_workers(struct worker
*worker
)
1609 struct global_cwq
*gcwq
= worker
->gcwq
;
1612 if (gcwq
->flags
& GCWQ_MANAGING_WORKERS
)
1615 gcwq
->flags
&= ~GCWQ_MANAGE_WORKERS
;
1616 gcwq
->flags
|= GCWQ_MANAGING_WORKERS
;
1619 * Destroy and then create so that may_start_working() is true
1622 ret
|= maybe_destroy_workers(gcwq
);
1623 ret
|= maybe_create_worker(gcwq
);
1625 gcwq
->flags
&= ~GCWQ_MANAGING_WORKERS
;
1628 * The trustee might be waiting to take over the manager
1629 * position, tell it we're done.
1631 if (unlikely(gcwq
->trustee
))
1632 wake_up_all(&gcwq
->trustee_wait
);
1638 * move_linked_works - move linked works to a list
1639 * @work: start of series of works to be scheduled
1640 * @head: target list to append @work to
1641 * @nextp: out paramter for nested worklist walking
1643 * Schedule linked works starting from @work to @head. Work series to
1644 * be scheduled starts at @work and includes any consecutive work with
1645 * WORK_STRUCT_LINKED set in its predecessor.
1647 * If @nextp is not NULL, it's updated to point to the next work of
1648 * the last scheduled work. This allows move_linked_works() to be
1649 * nested inside outer list_for_each_entry_safe().
1652 * spin_lock_irq(gcwq->lock).
1654 static void move_linked_works(struct work_struct
*work
, struct list_head
*head
,
1655 struct work_struct
**nextp
)
1657 struct work_struct
*n
;
1660 * Linked worklist will always end before the end of the list,
1661 * use NULL for list head.
1663 list_for_each_entry_safe_from(work
, n
, NULL
, entry
) {
1664 list_move_tail(&work
->entry
, head
);
1665 if (!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))
1670 * If we're already inside safe list traversal and have moved
1671 * multiple works to the scheduled queue, the next position
1672 * needs to be updated.
1678 static void cwq_activate_first_delayed(struct cpu_workqueue_struct
*cwq
)
1680 struct work_struct
*work
= list_first_entry(&cwq
->delayed_works
,
1681 struct work_struct
, entry
);
1682 struct list_head
*pos
= gcwq_determine_ins_pos(cwq
->gcwq
, cwq
);
1684 trace_workqueue_activate_work(work
);
1685 move_linked_works(work
, pos
, NULL
);
1686 __clear_bit(WORK_STRUCT_DELAYED_BIT
, work_data_bits(work
));
1691 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1692 * @cwq: cwq of interest
1693 * @color: color of work which left the queue
1694 * @delayed: for a delayed work
1696 * A work either has completed or is removed from pending queue,
1697 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1700 * spin_lock_irq(gcwq->lock).
1702 static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct
*cwq
, int color
,
1705 /* ignore uncolored works */
1706 if (color
== WORK_NO_COLOR
)
1709 cwq
->nr_in_flight
[color
]--;
1713 if (!list_empty(&cwq
->delayed_works
)) {
1714 /* one down, submit a delayed one */
1715 if (cwq
->nr_active
< cwq
->max_active
)
1716 cwq_activate_first_delayed(cwq
);
1720 /* is flush in progress and are we at the flushing tip? */
1721 if (likely(cwq
->flush_color
!= color
))
1724 /* are there still in-flight works? */
1725 if (cwq
->nr_in_flight
[color
])
1728 /* this cwq is done, clear flush_color */
1729 cwq
->flush_color
= -1;
1732 * If this was the last cwq, wake up the first flusher. It
1733 * will handle the rest.
1735 if (atomic_dec_and_test(&cwq
->wq
->nr_cwqs_to_flush
))
1736 complete(&cwq
->wq
->first_flusher
->done
);
1740 * process_one_work - process single work
1742 * @work: work to process
1744 * Process @work. This function contains all the logics necessary to
1745 * process a single work including synchronization against and
1746 * interaction with other workers on the same cpu, queueing and
1747 * flushing. As long as context requirement is met, any worker can
1748 * call this function to process a work.
1751 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1753 static void process_one_work(struct worker
*worker
, struct work_struct
*work
)
1754 __releases(&gcwq
->lock
)
1755 __acquires(&gcwq
->lock
)
1757 struct cpu_workqueue_struct
*cwq
= get_work_cwq(work
);
1758 struct global_cwq
*gcwq
= cwq
->gcwq
;
1759 struct hlist_head
*bwh
= busy_worker_head(gcwq
, work
);
1760 bool cpu_intensive
= cwq
->wq
->flags
& WQ_CPU_INTENSIVE
;
1761 work_func_t f
= work
->func
;
1763 struct worker
*collision
;
1764 #ifdef CONFIG_LOCKDEP
1766 * It is permissible to free the struct work_struct from
1767 * inside the function that is called from it, this we need to
1768 * take into account for lockdep too. To avoid bogus "held
1769 * lock freed" warnings as well as problems when looking into
1770 * work->lockdep_map, make a copy and use that here.
1772 struct lockdep_map lockdep_map
= work
->lockdep_map
;
1775 * A single work shouldn't be executed concurrently by
1776 * multiple workers on a single cpu. Check whether anyone is
1777 * already processing the work. If so, defer the work to the
1778 * currently executing one.
1780 collision
= __find_worker_executing_work(gcwq
, bwh
, work
);
1781 if (unlikely(collision
)) {
1782 move_linked_works(work
, &collision
->scheduled
, NULL
);
1786 /* claim and process */
1787 debug_work_deactivate(work
);
1788 hlist_add_head(&worker
->hentry
, bwh
);
1789 worker
->current_work
= work
;
1790 worker
->current_cwq
= cwq
;
1791 work_color
= get_work_color(work
);
1793 /* record the current cpu number in the work data and dequeue */
1794 set_work_cpu(work
, gcwq
->cpu
);
1795 list_del_init(&work
->entry
);
1798 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1799 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1801 if (unlikely(gcwq
->flags
& GCWQ_HIGHPRI_PENDING
)) {
1802 struct work_struct
*nwork
= list_first_entry(&gcwq
->worklist
,
1803 struct work_struct
, entry
);
1805 if (!list_empty(&gcwq
->worklist
) &&
1806 get_work_cwq(nwork
)->wq
->flags
& WQ_HIGHPRI
)
1807 wake_up_worker(gcwq
);
1809 gcwq
->flags
&= ~GCWQ_HIGHPRI_PENDING
;
1813 * CPU intensive works don't participate in concurrency
1814 * management. They're the scheduler's responsibility.
1816 if (unlikely(cpu_intensive
))
1817 worker_set_flags(worker
, WORKER_CPU_INTENSIVE
, true);
1819 spin_unlock_irq(&gcwq
->lock
);
1821 work_clear_pending(work
);
1822 lock_map_acquire(&cwq
->wq
->lockdep_map
);
1823 lock_map_acquire(&lockdep_map
);
1824 trace_workqueue_execute_start(work
);
1827 * While we must be careful to not use "work" after this, the trace
1828 * point will only record its address.
1830 trace_workqueue_execute_end(work
);
1831 lock_map_release(&lockdep_map
);
1832 lock_map_release(&cwq
->wq
->lockdep_map
);
1834 if (unlikely(in_atomic() || lockdep_depth(current
) > 0)) {
1835 printk(KERN_ERR
"BUG: workqueue leaked lock or atomic: "
1837 current
->comm
, preempt_count(), task_pid_nr(current
));
1838 printk(KERN_ERR
" last function: ");
1839 print_symbol("%s\n", (unsigned long)f
);
1840 debug_show_held_locks(current
);
1844 spin_lock_irq(&gcwq
->lock
);
1846 /* clear cpu intensive status */
1847 if (unlikely(cpu_intensive
))
1848 worker_clr_flags(worker
, WORKER_CPU_INTENSIVE
);
1850 /* we're done with it, release */
1851 hlist_del_init(&worker
->hentry
);
1852 worker
->current_work
= NULL
;
1853 worker
->current_cwq
= NULL
;
1854 cwq_dec_nr_in_flight(cwq
, work_color
, false);
1858 * process_scheduled_works - process scheduled works
1861 * Process all scheduled works. Please note that the scheduled list
1862 * may change while processing a work, so this function repeatedly
1863 * fetches a work from the top and executes it.
1866 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1869 static void process_scheduled_works(struct worker
*worker
)
1871 while (!list_empty(&worker
->scheduled
)) {
1872 struct work_struct
*work
= list_first_entry(&worker
->scheduled
,
1873 struct work_struct
, entry
);
1874 process_one_work(worker
, work
);
1879 * worker_thread - the worker thread function
1882 * The gcwq worker thread function. There's a single dynamic pool of
1883 * these per each cpu. These workers process all works regardless of
1884 * their specific target workqueue. The only exception is works which
1885 * belong to workqueues with a rescuer which will be explained in
1888 static int worker_thread(void *__worker
)
1890 struct worker
*worker
= __worker
;
1891 struct global_cwq
*gcwq
= worker
->gcwq
;
1893 /* tell the scheduler that this is a workqueue worker */
1894 worker
->task
->flags
|= PF_WQ_WORKER
;
1896 spin_lock_irq(&gcwq
->lock
);
1898 /* DIE can be set only while we're idle, checking here is enough */
1899 if (worker
->flags
& WORKER_DIE
) {
1900 spin_unlock_irq(&gcwq
->lock
);
1901 worker
->task
->flags
&= ~PF_WQ_WORKER
;
1905 worker_leave_idle(worker
);
1907 /* no more worker necessary? */
1908 if (!need_more_worker(gcwq
))
1911 /* do we need to manage? */
1912 if (unlikely(!may_start_working(gcwq
)) && manage_workers(worker
))
1916 * ->scheduled list can only be filled while a worker is
1917 * preparing to process a work or actually processing it.
1918 * Make sure nobody diddled with it while I was sleeping.
1920 BUG_ON(!list_empty(&worker
->scheduled
));
1923 * When control reaches this point, we're guaranteed to have
1924 * at least one idle worker or that someone else has already
1925 * assumed the manager role.
1927 worker_clr_flags(worker
, WORKER_PREP
);
1930 struct work_struct
*work
=
1931 list_first_entry(&gcwq
->worklist
,
1932 struct work_struct
, entry
);
1934 if (likely(!(*work_data_bits(work
) & WORK_STRUCT_LINKED
))) {
1935 /* optimization path, not strictly necessary */
1936 process_one_work(worker
, work
);
1937 if (unlikely(!list_empty(&worker
->scheduled
)))
1938 process_scheduled_works(worker
);
1940 move_linked_works(work
, &worker
->scheduled
, NULL
);
1941 process_scheduled_works(worker
);
1943 } while (keep_working(gcwq
));
1945 worker_set_flags(worker
, WORKER_PREP
, false);
1947 if (unlikely(need_to_manage_workers(gcwq
)) && manage_workers(worker
))
1951 * gcwq->lock is held and there's no work to process and no
1952 * need to manage, sleep. Workers are woken up only while
1953 * holding gcwq->lock or from local cpu, so setting the
1954 * current state before releasing gcwq->lock is enough to
1955 * prevent losing any event.
1957 worker_enter_idle(worker
);
1958 __set_current_state(TASK_INTERRUPTIBLE
);
1959 spin_unlock_irq(&gcwq
->lock
);
1965 * rescuer_thread - the rescuer thread function
1966 * @__wq: the associated workqueue
1968 * Workqueue rescuer thread function. There's one rescuer for each
1969 * workqueue which has WQ_RESCUER set.
1971 * Regular work processing on a gcwq may block trying to create a new
1972 * worker which uses GFP_KERNEL allocation which has slight chance of
1973 * developing into deadlock if some works currently on the same queue
1974 * need to be processed to satisfy the GFP_KERNEL allocation. This is
1975 * the problem rescuer solves.
1977 * When such condition is possible, the gcwq summons rescuers of all
1978 * workqueues which have works queued on the gcwq and let them process
1979 * those works so that forward progress can be guaranteed.
1981 * This should happen rarely.
1983 static int rescuer_thread(void *__wq
)
1985 struct workqueue_struct
*wq
= __wq
;
1986 struct worker
*rescuer
= wq
->rescuer
;
1987 struct list_head
*scheduled
= &rescuer
->scheduled
;
1988 bool is_unbound
= wq
->flags
& WQ_UNBOUND
;
1991 set_user_nice(current
, RESCUER_NICE_LEVEL
);
1993 set_current_state(TASK_INTERRUPTIBLE
);
1995 if (kthread_should_stop())
1999 * See whether any cpu is asking for help. Unbounded
2000 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2002 for_each_mayday_cpu(cpu
, wq
->mayday_mask
) {
2003 unsigned int tcpu
= is_unbound
? WORK_CPU_UNBOUND
: cpu
;
2004 struct cpu_workqueue_struct
*cwq
= get_cwq(tcpu
, wq
);
2005 struct global_cwq
*gcwq
= cwq
->gcwq
;
2006 struct work_struct
*work
, *n
;
2008 __set_current_state(TASK_RUNNING
);
2009 mayday_clear_cpu(cpu
, wq
->mayday_mask
);
2011 /* migrate to the target cpu if possible */
2012 rescuer
->gcwq
= gcwq
;
2013 worker_maybe_bind_and_lock(rescuer
);
2016 * Slurp in all works issued via this workqueue and
2019 BUG_ON(!list_empty(&rescuer
->scheduled
));
2020 list_for_each_entry_safe(work
, n
, &gcwq
->worklist
, entry
)
2021 if (get_work_cwq(work
) == cwq
)
2022 move_linked_works(work
, scheduled
, &n
);
2024 process_scheduled_works(rescuer
);
2025 spin_unlock_irq(&gcwq
->lock
);
2033 struct work_struct work
;
2034 struct completion done
;
2037 static void wq_barrier_func(struct work_struct
*work
)
2039 struct wq_barrier
*barr
= container_of(work
, struct wq_barrier
, work
);
2040 complete(&barr
->done
);
2044 * insert_wq_barrier - insert a barrier work
2045 * @cwq: cwq to insert barrier into
2046 * @barr: wq_barrier to insert
2047 * @target: target work to attach @barr to
2048 * @worker: worker currently executing @target, NULL if @target is not executing
2050 * @barr is linked to @target such that @barr is completed only after
2051 * @target finishes execution. Please note that the ordering
2052 * guarantee is observed only with respect to @target and on the local
2055 * Currently, a queued barrier can't be canceled. This is because
2056 * try_to_grab_pending() can't determine whether the work to be
2057 * grabbed is at the head of the queue and thus can't clear LINKED
2058 * flag of the previous work while there must be a valid next work
2059 * after a work with LINKED flag set.
2061 * Note that when @worker is non-NULL, @target may be modified
2062 * underneath us, so we can't reliably determine cwq from @target.
2065 * spin_lock_irq(gcwq->lock).
2067 static void insert_wq_barrier(struct cpu_workqueue_struct
*cwq
,
2068 struct wq_barrier
*barr
,
2069 struct work_struct
*target
, struct worker
*worker
)
2071 struct list_head
*head
;
2072 unsigned int linked
= 0;
2075 * debugobject calls are safe here even with gcwq->lock locked
2076 * as we know for sure that this will not trigger any of the
2077 * checks and call back into the fixup functions where we
2080 INIT_WORK_ON_STACK(&barr
->work
, wq_barrier_func
);
2081 __set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(&barr
->work
));
2082 init_completion(&barr
->done
);
2085 * If @target is currently being executed, schedule the
2086 * barrier to the worker; otherwise, put it after @target.
2089 head
= worker
->scheduled
.next
;
2091 unsigned long *bits
= work_data_bits(target
);
2093 head
= target
->entry
.next
;
2094 /* there can already be other linked works, inherit and set */
2095 linked
= *bits
& WORK_STRUCT_LINKED
;
2096 __set_bit(WORK_STRUCT_LINKED_BIT
, bits
);
2099 debug_work_activate(&barr
->work
);
2100 insert_work(cwq
, &barr
->work
, head
,
2101 work_color_to_flags(WORK_NO_COLOR
) | linked
);
2105 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2106 * @wq: workqueue being flushed
2107 * @flush_color: new flush color, < 0 for no-op
2108 * @work_color: new work color, < 0 for no-op
2110 * Prepare cwqs for workqueue flushing.
2112 * If @flush_color is non-negative, flush_color on all cwqs should be
2113 * -1. If no cwq has in-flight commands at the specified color, all
2114 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2115 * has in flight commands, its cwq->flush_color is set to
2116 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2117 * wakeup logic is armed and %true is returned.
2119 * The caller should have initialized @wq->first_flusher prior to
2120 * calling this function with non-negative @flush_color. If
2121 * @flush_color is negative, no flush color update is done and %false
2124 * If @work_color is non-negative, all cwqs should have the same
2125 * work_color which is previous to @work_color and all will be
2126 * advanced to @work_color.
2129 * mutex_lock(wq->flush_mutex).
2132 * %true if @flush_color >= 0 and there's something to flush. %false
2135 static bool flush_workqueue_prep_cwqs(struct workqueue_struct
*wq
,
2136 int flush_color
, int work_color
)
2141 if (flush_color
>= 0) {
2142 BUG_ON(atomic_read(&wq
->nr_cwqs_to_flush
));
2143 atomic_set(&wq
->nr_cwqs_to_flush
, 1);
2146 for_each_cwq_cpu(cpu
, wq
) {
2147 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2148 struct global_cwq
*gcwq
= cwq
->gcwq
;
2150 spin_lock_irq(&gcwq
->lock
);
2152 if (flush_color
>= 0) {
2153 BUG_ON(cwq
->flush_color
!= -1);
2155 if (cwq
->nr_in_flight
[flush_color
]) {
2156 cwq
->flush_color
= flush_color
;
2157 atomic_inc(&wq
->nr_cwqs_to_flush
);
2162 if (work_color
>= 0) {
2163 BUG_ON(work_color
!= work_next_color(cwq
->work_color
));
2164 cwq
->work_color
= work_color
;
2167 spin_unlock_irq(&gcwq
->lock
);
2170 if (flush_color
>= 0 && atomic_dec_and_test(&wq
->nr_cwqs_to_flush
))
2171 complete(&wq
->first_flusher
->done
);
2177 * flush_workqueue - ensure that any scheduled work has run to completion.
2178 * @wq: workqueue to flush
2180 * Forces execution of the workqueue and blocks until its completion.
2181 * This is typically used in driver shutdown handlers.
2183 * We sleep until all works which were queued on entry have been handled,
2184 * but we are not livelocked by new incoming ones.
2186 void flush_workqueue(struct workqueue_struct
*wq
)
2188 struct wq_flusher this_flusher
= {
2189 .list
= LIST_HEAD_INIT(this_flusher
.list
),
2191 .done
= COMPLETION_INITIALIZER_ONSTACK(this_flusher
.done
),
2195 lock_map_acquire(&wq
->lockdep_map
);
2196 lock_map_release(&wq
->lockdep_map
);
2198 mutex_lock(&wq
->flush_mutex
);
2201 * Start-to-wait phase
2203 next_color
= work_next_color(wq
->work_color
);
2205 if (next_color
!= wq
->flush_color
) {
2207 * Color space is not full. The current work_color
2208 * becomes our flush_color and work_color is advanced
2211 BUG_ON(!list_empty(&wq
->flusher_overflow
));
2212 this_flusher
.flush_color
= wq
->work_color
;
2213 wq
->work_color
= next_color
;
2215 if (!wq
->first_flusher
) {
2216 /* no flush in progress, become the first flusher */
2217 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2219 wq
->first_flusher
= &this_flusher
;
2221 if (!flush_workqueue_prep_cwqs(wq
, wq
->flush_color
,
2223 /* nothing to flush, done */
2224 wq
->flush_color
= next_color
;
2225 wq
->first_flusher
= NULL
;
2230 BUG_ON(wq
->flush_color
== this_flusher
.flush_color
);
2231 list_add_tail(&this_flusher
.list
, &wq
->flusher_queue
);
2232 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2236 * Oops, color space is full, wait on overflow queue.
2237 * The next flush completion will assign us
2238 * flush_color and transfer to flusher_queue.
2240 list_add_tail(&this_flusher
.list
, &wq
->flusher_overflow
);
2243 mutex_unlock(&wq
->flush_mutex
);
2245 wait_for_completion(&this_flusher
.done
);
2248 * Wake-up-and-cascade phase
2250 * First flushers are responsible for cascading flushes and
2251 * handling overflow. Non-first flushers can simply return.
2253 if (wq
->first_flusher
!= &this_flusher
)
2256 mutex_lock(&wq
->flush_mutex
);
2258 /* we might have raced, check again with mutex held */
2259 if (wq
->first_flusher
!= &this_flusher
)
2262 wq
->first_flusher
= NULL
;
2264 BUG_ON(!list_empty(&this_flusher
.list
));
2265 BUG_ON(wq
->flush_color
!= this_flusher
.flush_color
);
2268 struct wq_flusher
*next
, *tmp
;
2270 /* complete all the flushers sharing the current flush color */
2271 list_for_each_entry_safe(next
, tmp
, &wq
->flusher_queue
, list
) {
2272 if (next
->flush_color
!= wq
->flush_color
)
2274 list_del_init(&next
->list
);
2275 complete(&next
->done
);
2278 BUG_ON(!list_empty(&wq
->flusher_overflow
) &&
2279 wq
->flush_color
!= work_next_color(wq
->work_color
));
2281 /* this flush_color is finished, advance by one */
2282 wq
->flush_color
= work_next_color(wq
->flush_color
);
2284 /* one color has been freed, handle overflow queue */
2285 if (!list_empty(&wq
->flusher_overflow
)) {
2287 * Assign the same color to all overflowed
2288 * flushers, advance work_color and append to
2289 * flusher_queue. This is the start-to-wait
2290 * phase for these overflowed flushers.
2292 list_for_each_entry(tmp
, &wq
->flusher_overflow
, list
)
2293 tmp
->flush_color
= wq
->work_color
;
2295 wq
->work_color
= work_next_color(wq
->work_color
);
2297 list_splice_tail_init(&wq
->flusher_overflow
,
2298 &wq
->flusher_queue
);
2299 flush_workqueue_prep_cwqs(wq
, -1, wq
->work_color
);
2302 if (list_empty(&wq
->flusher_queue
)) {
2303 BUG_ON(wq
->flush_color
!= wq
->work_color
);
2308 * Need to flush more colors. Make the next flusher
2309 * the new first flusher and arm cwqs.
2311 BUG_ON(wq
->flush_color
== wq
->work_color
);
2312 BUG_ON(wq
->flush_color
!= next
->flush_color
);
2314 list_del_init(&next
->list
);
2315 wq
->first_flusher
= next
;
2317 if (flush_workqueue_prep_cwqs(wq
, wq
->flush_color
, -1))
2321 * Meh... this color is already done, clear first
2322 * flusher and repeat cascading.
2324 wq
->first_flusher
= NULL
;
2328 mutex_unlock(&wq
->flush_mutex
);
2330 EXPORT_SYMBOL_GPL(flush_workqueue
);
2332 static bool start_flush_work(struct work_struct
*work
, struct wq_barrier
*barr
,
2333 bool wait_executing
)
2335 struct worker
*worker
= NULL
;
2336 struct global_cwq
*gcwq
;
2337 struct cpu_workqueue_struct
*cwq
;
2340 gcwq
= get_work_gcwq(work
);
2344 spin_lock_irq(&gcwq
->lock
);
2345 if (!list_empty(&work
->entry
)) {
2347 * See the comment near try_to_grab_pending()->smp_rmb().
2348 * If it was re-queued to a different gcwq under us, we
2349 * are not going to wait.
2352 cwq
= get_work_cwq(work
);
2353 if (unlikely(!cwq
|| gcwq
!= cwq
->gcwq
))
2355 } else if (wait_executing
) {
2356 worker
= find_worker_executing_work(gcwq
, work
);
2359 cwq
= worker
->current_cwq
;
2363 insert_wq_barrier(cwq
, barr
, work
, worker
);
2364 spin_unlock_irq(&gcwq
->lock
);
2366 lock_map_acquire(&cwq
->wq
->lockdep_map
);
2367 lock_map_release(&cwq
->wq
->lockdep_map
);
2370 spin_unlock_irq(&gcwq
->lock
);
2375 * flush_work - wait for a work to finish executing the last queueing instance
2376 * @work: the work to flush
2378 * Wait until @work has finished execution. This function considers
2379 * only the last queueing instance of @work. If @work has been
2380 * enqueued across different CPUs on a non-reentrant workqueue or on
2381 * multiple workqueues, @work might still be executing on return on
2382 * some of the CPUs from earlier queueing.
2384 * If @work was queued only on a non-reentrant, ordered or unbound
2385 * workqueue, @work is guaranteed to be idle on return if it hasn't
2386 * been requeued since flush started.
2389 * %true if flush_work() waited for the work to finish execution,
2390 * %false if it was already idle.
2392 bool flush_work(struct work_struct
*work
)
2394 struct wq_barrier barr
;
2396 if (start_flush_work(work
, &barr
, true)) {
2397 wait_for_completion(&barr
.done
);
2398 destroy_work_on_stack(&barr
.work
);
2403 EXPORT_SYMBOL_GPL(flush_work
);
2405 static bool wait_on_cpu_work(struct global_cwq
*gcwq
, struct work_struct
*work
)
2407 struct wq_barrier barr
;
2408 struct worker
*worker
;
2410 spin_lock_irq(&gcwq
->lock
);
2412 worker
= find_worker_executing_work(gcwq
, work
);
2413 if (unlikely(worker
))
2414 insert_wq_barrier(worker
->current_cwq
, &barr
, work
, worker
);
2416 spin_unlock_irq(&gcwq
->lock
);
2418 if (unlikely(worker
)) {
2419 wait_for_completion(&barr
.done
);
2420 destroy_work_on_stack(&barr
.work
);
2426 static bool wait_on_work(struct work_struct
*work
)
2433 lock_map_acquire(&work
->lockdep_map
);
2434 lock_map_release(&work
->lockdep_map
);
2436 for_each_gcwq_cpu(cpu
)
2437 ret
|= wait_on_cpu_work(get_gcwq(cpu
), work
);
2442 * flush_work_sync - wait until a work has finished execution
2443 * @work: the work to flush
2445 * Wait until @work has finished execution. On return, it's
2446 * guaranteed that all queueing instances of @work which happened
2447 * before this function is called are finished. In other words, if
2448 * @work hasn't been requeued since this function was called, @work is
2449 * guaranteed to be idle on return.
2452 * %true if flush_work_sync() waited for the work to finish execution,
2453 * %false if it was already idle.
2455 bool flush_work_sync(struct work_struct
*work
)
2457 struct wq_barrier barr
;
2458 bool pending
, waited
;
2460 /* we'll wait for executions separately, queue barr only if pending */
2461 pending
= start_flush_work(work
, &barr
, false);
2463 /* wait for executions to finish */
2464 waited
= wait_on_work(work
);
2466 /* wait for the pending one */
2468 wait_for_completion(&barr
.done
);
2469 destroy_work_on_stack(&barr
.work
);
2472 return pending
|| waited
;
2474 EXPORT_SYMBOL_GPL(flush_work_sync
);
2477 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2478 * so this work can't be re-armed in any way.
2480 static int try_to_grab_pending(struct work_struct
*work
)
2482 struct global_cwq
*gcwq
;
2485 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT
, work_data_bits(work
)))
2489 * The queueing is in progress, or it is already queued. Try to
2490 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2492 gcwq
= get_work_gcwq(work
);
2496 spin_lock_irq(&gcwq
->lock
);
2497 if (!list_empty(&work
->entry
)) {
2499 * This work is queued, but perhaps we locked the wrong gcwq.
2500 * In that case we must see the new value after rmb(), see
2501 * insert_work()->wmb().
2504 if (gcwq
== get_work_gcwq(work
)) {
2505 debug_work_deactivate(work
);
2506 list_del_init(&work
->entry
);
2507 cwq_dec_nr_in_flight(get_work_cwq(work
),
2508 get_work_color(work
),
2509 *work_data_bits(work
) & WORK_STRUCT_DELAYED
);
2513 spin_unlock_irq(&gcwq
->lock
);
2518 static bool __cancel_work_timer(struct work_struct
*work
,
2519 struct timer_list
* timer
)
2524 ret
= (timer
&& likely(del_timer(timer
)));
2526 ret
= try_to_grab_pending(work
);
2528 } while (unlikely(ret
< 0));
2530 clear_work_data(work
);
2535 * cancel_work_sync - cancel a work and wait for it to finish
2536 * @work: the work to cancel
2538 * Cancel @work and wait for its execution to finish. This function
2539 * can be used even if the work re-queues itself or migrates to
2540 * another workqueue. On return from this function, @work is
2541 * guaranteed to be not pending or executing on any CPU.
2543 * cancel_work_sync(&delayed_work->work) must not be used for
2544 * delayed_work's. Use cancel_delayed_work_sync() instead.
2546 * The caller must ensure that the workqueue on which @work was last
2547 * queued can't be destroyed before this function returns.
2550 * %true if @work was pending, %false otherwise.
2552 bool cancel_work_sync(struct work_struct
*work
)
2554 return __cancel_work_timer(work
, NULL
);
2556 EXPORT_SYMBOL_GPL(cancel_work_sync
);
2559 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2560 * @dwork: the delayed work to flush
2562 * Delayed timer is cancelled and the pending work is queued for
2563 * immediate execution. Like flush_work(), this function only
2564 * considers the last queueing instance of @dwork.
2567 * %true if flush_work() waited for the work to finish execution,
2568 * %false if it was already idle.
2570 bool flush_delayed_work(struct delayed_work
*dwork
)
2572 if (del_timer_sync(&dwork
->timer
))
2573 __queue_work(raw_smp_processor_id(),
2574 get_work_cwq(&dwork
->work
)->wq
, &dwork
->work
);
2575 return flush_work(&dwork
->work
);
2577 EXPORT_SYMBOL(flush_delayed_work
);
2580 * flush_delayed_work_sync - wait for a dwork to finish
2581 * @dwork: the delayed work to flush
2583 * Delayed timer is cancelled and the pending work is queued for
2584 * execution immediately. Other than timer handling, its behavior
2585 * is identical to flush_work_sync().
2588 * %true if flush_work_sync() waited for the work to finish execution,
2589 * %false if it was already idle.
2591 bool flush_delayed_work_sync(struct delayed_work
*dwork
)
2593 if (del_timer_sync(&dwork
->timer
))
2594 __queue_work(raw_smp_processor_id(),
2595 get_work_cwq(&dwork
->work
)->wq
, &dwork
->work
);
2596 return flush_work_sync(&dwork
->work
);
2598 EXPORT_SYMBOL(flush_delayed_work_sync
);
2601 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2602 * @dwork: the delayed work cancel
2604 * This is cancel_work_sync() for delayed works.
2607 * %true if @dwork was pending, %false otherwise.
2609 bool cancel_delayed_work_sync(struct delayed_work
*dwork
)
2611 return __cancel_work_timer(&dwork
->work
, &dwork
->timer
);
2613 EXPORT_SYMBOL(cancel_delayed_work_sync
);
2616 * schedule_work - put work task in global workqueue
2617 * @work: job to be done
2619 * Returns zero if @work was already on the kernel-global workqueue and
2620 * non-zero otherwise.
2622 * This puts a job in the kernel-global workqueue if it was not already
2623 * queued and leaves it in the same position on the kernel-global
2624 * workqueue otherwise.
2626 int schedule_work(struct work_struct
*work
)
2628 return queue_work(system_wq
, work
);
2630 EXPORT_SYMBOL(schedule_work
);
2633 * schedule_work_on - put work task on a specific cpu
2634 * @cpu: cpu to put the work task on
2635 * @work: job to be done
2637 * This puts a job on a specific cpu
2639 int schedule_work_on(int cpu
, struct work_struct
*work
)
2641 return queue_work_on(cpu
, system_wq
, work
);
2643 EXPORT_SYMBOL(schedule_work_on
);
2646 * schedule_delayed_work - put work task in global workqueue after delay
2647 * @dwork: job to be done
2648 * @delay: number of jiffies to wait or 0 for immediate execution
2650 * After waiting for a given time this puts a job in the kernel-global
2653 int schedule_delayed_work(struct delayed_work
*dwork
,
2654 unsigned long delay
)
2656 return queue_delayed_work(system_wq
, dwork
, delay
);
2658 EXPORT_SYMBOL(schedule_delayed_work
);
2661 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2663 * @dwork: job to be done
2664 * @delay: number of jiffies to wait
2666 * After waiting for a given time this puts a job in the kernel-global
2667 * workqueue on the specified CPU.
2669 int schedule_delayed_work_on(int cpu
,
2670 struct delayed_work
*dwork
, unsigned long delay
)
2672 return queue_delayed_work_on(cpu
, system_wq
, dwork
, delay
);
2674 EXPORT_SYMBOL(schedule_delayed_work_on
);
2677 * schedule_on_each_cpu - call a function on each online CPU from keventd
2678 * @func: the function to call
2680 * Returns zero on success.
2681 * Returns -ve errno on failure.
2683 * schedule_on_each_cpu() is very slow.
2685 int schedule_on_each_cpu(work_func_t func
)
2688 struct work_struct __percpu
*works
;
2690 works
= alloc_percpu(struct work_struct
);
2696 for_each_online_cpu(cpu
) {
2697 struct work_struct
*work
= per_cpu_ptr(works
, cpu
);
2699 INIT_WORK(work
, func
);
2700 schedule_work_on(cpu
, work
);
2703 for_each_online_cpu(cpu
)
2704 flush_work(per_cpu_ptr(works
, cpu
));
2712 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2714 * Forces execution of the kernel-global workqueue and blocks until its
2717 * Think twice before calling this function! It's very easy to get into
2718 * trouble if you don't take great care. Either of the following situations
2719 * will lead to deadlock:
2721 * One of the work items currently on the workqueue needs to acquire
2722 * a lock held by your code or its caller.
2724 * Your code is running in the context of a work routine.
2726 * They will be detected by lockdep when they occur, but the first might not
2727 * occur very often. It depends on what work items are on the workqueue and
2728 * what locks they need, which you have no control over.
2730 * In most situations flushing the entire workqueue is overkill; you merely
2731 * need to know that a particular work item isn't queued and isn't running.
2732 * In such cases you should use cancel_delayed_work_sync() or
2733 * cancel_work_sync() instead.
2735 void flush_scheduled_work(void)
2737 flush_workqueue(system_wq
);
2739 EXPORT_SYMBOL(flush_scheduled_work
);
2742 * execute_in_process_context - reliably execute the routine with user context
2743 * @fn: the function to execute
2744 * @ew: guaranteed storage for the execute work structure (must
2745 * be available when the work executes)
2747 * Executes the function immediately if process context is available,
2748 * otherwise schedules the function for delayed execution.
2750 * Returns: 0 - function was executed
2751 * 1 - function was scheduled for execution
2753 int execute_in_process_context(work_func_t fn
, struct execute_work
*ew
)
2755 if (!in_interrupt()) {
2760 INIT_WORK(&ew
->work
, fn
);
2761 schedule_work(&ew
->work
);
2765 EXPORT_SYMBOL_GPL(execute_in_process_context
);
2767 int keventd_up(void)
2769 return system_wq
!= NULL
;
2772 static int alloc_cwqs(struct workqueue_struct
*wq
)
2775 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2776 * Make sure that the alignment isn't lower than that of
2777 * unsigned long long.
2779 const size_t size
= sizeof(struct cpu_workqueue_struct
);
2780 const size_t align
= max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS
,
2781 __alignof__(unsigned long long));
2783 bool percpu
= !(wq
->flags
& WQ_UNBOUND
);
2785 bool percpu
= false;
2789 wq
->cpu_wq
.pcpu
= __alloc_percpu(size
, align
);
2794 * Allocate enough room to align cwq and put an extra
2795 * pointer at the end pointing back to the originally
2796 * allocated pointer which will be used for free.
2798 ptr
= kzalloc(size
+ align
+ sizeof(void *), GFP_KERNEL
);
2800 wq
->cpu_wq
.single
= PTR_ALIGN(ptr
, align
);
2801 *(void **)(wq
->cpu_wq
.single
+ 1) = ptr
;
2805 /* just in case, make sure it's actually aligned */
2806 BUG_ON(!IS_ALIGNED(wq
->cpu_wq
.v
, align
));
2807 return wq
->cpu_wq
.v
? 0 : -ENOMEM
;
2810 static void free_cwqs(struct workqueue_struct
*wq
)
2813 bool percpu
= !(wq
->flags
& WQ_UNBOUND
);
2815 bool percpu
= false;
2819 free_percpu(wq
->cpu_wq
.pcpu
);
2820 else if (wq
->cpu_wq
.single
) {
2821 /* the pointer to free is stored right after the cwq */
2822 kfree(*(void **)(wq
->cpu_wq
.single
+ 1));
2826 static int wq_clamp_max_active(int max_active
, unsigned int flags
,
2829 int lim
= flags
& WQ_UNBOUND
? WQ_UNBOUND_MAX_ACTIVE
: WQ_MAX_ACTIVE
;
2831 if (max_active
< 1 || max_active
> lim
)
2832 printk(KERN_WARNING
"workqueue: max_active %d requested for %s "
2833 "is out of range, clamping between %d and %d\n",
2834 max_active
, name
, 1, lim
);
2836 return clamp_val(max_active
, 1, lim
);
2839 struct workqueue_struct
*__alloc_workqueue_key(const char *name
,
2842 struct lock_class_key
*key
,
2843 const char *lock_name
)
2845 struct workqueue_struct
*wq
;
2849 * Unbound workqueues aren't concurrency managed and should be
2850 * dispatched to workers immediately.
2852 if (flags
& WQ_UNBOUND
)
2853 flags
|= WQ_HIGHPRI
;
2855 max_active
= max_active
?: WQ_DFL_ACTIVE
;
2856 max_active
= wq_clamp_max_active(max_active
, flags
, name
);
2858 wq
= kzalloc(sizeof(*wq
), GFP_KERNEL
);
2863 wq
->saved_max_active
= max_active
;
2864 mutex_init(&wq
->flush_mutex
);
2865 atomic_set(&wq
->nr_cwqs_to_flush
, 0);
2866 INIT_LIST_HEAD(&wq
->flusher_queue
);
2867 INIT_LIST_HEAD(&wq
->flusher_overflow
);
2870 lockdep_init_map(&wq
->lockdep_map
, lock_name
, key
, 0);
2871 INIT_LIST_HEAD(&wq
->list
);
2873 if (alloc_cwqs(wq
) < 0)
2876 for_each_cwq_cpu(cpu
, wq
) {
2877 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2878 struct global_cwq
*gcwq
= get_gcwq(cpu
);
2880 BUG_ON((unsigned long)cwq
& WORK_STRUCT_FLAG_MASK
);
2883 cwq
->flush_color
= -1;
2884 cwq
->max_active
= max_active
;
2885 INIT_LIST_HEAD(&cwq
->delayed_works
);
2888 if (flags
& WQ_RESCUER
) {
2889 struct worker
*rescuer
;
2891 if (!alloc_mayday_mask(&wq
->mayday_mask
, GFP_KERNEL
))
2894 wq
->rescuer
= rescuer
= alloc_worker();
2898 rescuer
->task
= kthread_create(rescuer_thread
, wq
, "%s", name
);
2899 if (IS_ERR(rescuer
->task
))
2902 rescuer
->task
->flags
|= PF_THREAD_BOUND
;
2903 wake_up_process(rescuer
->task
);
2907 * workqueue_lock protects global freeze state and workqueues
2908 * list. Grab it, set max_active accordingly and add the new
2909 * workqueue to workqueues list.
2911 spin_lock(&workqueue_lock
);
2913 if (workqueue_freezing
&& wq
->flags
& WQ_FREEZEABLE
)
2914 for_each_cwq_cpu(cpu
, wq
)
2915 get_cwq(cpu
, wq
)->max_active
= 0;
2917 list_add(&wq
->list
, &workqueues
);
2919 spin_unlock(&workqueue_lock
);
2925 free_mayday_mask(wq
->mayday_mask
);
2931 EXPORT_SYMBOL_GPL(__alloc_workqueue_key
);
2934 * destroy_workqueue - safely terminate a workqueue
2935 * @wq: target workqueue
2937 * Safely destroy a workqueue. All work currently pending will be done first.
2939 void destroy_workqueue(struct workqueue_struct
*wq
)
2943 wq
->flags
|= WQ_DYING
;
2944 flush_workqueue(wq
);
2947 * wq list is used to freeze wq, remove from list after
2948 * flushing is complete in case freeze races us.
2950 spin_lock(&workqueue_lock
);
2951 list_del(&wq
->list
);
2952 spin_unlock(&workqueue_lock
);
2955 for_each_cwq_cpu(cpu
, wq
) {
2956 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
2959 for (i
= 0; i
< WORK_NR_COLORS
; i
++)
2960 BUG_ON(cwq
->nr_in_flight
[i
]);
2961 BUG_ON(cwq
->nr_active
);
2962 BUG_ON(!list_empty(&cwq
->delayed_works
));
2965 if (wq
->flags
& WQ_RESCUER
) {
2966 kthread_stop(wq
->rescuer
->task
);
2967 free_mayday_mask(wq
->mayday_mask
);
2974 EXPORT_SYMBOL_GPL(destroy_workqueue
);
2977 * workqueue_set_max_active - adjust max_active of a workqueue
2978 * @wq: target workqueue
2979 * @max_active: new max_active value.
2981 * Set max_active of @wq to @max_active.
2984 * Don't call from IRQ context.
2986 void workqueue_set_max_active(struct workqueue_struct
*wq
, int max_active
)
2990 max_active
= wq_clamp_max_active(max_active
, wq
->flags
, wq
->name
);
2992 spin_lock(&workqueue_lock
);
2994 wq
->saved_max_active
= max_active
;
2996 for_each_cwq_cpu(cpu
, wq
) {
2997 struct global_cwq
*gcwq
= get_gcwq(cpu
);
2999 spin_lock_irq(&gcwq
->lock
);
3001 if (!(wq
->flags
& WQ_FREEZEABLE
) ||
3002 !(gcwq
->flags
& GCWQ_FREEZING
))
3003 get_cwq(gcwq
->cpu
, wq
)->max_active
= max_active
;
3005 spin_unlock_irq(&gcwq
->lock
);
3008 spin_unlock(&workqueue_lock
);
3010 EXPORT_SYMBOL_GPL(workqueue_set_max_active
);
3013 * workqueue_congested - test whether a workqueue is congested
3014 * @cpu: CPU in question
3015 * @wq: target workqueue
3017 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3018 * no synchronization around this function and the test result is
3019 * unreliable and only useful as advisory hints or for debugging.
3022 * %true if congested, %false otherwise.
3024 bool workqueue_congested(unsigned int cpu
, struct workqueue_struct
*wq
)
3026 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3028 return !list_empty(&cwq
->delayed_works
);
3030 EXPORT_SYMBOL_GPL(workqueue_congested
);
3033 * work_cpu - return the last known associated cpu for @work
3034 * @work: the work of interest
3037 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3039 unsigned int work_cpu(struct work_struct
*work
)
3041 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3043 return gcwq
? gcwq
->cpu
: WORK_CPU_NONE
;
3045 EXPORT_SYMBOL_GPL(work_cpu
);
3048 * work_busy - test whether a work is currently pending or running
3049 * @work: the work to be tested
3051 * Test whether @work is currently pending or running. There is no
3052 * synchronization around this function and the test result is
3053 * unreliable and only useful as advisory hints or for debugging.
3054 * Especially for reentrant wqs, the pending state might hide the
3058 * OR'd bitmask of WORK_BUSY_* bits.
3060 unsigned int work_busy(struct work_struct
*work
)
3062 struct global_cwq
*gcwq
= get_work_gcwq(work
);
3063 unsigned long flags
;
3064 unsigned int ret
= 0;
3069 spin_lock_irqsave(&gcwq
->lock
, flags
);
3071 if (work_pending(work
))
3072 ret
|= WORK_BUSY_PENDING
;
3073 if (find_worker_executing_work(gcwq
, work
))
3074 ret
|= WORK_BUSY_RUNNING
;
3076 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
3080 EXPORT_SYMBOL_GPL(work_busy
);
3085 * There are two challenges in supporting CPU hotplug. Firstly, there
3086 * are a lot of assumptions on strong associations among work, cwq and
3087 * gcwq which make migrating pending and scheduled works very
3088 * difficult to implement without impacting hot paths. Secondly,
3089 * gcwqs serve mix of short, long and very long running works making
3090 * blocked draining impractical.
3092 * This is solved by allowing a gcwq to be detached from CPU, running
3093 * it with unbound (rogue) workers and allowing it to be reattached
3094 * later if the cpu comes back online. A separate thread is created
3095 * to govern a gcwq in such state and is called the trustee of the
3098 * Trustee states and their descriptions.
3100 * START Command state used on startup. On CPU_DOWN_PREPARE, a
3101 * new trustee is started with this state.
3103 * IN_CHARGE Once started, trustee will enter this state after
3104 * assuming the manager role and making all existing
3105 * workers rogue. DOWN_PREPARE waits for trustee to
3106 * enter this state. After reaching IN_CHARGE, trustee
3107 * tries to execute the pending worklist until it's empty
3108 * and the state is set to BUTCHER, or the state is set
3111 * BUTCHER Command state which is set by the cpu callback after
3112 * the cpu has went down. Once this state is set trustee
3113 * knows that there will be no new works on the worklist
3114 * and once the worklist is empty it can proceed to
3115 * killing idle workers.
3117 * RELEASE Command state which is set by the cpu callback if the
3118 * cpu down has been canceled or it has come online
3119 * again. After recognizing this state, trustee stops
3120 * trying to drain or butcher and clears ROGUE, rebinds
3121 * all remaining workers back to the cpu and releases
3124 * DONE Trustee will enter this state after BUTCHER or RELEASE
3127 * trustee CPU draining
3128 * took over down complete
3129 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3131 * | CPU is back online v return workers |
3132 * ----------------> RELEASE --------------
3136 * trustee_wait_event_timeout - timed event wait for trustee
3137 * @cond: condition to wait for
3138 * @timeout: timeout in jiffies
3140 * wait_event_timeout() for trustee to use. Handles locking and
3141 * checks for RELEASE request.
3144 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3145 * multiple times. To be used by trustee.
3148 * Positive indicating left time if @cond is satisfied, 0 if timed
3149 * out, -1 if canceled.
3151 #define trustee_wait_event_timeout(cond, timeout) ({ \
3152 long __ret = (timeout); \
3153 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3155 spin_unlock_irq(&gcwq->lock); \
3156 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3157 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3159 spin_lock_irq(&gcwq->lock); \
3161 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3165 * trustee_wait_event - event wait for trustee
3166 * @cond: condition to wait for
3168 * wait_event() for trustee to use. Automatically handles locking and
3169 * checks for CANCEL request.
3172 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3173 * multiple times. To be used by trustee.
3176 * 0 if @cond is satisfied, -1 if canceled.
3178 #define trustee_wait_event(cond) ({ \
3180 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3181 __ret1 < 0 ? -1 : 0; \
3184 static int __cpuinit
trustee_thread(void *__gcwq
)
3186 struct global_cwq
*gcwq
= __gcwq
;
3187 struct worker
*worker
;
3188 struct work_struct
*work
;
3189 struct hlist_node
*pos
;
3193 BUG_ON(gcwq
->cpu
!= smp_processor_id());
3195 spin_lock_irq(&gcwq
->lock
);
3197 * Claim the manager position and make all workers rogue.
3198 * Trustee must be bound to the target cpu and can't be
3201 BUG_ON(gcwq
->cpu
!= smp_processor_id());
3202 rc
= trustee_wait_event(!(gcwq
->flags
& GCWQ_MANAGING_WORKERS
));
3205 gcwq
->flags
|= GCWQ_MANAGING_WORKERS
;
3207 list_for_each_entry(worker
, &gcwq
->idle_list
, entry
)
3208 worker
->flags
|= WORKER_ROGUE
;
3210 for_each_busy_worker(worker
, i
, pos
, gcwq
)
3211 worker
->flags
|= WORKER_ROGUE
;
3214 * Call schedule() so that we cross rq->lock and thus can
3215 * guarantee sched callbacks see the rogue flag. This is
3216 * necessary as scheduler callbacks may be invoked from other
3219 spin_unlock_irq(&gcwq
->lock
);
3221 spin_lock_irq(&gcwq
->lock
);
3224 * Sched callbacks are disabled now. Zap nr_running. After
3225 * this, nr_running stays zero and need_more_worker() and
3226 * keep_working() are always true as long as the worklist is
3229 atomic_set(get_gcwq_nr_running(gcwq
->cpu
), 0);
3231 spin_unlock_irq(&gcwq
->lock
);
3232 del_timer_sync(&gcwq
->idle_timer
);
3233 spin_lock_irq(&gcwq
->lock
);
3236 * We're now in charge. Notify and proceed to drain. We need
3237 * to keep the gcwq running during the whole CPU down
3238 * procedure as other cpu hotunplug callbacks may need to
3239 * flush currently running tasks.
3241 gcwq
->trustee_state
= TRUSTEE_IN_CHARGE
;
3242 wake_up_all(&gcwq
->trustee_wait
);
3245 * The original cpu is in the process of dying and may go away
3246 * anytime now. When that happens, we and all workers would
3247 * be migrated to other cpus. Try draining any left work. We
3248 * want to get it over with ASAP - spam rescuers, wake up as
3249 * many idlers as necessary and create new ones till the
3250 * worklist is empty. Note that if the gcwq is frozen, there
3251 * may be frozen works in freezeable cwqs. Don't declare
3252 * completion while frozen.
3254 while (gcwq
->nr_workers
!= gcwq
->nr_idle
||
3255 gcwq
->flags
& GCWQ_FREEZING
||
3256 gcwq
->trustee_state
== TRUSTEE_IN_CHARGE
) {
3259 list_for_each_entry(work
, &gcwq
->worklist
, entry
) {
3264 list_for_each_entry(worker
, &gcwq
->idle_list
, entry
) {
3267 wake_up_process(worker
->task
);
3270 if (need_to_create_worker(gcwq
)) {
3271 spin_unlock_irq(&gcwq
->lock
);
3272 worker
= create_worker(gcwq
, false);
3273 spin_lock_irq(&gcwq
->lock
);
3275 worker
->flags
|= WORKER_ROGUE
;
3276 start_worker(worker
);
3280 /* give a breather */
3281 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN
) < 0)
3286 * Either all works have been scheduled and cpu is down, or
3287 * cpu down has already been canceled. Wait for and butcher
3288 * all workers till we're canceled.
3291 rc
= trustee_wait_event(!list_empty(&gcwq
->idle_list
));
3292 while (!list_empty(&gcwq
->idle_list
))
3293 destroy_worker(list_first_entry(&gcwq
->idle_list
,
3294 struct worker
, entry
));
3295 } while (gcwq
->nr_workers
&& rc
>= 0);
3298 * At this point, either draining has completed and no worker
3299 * is left, or cpu down has been canceled or the cpu is being
3300 * brought back up. There shouldn't be any idle one left.
3301 * Tell the remaining busy ones to rebind once it finishes the
3302 * currently scheduled works by scheduling the rebind_work.
3304 WARN_ON(!list_empty(&gcwq
->idle_list
));
3306 for_each_busy_worker(worker
, i
, pos
, gcwq
) {
3307 struct work_struct
*rebind_work
= &worker
->rebind_work
;
3310 * Rebind_work may race with future cpu hotplug
3311 * operations. Use a separate flag to mark that
3312 * rebinding is scheduled.
3314 worker
->flags
|= WORKER_REBIND
;
3315 worker
->flags
&= ~WORKER_ROGUE
;
3317 /* queue rebind_work, wq doesn't matter, use the default one */
3318 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT
,
3319 work_data_bits(rebind_work
)))
3322 debug_work_activate(rebind_work
);
3323 insert_work(get_cwq(gcwq
->cpu
, system_wq
), rebind_work
,
3324 worker
->scheduled
.next
,
3325 work_color_to_flags(WORK_NO_COLOR
));
3328 /* relinquish manager role */
3329 gcwq
->flags
&= ~GCWQ_MANAGING_WORKERS
;
3331 /* notify completion */
3332 gcwq
->trustee
= NULL
;
3333 gcwq
->trustee_state
= TRUSTEE_DONE
;
3334 wake_up_all(&gcwq
->trustee_wait
);
3335 spin_unlock_irq(&gcwq
->lock
);
3340 * wait_trustee_state - wait for trustee to enter the specified state
3341 * @gcwq: gcwq the trustee of interest belongs to
3342 * @state: target state to wait for
3344 * Wait for the trustee to reach @state. DONE is already matched.
3347 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3348 * multiple times. To be used by cpu_callback.
3350 static void __cpuinit
wait_trustee_state(struct global_cwq
*gcwq
, int state
)
3351 __releases(&gcwq
->lock
)
3352 __acquires(&gcwq
->lock
)
3354 if (!(gcwq
->trustee_state
== state
||
3355 gcwq
->trustee_state
== TRUSTEE_DONE
)) {
3356 spin_unlock_irq(&gcwq
->lock
);
3357 __wait_event(gcwq
->trustee_wait
,
3358 gcwq
->trustee_state
== state
||
3359 gcwq
->trustee_state
== TRUSTEE_DONE
);
3360 spin_lock_irq(&gcwq
->lock
);
3364 static int __devinit
workqueue_cpu_callback(struct notifier_block
*nfb
,
3365 unsigned long action
,
3368 unsigned int cpu
= (unsigned long)hcpu
;
3369 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3370 struct task_struct
*new_trustee
= NULL
;
3371 struct worker
*uninitialized_var(new_worker
);
3372 unsigned long flags
;
3374 action
&= ~CPU_TASKS_FROZEN
;
3377 case CPU_DOWN_PREPARE
:
3378 new_trustee
= kthread_create(trustee_thread
, gcwq
,
3379 "workqueue_trustee/%d\n", cpu
);
3380 if (IS_ERR(new_trustee
))
3381 return notifier_from_errno(PTR_ERR(new_trustee
));
3382 kthread_bind(new_trustee
, cpu
);
3384 case CPU_UP_PREPARE
:
3385 BUG_ON(gcwq
->first_idle
);
3386 new_worker
= create_worker(gcwq
, false);
3389 kthread_stop(new_trustee
);
3394 /* some are called w/ irq disabled, don't disturb irq status */
3395 spin_lock_irqsave(&gcwq
->lock
, flags
);
3398 case CPU_DOWN_PREPARE
:
3399 /* initialize trustee and tell it to acquire the gcwq */
3400 BUG_ON(gcwq
->trustee
|| gcwq
->trustee_state
!= TRUSTEE_DONE
);
3401 gcwq
->trustee
= new_trustee
;
3402 gcwq
->trustee_state
= TRUSTEE_START
;
3403 wake_up_process(gcwq
->trustee
);
3404 wait_trustee_state(gcwq
, TRUSTEE_IN_CHARGE
);
3406 case CPU_UP_PREPARE
:
3407 BUG_ON(gcwq
->first_idle
);
3408 gcwq
->first_idle
= new_worker
;
3413 * Before this, the trustee and all workers except for
3414 * the ones which are still executing works from
3415 * before the last CPU down must be on the cpu. After
3416 * this, they'll all be diasporas.
3418 gcwq
->flags
|= GCWQ_DISASSOCIATED
;
3422 gcwq
->trustee_state
= TRUSTEE_BUTCHER
;
3424 case CPU_UP_CANCELED
:
3425 destroy_worker(gcwq
->first_idle
);
3426 gcwq
->first_idle
= NULL
;
3429 case CPU_DOWN_FAILED
:
3431 gcwq
->flags
&= ~GCWQ_DISASSOCIATED
;
3432 if (gcwq
->trustee_state
!= TRUSTEE_DONE
) {
3433 gcwq
->trustee_state
= TRUSTEE_RELEASE
;
3434 wake_up_process(gcwq
->trustee
);
3435 wait_trustee_state(gcwq
, TRUSTEE_DONE
);
3439 * Trustee is done and there might be no worker left.
3440 * Put the first_idle in and request a real manager to
3443 spin_unlock_irq(&gcwq
->lock
);
3444 kthread_bind(gcwq
->first_idle
->task
, cpu
);
3445 spin_lock_irq(&gcwq
->lock
);
3446 gcwq
->flags
|= GCWQ_MANAGE_WORKERS
;
3447 start_worker(gcwq
->first_idle
);
3448 gcwq
->first_idle
= NULL
;
3452 spin_unlock_irqrestore(&gcwq
->lock
, flags
);
3454 return notifier_from_errno(0);
3459 struct work_for_cpu
{
3460 struct completion completion
;
3466 static int do_work_for_cpu(void *_wfc
)
3468 struct work_for_cpu
*wfc
= _wfc
;
3469 wfc
->ret
= wfc
->fn(wfc
->arg
);
3470 complete(&wfc
->completion
);
3475 * work_on_cpu - run a function in user context on a particular cpu
3476 * @cpu: the cpu to run on
3477 * @fn: the function to run
3478 * @arg: the function arg
3480 * This will return the value @fn returns.
3481 * It is up to the caller to ensure that the cpu doesn't go offline.
3482 * The caller must not hold any locks which would prevent @fn from completing.
3484 long work_on_cpu(unsigned int cpu
, long (*fn
)(void *), void *arg
)
3486 struct task_struct
*sub_thread
;
3487 struct work_for_cpu wfc
= {
3488 .completion
= COMPLETION_INITIALIZER_ONSTACK(wfc
.completion
),
3493 sub_thread
= kthread_create(do_work_for_cpu
, &wfc
, "work_for_cpu");
3494 if (IS_ERR(sub_thread
))
3495 return PTR_ERR(sub_thread
);
3496 kthread_bind(sub_thread
, cpu
);
3497 wake_up_process(sub_thread
);
3498 wait_for_completion(&wfc
.completion
);
3501 EXPORT_SYMBOL_GPL(work_on_cpu
);
3502 #endif /* CONFIG_SMP */
3504 #ifdef CONFIG_FREEZER
3507 * freeze_workqueues_begin - begin freezing workqueues
3509 * Start freezing workqueues. After this function returns, all
3510 * freezeable workqueues will queue new works to their frozen_works
3511 * list instead of gcwq->worklist.
3514 * Grabs and releases workqueue_lock and gcwq->lock's.
3516 void freeze_workqueues_begin(void)
3520 spin_lock(&workqueue_lock
);
3522 BUG_ON(workqueue_freezing
);
3523 workqueue_freezing
= true;
3525 for_each_gcwq_cpu(cpu
) {
3526 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3527 struct workqueue_struct
*wq
;
3529 spin_lock_irq(&gcwq
->lock
);
3531 BUG_ON(gcwq
->flags
& GCWQ_FREEZING
);
3532 gcwq
->flags
|= GCWQ_FREEZING
;
3534 list_for_each_entry(wq
, &workqueues
, list
) {
3535 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3537 if (cwq
&& wq
->flags
& WQ_FREEZEABLE
)
3538 cwq
->max_active
= 0;
3541 spin_unlock_irq(&gcwq
->lock
);
3544 spin_unlock(&workqueue_lock
);
3548 * freeze_workqueues_busy - are freezeable workqueues still busy?
3550 * Check whether freezing is complete. This function must be called
3551 * between freeze_workqueues_begin() and thaw_workqueues().
3554 * Grabs and releases workqueue_lock.
3557 * %true if some freezeable workqueues are still busy. %false if
3558 * freezing is complete.
3560 bool freeze_workqueues_busy(void)
3565 spin_lock(&workqueue_lock
);
3567 BUG_ON(!workqueue_freezing
);
3569 for_each_gcwq_cpu(cpu
) {
3570 struct workqueue_struct
*wq
;
3572 * nr_active is monotonically decreasing. It's safe
3573 * to peek without lock.
3575 list_for_each_entry(wq
, &workqueues
, list
) {
3576 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3578 if (!cwq
|| !(wq
->flags
& WQ_FREEZEABLE
))
3581 BUG_ON(cwq
->nr_active
< 0);
3582 if (cwq
->nr_active
) {
3589 spin_unlock(&workqueue_lock
);
3594 * thaw_workqueues - thaw workqueues
3596 * Thaw workqueues. Normal queueing is restored and all collected
3597 * frozen works are transferred to their respective gcwq worklists.
3600 * Grabs and releases workqueue_lock and gcwq->lock's.
3602 void thaw_workqueues(void)
3606 spin_lock(&workqueue_lock
);
3608 if (!workqueue_freezing
)
3611 for_each_gcwq_cpu(cpu
) {
3612 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3613 struct workqueue_struct
*wq
;
3615 spin_lock_irq(&gcwq
->lock
);
3617 BUG_ON(!(gcwq
->flags
& GCWQ_FREEZING
));
3618 gcwq
->flags
&= ~GCWQ_FREEZING
;
3620 list_for_each_entry(wq
, &workqueues
, list
) {
3621 struct cpu_workqueue_struct
*cwq
= get_cwq(cpu
, wq
);
3623 if (!cwq
|| !(wq
->flags
& WQ_FREEZEABLE
))
3626 /* restore max_active and repopulate worklist */
3627 cwq
->max_active
= wq
->saved_max_active
;
3629 while (!list_empty(&cwq
->delayed_works
) &&
3630 cwq
->nr_active
< cwq
->max_active
)
3631 cwq_activate_first_delayed(cwq
);
3634 wake_up_worker(gcwq
);
3636 spin_unlock_irq(&gcwq
->lock
);
3639 workqueue_freezing
= false;
3641 spin_unlock(&workqueue_lock
);
3643 #endif /* CONFIG_FREEZER */
3645 static int __init
init_workqueues(void)
3650 cpu_notifier(workqueue_cpu_callback
, CPU_PRI_WORKQUEUE
);
3652 /* initialize gcwqs */
3653 for_each_gcwq_cpu(cpu
) {
3654 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3656 spin_lock_init(&gcwq
->lock
);
3657 INIT_LIST_HEAD(&gcwq
->worklist
);
3659 gcwq
->flags
|= GCWQ_DISASSOCIATED
;
3661 INIT_LIST_HEAD(&gcwq
->idle_list
);
3662 for (i
= 0; i
< BUSY_WORKER_HASH_SIZE
; i
++)
3663 INIT_HLIST_HEAD(&gcwq
->busy_hash
[i
]);
3665 init_timer_deferrable(&gcwq
->idle_timer
);
3666 gcwq
->idle_timer
.function
= idle_worker_timeout
;
3667 gcwq
->idle_timer
.data
= (unsigned long)gcwq
;
3669 setup_timer(&gcwq
->mayday_timer
, gcwq_mayday_timeout
,
3670 (unsigned long)gcwq
);
3672 ida_init(&gcwq
->worker_ida
);
3674 gcwq
->trustee_state
= TRUSTEE_DONE
;
3675 init_waitqueue_head(&gcwq
->trustee_wait
);
3678 /* create the initial worker */
3679 for_each_online_gcwq_cpu(cpu
) {
3680 struct global_cwq
*gcwq
= get_gcwq(cpu
);
3681 struct worker
*worker
;
3683 if (cpu
!= WORK_CPU_UNBOUND
)
3684 gcwq
->flags
&= ~GCWQ_DISASSOCIATED
;
3685 worker
= create_worker(gcwq
, true);
3687 spin_lock_irq(&gcwq
->lock
);
3688 start_worker(worker
);
3689 spin_unlock_irq(&gcwq
->lock
);
3692 system_wq
= alloc_workqueue("events", 0, 0);
3693 system_long_wq
= alloc_workqueue("events_long", 0, 0);
3694 system_nrt_wq
= alloc_workqueue("events_nrt", WQ_NON_REENTRANT
, 0);
3695 system_unbound_wq
= alloc_workqueue("events_unbound", WQ_UNBOUND
,
3696 WQ_UNBOUND_MAX_ACTIVE
);
3697 BUG_ON(!system_wq
|| !system_long_wq
|| !system_nrt_wq
);
3700 early_initcall(init_workqueues
);